[quake3-commits] r1926 - in trunk: . code code/jpeg-8c code/renderer

DONOTREPLY at icculus.org DONOTREPLY at icculus.org
Sat Mar 12 11:45:19 EST 2011


Author: thilo
Date: 2011-03-12 11:45:15 -0500 (Sat, 12 Mar 2011)
New Revision: 1926

Added:
   trunk/code/jpeg-8c/
   trunk/code/jpeg-8c/README
   trunk/code/jpeg-8c/jaricom.c
   trunk/code/jpeg-8c/jcapimin.c
   trunk/code/jpeg-8c/jcapistd.c
   trunk/code/jpeg-8c/jcarith.c
   trunk/code/jpeg-8c/jccoefct.c
   trunk/code/jpeg-8c/jccolor.c
   trunk/code/jpeg-8c/jcdctmgr.c
   trunk/code/jpeg-8c/jchuff.c
   trunk/code/jpeg-8c/jcinit.c
   trunk/code/jpeg-8c/jcmainct.c
   trunk/code/jpeg-8c/jcmarker.c
   trunk/code/jpeg-8c/jcmaster.c
   trunk/code/jpeg-8c/jcomapi.c
   trunk/code/jpeg-8c/jcparam.c
   trunk/code/jpeg-8c/jcprepct.c
   trunk/code/jpeg-8c/jcsample.c
   trunk/code/jpeg-8c/jctrans.c
   trunk/code/jpeg-8c/jdapimin.c
   trunk/code/jpeg-8c/jdapistd.c
   trunk/code/jpeg-8c/jdarith.c
   trunk/code/jpeg-8c/jdatadst.c
   trunk/code/jpeg-8c/jdatasrc.c
   trunk/code/jpeg-8c/jdcoefct.c
   trunk/code/jpeg-8c/jdcolor.c
   trunk/code/jpeg-8c/jdct.h
   trunk/code/jpeg-8c/jddctmgr.c
   trunk/code/jpeg-8c/jdhuff.c
   trunk/code/jpeg-8c/jdinput.c
   trunk/code/jpeg-8c/jdmainct.c
   trunk/code/jpeg-8c/jdmarker.c
   trunk/code/jpeg-8c/jdmaster.c
   trunk/code/jpeg-8c/jdmerge.c
   trunk/code/jpeg-8c/jdpostct.c
   trunk/code/jpeg-8c/jdsample.c
   trunk/code/jpeg-8c/jdtrans.c
   trunk/code/jpeg-8c/jerror.c
   trunk/code/jpeg-8c/jerror.h
   trunk/code/jpeg-8c/jfdctflt.c
   trunk/code/jpeg-8c/jfdctfst.c
   trunk/code/jpeg-8c/jfdctint.c
   trunk/code/jpeg-8c/jidctflt.c
   trunk/code/jpeg-8c/jidctfst.c
   trunk/code/jpeg-8c/jidctint.c
   trunk/code/jpeg-8c/jinclude.h
   trunk/code/jpeg-8c/jmemmgr.c
   trunk/code/jpeg-8c/jmemnobs.c
   trunk/code/jpeg-8c/jmemsys.h
   trunk/code/jpeg-8c/jmorecfg.h
   trunk/code/jpeg-8c/jpegint.h
   trunk/code/jpeg-8c/jpeglib.h
   trunk/code/jpeg-8c/jquant1.c
   trunk/code/jpeg-8c/jquant2.c
   trunk/code/jpeg-8c/jutils.c
   trunk/code/jpeg-8c/jversion.h
Removed:
   trunk/code/jpeg-6b/
Modified:
   trunk/Makefile
   trunk/README
   trunk/code/renderer/tr_image_jpg.c
Log:
- Migrate to jpeg-8c
- Add possibility to link against system libjpeg


Modified: trunk/Makefile
===================================================================
--- trunk/Makefile	2011-03-11 14:40:07 UTC (rev 1925)
+++ trunk/Makefile	2011-03-12 16:45:15 UTC (rev 1926)
@@ -151,6 +151,10 @@
 USE_INTERNAL_ZLIB=1
 endif
 
+ifndef USE_INTERNAL_JPEG
+USE_INTERNAL_JPEG=1
+endif
+
 ifndef USE_LOCAL_HEADERS
 USE_LOCAL_HEADERS=1
 endif
@@ -176,7 +180,7 @@
 NDIR=$(MOUNT_DIR)/null
 UIDIR=$(MOUNT_DIR)/ui
 Q3UIDIR=$(MOUNT_DIR)/q3_ui
-JPDIR=$(MOUNT_DIR)/jpeg-6b
+JPDIR=$(MOUNT_DIR)/jpeg-8c
 SPEEXDIR=$(MOUNT_DIR)/libspeex
 ZDIR=$(MOUNT_DIR)/zlib
 Q3ASMDIR=$(MOUNT_DIR)/tools/asm
@@ -908,6 +912,13 @@
   LIBS += -lz
 endif
 
+ifeq ($(USE_INTERNAL_JPEG),1)
+  BASE_CFLAGS += -DUSE_INTERNAL_JPEG
+  BASE_CFLAGS += -I$(JPDIR)
+else
+  LIBS += -ljpeg
+endif
+
 ifdef DEFAULT_BASEDIR
   BASE_CFLAGS += -DDEFAULT_BASEDIR=\\\"$(DEFAULT_BASEDIR)\\\"
 endif
@@ -1398,43 +1409,6 @@
   $(B)/client/l_precomp.o \
   $(B)/client/l_script.o \
   $(B)/client/l_struct.o \
-  \
-  $(B)/client/jcapimin.o \
-  $(B)/client/jcapistd.o \
-  $(B)/client/jccoefct.o  \
-  $(B)/client/jccolor.o \
-  $(B)/client/jcdctmgr.o \
-  $(B)/client/jchuff.o   \
-  $(B)/client/jcinit.o \
-  $(B)/client/jcmainct.o \
-  $(B)/client/jcmarker.o \
-  $(B)/client/jcmaster.o \
-  $(B)/client/jcomapi.o \
-  $(B)/client/jcparam.o \
-  $(B)/client/jcphuff.o \
-  $(B)/client/jcprepct.o \
-  $(B)/client/jcsample.o \
-  $(B)/client/jdapimin.o \
-  $(B)/client/jdapistd.o \
-  $(B)/client/jdatasrc.o \
-  $(B)/client/jdcoefct.o \
-  $(B)/client/jdcolor.o \
-  $(B)/client/jddctmgr.o \
-  $(B)/client/jdhuff.o \
-  $(B)/client/jdinput.o \
-  $(B)/client/jdmainct.o \
-  $(B)/client/jdmarker.o \
-  $(B)/client/jdmaster.o \
-  $(B)/client/jdpostct.o \
-  $(B)/client/jdsample.o \
-  $(B)/client/jdtrans.o \
-  $(B)/client/jerror.o \
-  $(B)/client/jfdctflt.o \
-  $(B)/client/jidctflt.o \
-  $(B)/client/jmemmgr.o \
-  $(B)/client/jmemnobs.o \
-  $(B)/client/jutils.o \
-  \
   $(B)/client/tr_animation.o \
   $(B)/client/tr_backend.o \
   $(B)/client/tr_bsp.o \
@@ -1472,6 +1446,56 @@
   $(B)/client/con_log.o \
   $(B)/client/sys_main.o
 
+ifneq ($(USE_INTERNAL_JPEG),0)
+  Q3OBJ += \
+    $(B)/client/jaricom.o \
+    $(B)/client/jcapimin.o \
+    $(B)/client/jcapistd.o \
+    $(B)/client/jcarith.o \
+    $(B)/client/jccoefct.o  \
+    $(B)/client/jccolor.o \
+    $(B)/client/jcdctmgr.o \
+    $(B)/client/jchuff.o   \
+    $(B)/client/jcinit.o \
+    $(B)/client/jcmainct.o \
+    $(B)/client/jcmarker.o \
+    $(B)/client/jcmaster.o \
+    $(B)/client/jcomapi.o \
+    $(B)/client/jcparam.o \
+    $(B)/client/jcprepct.o \
+    $(B)/client/jcsample.o \
+    $(B)/client/jctrans.o \
+    $(B)/client/jdapimin.o \
+    $(B)/client/jdapistd.o \
+    $(B)/client/jdarith.o \
+    $(B)/client/jdatadst.o \
+    $(B)/client/jdatasrc.o \
+    $(B)/client/jdcoefct.o \
+    $(B)/client/jdcolor.o \
+    $(B)/client/jddctmgr.o \
+    $(B)/client/jdhuff.o \
+    $(B)/client/jdinput.o \
+    $(B)/client/jdmainct.o \
+    $(B)/client/jdmarker.o \
+    $(B)/client/jdmaster.o \
+    $(B)/client/jdmerge.o \
+    $(B)/client/jdpostct.o \
+    $(B)/client/jdsample.o \
+    $(B)/client/jdtrans.o \
+    $(B)/client/jerror.o \
+    $(B)/client/jfdctflt.o \
+    $(B)/client/jfdctfst.o \
+    $(B)/client/jfdctint.o \
+    $(B)/client/jidctflt.o \
+    $(B)/client/jidctfst.o \
+    $(B)/client/jidctint.o \
+    $(B)/client/jmemmgr.o \
+    $(B)/client/jmemnobs.o \
+    $(B)/client/jquant1.o \
+    $(B)/client/jquant2.o \
+    $(B)/client/jutils.o
+endif
+
 ifeq ($(ARCH),i386)
   Q3OBJ += \
     $(B)/client/snd_mixa.o \

Modified: trunk/README
===================================================================
--- trunk/README	2011-03-11 14:40:07 UTC (rev 1925)
+++ trunk/README	2011-03-12 16:45:15 UTC (rev 1926)
@@ -77,23 +77,30 @@
 The following variables may be set, either on the command line or in
 Makefile.local:
 
-  CFLAGS            - use this for custom CFLAGS
-  V                 - set to show cc command line when building
-  DEFAULT_BASEDIR   - extra path to search for baseq3 and such
-  BUILD_SERVER      - build the 'ioq3ded' server binary
-  BUILD_CLIENT      - build the 'ioquake3' client binary
-  BUILD_CLIENT_SMP  - build the 'ioquake3-smp' client binary
-  BUILD_GAME_SO     - build the game shared libraries
-  BUILD_GAME_QVM    - build the game qvms
-  BUILD_STANDALONE  - build binaries suited for stand-alone games
-  USE_OPENAL        - use OpenAL where available
-  USE_OPENAL_DLOPEN - link with OpenAL at runtime
-  USE_CURL          - use libcurl for http/ftp download support
-  USE_CURL_DLOPEN   - link with libcurl at runtime
-  USE_CODEC_VORBIS  - enable Ogg Vorbis support
-  USE_LOCAL_HEADERS - use headers local to ioq3 instead of system ones
-  COPYDIR           - the target installation directory
-  TEMPDIR           - specify user defined directory for temp files
+  CFLAGS             - use this for custom CFLAGS
+  V                  - set to show cc command line when building
+  DEFAULT_BASEDIR    - extra path to search for baseq3 and such
+  BUILD_SERVER       - build the 'ioq3ded' server binary
+  BUILD_CLIENT       - build the 'ioquake3' client binary
+  BUILD_CLIENT_SMP   - build the 'ioquake3-smp' client binary
+  BUILD_GAME_SO      - build the game shared libraries
+  BUILD_GAME_QVM     - build the game qvms
+  BUILD_STANDALONE   - build binaries suited for stand-alone games
+  USE_OPENAL         - use OpenAL where available
+  USE_OPENAL_DLOPEN  - link with OpenAL at runtime
+  USE_CURL           - use libcurl for http/ftp download support
+  USE_CURL_DLOPEN    - link with libcurl at runtime
+  USE_CODEC_VORBIS   - enable Ogg Vorbis support
+  USE_MUMBLE         - enable Mumble support
+  USE_VOIP           - enable built-in VoIP support
+  USE_INTERNAL_SPEEX - build internal speex library instead of dynamically
+                       linking against system libspeex
+  USE_INTERNAL_ZLIB  - build and link against internal zlib
+  USE_INTERNAL_JPEG  - build and link against internal JPEG library
+  USE_LOCAL_HEADERS  - use headers local to ioq3 instead of system ones
+  DEBUG_CFLAGS       - C compiler flags to use for building debug version
+  COPYDIR            - the target installation directory
+  TEMPDIR            - specify user defined directory for temp files
 
 The defaults for these variables differ depending on the target platform.
 

Added: trunk/code/jpeg-8c/README
===================================================================
--- trunk/code/jpeg-8c/README	                        (rev 0)
+++ trunk/code/jpeg-8c/README	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,326 @@
+The Independent JPEG Group's JPEG software
+==========================================
+
+README for release 8c of 16-Jan-2011
+====================================
+
+This distribution contains the eighth public release of the Independent JPEG
+Group's free JPEG software.  You are welcome to redistribute this software and
+to use it for any purpose, subject to the conditions under LEGAL ISSUES, below.
+
+This software is the work of Tom Lane, Guido Vollbeding, Philip Gladstone,
+Bill Allombert, Jim Boucher, Lee Crocker, Bob Friesenhahn, Ben Jackson,
+Julian Minguillon, Luis Ortiz, George Phillips, Davide Rossi, Ge' Weijers,
+and other members of the Independent JPEG Group.
+
+IJG is not affiliated with the official ISO JPEG standards committee.
+
+
+DOCUMENTATION ROADMAP
+=====================
+
+This file contains the following sections:
+
+OVERVIEW            General description of JPEG and the IJG software.
+LEGAL ISSUES        Copyright, lack of warranty, terms of distribution.
+REFERENCES          Where to learn more about JPEG.
+ARCHIVE LOCATIONS   Where to find newer versions of this software.
+ACKNOWLEDGMENTS     Special thanks.
+FILE FORMAT WARS    Software *not* to get.
+TO DO               Plans for future IJG releases.
+
+Other documentation files in the distribution are:
+
+User documentation:
+  install.txt       How to configure and install the IJG software.
+  usage.txt         Usage instructions for cjpeg, djpeg, jpegtran,
+                    rdjpgcom, and wrjpgcom.
+  *.1               Unix-style man pages for programs (same info as usage.txt).
+  wizard.txt        Advanced usage instructions for JPEG wizards only.
+  change.log        Version-to-version change highlights.
+Programmer and internal documentation:
+  libjpeg.txt       How to use the JPEG library in your own programs.
+  example.c         Sample code for calling the JPEG library.
+  structure.txt     Overview of the JPEG library's internal structure.
+  filelist.txt      Road map of IJG files.
+  coderules.txt     Coding style rules --- please read if you contribute code.
+
+Please read at least the files install.txt and usage.txt.  Some information
+can also be found in the JPEG FAQ (Frequently Asked Questions) article.  See
+ARCHIVE LOCATIONS below to find out where to obtain the FAQ article.
+
+If you want to understand how the JPEG code works, we suggest reading one or
+more of the REFERENCES, then looking at the documentation files (in roughly
+the order listed) before diving into the code.
+
+
+OVERVIEW
+========
+
+This package contains C software to implement JPEG image encoding, decoding,
+and transcoding.  JPEG (pronounced "jay-peg") is a standardized compression
+method for full-color and gray-scale images.
+
+This software implements JPEG baseline, extended-sequential, and progressive
+compression processes.  Provision is made for supporting all variants of these
+processes, although some uncommon parameter settings aren't implemented yet.
+We have made no provision for supporting the hierarchical or lossless
+processes defined in the standard.
+
+We provide a set of library routines for reading and writing JPEG image files,
+plus two sample applications "cjpeg" and "djpeg", which use the library to
+perform conversion between JPEG and some other popular image file formats.
+The library is intended to be reused in other applications.
+
+In order to support file conversion and viewing software, we have included
+considerable functionality beyond the bare JPEG coding/decoding capability;
+for example, the color quantization modules are not strictly part of JPEG
+decoding, but they are essential for output to colormapped file formats or
+colormapped displays.  These extra functions can be compiled out of the
+library if not required for a particular application.
+
+We have also included "jpegtran", a utility for lossless transcoding between
+different JPEG processes, and "rdjpgcom" and "wrjpgcom", two simple
+applications for inserting and extracting textual comments in JFIF files.
+
+The emphasis in designing this software has been on achieving portability and
+flexibility, while also making it fast enough to be useful.  In particular,
+the software is not intended to be read as a tutorial on JPEG.  (See the
+REFERENCES section for introductory material.)  Rather, it is intended to
+be reliable, portable, industrial-strength code.  We do not claim to have
+achieved that goal in every aspect of the software, but we strive for it.
+
+We welcome the use of this software as a component of commercial products.
+No royalty is required, but we do ask for an acknowledgement in product
+documentation, as described under LEGAL ISSUES.
+
+
+LEGAL ISSUES
+============
+
+In plain English:
+
+1. We don't promise that this software works.  (But if you find any bugs,
+   please let us know!)
+2. You can use this software for whatever you want.  You don't have to pay us.
+3. You may not pretend that you wrote this software.  If you use it in a
+   program, you must acknowledge somewhere in your documentation that
+   you've used the IJG code.
+
+In legalese:
+
+The authors make NO WARRANTY or representation, either express or implied,
+with respect to this software, its quality, accuracy, merchantability, or
+fitness for a particular purpose.  This software is provided "AS IS", and you,
+its user, assume the entire risk as to its quality and accuracy.
+
+This software is copyright (C) 1991-2011, Thomas G. Lane, Guido Vollbeding.
+All Rights Reserved except as specified below.
+
+Permission is hereby granted to use, copy, modify, and distribute this
+software (or portions thereof) for any purpose, without fee, subject to these
+conditions:
+(1) If any part of the source code for this software is distributed, then this
+README file must be included, with this copyright and no-warranty notice
+unaltered; and any additions, deletions, or changes to the original files
+must be clearly indicated in accompanying documentation.
+(2) If only executable code is distributed, then the accompanying
+documentation must state that "this software is based in part on the work of
+the Independent JPEG Group".
+(3) Permission for use of this software is granted only if the user accepts
+full responsibility for any undesirable consequences; the authors accept
+NO LIABILITY for damages of any kind.
+
+These conditions apply to any software derived from or based on the IJG code,
+not just to the unmodified library.  If you use our work, you ought to
+acknowledge us.
+
+Permission is NOT granted for the use of any IJG author's name or company name
+in advertising or publicity relating to this software or products derived from
+it.  This software may be referred to only as "the Independent JPEG Group's
+software".
+
+We specifically permit and encourage the use of this software as the basis of
+commercial products, provided that all warranty or liability claims are
+assumed by the product vendor.
+
+
+ansi2knr.c is included in this distribution by permission of L. Peter Deutsch,
+sole proprietor of its copyright holder, Aladdin Enterprises of Menlo Park, CA.
+ansi2knr.c is NOT covered by the above copyright and conditions, but instead
+by the usual distribution terms of the Free Software Foundation; principally,
+that you must include source code if you redistribute it.  (See the file
+ansi2knr.c for full details.)  However, since ansi2knr.c is not needed as part
+of any program generated from the IJG code, this does not limit you more than
+the foregoing paragraphs do.
+
+The Unix configuration script "configure" was produced with GNU Autoconf.
+It is copyright by the Free Software Foundation but is freely distributable.
+The same holds for its supporting scripts (config.guess, config.sub,
+ltmain.sh).  Another support script, install-sh, is copyright by X Consortium
+but is also freely distributable.
+
+The IJG distribution formerly included code to read and write GIF files.
+To avoid entanglement with the Unisys LZW patent, GIF reading support has
+been removed altogether, and the GIF writer has been simplified to produce
+"uncompressed GIFs".  This technique does not use the LZW algorithm; the
+resulting GIF files are larger than usual, but are readable by all standard
+GIF decoders.
+
+We are required to state that
+    "The Graphics Interchange Format(c) is the Copyright property of
+    CompuServe Incorporated.  GIF(sm) is a Service Mark property of
+    CompuServe Incorporated."
+
+
+REFERENCES
+==========
+
+We recommend reading one or more of these references before trying to
+understand the innards of the JPEG software.
+
+The best short technical introduction to the JPEG compression algorithm is
+	Wallace, Gregory K.  "The JPEG Still Picture Compression Standard",
+	Communications of the ACM, April 1991 (vol. 34 no. 4), pp. 30-44.
+(Adjacent articles in that issue discuss MPEG motion picture compression,
+applications of JPEG, and related topics.)  If you don't have the CACM issue
+handy, a PostScript file containing a revised version of Wallace's article is
+available at http://www.ijg.org/files/wallace.ps.gz.  The file (actually
+a preprint for an article that appeared in IEEE Trans. Consumer Electronics)
+omits the sample images that appeared in CACM, but it includes corrections
+and some added material.  Note: the Wallace article is copyright ACM and IEEE,
+and it may not be used for commercial purposes.
+
+A somewhat less technical, more leisurely introduction to JPEG can be found in
+"The Data Compression Book" by Mark Nelson and Jean-loup Gailly, published by
+M&T Books (New York), 2nd ed. 1996, ISBN 1-55851-434-1.  This book provides
+good explanations and example C code for a multitude of compression methods
+including JPEG.  It is an excellent source if you are comfortable reading C
+code but don't know much about data compression in general.  The book's JPEG
+sample code is far from industrial-strength, but when you are ready to look
+at a full implementation, you've got one here...
+
+The best currently available description of JPEG is the textbook "JPEG Still
+Image Data Compression Standard" by William B. Pennebaker and Joan L.
+Mitchell, published by Van Nostrand Reinhold, 1993, ISBN 0-442-01272-1.
+Price US$59.95, 638 pp.  The book includes the complete text of the ISO JPEG
+standards (DIS 10918-1 and draft DIS 10918-2).
+Although this is by far the most detailed and comprehensive exposition of
+JPEG publicly available, we point out that it is still missing an explanation
+of the most essential properties and algorithms of the underlying DCT
+technology.
+If you think that you know about DCT-based JPEG after reading this book,
+then you are in delusion.  The real fundamentals and corresponding potential
+of DCT-based JPEG are not publicly known so far, and that is the reason for
+all the mistaken developments taking place in the image coding domain.
+
+The original JPEG standard is divided into two parts, Part 1 being the actual
+specification, while Part 2 covers compliance testing methods.  Part 1 is
+titled "Digital Compression and Coding of Continuous-tone Still Images,
+Part 1: Requirements and guidelines" and has document numbers ISO/IEC IS
+10918-1, ITU-T T.81.  Part 2 is titled "Digital Compression and Coding of
+Continuous-tone Still Images, Part 2: Compliance testing" and has document
+numbers ISO/IEC IS 10918-2, ITU-T T.83.
+IJG JPEG 8 introduces an implementation of the JPEG SmartScale extension
+which is specified in a contributed document at ITU and ISO with title "ITU-T
+JPEG-Plus Proposal for Extending ITU-T T.81 for Advanced Image Coding", April
+2006, Geneva, Switzerland.  The latest version of the document is Revision 3.
+
+The JPEG standard does not specify all details of an interchangeable file
+format.  For the omitted details we follow the "JFIF" conventions, revision
+1.02.  JFIF 1.02 has been adopted as an Ecma International Technical Report
+and thus received a formal publication status.  It is available as a free
+download in PDF format from
+http://www.ecma-international.org/publications/techreports/E-TR-098.htm.
+A PostScript version of the JFIF document is available at
+http://www.ijg.org/files/jfif.ps.gz.  There is also a plain text version at
+http://www.ijg.org/files/jfif.txt.gz, but it is missing the figures.
+
+The TIFF 6.0 file format specification can be obtained by FTP from
+ftp://ftp.sgi.com/graphics/tiff/TIFF6.ps.gz.  The JPEG incorporation scheme
+found in the TIFF 6.0 spec of 3-June-92 has a number of serious problems.
+IJG does not recommend use of the TIFF 6.0 design (TIFF Compression tag 6).
+Instead, we recommend the JPEG design proposed by TIFF Technical Note #2
+(Compression tag 7).  Copies of this Note can be obtained from
+http://www.ijg.org/files/.  It is expected that the next revision
+of the TIFF spec will replace the 6.0 JPEG design with the Note's design.
+Although IJG's own code does not support TIFF/JPEG, the free libtiff library
+uses our library to implement TIFF/JPEG per the Note.
+
+
+ARCHIVE LOCATIONS
+=================
+
+The "official" archive site for this software is www.ijg.org.
+The most recent released version can always be found there in
+directory "files".  This particular version will be archived as
+http://www.ijg.org/files/jpegsrc.v8c.tar.gz, and in Windows-compatible
+"zip" archive format as http://www.ijg.org/files/jpegsr8c.zip.
+
+The JPEG FAQ (Frequently Asked Questions) article is a source of some
+general information about JPEG.
+It is available on the World Wide Web at http://www.faqs.org/faqs/jpeg-faq/
+and other news.answers archive sites, including the official news.answers
+archive at rtfm.mit.edu: ftp://rtfm.mit.edu/pub/usenet/news.answers/jpeg-faq/.
+If you don't have Web or FTP access, send e-mail to mail-server at rtfm.mit.edu
+with body
+	send usenet/news.answers/jpeg-faq/part1
+	send usenet/news.answers/jpeg-faq/part2
+
+
+ACKNOWLEDGMENTS
+===============
+
+Thank to Juergen Bruder for providing me with a copy of the common DCT
+algorithm article, only to find out that I had come to the same result
+in a more direct and comprehensible way with a more generative approach.
+
+Thank to Istvan Sebestyen and Joan L. Mitchell for inviting me to the
+ITU JPEG (Study Group 16) meeting in Geneva, Switzerland.
+
+Thank to Thomas Wiegand and Gary Sullivan for inviting me to the
+Joint Video Team (MPEG & ITU) meeting in Geneva, Switzerland.
+
+Thank to John Korejwa and Massimo Ballerini for inviting me to
+fruitful consultations in Boston, MA and Milan, Italy.
+
+Thank to Hendrik Elstner, Roland Fassauer, Simone Zuck, Guenther
+Maier-Gerber, Walter Stoeber, Fred Schmitz, and Norbert Braunagel
+for corresponding business development.
+
+Thank to Nico Zschach and Dirk Stelling of the technical support team
+at the Digital Images company in Halle for providing me with extra
+equipment for configuration tests.
+
+Thank to Richard F. Lyon (then of Foveon Inc.) for fruitful
+communication about JPEG configuration in Sigma Photo Pro software.
+
+Thank to Andrew Finkenstadt for hosting the ijg.org site.
+
+Last but not least special thank to Thomas G. Lane for the original
+design and development of this singular software package.
+
+
+FILE FORMAT WARS
+================
+
+The ISO JPEG standards committee actually promotes different formats like
+"JPEG 2000" or "JPEG XR" which are incompatible with original DCT-based
+JPEG and which are based on faulty technologies.  IJG therefore does not
+and will not support such momentary mistakes (see REFERENCES).
+We have little or no sympathy for the promotion of these formats.  Indeed,
+one of the original reasons for developing this free software was to help
+force convergence on common, interoperable format standards for JPEG files.
+Don't use an incompatible file format!
+(In any case, our decoder will remain capable of reading existing JPEG
+image files indefinitely.)
+
+
+TO DO
+=====
+
+Version 8 is the first release of a new generation JPEG standard
+to overcome the limitations of the original JPEG specification.
+More features are being prepared for coming releases...
+
+Please send bug reports, offers of help, etc. to jpeg-info at uc.ag.

Added: trunk/code/jpeg-8c/jaricom.c
===================================================================
--- trunk/code/jpeg-8c/jaricom.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jaricom.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,153 @@
+/*
+ * jaricom.c
+ *
+ * Developed 1997-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains probability estimation tables for common use in
+ * arithmetic entropy encoding and decoding routines.
+ *
+ * This data represents Table D.2 in the JPEG spec (ISO/IEC IS 10918-1
+ * and CCITT Recommendation ITU-T T.81) and Table 24 in the JBIG spec
+ * (ISO/IEC IS 11544 and CCITT Recommendation ITU-T T.82).
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+/* The following #define specifies the packing of the four components
+ * into the compact INT32 representation.
+ * Note that this formula must match the actual arithmetic encoder
+ * and decoder implementation.  The implementation has to be changed
+ * if this formula is changed.
+ * The current organization is leaned on Markus Kuhn's JBIG
+ * implementation (jbig_tab.c).
+ */
+
+#define V(i,a,b,c,d) (((INT32)a << 16) | ((INT32)c << 8) | ((INT32)d << 7) | b)
+
+const INT32 jpeg_aritab[113+1] = {
+/*
+ * Index, Qe_Value, Next_Index_LPS, Next_Index_MPS, Switch_MPS
+ */
+  V(   0, 0x5a1d,   1,   1, 1 ),
+  V(   1, 0x2586,  14,   2, 0 ),
+  V(   2, 0x1114,  16,   3, 0 ),
+  V(   3, 0x080b,  18,   4, 0 ),
+  V(   4, 0x03d8,  20,   5, 0 ),
+  V(   5, 0x01da,  23,   6, 0 ),
+  V(   6, 0x00e5,  25,   7, 0 ),
+  V(   7, 0x006f,  28,   8, 0 ),
+  V(   8, 0x0036,  30,   9, 0 ),
+  V(   9, 0x001a,  33,  10, 0 ),
+  V(  10, 0x000d,  35,  11, 0 ),
+  V(  11, 0x0006,   9,  12, 0 ),
+  V(  12, 0x0003,  10,  13, 0 ),
+  V(  13, 0x0001,  12,  13, 0 ),
+  V(  14, 0x5a7f,  15,  15, 1 ),
+  V(  15, 0x3f25,  36,  16, 0 ),
+  V(  16, 0x2cf2,  38,  17, 0 ),
+  V(  17, 0x207c,  39,  18, 0 ),
+  V(  18, 0x17b9,  40,  19, 0 ),
+  V(  19, 0x1182,  42,  20, 0 ),
+  V(  20, 0x0cef,  43,  21, 0 ),
+  V(  21, 0x09a1,  45,  22, 0 ),
+  V(  22, 0x072f,  46,  23, 0 ),
+  V(  23, 0x055c,  48,  24, 0 ),
+  V(  24, 0x0406,  49,  25, 0 ),
+  V(  25, 0x0303,  51,  26, 0 ),
+  V(  26, 0x0240,  52,  27, 0 ),
+  V(  27, 0x01b1,  54,  28, 0 ),
+  V(  28, 0x0144,  56,  29, 0 ),
+  V(  29, 0x00f5,  57,  30, 0 ),
+  V(  30, 0x00b7,  59,  31, 0 ),
+  V(  31, 0x008a,  60,  32, 0 ),
+  V(  32, 0x0068,  62,  33, 0 ),
+  V(  33, 0x004e,  63,  34, 0 ),
+  V(  34, 0x003b,  32,  35, 0 ),
+  V(  35, 0x002c,  33,   9, 0 ),
+  V(  36, 0x5ae1,  37,  37, 1 ),
+  V(  37, 0x484c,  64,  38, 0 ),
+  V(  38, 0x3a0d,  65,  39, 0 ),
+  V(  39, 0x2ef1,  67,  40, 0 ),
+  V(  40, 0x261f,  68,  41, 0 ),
+  V(  41, 0x1f33,  69,  42, 0 ),
+  V(  42, 0x19a8,  70,  43, 0 ),
+  V(  43, 0x1518,  72,  44, 0 ),
+  V(  44, 0x1177,  73,  45, 0 ),
+  V(  45, 0x0e74,  74,  46, 0 ),
+  V(  46, 0x0bfb,  75,  47, 0 ),
+  V(  47, 0x09f8,  77,  48, 0 ),
+  V(  48, 0x0861,  78,  49, 0 ),
+  V(  49, 0x0706,  79,  50, 0 ),
+  V(  50, 0x05cd,  48,  51, 0 ),
+  V(  51, 0x04de,  50,  52, 0 ),
+  V(  52, 0x040f,  50,  53, 0 ),
+  V(  53, 0x0363,  51,  54, 0 ),
+  V(  54, 0x02d4,  52,  55, 0 ),
+  V(  55, 0x025c,  53,  56, 0 ),
+  V(  56, 0x01f8,  54,  57, 0 ),
+  V(  57, 0x01a4,  55,  58, 0 ),
+  V(  58, 0x0160,  56,  59, 0 ),
+  V(  59, 0x0125,  57,  60, 0 ),
+  V(  60, 0x00f6,  58,  61, 0 ),
+  V(  61, 0x00cb,  59,  62, 0 ),
+  V(  62, 0x00ab,  61,  63, 0 ),
+  V(  63, 0x008f,  61,  32, 0 ),
+  V(  64, 0x5b12,  65,  65, 1 ),
+  V(  65, 0x4d04,  80,  66, 0 ),
+  V(  66, 0x412c,  81,  67, 0 ),
+  V(  67, 0x37d8,  82,  68, 0 ),
+  V(  68, 0x2fe8,  83,  69, 0 ),
+  V(  69, 0x293c,  84,  70, 0 ),
+  V(  70, 0x2379,  86,  71, 0 ),
+  V(  71, 0x1edf,  87,  72, 0 ),
+  V(  72, 0x1aa9,  87,  73, 0 ),
+  V(  73, 0x174e,  72,  74, 0 ),
+  V(  74, 0x1424,  72,  75, 0 ),
+  V(  75, 0x119c,  74,  76, 0 ),
+  V(  76, 0x0f6b,  74,  77, 0 ),
+  V(  77, 0x0d51,  75,  78, 0 ),
+  V(  78, 0x0bb6,  77,  79, 0 ),
+  V(  79, 0x0a40,  77,  48, 0 ),
+  V(  80, 0x5832,  80,  81, 1 ),
+  V(  81, 0x4d1c,  88,  82, 0 ),
+  V(  82, 0x438e,  89,  83, 0 ),
+  V(  83, 0x3bdd,  90,  84, 0 ),
+  V(  84, 0x34ee,  91,  85, 0 ),
+  V(  85, 0x2eae,  92,  86, 0 ),
+  V(  86, 0x299a,  93,  87, 0 ),
+  V(  87, 0x2516,  86,  71, 0 ),
+  V(  88, 0x5570,  88,  89, 1 ),
+  V(  89, 0x4ca9,  95,  90, 0 ),
+  V(  90, 0x44d9,  96,  91, 0 ),
+  V(  91, 0x3e22,  97,  92, 0 ),
+  V(  92, 0x3824,  99,  93, 0 ),
+  V(  93, 0x32b4,  99,  94, 0 ),
+  V(  94, 0x2e17,  93,  86, 0 ),
+  V(  95, 0x56a8,  95,  96, 1 ),
+  V(  96, 0x4f46, 101,  97, 0 ),
+  V(  97, 0x47e5, 102,  98, 0 ),
+  V(  98, 0x41cf, 103,  99, 0 ),
+  V(  99, 0x3c3d, 104, 100, 0 ),
+  V( 100, 0x375e,  99,  93, 0 ),
+  V( 101, 0x5231, 105, 102, 0 ),
+  V( 102, 0x4c0f, 106, 103, 0 ),
+  V( 103, 0x4639, 107, 104, 0 ),
+  V( 104, 0x415e, 103,  99, 0 ),
+  V( 105, 0x5627, 105, 106, 1 ),
+  V( 106, 0x50e7, 108, 107, 0 ),
+  V( 107, 0x4b85, 109, 103, 0 ),
+  V( 108, 0x5597, 110, 109, 0 ),
+  V( 109, 0x504f, 111, 107, 0 ),
+  V( 110, 0x5a10, 110, 111, 1 ),
+  V( 111, 0x5522, 112, 109, 0 ),
+  V( 112, 0x59eb, 112, 111, 1 ),
+/*
+ * This last entry is used for fixed probability estimate of 0.5
+ * as recommended in Section 10.3 Table 5 of ITU-T Rec. T.851.
+ */
+  V( 113, 0x5a1d, 113, 113, 0 )
+};

Added: trunk/code/jpeg-8c/jcapimin.c
===================================================================
--- trunk/code/jpeg-8c/jcapimin.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcapimin.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,288 @@
+/*
+ * jcapimin.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * Modified 2003-2010 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the compression half
+ * of the JPEG library.  These are the "minimum" API routines that may be
+ * needed in either the normal full-compression case or the transcoding-only
+ * case.
+ *
+ * Most of the routines intended to be called directly by an application
+ * are in this file or in jcapistd.c.  But also see jcparam.c for
+ * parameter-setup helper routines, jcomapi.c for routines shared by
+ * compression and decompression, and jctrans.c for the transcoding case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Initialization of a JPEG compression object.
+ * The error manager must already be set up (in case memory manager fails).
+ */
+
+GLOBAL(void)
+jpeg_CreateCompress (j_compress_ptr cinfo, int version, size_t structsize)
+{
+  int i;
+
+  /* Guard against version mismatches between library and caller. */
+  cinfo->mem = NULL;		/* so jpeg_destroy knows mem mgr not called */
+  if (version != JPEG_LIB_VERSION)
+    ERREXIT2(cinfo, JERR_BAD_LIB_VERSION, JPEG_LIB_VERSION, version);
+  if (structsize != SIZEOF(struct jpeg_compress_struct))
+    ERREXIT2(cinfo, JERR_BAD_STRUCT_SIZE, 
+	     (int) SIZEOF(struct jpeg_compress_struct), (int) structsize);
+
+  /* For debugging purposes, we zero the whole master structure.
+   * But the application has already set the err pointer, and may have set
+   * client_data, so we have to save and restore those fields.
+   * Note: if application hasn't set client_data, tools like Purify may
+   * complain here.
+   */
+  {
+    struct jpeg_error_mgr * err = cinfo->err;
+    void * client_data = cinfo->client_data; /* ignore Purify complaint here */
+    MEMZERO(cinfo, SIZEOF(struct jpeg_compress_struct));
+    cinfo->err = err;
+    cinfo->client_data = client_data;
+  }
+  cinfo->is_decompressor = FALSE;
+
+  /* Initialize a memory manager instance for this object */
+  jinit_memory_mgr((j_common_ptr) cinfo);
+
+  /* Zero out pointers to permanent structures. */
+  cinfo->progress = NULL;
+  cinfo->dest = NULL;
+
+  cinfo->comp_info = NULL;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    cinfo->quant_tbl_ptrs[i] = NULL;
+    cinfo->q_scale_factor[i] = 100;
+  }
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    cinfo->dc_huff_tbl_ptrs[i] = NULL;
+    cinfo->ac_huff_tbl_ptrs[i] = NULL;
+  }
+
+  /* Must do it here for emit_dqt in case jpeg_write_tables is used */
+  cinfo->block_size = DCTSIZE;
+  cinfo->natural_order = jpeg_natural_order;
+  cinfo->lim_Se = DCTSIZE2-1;
+
+  cinfo->script_space = NULL;
+
+  cinfo->input_gamma = 1.0;	/* in case application forgets */
+
+  /* OK, I'm ready */
+  cinfo->global_state = CSTATE_START;
+}
+
+
+/*
+ * Destruction of a JPEG compression object
+ */
+
+GLOBAL(void)
+jpeg_destroy_compress (j_compress_ptr cinfo)
+{
+  jpeg_destroy((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Abort processing of a JPEG compression operation,
+ * but don't destroy the object itself.
+ */
+
+GLOBAL(void)
+jpeg_abort_compress (j_compress_ptr cinfo)
+{
+  jpeg_abort((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Forcibly suppress or un-suppress all quantization and Huffman tables.
+ * Marks all currently defined tables as already written (if suppress)
+ * or not written (if !suppress).  This will control whether they get emitted
+ * by a subsequent jpeg_start_compress call.
+ *
+ * This routine is exported for use by applications that want to produce
+ * abbreviated JPEG datastreams.  It logically belongs in jcparam.c, but
+ * since it is called by jpeg_start_compress, we put it here --- otherwise
+ * jcparam.o would be linked whether the application used it or not.
+ */
+
+GLOBAL(void)
+jpeg_suppress_tables (j_compress_ptr cinfo, boolean suppress)
+{
+  int i;
+  JQUANT_TBL * qtbl;
+  JHUFF_TBL * htbl;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    if ((qtbl = cinfo->quant_tbl_ptrs[i]) != NULL)
+      qtbl->sent_table = suppress;
+  }
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    if ((htbl = cinfo->dc_huff_tbl_ptrs[i]) != NULL)
+      htbl->sent_table = suppress;
+    if ((htbl = cinfo->ac_huff_tbl_ptrs[i]) != NULL)
+      htbl->sent_table = suppress;
+  }
+}
+
+
+/*
+ * Finish JPEG compression.
+ *
+ * If a multipass operating mode was selected, this may do a great deal of
+ * work including most of the actual output.
+ */
+
+GLOBAL(void)
+jpeg_finish_compress (j_compress_ptr cinfo)
+{
+  JDIMENSION iMCU_row;
+
+  if (cinfo->global_state == CSTATE_SCANNING ||
+      cinfo->global_state == CSTATE_RAW_OK) {
+    /* Terminate first pass */
+    if (cinfo->next_scanline < cinfo->image_height)
+      ERREXIT(cinfo, JERR_TOO_LITTLE_DATA);
+    (*cinfo->master->finish_pass) (cinfo);
+  } else if (cinfo->global_state != CSTATE_WRCOEFS)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Perform any remaining passes */
+  while (! cinfo->master->is_last_pass) {
+    (*cinfo->master->prepare_for_pass) (cinfo);
+    for (iMCU_row = 0; iMCU_row < cinfo->total_iMCU_rows; iMCU_row++) {
+      if (cinfo->progress != NULL) {
+	cinfo->progress->pass_counter = (long) iMCU_row;
+	cinfo->progress->pass_limit = (long) cinfo->total_iMCU_rows;
+	(*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+      }
+      /* We bypass the main controller and invoke coef controller directly;
+       * all work is being done from the coefficient buffer.
+       */
+      if (! (*cinfo->coef->compress_data) (cinfo, (JSAMPIMAGE) NULL))
+	ERREXIT(cinfo, JERR_CANT_SUSPEND);
+    }
+    (*cinfo->master->finish_pass) (cinfo);
+  }
+  /* Write EOI, do final cleanup */
+  (*cinfo->marker->write_file_trailer) (cinfo);
+  (*cinfo->dest->term_destination) (cinfo);
+  /* We can use jpeg_abort to release memory and reset global_state */
+  jpeg_abort((j_common_ptr) cinfo);
+}
+
+
+/*
+ * Write a special marker.
+ * This is only recommended for writing COM or APPn markers.
+ * Must be called after jpeg_start_compress() and before
+ * first call to jpeg_write_scanlines() or jpeg_write_raw_data().
+ */
+
+GLOBAL(void)
+jpeg_write_marker (j_compress_ptr cinfo, int marker,
+		   const JOCTET *dataptr, unsigned int datalen)
+{
+  JMETHOD(void, write_marker_byte, (j_compress_ptr info, int val));
+
+  if (cinfo->next_scanline != 0 ||
+      (cinfo->global_state != CSTATE_SCANNING &&
+       cinfo->global_state != CSTATE_RAW_OK &&
+       cinfo->global_state != CSTATE_WRCOEFS))
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  (*cinfo->marker->write_marker_header) (cinfo, marker, datalen);
+  write_marker_byte = cinfo->marker->write_marker_byte;	/* copy for speed */
+  while (datalen--) {
+    (*write_marker_byte) (cinfo, *dataptr);
+    dataptr++;
+  }
+}
+
+/* Same, but piecemeal. */
+
+GLOBAL(void)
+jpeg_write_m_header (j_compress_ptr cinfo, int marker, unsigned int datalen)
+{
+  if (cinfo->next_scanline != 0 ||
+      (cinfo->global_state != CSTATE_SCANNING &&
+       cinfo->global_state != CSTATE_RAW_OK &&
+       cinfo->global_state != CSTATE_WRCOEFS))
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  (*cinfo->marker->write_marker_header) (cinfo, marker, datalen);
+}
+
+GLOBAL(void)
+jpeg_write_m_byte (j_compress_ptr cinfo, int val)
+{
+  (*cinfo->marker->write_marker_byte) (cinfo, val);
+}
+
+
+/*
+ * Alternate compression function: just write an abbreviated table file.
+ * Before calling this, all parameters and a data destination must be set up.
+ *
+ * To produce a pair of files containing abbreviated tables and abbreviated
+ * image data, one would proceed as follows:
+ *
+ *		initialize JPEG object
+ *		set JPEG parameters
+ *		set destination to table file
+ *		jpeg_write_tables(cinfo);
+ *		set destination to image file
+ *		jpeg_start_compress(cinfo, FALSE);
+ *		write data...
+ *		jpeg_finish_compress(cinfo);
+ *
+ * jpeg_write_tables has the side effect of marking all tables written
+ * (same as jpeg_suppress_tables(..., TRUE)).  Thus a subsequent start_compress
+ * will not re-emit the tables unless it is passed write_all_tables=TRUE.
+ */
+
+GLOBAL(void)
+jpeg_write_tables (j_compress_ptr cinfo)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Initialize the marker writer ... bit of a crock to do it here. */
+  jinit_marker_writer(cinfo);
+  /* Write them tables! */
+  (*cinfo->marker->write_tables_only) (cinfo);
+  /* And clean up. */
+  (*cinfo->dest->term_destination) (cinfo);
+  /*
+   * In library releases up through v6a, we called jpeg_abort() here to free
+   * any working memory allocated by the destination manager and marker
+   * writer.  Some applications had a problem with that: they allocated space
+   * of their own from the library memory manager, and didn't want it to go
+   * away during write_tables.  So now we do nothing.  This will cause a
+   * memory leak if an app calls write_tables repeatedly without doing a full
+   * compression cycle or otherwise resetting the JPEG object.  However, that
+   * seems less bad than unexpectedly freeing memory in the normal case.
+   * An app that prefers the old behavior can call jpeg_abort for itself after
+   * each call to jpeg_write_tables().
+   */
+}

Added: trunk/code/jpeg-8c/jcapistd.c
===================================================================
--- trunk/code/jpeg-8c/jcapistd.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcapistd.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,161 @@
+/*
+ * jcapistd.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the compression half
+ * of the JPEG library.  These are the "standard" API routines that are
+ * used in the normal full-compression case.  They are not used by a
+ * transcoding-only application.  Note that if an application links in
+ * jpeg_start_compress, it will end up linking in the entire compressor.
+ * We thus must separate this file from jcapimin.c to avoid linking the
+ * whole compression library into a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Compression initialization.
+ * Before calling this, all parameters and a data destination must be set up.
+ *
+ * We require a write_all_tables parameter as a failsafe check when writing
+ * multiple datastreams from the same compression object.  Since prior runs
+ * will have left all the tables marked sent_table=TRUE, a subsequent run
+ * would emit an abbreviated stream (no tables) by default.  This may be what
+ * is wanted, but for safety's sake it should not be the default behavior:
+ * programmers should have to make a deliberate choice to emit abbreviated
+ * images.  Therefore the documentation and examples should encourage people
+ * to pass write_all_tables=TRUE; then it will take active thought to do the
+ * wrong thing.
+ */
+
+GLOBAL(void)
+jpeg_start_compress (j_compress_ptr cinfo, boolean write_all_tables)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (write_all_tables)
+    jpeg_suppress_tables(cinfo, FALSE);	/* mark all tables to be written */
+
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Perform master selection of active modules */
+  jinit_compress_master(cinfo);
+  /* Set up for the first pass */
+  (*cinfo->master->prepare_for_pass) (cinfo);
+  /* Ready for application to drive first pass through jpeg_write_scanlines
+   * or jpeg_write_raw_data.
+   */
+  cinfo->next_scanline = 0;
+  cinfo->global_state = (cinfo->raw_data_in ? CSTATE_RAW_OK : CSTATE_SCANNING);
+}
+
+
+/*
+ * Write some scanlines of data to the JPEG compressor.
+ *
+ * The return value will be the number of lines actually written.
+ * This should be less than the supplied num_lines only in case that
+ * the data destination module has requested suspension of the compressor,
+ * or if more than image_height scanlines are passed in.
+ *
+ * Note: we warn about excess calls to jpeg_write_scanlines() since
+ * this likely signals an application programmer error.  However,
+ * excess scanlines passed in the last valid call are *silently* ignored,
+ * so that the application need not adjust num_lines for end-of-image
+ * when using a multiple-scanline buffer.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_write_scanlines (j_compress_ptr cinfo, JSAMPARRAY scanlines,
+		      JDIMENSION num_lines)
+{
+  JDIMENSION row_ctr, rows_left;
+
+  if (cinfo->global_state != CSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->next_scanline >= cinfo->image_height)
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->next_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->image_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Give master control module another chance if this is first call to
+   * jpeg_write_scanlines.  This lets output of the frame/scan headers be
+   * delayed so that application can write COM, etc, markers between
+   * jpeg_start_compress and jpeg_write_scanlines.
+   */
+  if (cinfo->master->call_pass_startup)
+    (*cinfo->master->pass_startup) (cinfo);
+
+  /* Ignore any extra scanlines at bottom of image. */
+  rows_left = cinfo->image_height - cinfo->next_scanline;
+  if (num_lines > rows_left)
+    num_lines = rows_left;
+
+  row_ctr = 0;
+  (*cinfo->main->process_data) (cinfo, scanlines, &row_ctr, num_lines);
+  cinfo->next_scanline += row_ctr;
+  return row_ctr;
+}
+
+
+/*
+ * Alternate entry point to write raw data.
+ * Processes exactly one iMCU row per call, unless suspended.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_write_raw_data (j_compress_ptr cinfo, JSAMPIMAGE data,
+		     JDIMENSION num_lines)
+{
+  JDIMENSION lines_per_iMCU_row;
+
+  if (cinfo->global_state != CSTATE_RAW_OK)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->next_scanline >= cinfo->image_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->next_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->image_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Give master control module another chance if this is first call to
+   * jpeg_write_raw_data.  This lets output of the frame/scan headers be
+   * delayed so that application can write COM, etc, markers between
+   * jpeg_start_compress and jpeg_write_raw_data.
+   */
+  if (cinfo->master->call_pass_startup)
+    (*cinfo->master->pass_startup) (cinfo);
+
+  /* Verify that at least one iMCU row has been passed. */
+  lines_per_iMCU_row = cinfo->max_v_samp_factor * DCTSIZE;
+  if (num_lines < lines_per_iMCU_row)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Directly compress the row. */
+  if (! (*cinfo->coef->compress_data) (cinfo, data)) {
+    /* If compressor did not consume the whole row, suspend processing. */
+    return 0;
+  }
+
+  /* OK, we processed one iMCU row. */
+  cinfo->next_scanline += lines_per_iMCU_row;
+  return lines_per_iMCU_row;
+}

Added: trunk/code/jpeg-8c/jcarith.c
===================================================================
--- trunk/code/jpeg-8c/jcarith.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcarith.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,934 @@
+/*
+ * jcarith.c
+ *
+ * Developed 1997-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains portable arithmetic entropy encoding routines for JPEG
+ * (implementing the ISO/IEC IS 10918-1 and CCITT Recommendation ITU-T T.81).
+ *
+ * Both sequential and progressive modes are supported in this single module.
+ *
+ * Suspension is not currently supported in this module.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Expanded entropy encoder object for arithmetic encoding. */
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  INT32 c; /* C register, base of coding interval, layout as in sec. D.1.3 */
+  INT32 a;               /* A register, normalized size of coding interval */
+  INT32 sc;        /* counter for stacked 0xFF values which might overflow */
+  INT32 zc;          /* counter for pending 0x00 output values which might *
+                          * be discarded at the end ("Pacman" termination) */
+  int ct;  /* bit shift counter, determines when next byte will be written */
+  int buffer;                /* buffer for most recent output byte != 0xFF */
+
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+  int dc_context[MAX_COMPS_IN_SCAN]; /* context index for DC conditioning */
+
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+  int next_restart_num;		/* next restart number to write (0-7) */
+
+  /* Pointers to statistics areas (these workspaces have image lifespan) */
+  unsigned char * dc_stats[NUM_ARITH_TBLS];
+  unsigned char * ac_stats[NUM_ARITH_TBLS];
+
+  /* Statistics bin for coding with fixed probability 0.5 */
+  unsigned char fixed_bin[4];
+} arith_entropy_encoder;
+
+typedef arith_entropy_encoder * arith_entropy_ptr;
+
+/* The following two definitions specify the allocation chunk size
+ * for the statistics area.
+ * According to sections F.1.4.4.1.3 and F.1.4.4.2, we need at least
+ * 49 statistics bins for DC, and 245 statistics bins for AC coding.
+ *
+ * We use a compact representation with 1 byte per statistics bin,
+ * thus the numbers directly represent byte sizes.
+ * This 1 byte per statistics bin contains the meaning of the MPS
+ * (more probable symbol) in the highest bit (mask 0x80), and the
+ * index into the probability estimation state machine table
+ * in the lower bits (mask 0x7F).
+ */
+
+#define DC_STAT_BINS 64
+#define AC_STAT_BINS 256
+
+/* NOTE: Uncomment the following #define if you want to use the
+ * given formula for calculating the AC conditioning parameter Kx
+ * for spectral selection progressive coding in section G.1.3.2
+ * of the spec (Kx = Kmin + SRL (8 + Se - Kmin) 4).
+ * Although the spec and P&M authors claim that this "has proven
+ * to give good results for 8 bit precision samples", I'm not
+ * convinced yet that this is really beneficial.
+ * Early tests gave only very marginal compression enhancements
+ * (a few - around 5 or so - bytes even for very large files),
+ * which would turn out rather negative if we'd suppress the
+ * DAC (Define Arithmetic Conditioning) marker segments for
+ * the default parameters in the future.
+ * Note that currently the marker writing module emits 12-byte
+ * DAC segments for a full-component scan in a color image.
+ * This is not worth worrying about IMHO. However, since the
+ * spec defines the default values to be used if the tables
+ * are omitted (unlike Huffman tables, which are required
+ * anyway), one might optimize this behaviour in the future,
+ * and then it would be disadvantageous to use custom tables if
+ * they don't provide sufficient gain to exceed the DAC size.
+ *
+ * On the other hand, I'd consider it as a reasonable result
+ * that the conditioning has no significant influence on the
+ * compression performance. This means that the basic
+ * statistical model is already rather stable.
+ *
+ * Thus, at the moment, we use the default conditioning values
+ * anyway, and do not use the custom formula.
+ *
+#define CALCULATE_SPECTRAL_CONDITIONING
+ */
+
+/* IRIGHT_SHIFT is like RIGHT_SHIFT, but works on int rather than INT32.
+ * We assume that int right shift is unsigned if INT32 right shift is,
+ * which should be safe.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS	int ishift_temp;
+#define IRIGHT_SHIFT(x,shft)  \
+	((ishift_temp = (x)) < 0 ? \
+	 (ishift_temp >> (shft)) | ((~0) << (16-(shft))) : \
+	 (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+
+LOCAL(void)
+emit_byte (int val, j_compress_ptr cinfo)
+/* Write next output byte; we do not support suspension in this module. */
+{
+  struct jpeg_destination_mgr * dest = cinfo->dest;
+
+  *dest->next_output_byte++ = (JOCTET) val;
+  if (--dest->free_in_buffer == 0)
+    if (! (*dest->empty_output_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+}
+
+
+/*
+ * Finish up at the end of an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+finish_pass (j_compress_ptr cinfo)
+{
+  arith_entropy_ptr e = (arith_entropy_ptr) cinfo->entropy;
+  INT32 temp;
+
+  /* Section D.1.8: Termination of encoding */
+
+  /* Find the e->c in the coding interval with the largest
+   * number of trailing zero bits */
+  if ((temp = (e->a - 1 + e->c) & 0xFFFF0000L) < e->c)
+    e->c = temp + 0x8000L;
+  else
+    e->c = temp;
+  /* Send remaining bytes to output */
+  e->c <<= e->ct;
+  if (e->c & 0xF8000000L) {
+    /* One final overflow has to be handled */
+    if (e->buffer >= 0) {
+      if (e->zc)
+	do emit_byte(0x00, cinfo);
+	while (--e->zc);
+      emit_byte(e->buffer + 1, cinfo);
+      if (e->buffer + 1 == 0xFF)
+	emit_byte(0x00, cinfo);
+    }
+    e->zc += e->sc;  /* carry-over converts stacked 0xFF bytes to 0x00 */
+    e->sc = 0;
+  } else {
+    if (e->buffer == 0)
+      ++e->zc;
+    else if (e->buffer >= 0) {
+      if (e->zc)
+	do emit_byte(0x00, cinfo);
+	while (--e->zc);
+      emit_byte(e->buffer, cinfo);
+    }
+    if (e->sc) {
+      if (e->zc)
+	do emit_byte(0x00, cinfo);
+	while (--e->zc);
+      do {
+	emit_byte(0xFF, cinfo);
+	emit_byte(0x00, cinfo);
+      } while (--e->sc);
+    }
+  }
+  /* Output final bytes only if they are not 0x00 */
+  if (e->c & 0x7FFF800L) {
+    if (e->zc)  /* output final pending zero bytes */
+      do emit_byte(0x00, cinfo);
+      while (--e->zc);
+    emit_byte((e->c >> 19) & 0xFF, cinfo);
+    if (((e->c >> 19) & 0xFF) == 0xFF)
+      emit_byte(0x00, cinfo);
+    if (e->c & 0x7F800L) {
+      emit_byte((e->c >> 11) & 0xFF, cinfo);
+      if (((e->c >> 11) & 0xFF) == 0xFF)
+	emit_byte(0x00, cinfo);
+    }
+  }
+}
+
+
+/*
+ * The core arithmetic encoding routine (common in JPEG and JBIG).
+ * This needs to go as fast as possible.
+ * Machine-dependent optimization facilities
+ * are not utilized in this portable implementation.
+ * However, this code should be fairly efficient and
+ * may be a good base for further optimizations anyway.
+ *
+ * Parameter 'val' to be encoded may be 0 or 1 (binary decision).
+ *
+ * Note: I've added full "Pacman" termination support to the
+ * byte output routines, which is equivalent to the optional
+ * Discard_final_zeros procedure (Figure D.15) in the spec.
+ * Thus, we always produce the shortest possible output
+ * stream compliant to the spec (no trailing zero bytes,
+ * except for FF stuffing).
+ *
+ * I've also introduced a new scheme for accessing
+ * the probability estimation state machine table,
+ * derived from Markus Kuhn's JBIG implementation.
+ */
+
+LOCAL(void)
+arith_encode (j_compress_ptr cinfo, unsigned char *st, int val) 
+{
+  register arith_entropy_ptr e = (arith_entropy_ptr) cinfo->entropy;
+  register unsigned char nl, nm;
+  register INT32 qe, temp;
+  register int sv;
+
+  /* Fetch values from our compact representation of Table D.2:
+   * Qe values and probability estimation state machine
+   */
+  sv = *st;
+  qe = jpeg_aritab[sv & 0x7F];	/* => Qe_Value */
+  nl = qe & 0xFF; qe >>= 8;	/* Next_Index_LPS + Switch_MPS */
+  nm = qe & 0xFF; qe >>= 8;	/* Next_Index_MPS */
+
+  /* Encode & estimation procedures per sections D.1.4 & D.1.5 */
+  e->a -= qe;
+  if (val != (sv >> 7)) {
+    /* Encode the less probable symbol */
+    if (e->a >= qe) {
+      /* If the interval size (qe) for the less probable symbol (LPS)
+       * is larger than the interval size for the MPS, then exchange
+       * the two symbols for coding efficiency, otherwise code the LPS
+       * as usual: */
+      e->c += e->a;
+      e->a = qe;
+    }
+    *st = (sv & 0x80) ^ nl;	/* Estimate_after_LPS */
+  } else {
+    /* Encode the more probable symbol */
+    if (e->a >= 0x8000L)
+      return;  /* A >= 0x8000 -> ready, no renormalization required */
+    if (e->a < qe) {
+      /* If the interval size (qe) for the less probable symbol (LPS)
+       * is larger than the interval size for the MPS, then exchange
+       * the two symbols for coding efficiency: */
+      e->c += e->a;
+      e->a = qe;
+    }
+    *st = (sv & 0x80) ^ nm;	/* Estimate_after_MPS */
+  }
+
+  /* Renormalization & data output per section D.1.6 */
+  do {
+    e->a <<= 1;
+    e->c <<= 1;
+    if (--e->ct == 0) {
+      /* Another byte is ready for output */
+      temp = e->c >> 19;
+      if (temp > 0xFF) {
+	/* Handle overflow over all stacked 0xFF bytes */
+	if (e->buffer >= 0) {
+	  if (e->zc)
+	    do emit_byte(0x00, cinfo);
+	    while (--e->zc);
+	  emit_byte(e->buffer + 1, cinfo);
+	  if (e->buffer + 1 == 0xFF)
+	    emit_byte(0x00, cinfo);
+	}
+	e->zc += e->sc;  /* carry-over converts stacked 0xFF bytes to 0x00 */
+	e->sc = 0;
+	/* Note: The 3 spacer bits in the C register guarantee
+	 * that the new buffer byte can't be 0xFF here
+	 * (see page 160 in the P&M JPEG book). */
+	e->buffer = temp & 0xFF;  /* new output byte, might overflow later */
+      } else if (temp == 0xFF) {
+	++e->sc;  /* stack 0xFF byte (which might overflow later) */
+      } else {
+	/* Output all stacked 0xFF bytes, they will not overflow any more */
+	if (e->buffer == 0)
+	  ++e->zc;
+	else if (e->buffer >= 0) {
+	  if (e->zc)
+	    do emit_byte(0x00, cinfo);
+	    while (--e->zc);
+	  emit_byte(e->buffer, cinfo);
+	}
+	if (e->sc) {
+	  if (e->zc)
+	    do emit_byte(0x00, cinfo);
+	    while (--e->zc);
+	  do {
+	    emit_byte(0xFF, cinfo);
+	    emit_byte(0x00, cinfo);
+	  } while (--e->sc);
+	}
+	e->buffer = temp & 0xFF;  /* new output byte (can still overflow) */
+      }
+      e->c &= 0x7FFFFL;
+      e->ct += 8;
+    }
+  } while (e->a < 0x8000L);
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(void)
+emit_restart (j_compress_ptr cinfo, int restart_num)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  int ci;
+  jpeg_component_info * compptr;
+
+  finish_pass(cinfo);
+
+  emit_byte(0xFF, cinfo);
+  emit_byte(JPEG_RST0 + restart_num, cinfo);
+
+  /* Re-initialize statistics areas */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* DC needs no table for refinement scan */
+    if (cinfo->Ss == 0 && cinfo->Ah == 0) {
+      MEMZERO(entropy->dc_stats[compptr->dc_tbl_no], DC_STAT_BINS);
+      /* Reset DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    /* AC needs no table when not present */
+    if (cinfo->Se) {
+      MEMZERO(entropy->ac_stats[compptr->ac_tbl_no], AC_STAT_BINS);
+    }
+  }
+
+  /* Reset arithmetic encoding variables */
+  entropy->c = 0;
+  entropy->a = 0x10000L;
+  entropy->sc = 0;
+  entropy->zc = 0;
+  entropy->ct = 11;
+  entropy->buffer = -1;  /* empty */
+}
+
+
+/*
+ * MCU encoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl;
+  int v, v2, m;
+  ISHIFT_TEMPS
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    tbl = cinfo->cur_comp_info[ci]->dc_tbl_no;
+
+    /* Compute the DC value after the required point transform by Al.
+     * This is simply an arithmetic right shift.
+     */
+    m = IRIGHT_SHIFT((int) ((*block)[0]), cinfo->Al);
+
+    /* Sections F.1.4.1 & F.1.4.4.1: Encoding of DC coefficients */
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.4: Encode_DC_DIFF */
+    if ((v = m - entropy->last_dc_val[ci]) == 0) {
+      arith_encode(cinfo, st, 0);
+      entropy->dc_context[ci] = 0;	/* zero diff category */
+    } else {
+      entropy->last_dc_val[ci] = m;
+      arith_encode(cinfo, st, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      if (v > 0) {
+	arith_encode(cinfo, st + 1, 0);	/* Table F.4: SS = S0 + 1 */
+	st += 2;			/* Table F.4: SP = S0 + 2 */
+	entropy->dc_context[ci] = 4;	/* small positive diff category */
+      } else {
+	v = -v;
+	arith_encode(cinfo, st + 1, 1);	/* Table F.4: SS = S0 + 1 */
+	st += 3;			/* Table F.4: SN = S0 + 3 */
+	entropy->dc_context[ci] = 8;	/* small negative diff category */
+      }
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+	arith_encode(cinfo, st, 1);
+	m = 1;
+	v2 = v;
+	st = entropy->dc_stats[tbl] + 20; /* Table F.4: X1 = 20 */
+	while (v2 >>= 1) {
+	  arith_encode(cinfo, st, 1);
+	  m <<= 1;
+	  st += 1;
+	}
+      }
+      arith_encode(cinfo, st, 0);
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int) ((1L << cinfo->arith_dc_L[tbl]) >> 1))
+	entropy->dc_context[ci] = 0;	/* zero diff category */
+      else if (m > (int) ((1L << cinfo->arith_dc_U[tbl]) >> 1))
+	entropy->dc_context[ci] += 8;	/* large diff category */
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, k, ke;
+  int v, v2, m;
+  const int * natural_order;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  natural_order = cinfo->natural_order;
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Sections F.1.4.2 & F.1.4.4.2: Encoding of AC coefficients */
+
+  /* Establish EOB (end-of-block) index */
+  for (ke = cinfo->Se; ke > 0; ke--)
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if ((v = (*block)[natural_order[ke]]) >= 0) {
+      if (v >>= cinfo->Al) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Al) break;
+    }
+
+  /* Figure F.5: Encode_AC_Coefficients */
+  for (k = cinfo->Ss; k <= ke; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 0);		/* EOB decision */
+    for (;;) {
+      if ((v = (*block)[natural_order[k]]) >= 0) {
+	if (v >>= cinfo->Al) {
+	  arith_encode(cinfo, st + 1, 1);
+	  arith_encode(cinfo, entropy->fixed_bin, 0);
+	  break;
+	}
+      } else {
+	v = -v;
+	if (v >>= cinfo->Al) {
+	  arith_encode(cinfo, st + 1, 1);
+	  arith_encode(cinfo, entropy->fixed_bin, 1);
+	  break;
+	}
+      }
+      arith_encode(cinfo, st + 1, 0); st += 3; k++;
+    }
+    st += 2;
+    /* Figure F.8: Encoding the magnitude category of v */
+    m = 0;
+    if (v -= 1) {
+      arith_encode(cinfo, st, 1);
+      m = 1;
+      v2 = v;
+      if (v2 >>= 1) {
+	arith_encode(cinfo, st, 1);
+	m <<= 1;
+	st = entropy->ac_stats[tbl] +
+	     (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+	while (v2 >>= 1) {
+	  arith_encode(cinfo, st, 1);
+	  m <<= 1;
+	  st += 1;
+	}
+      }
+    }
+    arith_encode(cinfo, st, 0);
+    /* Figure F.9: Encoding the magnitude bit pattern of v */
+    st += 14;
+    while (m >>= 1)
+      arith_encode(cinfo, st, (m & v) ? 1 : 0);
+  }
+  /* Encode EOB decision only if k <= cinfo->Se */
+  if (k <= cinfo->Se) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for DC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  unsigned char *st;
+  int Al, blkn;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  st = entropy->fixed_bin;	/* use fixed probability estimation */
+  Al = cinfo->Al;
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    /* We simply emit the Al'th bit of the DC coefficient value. */
+    arith_encode(cinfo, st, (MCU_data[blkn][0][0] >> Al) & 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, k, ke, kex;
+  int v;
+  const int * natural_order;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  natural_order = cinfo->natural_order;
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Section G.1.3.3: Encoding of AC coefficients */
+
+  /* Establish EOB (end-of-block) index */
+  for (ke = cinfo->Se; ke > 0; ke--)
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if ((v = (*block)[natural_order[ke]]) >= 0) {
+      if (v >>= cinfo->Al) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Al) break;
+    }
+
+  /* Establish EOBx (previous stage end-of-block) index */
+  for (kex = ke; kex > 0; kex--)
+    if ((v = (*block)[natural_order[kex]]) >= 0) {
+      if (v >>= cinfo->Ah) break;
+    } else {
+      v = -v;
+      if (v >>= cinfo->Ah) break;
+    }
+
+  /* Figure G.10: Encode_AC_Coefficients_SA */
+  for (k = cinfo->Ss; k <= ke; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (k > kex)
+      arith_encode(cinfo, st, 0);	/* EOB decision */
+    for (;;) {
+      if ((v = (*block)[natural_order[k]]) >= 0) {
+	if (v >>= cinfo->Al) {
+	  if (v >> 1)			/* previously nonzero coef */
+	    arith_encode(cinfo, st + 2, (v & 1));
+	  else {			/* newly nonzero coef */
+	    arith_encode(cinfo, st + 1, 1);
+	    arith_encode(cinfo, entropy->fixed_bin, 0);
+	  }
+	  break;
+	}
+      } else {
+	v = -v;
+	if (v >>= cinfo->Al) {
+	  if (v >> 1)			/* previously nonzero coef */
+	    arith_encode(cinfo, st + 2, (v & 1));
+	  else {			/* newly nonzero coef */
+	    arith_encode(cinfo, st + 1, 1);
+	    arith_encode(cinfo, entropy->fixed_bin, 1);
+	  }
+	  break;
+	}
+      }
+      arith_encode(cinfo, st + 1, 0); st += 3; k++;
+    }
+  }
+  /* Encode EOB decision only if k <= cinfo->Se */
+  if (k <= cinfo->Se) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    arith_encode(cinfo, st, 1);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Encode and output one MCU's worth of arithmetic-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+encode_mcu (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  jpeg_component_info * compptr;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, k, ke;
+  int v, v2, m;
+  const int * natural_order;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      emit_restart(cinfo, entropy->next_restart_num);
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  natural_order = cinfo->natural_order;
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Sections F.1.4.1 & F.1.4.4.1: Encoding of DC coefficients */
+
+    tbl = compptr->dc_tbl_no;
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.4: Encode_DC_DIFF */
+    if ((v = (*block)[0] - entropy->last_dc_val[ci]) == 0) {
+      arith_encode(cinfo, st, 0);
+      entropy->dc_context[ci] = 0;	/* zero diff category */
+    } else {
+      entropy->last_dc_val[ci] = (*block)[0];
+      arith_encode(cinfo, st, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      if (v > 0) {
+	arith_encode(cinfo, st + 1, 0);	/* Table F.4: SS = S0 + 1 */
+	st += 2;			/* Table F.4: SP = S0 + 2 */
+	entropy->dc_context[ci] = 4;	/* small positive diff category */
+      } else {
+	v = -v;
+	arith_encode(cinfo, st + 1, 1);	/* Table F.4: SS = S0 + 1 */
+	st += 3;			/* Table F.4: SN = S0 + 3 */
+	entropy->dc_context[ci] = 8;	/* small negative diff category */
+      }
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+	arith_encode(cinfo, st, 1);
+	m = 1;
+	v2 = v;
+	st = entropy->dc_stats[tbl] + 20; /* Table F.4: X1 = 20 */
+	while (v2 >>= 1) {
+	  arith_encode(cinfo, st, 1);
+	  m <<= 1;
+	  st += 1;
+	}
+      }
+      arith_encode(cinfo, st, 0);
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int) ((1L << cinfo->arith_dc_L[tbl]) >> 1))
+	entropy->dc_context[ci] = 0;	/* zero diff category */
+      else if (m > (int) ((1L << cinfo->arith_dc_U[tbl]) >> 1))
+	entropy->dc_context[ci] += 8;	/* large diff category */
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+
+    /* Sections F.1.4.2 & F.1.4.4.2: Encoding of AC coefficients */
+
+    tbl = compptr->ac_tbl_no;
+
+    /* Establish EOB (end-of-block) index */
+    for (ke = cinfo->lim_Se; ke > 0; ke--)
+      if ((*block)[natural_order[ke]]) break;
+
+    /* Figure F.5: Encode_AC_Coefficients */
+    for (k = 1; k <= ke; k++) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      arith_encode(cinfo, st, 0);	/* EOB decision */
+      while ((v = (*block)[natural_order[k]]) == 0) {
+	arith_encode(cinfo, st + 1, 0); st += 3; k++;
+      }
+      arith_encode(cinfo, st + 1, 1);
+      /* Figure F.6: Encoding nonzero value v */
+      /* Figure F.7: Encoding the sign of v */
+      if (v > 0) {
+	arith_encode(cinfo, entropy->fixed_bin, 0);
+      } else {
+	v = -v;
+	arith_encode(cinfo, entropy->fixed_bin, 1);
+      }
+      st += 2;
+      /* Figure F.8: Encoding the magnitude category of v */
+      m = 0;
+      if (v -= 1) {
+	arith_encode(cinfo, st, 1);
+	m = 1;
+	v2 = v;
+	if (v2 >>= 1) {
+	  arith_encode(cinfo, st, 1);
+	  m <<= 1;
+	  st = entropy->ac_stats[tbl] +
+	       (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+	  while (v2 >>= 1) {
+	    arith_encode(cinfo, st, 1);
+	    m <<= 1;
+	    st += 1;
+	  }
+	}
+      }
+      arith_encode(cinfo, st, 0);
+      /* Figure F.9: Encoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	arith_encode(cinfo, st, (m & v) ? 1 : 0);
+    }
+    /* Encode EOB decision only if k <= cinfo->lim_Se */
+    if (k <= cinfo->lim_Se) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      arith_encode(cinfo, st, 1);
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Initialize for an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass (j_compress_ptr cinfo, boolean gather_statistics)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+
+  if (gather_statistics)
+    /* Make sure to avoid that in the master control logic!
+     * We are fully adaptive here and need no extra
+     * statistics gathering pass!
+     */
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+
+  /* We assume jcmaster.c already validated the progressive scan parameters. */
+
+  /* Select execution routines */
+  if (cinfo->progressive_mode) {
+    if (cinfo->Ah == 0) {
+      if (cinfo->Ss == 0)
+	entropy->pub.encode_mcu = encode_mcu_DC_first;
+      else
+	entropy->pub.encode_mcu = encode_mcu_AC_first;
+    } else {
+      if (cinfo->Ss == 0)
+	entropy->pub.encode_mcu = encode_mcu_DC_refine;
+      else
+	entropy->pub.encode_mcu = encode_mcu_AC_refine;
+    }
+  } else
+    entropy->pub.encode_mcu = encode_mcu;
+
+  /* Allocate & initialize requested statistics areas */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* DC needs no table for refinement scan */
+    if (cinfo->Ss == 0 && cinfo->Ah == 0) {
+      tbl = compptr->dc_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+	ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->dc_stats[tbl] == NULL)
+	entropy->dc_stats[tbl] = (unsigned char *) (*cinfo->mem->alloc_small)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, DC_STAT_BINS);
+      MEMZERO(entropy->dc_stats[tbl], DC_STAT_BINS);
+      /* Initialize DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    /* AC needs no table when not present */
+    if (cinfo->Se) {
+      tbl = compptr->ac_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+	ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->ac_stats[tbl] == NULL)
+	entropy->ac_stats[tbl] = (unsigned char *) (*cinfo->mem->alloc_small)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, AC_STAT_BINS);
+      MEMZERO(entropy->ac_stats[tbl], AC_STAT_BINS);
+#ifdef CALCULATE_SPECTRAL_CONDITIONING
+      if (cinfo->progressive_mode)
+	/* Section G.1.3.2: Set appropriate arithmetic conditioning value Kx */
+	cinfo->arith_ac_K[tbl] = cinfo->Ss + ((8 + cinfo->Se - cinfo->Ss) >> 4);
+#endif
+    }
+  }
+
+  /* Initialize arithmetic encoding variables */
+  entropy->c = 0;
+  entropy->a = 0x10000L;
+  entropy->sc = 0;
+  entropy->zc = 0;
+  entropy->ct = 11;
+  entropy->buffer = -1;  /* empty */
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/*
+ * Module initialization routine for arithmetic entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_arith_encoder (j_compress_ptr cinfo)
+{
+  arith_entropy_ptr entropy;
+  int i;
+
+  entropy = (arith_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(arith_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
+  entropy->pub.start_pass = start_pass;
+  entropy->pub.finish_pass = finish_pass;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    entropy->dc_stats[i] = NULL;
+    entropy->ac_stats[i] = NULL;
+  }
+
+  /* Initialize index for fixed probability estimation */
+  entropy->fixed_bin[0] = 113;
+}

Added: trunk/code/jpeg-8c/jccoefct.c
===================================================================
--- trunk/code/jpeg-8c/jccoefct.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jccoefct.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,453 @@
+/*
+ * jccoefct.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the coefficient buffer controller for compression.
+ * This controller is the top level of the JPEG compressor proper.
+ * The coefficient buffer lies between forward-DCT and entropy encoding steps.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* We use a full-image coefficient buffer when doing Huffman optimization,
+ * and also for writing multiple-scan JPEG files.  In all cases, the DCT
+ * step is run during the first pass, and subsequent passes need only read
+ * the buffered coefficients.
+ */
+#ifdef ENTROPY_OPT_SUPPORTED
+#define FULL_COEF_BUFFER_SUPPORTED
+#else
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+#define FULL_COEF_BUFFER_SUPPORTED
+#endif
+#endif
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_coef_controller pub; /* public fields */
+
+  JDIMENSION iMCU_row_num;	/* iMCU row # within image */
+  JDIMENSION mcu_ctr;		/* counts MCUs processed in current row */
+  int MCU_vert_offset;		/* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;	/* number of such rows needed */
+
+  /* For single-pass compression, it's sufficient to buffer just one MCU
+   * (although this may prove a bit slow in practice).  We allocate a
+   * workspace of C_MAX_BLOCKS_IN_MCU coefficient blocks, and reuse it for each
+   * MCU constructed and sent.  (On 80x86, the workspace is FAR even though
+   * it's not really very big; this is to keep the module interfaces unchanged
+   * when a large coefficient buffer is necessary.)
+   * In multi-pass modes, this array points to the current MCU's blocks
+   * within the virtual arrays.
+   */
+  JBLOCKROW MCU_buffer[C_MAX_BLOCKS_IN_MCU];
+
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+} my_coef_controller;
+
+typedef my_coef_controller * my_coef_ptr;
+
+
+/* Forward declarations */
+METHODDEF(boolean) compress_data
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+METHODDEF(boolean) compress_first_pass
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+METHODDEF(boolean) compress_output
+    JPP((j_compress_ptr cinfo, JSAMPIMAGE input_buf));
+#endif
+
+
+LOCAL(void)
+start_iMCU_row (j_compress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row */
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (coef->iMCU_row_num < (cinfo->total_iMCU_rows-1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->mcu_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_coef (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  coef->iMCU_row_num = 0;
+  start_iMCU_row(cinfo);
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (coef->whole_image[0] != NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_data;
+    break;
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+  case JBUF_SAVE_AND_PASS:
+    if (coef->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_first_pass;
+    break;
+  case JBUF_CRANK_DEST:
+    if (coef->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    coef->pub.compress_data = compress_output;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data in the single-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the image.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf contains a plane for each component in image,
+ * which we index according to the component's SOF position.
+ */
+
+METHODDEF(boolean)
+compress_data (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, bi, ci, yindex, yoffset, blockcnt;
+  JDIMENSION ypos, xpos;
+  jpeg_component_info *compptr;
+  forward_DCT_ptr forward_DCT;
+
+  /* Loop to write as much as one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num <= last_MCU_col;
+	 MCU_col_num++) {
+      /* Determine where data comes from in input_buf and do the DCT thing.
+       * Each call on forward_DCT processes a horizontal row of DCT blocks
+       * as wide as an MCU; we rely on having allocated the MCU_buffer[] blocks
+       * sequentially.  Dummy blocks at the right or bottom edge are filled in
+       * specially.  The data in them does not matter for image reconstruction,
+       * so we fill them with values that will encode to the smallest amount of
+       * data, viz: all zeroes in the AC entries, DC entries equal to previous
+       * block's DC value.  (Thanks to Thomas Kinsman for this idea.)
+       */
+      blkn = 0;
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	forward_DCT = cinfo->fdct->forward_DCT[compptr->component_index];
+	blockcnt = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
+						: compptr->last_col_width;
+	xpos = MCU_col_num * compptr->MCU_sample_width;
+	ypos = yoffset * compptr->DCT_v_scaled_size;
+	/* ypos == (yoffset+yindex) * DCTSIZE */
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  if (coef->iMCU_row_num < last_iMCU_row ||
+	      yoffset+yindex < compptr->last_row_height) {
+	    (*forward_DCT) (cinfo, compptr,
+			    input_buf[compptr->component_index],
+			    coef->MCU_buffer[blkn],
+			    ypos, xpos, (JDIMENSION) blockcnt);
+	    if (blockcnt < compptr->MCU_width) {
+	      /* Create some dummy blocks at the right edge of the image. */
+	      jzero_far((void FAR *) coef->MCU_buffer[blkn + blockcnt],
+			(compptr->MCU_width - blockcnt) * SIZEOF(JBLOCK));
+	      for (bi = blockcnt; bi < compptr->MCU_width; bi++) {
+		coef->MCU_buffer[blkn+bi][0][0] = coef->MCU_buffer[blkn+bi-1][0][0];
+	      }
+	    }
+	  } else {
+	    /* Create a row of dummy blocks at the bottom of the image. */
+	    jzero_far((void FAR *) coef->MCU_buffer[blkn],
+		      compptr->MCU_width * SIZEOF(JBLOCK));
+	    for (bi = 0; bi < compptr->MCU_width; bi++) {
+	      coef->MCU_buffer[blkn+bi][0][0] = coef->MCU_buffer[blkn-1][0][0];
+	    }
+	  }
+	  blkn += compptr->MCU_width;
+	  ypos += compptr->DCT_v_scaled_size;
+	}
+      }
+      /* Try to write the MCU.  In event of a suspension failure, we will
+       * re-DCT the MCU on restart (a bit inefficient, could be fixed...)
+       */
+      if (! (*cinfo->entropy->encode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->mcu_ctr = MCU_col_num;
+	return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+
+/*
+ * Process some data in the first pass of a multi-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the image.
+ * This amount of data is read from the source buffer, DCT'd and quantized,
+ * and saved into the virtual arrays.  We also generate suitable dummy blocks
+ * as needed at the right and lower edges.  (The dummy blocks are constructed
+ * in the virtual arrays, which have been padded appropriately.)  This makes
+ * it possible for subsequent passes not to worry about real vs. dummy blocks.
+ *
+ * We must also emit the data to the entropy encoder.  This is conveniently
+ * done by calling compress_output() after we've loaded the current strip
+ * of the virtual arrays.
+ *
+ * NB: input_buf contains a plane for each component in image.  All
+ * components are DCT'd and loaded into the virtual arrays in this pass.
+ * However, it may be that only a subset of the components are emitted to
+ * the entropy encoder during this first pass; be careful about looking
+ * at the scan-dependent variables (MCU dimensions, etc).
+ */
+
+METHODDEF(boolean)
+compress_first_pass (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION blocks_across, MCUs_across, MCUindex;
+  int bi, ci, h_samp_factor, block_row, block_rows, ndummy;
+  JCOEF lastDC;
+  jpeg_component_info *compptr;
+  JBLOCKARRAY buffer;
+  JBLOCKROW thisblockrow, lastblockrow;
+  forward_DCT_ptr forward_DCT;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Align the virtual buffer for this component. */
+    buffer = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[ci],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, TRUE);
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (coef->iMCU_row_num < last_iMCU_row)
+      block_rows = compptr->v_samp_factor;
+    else {
+      /* NB: can't use last_row_height here, since may not be set! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+    }
+    blocks_across = compptr->width_in_blocks;
+    h_samp_factor = compptr->h_samp_factor;
+    /* Count number of dummy blocks to be added at the right margin. */
+    ndummy = (int) (blocks_across % h_samp_factor);
+    if (ndummy > 0)
+      ndummy = h_samp_factor - ndummy;
+    forward_DCT = cinfo->fdct->forward_DCT[ci];
+    /* Perform DCT for all non-dummy blocks in this iMCU row.  Each call
+     * on forward_DCT processes a complete horizontal row of DCT blocks.
+     */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      thisblockrow = buffer[block_row];
+      (*forward_DCT) (cinfo, compptr, input_buf[ci], thisblockrow,
+		      (JDIMENSION) (block_row * compptr->DCT_v_scaled_size),
+		      (JDIMENSION) 0, blocks_across);
+      if (ndummy > 0) {
+	/* Create dummy blocks at the right edge of the image. */
+	thisblockrow += blocks_across; /* => first dummy block */
+	jzero_far((void FAR *) thisblockrow, ndummy * SIZEOF(JBLOCK));
+	lastDC = thisblockrow[-1][0];
+	for (bi = 0; bi < ndummy; bi++) {
+	  thisblockrow[bi][0] = lastDC;
+	}
+      }
+    }
+    /* If at end of image, create dummy block rows as needed.
+     * The tricky part here is that within each MCU, we want the DC values
+     * of the dummy blocks to match the last real block's DC value.
+     * This squeezes a few more bytes out of the resulting file...
+     */
+    if (coef->iMCU_row_num == last_iMCU_row) {
+      blocks_across += ndummy;	/* include lower right corner */
+      MCUs_across = blocks_across / h_samp_factor;
+      for (block_row = block_rows; block_row < compptr->v_samp_factor;
+	   block_row++) {
+	thisblockrow = buffer[block_row];
+	lastblockrow = buffer[block_row-1];
+	jzero_far((void FAR *) thisblockrow,
+		  (size_t) (blocks_across * SIZEOF(JBLOCK)));
+	for (MCUindex = 0; MCUindex < MCUs_across; MCUindex++) {
+	  lastDC = lastblockrow[h_samp_factor-1][0];
+	  for (bi = 0; bi < h_samp_factor; bi++) {
+	    thisblockrow[bi][0] = lastDC;
+	  }
+	  thisblockrow += h_samp_factor; /* advance to next MCU in row */
+	  lastblockrow += h_samp_factor;
+	}
+      }
+    }
+  }
+  /* NB: compress_output will increment iMCU_row_num if successful.
+   * A suspension return will result in redoing all the work above next time.
+   */
+
+  /* Emit data to the entropy encoder, sharing code with subsequent passes */
+  return compress_output(cinfo, input_buf);
+}
+
+
+/*
+ * Process some data in subsequent passes of a multi-pass case.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the scan.
+ * The data is obtained from the virtual arrays and fed to the entropy coder.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf is ignored; it is likely to be a NULL pointer.
+ */
+
+METHODDEF(boolean)
+compress_output (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  int blkn, ci, xindex, yindex, yoffset;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan.
+   * NB: during first pass, this is safe only because the buffers will
+   * already be aligned properly, so jmemmgr.c won't need to do any I/O.
+   */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, FALSE);
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num < cinfo->MCUs_per_row;
+	 MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	start_col = MCU_col_num * compptr->MCU_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
+	  for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
+	    coef->MCU_buffer[blkn++] = buffer_ptr++;
+	  }
+	}
+      }
+      /* Try to write the MCU. */
+      if (! (*cinfo->entropy->encode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->mcu_ctr = MCU_col_num;
+	return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+#endif /* FULL_COEF_BUFFER_SUPPORTED */
+
+
+/*
+ * Initialize coefficient buffer controller.
+ */
+
+GLOBAL(void)
+jinit_c_coef_controller (j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_coef_ptr coef;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_coef_controller));
+  cinfo->coef = (struct jpeg_c_coef_controller *) coef;
+  coef->pub.start_pass = start_pass_coef;
+
+  /* Create the coefficient buffer. */
+  if (need_full_buffer) {
+#ifdef FULL_COEF_BUFFER_SUPPORTED
+    /* Allocate a full-image virtual array for each component, */
+    /* padded to a multiple of samp_factor DCT blocks in each direction. */
+    int ci;
+    jpeg_component_info *compptr;
+
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 (JDIMENSION) jround_up((long) compptr->width_in_blocks,
+				(long) compptr->h_samp_factor),
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor),
+	 (JDIMENSION) compptr->v_samp_factor);
+    }
+#else
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+  } else {
+    /* We only need a single-MCU buffer. */
+    JBLOCKROW buffer;
+    int i;
+
+    buffer = (JBLOCKROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  C_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+    for (i = 0; i < C_MAX_BLOCKS_IN_MCU; i++) {
+      coef->MCU_buffer[i] = buffer + i;
+    }
+    coef->whole_image[0] = NULL; /* flag for no virtual arrays */
+  }
+}

Added: trunk/code/jpeg-8c/jccolor.c
===================================================================
--- trunk/code/jpeg-8c/jccolor.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jccolor.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,459 @@
+/*
+ * jccolor.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains input colorspace conversion routines.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_converter pub; /* public fields */
+
+  /* Private state for RGB->YCC conversion */
+  INT32 * rgb_ycc_tab;		/* => table for RGB to YCbCr conversion */
+} my_color_converter;
+
+typedef my_color_converter * my_cconvert_ptr;
+
+
+/**************** RGB -> YCbCr conversion: most common case **************/
+
+/*
+ * YCbCr is defined per CCIR 601-1, except that Cb and Cr are
+ * normalized to the range 0..MAXJSAMPLE rather than -0.5 .. 0.5.
+ * The conversion equations to be implemented are therefore
+ *	Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
+ *	Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B  + CENTERJSAMPLE
+ *	Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B  + CENTERJSAMPLE
+ * (These numbers are derived from TIFF 6.0 section 21, dated 3-June-92.)
+ * Note: older versions of the IJG code used a zero offset of MAXJSAMPLE/2,
+ * rather than CENTERJSAMPLE, for Cb and Cr.  This gave equal positive and
+ * negative swings for Cb/Cr, but meant that grayscale values (Cb=Cr=0)
+ * were not represented exactly.  Now we sacrifice exact representation of
+ * maximum red and maximum blue in order to get exact grayscales.
+ *
+ * To avoid floating-point arithmetic, we represent the fractional constants
+ * as integers scaled up by 2^16 (about 4 digits precision); we have to divide
+ * the products by 2^16, with appropriate rounding, to get the correct answer.
+ *
+ * For even more speed, we avoid doing any multiplications in the inner loop
+ * by precalculating the constants times R,G,B for all possible values.
+ * For 8-bit JSAMPLEs this is very reasonable (only 256 entries per table);
+ * for 12-bit samples it is still acceptable.  It's not very reasonable for
+ * 16-bit samples, but if you want lossless storage you shouldn't be changing
+ * colorspace anyway.
+ * The CENTERJSAMPLE offsets and the rounding fudge-factor of 0.5 are included
+ * in the tables to save adding them separately in the inner loop.
+ */
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define CBCR_OFFSET	((INT32) CENTERJSAMPLE << SCALEBITS)
+#define ONE_HALF	((INT32) 1 << (SCALEBITS-1))
+#define FIX(x)		((INT32) ((x) * (1L<<SCALEBITS) + 0.5))
+
+/* We allocate one big table and divide it up into eight parts, instead of
+ * doing eight alloc_small requests.  This lets us use a single table base
+ * address, which can be held in a register in the inner loops on many
+ * machines (more than can hold all eight addresses, anyway).
+ */
+
+#define R_Y_OFF		0			/* offset to R => Y section */
+#define G_Y_OFF		(1*(MAXJSAMPLE+1))	/* offset to G => Y section */
+#define B_Y_OFF		(2*(MAXJSAMPLE+1))	/* etc. */
+#define R_CB_OFF	(3*(MAXJSAMPLE+1))
+#define G_CB_OFF	(4*(MAXJSAMPLE+1))
+#define B_CB_OFF	(5*(MAXJSAMPLE+1))
+#define R_CR_OFF	B_CB_OFF		/* B=>Cb, R=>Cr are the same */
+#define G_CR_OFF	(6*(MAXJSAMPLE+1))
+#define B_CR_OFF	(7*(MAXJSAMPLE+1))
+#define TABLE_SIZE	(8*(MAXJSAMPLE+1))
+
+
+/*
+ * Initialize for RGB->YCC colorspace conversion.
+ */
+
+METHODDEF(void)
+rgb_ycc_start (j_compress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  INT32 * rgb_ycc_tab;
+  INT32 i;
+
+  /* Allocate and fill in the conversion tables. */
+  cconvert->rgb_ycc_tab = rgb_ycc_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(TABLE_SIZE * SIZEOF(INT32)));
+
+  for (i = 0; i <= MAXJSAMPLE; i++) {
+    rgb_ycc_tab[i+R_Y_OFF] = FIX(0.29900) * i;
+    rgb_ycc_tab[i+G_Y_OFF] = FIX(0.58700) * i;
+    rgb_ycc_tab[i+B_Y_OFF] = FIX(0.11400) * i     + ONE_HALF;
+    rgb_ycc_tab[i+R_CB_OFF] = (-FIX(0.16874)) * i;
+    rgb_ycc_tab[i+G_CB_OFF] = (-FIX(0.33126)) * i;
+    /* We use a rounding fudge-factor of 0.5-epsilon for Cb and Cr.
+     * This ensures that the maximum output will round to MAXJSAMPLE
+     * not MAXJSAMPLE+1, and thus that we don't have to range-limit.
+     */
+    rgb_ycc_tab[i+B_CB_OFF] = FIX(0.50000) * i    + CBCR_OFFSET + ONE_HALF-1;
+/*  B=>Cb and R=>Cr tables are the same
+    rgb_ycc_tab[i+R_CR_OFF] = FIX(0.50000) * i    + CBCR_OFFSET + ONE_HALF-1;
+*/
+    rgb_ycc_tab[i+G_CR_OFF] = (-FIX(0.41869)) * i;
+    rgb_ycc_tab[i+B_CR_OFF] = (-FIX(0.08131)) * i;
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ *
+ * Note that we change from the application's interleaved-pixel format
+ * to our internal noninterleaved, one-plane-per-component format.
+ * The input buffer is therefore three times as wide as the output buffer.
+ *
+ * A starting row offset is provided only for the output buffer.  The caller
+ * can easily adjust the passed input_buf value to accommodate any row
+ * offset required on that side.
+ */
+
+METHODDEF(void)
+rgb_ycc_convert (j_compress_ptr cinfo,
+		 JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		 JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int r, g, b;
+  register INT32 * ctab = cconvert->rgb_ycc_tab;
+  register JSAMPROW inptr;
+  register JSAMPROW outptr0, outptr1, outptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = GETJSAMPLE(inptr[RGB_RED]);
+      g = GETJSAMPLE(inptr[RGB_GREEN]);
+      b = GETJSAMPLE(inptr[RGB_BLUE]);
+      inptr += RGB_PIXELSIZE;
+      /* If the inputs are 0..MAXJSAMPLE, the outputs of these equations
+       * must be too; we do not need an explicit range-limiting operation.
+       * Hence the value being shifted is never negative, and we don't
+       * need the general RIGHT_SHIFT macro.
+       */
+      /* Y */
+      outptr0[col] = (JSAMPLE)
+		((ctab[r+R_Y_OFF] + ctab[g+G_Y_OFF] + ctab[b+B_Y_OFF])
+		 >> SCALEBITS);
+      /* Cb */
+      outptr1[col] = (JSAMPLE)
+		((ctab[r+R_CB_OFF] + ctab[g+G_CB_OFF] + ctab[b+B_CB_OFF])
+		 >> SCALEBITS);
+      /* Cr */
+      outptr2[col] = (JSAMPLE)
+		((ctab[r+R_CR_OFF] + ctab[g+G_CR_OFF] + ctab[b+B_CR_OFF])
+		 >> SCALEBITS);
+    }
+  }
+}
+
+
+/**************** Cases other than RGB -> YCbCr **************/
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles RGB->grayscale conversion, which is the same
+ * as the RGB->Y portion of RGB->YCbCr.
+ * We assume rgb_ycc_start has been called (we only use the Y tables).
+ */
+
+METHODDEF(void)
+rgb_gray_convert (j_compress_ptr cinfo,
+		  JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		  JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int r, g, b;
+  register INT32 * ctab = cconvert->rgb_ycc_tab;
+  register JSAMPROW inptr;
+  register JSAMPROW outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = GETJSAMPLE(inptr[RGB_RED]);
+      g = GETJSAMPLE(inptr[RGB_GREEN]);
+      b = GETJSAMPLE(inptr[RGB_BLUE]);
+      inptr += RGB_PIXELSIZE;
+      /* Y */
+      outptr[col] = (JSAMPLE)
+		((ctab[r+R_Y_OFF] + ctab[g+G_Y_OFF] + ctab[b+B_Y_OFF])
+		 >> SCALEBITS);
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles Adobe-style CMYK->YCCK conversion,
+ * where we convert R=1-C, G=1-M, and B=1-Y to YCbCr using the same
+ * conversion as above, while passing K (black) unchanged.
+ * We assume rgb_ycc_start has been called.
+ */
+
+METHODDEF(void)
+cmyk_ycck_convert (j_compress_ptr cinfo,
+		   JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		   JDIMENSION output_row, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int r, g, b;
+  register INT32 * ctab = cconvert->rgb_ycc_tab;
+  register JSAMPROW inptr;
+  register JSAMPROW outptr0, outptr1, outptr2, outptr3;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr0 = output_buf[0][output_row];
+    outptr1 = output_buf[1][output_row];
+    outptr2 = output_buf[2][output_row];
+    outptr3 = output_buf[3][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      r = MAXJSAMPLE - GETJSAMPLE(inptr[0]);
+      g = MAXJSAMPLE - GETJSAMPLE(inptr[1]);
+      b = MAXJSAMPLE - GETJSAMPLE(inptr[2]);
+      /* K passes through as-is */
+      outptr3[col] = inptr[3];	/* don't need GETJSAMPLE here */
+      inptr += 4;
+      /* If the inputs are 0..MAXJSAMPLE, the outputs of these equations
+       * must be too; we do not need an explicit range-limiting operation.
+       * Hence the value being shifted is never negative, and we don't
+       * need the general RIGHT_SHIFT macro.
+       */
+      /* Y */
+      outptr0[col] = (JSAMPLE)
+		((ctab[r+R_Y_OFF] + ctab[g+G_Y_OFF] + ctab[b+B_Y_OFF])
+		 >> SCALEBITS);
+      /* Cb */
+      outptr1[col] = (JSAMPLE)
+		((ctab[r+R_CB_OFF] + ctab[g+G_CB_OFF] + ctab[b+B_CB_OFF])
+		 >> SCALEBITS);
+      /* Cr */
+      outptr2[col] = (JSAMPLE)
+		((ctab[r+R_CR_OFF] + ctab[g+G_CR_OFF] + ctab[b+B_CR_OFF])
+		 >> SCALEBITS);
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles grayscale output with no conversion.
+ * The source can be either plain grayscale or YCbCr (since Y == gray).
+ */
+
+METHODDEF(void)
+grayscale_convert (j_compress_ptr cinfo,
+		   JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+		   JDIMENSION output_row, int num_rows)
+{
+  register JSAMPROW inptr;
+  register JSAMPROW outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->image_width;
+  int instride = cinfo->input_components;
+
+  while (--num_rows >= 0) {
+    inptr = *input_buf++;
+    outptr = output_buf[0][output_row];
+    output_row++;
+    for (col = 0; col < num_cols; col++) {
+      outptr[col] = inptr[0];	/* don't need GETJSAMPLE() here */
+      inptr += instride;
+    }
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the JPEG colorspace.
+ * This version handles multi-component colorspaces without conversion.
+ * We assume input_components == num_components.
+ */
+
+METHODDEF(void)
+null_convert (j_compress_ptr cinfo,
+	      JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
+	      JDIMENSION output_row, int num_rows)
+{
+  register JSAMPROW inptr;
+  register JSAMPROW outptr;
+  register JDIMENSION col;
+  register int ci;
+  int nc = cinfo->num_components;
+  JDIMENSION num_cols = cinfo->image_width;
+
+  while (--num_rows >= 0) {
+    /* It seems fastest to make a separate pass for each component. */
+    for (ci = 0; ci < nc; ci++) {
+      inptr = *input_buf;
+      outptr = output_buf[ci][output_row];
+      for (col = 0; col < num_cols; col++) {
+	outptr[col] = inptr[ci]; /* don't need GETJSAMPLE() here */
+	inptr += nc;
+      }
+    }
+    input_buf++;
+    output_row++;
+  }
+}
+
+
+/*
+ * Empty method for start_pass.
+ */
+
+METHODDEF(void)
+null_method (j_compress_ptr cinfo)
+{
+  /* no work needed */
+}
+
+
+/*
+ * Module initialization routine for input colorspace conversion.
+ */
+
+GLOBAL(void)
+jinit_color_converter (j_compress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert;
+
+  cconvert = (my_cconvert_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_color_converter));
+  cinfo->cconvert = (struct jpeg_color_converter *) cconvert;
+  /* set start_pass to null method until we find out differently */
+  cconvert->pub.start_pass = null_method;
+
+  /* Make sure input_components agrees with in_color_space */
+  switch (cinfo->in_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->input_components != 1)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_RGB:
+#if RGB_PIXELSIZE != 3
+    if (cinfo->input_components != RGB_PIXELSIZE)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+#endif /* else share code with YCbCr */
+
+  case JCS_YCbCr:
+    if (cinfo->input_components != 3)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  case JCS_CMYK:
+  case JCS_YCCK:
+    if (cinfo->input_components != 4)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+
+  default:			/* JCS_UNKNOWN can be anything */
+    if (cinfo->input_components < 1)
+      ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+    break;
+  }
+
+  /* Check num_components, set conversion method based on requested space */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->num_components != 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_GRAYSCALE)
+      cconvert->pub.color_convert = grayscale_convert;
+    else if (cinfo->in_color_space == JCS_RGB) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub.color_convert = rgb_gray_convert;
+    } else if (cinfo->in_color_space == JCS_YCbCr)
+      cconvert->pub.color_convert = grayscale_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_RGB && RGB_PIXELSIZE == 3)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_YCbCr:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_RGB) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub.color_convert = rgb_ycc_convert;
+    } else if (cinfo->in_color_space == JCS_YCbCr)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_CMYK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_CMYK)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_YCCK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    if (cinfo->in_color_space == JCS_CMYK) {
+      cconvert->pub.start_pass = rgb_ycc_start;
+      cconvert->pub.color_convert = cmyk_ycck_convert;
+    } else if (cinfo->in_color_space == JCS_YCCK)
+      cconvert->pub.color_convert = null_convert;
+    else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  default:			/* allow null conversion of JCS_UNKNOWN */
+    if (cinfo->jpeg_color_space != cinfo->in_color_space ||
+	cinfo->num_components != cinfo->input_components)
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    cconvert->pub.color_convert = null_convert;
+    break;
+  }
+}

Added: trunk/code/jpeg-8c/jcdctmgr.c
===================================================================
--- trunk/code/jpeg-8c/jcdctmgr.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcdctmgr.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,482 @@
+/*
+ * jcdctmgr.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the forward-DCT management logic.
+ * This code selects a particular DCT implementation to be used,
+ * and it performs related housekeeping chores including coefficient
+ * quantization.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+
+/* Private subobject for this module */
+
+typedef struct {
+  struct jpeg_forward_dct pub;	/* public fields */
+
+  /* Pointer to the DCT routine actually in use */
+  forward_DCT_method_ptr do_dct[MAX_COMPONENTS];
+
+  /* The actual post-DCT divisors --- not identical to the quant table
+   * entries, because of scaling (especially for an unnormalized DCT).
+   * Each table is given in normal array order.
+   */
+  DCTELEM * divisors[NUM_QUANT_TBLS];
+
+#ifdef DCT_FLOAT_SUPPORTED
+  /* Same as above for the floating-point case. */
+  float_DCT_method_ptr do_float_dct[MAX_COMPONENTS];
+  FAST_FLOAT * float_divisors[NUM_QUANT_TBLS];
+#endif
+} my_fdct_controller;
+
+typedef my_fdct_controller * my_fdct_ptr;
+
+
+/* The current scaled-DCT routines require ISLOW-style divisor tables,
+ * so be sure to compile that code if either ISLOW or SCALING is requested.
+ */
+#ifdef DCT_ISLOW_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#else
+#ifdef DCT_SCALING_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#endif
+#endif
+
+
+/*
+ * Perform forward DCT on one or more blocks of a component.
+ *
+ * The input samples are taken from the sample_data[] array starting at
+ * position start_row/start_col, and moving to the right for any additional
+ * blocks. The quantized coefficients are returned in coef_blocks[].
+ */
+
+METHODDEF(void)
+forward_DCT (j_compress_ptr cinfo, jpeg_component_info * compptr,
+	     JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+	     JDIMENSION start_row, JDIMENSION start_col,
+	     JDIMENSION num_blocks)
+/* This version is used for integer DCT implementations. */
+{
+  /* This routine is heavily used, so it's worth coding it tightly. */
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  forward_DCT_method_ptr do_dct = fdct->do_dct[compptr->component_index];
+  DCTELEM * divisors = fdct->divisors[compptr->quant_tbl_no];
+  DCTELEM workspace[DCTSIZE2];	/* work area for FDCT subroutine */
+  JDIMENSION bi;
+
+  sample_data += start_row;	/* fold in the vertical offset once */
+
+  for (bi = 0; bi < num_blocks; bi++, start_col += compptr->DCT_h_scaled_size) {
+    /* Perform the DCT */
+    (*do_dct) (workspace, sample_data, start_col);
+
+    /* Quantize/descale the coefficients, and store into coef_blocks[] */
+    { register DCTELEM temp, qval;
+      register int i;
+      register JCOEFPTR output_ptr = coef_blocks[bi];
+
+      for (i = 0; i < DCTSIZE2; i++) {
+	qval = divisors[i];
+	temp = workspace[i];
+	/* Divide the coefficient value by qval, ensuring proper rounding.
+	 * Since C does not specify the direction of rounding for negative
+	 * quotients, we have to force the dividend positive for portability.
+	 *
+	 * In most files, at least half of the output values will be zero
+	 * (at default quantization settings, more like three-quarters...)
+	 * so we should ensure that this case is fast.  On many machines,
+	 * a comparison is enough cheaper than a divide to make a special test
+	 * a win.  Since both inputs will be nonnegative, we need only test
+	 * for a < b to discover whether a/b is 0.
+	 * If your machine's division is fast enough, define FAST_DIVIDE.
+	 */
+#ifdef FAST_DIVIDE
+#define DIVIDE_BY(a,b)	a /= b
+#else
+#define DIVIDE_BY(a,b)	if (a >= b) a /= b; else a = 0
+#endif
+	if (temp < 0) {
+	  temp = -temp;
+	  temp += qval>>1;	/* for rounding */
+	  DIVIDE_BY(temp, qval);
+	  temp = -temp;
+	} else {
+	  temp += qval>>1;	/* for rounding */
+	  DIVIDE_BY(temp, qval);
+	}
+	output_ptr[i] = (JCOEF) temp;
+      }
+    }
+  }
+}
+
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+METHODDEF(void)
+forward_DCT_float (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		   JSAMPARRAY sample_data, JBLOCKROW coef_blocks,
+		   JDIMENSION start_row, JDIMENSION start_col,
+		   JDIMENSION num_blocks)
+/* This version is used for floating-point DCT implementations. */
+{
+  /* This routine is heavily used, so it's worth coding it tightly. */
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  float_DCT_method_ptr do_dct = fdct->do_float_dct[compptr->component_index];
+  FAST_FLOAT * divisors = fdct->float_divisors[compptr->quant_tbl_no];
+  FAST_FLOAT workspace[DCTSIZE2]; /* work area for FDCT subroutine */
+  JDIMENSION bi;
+
+  sample_data += start_row;	/* fold in the vertical offset once */
+
+  for (bi = 0; bi < num_blocks; bi++, start_col += compptr->DCT_h_scaled_size) {
+    /* Perform the DCT */
+    (*do_dct) (workspace, sample_data, start_col);
+
+    /* Quantize/descale the coefficients, and store into coef_blocks[] */
+    { register FAST_FLOAT temp;
+      register int i;
+      register JCOEFPTR output_ptr = coef_blocks[bi];
+
+      for (i = 0; i < DCTSIZE2; i++) {
+	/* Apply the quantization and scaling factor */
+	temp = workspace[i] * divisors[i];
+	/* Round to nearest integer.
+	 * Since C does not specify the direction of rounding for negative
+	 * quotients, we have to force the dividend positive for portability.
+	 * The maximum coefficient size is +-16K (for 12-bit data), so this
+	 * code should work for either 16-bit or 32-bit ints.
+	 */
+	output_ptr[i] = (JCOEF) ((int) (temp + (FAST_FLOAT) 16384.5) - 16384);
+      }
+    }
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */
+
+
+/*
+ * Initialize for a processing pass.
+ * Verify that all referenced Q-tables are present, and set up
+ * the divisor table for each one.
+ * In the current implementation, DCT of all components is done during
+ * the first pass, even if only some components will be output in the
+ * first scan.  Hence all components should be examined here.
+ */
+
+METHODDEF(void)
+start_pass_fdctmgr (j_compress_ptr cinfo)
+{
+  my_fdct_ptr fdct = (my_fdct_ptr) cinfo->fdct;
+  int ci, qtblno, i;
+  jpeg_component_info *compptr;
+  int method = 0;
+  JQUANT_TBL * qtbl;
+  DCTELEM * dtbl;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Select the proper DCT routine for this component's scaling */
+    switch ((compptr->DCT_h_scaled_size << 8) + compptr->DCT_v_scaled_size) {
+#ifdef DCT_SCALING_SUPPORTED
+    case ((1 << 8) + 1):
+      fdct->do_dct[ci] = jpeg_fdct_1x1;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((2 << 8) + 2):
+      fdct->do_dct[ci] = jpeg_fdct_2x2;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((3 << 8) + 3):
+      fdct->do_dct[ci] = jpeg_fdct_3x3;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((4 << 8) + 4):
+      fdct->do_dct[ci] = jpeg_fdct_4x4;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((5 << 8) + 5):
+      fdct->do_dct[ci] = jpeg_fdct_5x5;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((6 << 8) + 6):
+      fdct->do_dct[ci] = jpeg_fdct_6x6;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((7 << 8) + 7):
+      fdct->do_dct[ci] = jpeg_fdct_7x7;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((9 << 8) + 9):
+      fdct->do_dct[ci] = jpeg_fdct_9x9;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((10 << 8) + 10):
+      fdct->do_dct[ci] = jpeg_fdct_10x10;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((11 << 8) + 11):
+      fdct->do_dct[ci] = jpeg_fdct_11x11;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((12 << 8) + 12):
+      fdct->do_dct[ci] = jpeg_fdct_12x12;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((13 << 8) + 13):
+      fdct->do_dct[ci] = jpeg_fdct_13x13;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((14 << 8) + 14):
+      fdct->do_dct[ci] = jpeg_fdct_14x14;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((15 << 8) + 15):
+      fdct->do_dct[ci] = jpeg_fdct_15x15;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((16 << 8) + 16):
+      fdct->do_dct[ci] = jpeg_fdct_16x16;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((16 << 8) + 8):
+      fdct->do_dct[ci] = jpeg_fdct_16x8;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((14 << 8) + 7):
+      fdct->do_dct[ci] = jpeg_fdct_14x7;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((12 << 8) + 6):
+      fdct->do_dct[ci] = jpeg_fdct_12x6;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((10 << 8) + 5):
+      fdct->do_dct[ci] = jpeg_fdct_10x5;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((8 << 8) + 4):
+      fdct->do_dct[ci] = jpeg_fdct_8x4;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((6 << 8) + 3):
+      fdct->do_dct[ci] = jpeg_fdct_6x3;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((4 << 8) + 2):
+      fdct->do_dct[ci] = jpeg_fdct_4x2;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((2 << 8) + 1):
+      fdct->do_dct[ci] = jpeg_fdct_2x1;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((8 << 8) + 16):
+      fdct->do_dct[ci] = jpeg_fdct_8x16;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((7 << 8) + 14):
+      fdct->do_dct[ci] = jpeg_fdct_7x14;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((6 << 8) + 12):
+      fdct->do_dct[ci] = jpeg_fdct_6x12;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((5 << 8) + 10):
+      fdct->do_dct[ci] = jpeg_fdct_5x10;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((4 << 8) + 8):
+      fdct->do_dct[ci] = jpeg_fdct_4x8;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((3 << 8) + 6):
+      fdct->do_dct[ci] = jpeg_fdct_3x6;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((2 << 8) + 4):
+      fdct->do_dct[ci] = jpeg_fdct_2x4;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+    case ((1 << 8) + 2):
+      fdct->do_dct[ci] = jpeg_fdct_1x2;
+      method = JDCT_ISLOW;	/* jfdctint uses islow-style table */
+      break;
+#endif
+    case ((DCTSIZE << 8) + DCTSIZE):
+      switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+      case JDCT_ISLOW:
+	fdct->do_dct[ci] = jpeg_fdct_islow;
+	method = JDCT_ISLOW;
+	break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+      case JDCT_IFAST:
+	fdct->do_dct[ci] = jpeg_fdct_ifast;
+	method = JDCT_IFAST;
+	break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+      case JDCT_FLOAT:
+	fdct->do_float_dct[ci] = jpeg_fdct_float;
+	method = JDCT_FLOAT;
+	break;
+#endif
+      default:
+	ERREXIT(cinfo, JERR_NOT_COMPILED);
+	break;
+      }
+      break;
+    default:
+      ERREXIT2(cinfo, JERR_BAD_DCTSIZE,
+	       compptr->DCT_h_scaled_size, compptr->DCT_v_scaled_size);
+      break;
+    }
+    qtblno = compptr->quant_tbl_no;
+    /* Make sure specified quantization table is present */
+    if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
+	cinfo->quant_tbl_ptrs[qtblno] == NULL)
+      ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
+    qtbl = cinfo->quant_tbl_ptrs[qtblno];
+    /* Compute divisors for this quant table */
+    /* We may do this more than once for same table, but it's not a big deal */
+    switch (method) {
+#ifdef PROVIDE_ISLOW_TABLES
+    case JDCT_ISLOW:
+      /* For LL&M IDCT method, divisors are equal to raw quantization
+       * coefficients multiplied by 8 (to counteract scaling).
+       */
+      if (fdct->divisors[qtblno] == NULL) {
+	fdct->divisors[qtblno] = (DCTELEM *)
+	  (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				      DCTSIZE2 * SIZEOF(DCTELEM));
+      }
+      dtbl = fdct->divisors[qtblno];
+      for (i = 0; i < DCTSIZE2; i++) {
+	dtbl[i] = ((DCTELEM) qtbl->quantval[i]) << 3;
+      }
+      fdct->pub.forward_DCT[ci] = forward_DCT;
+      break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+    case JDCT_IFAST:
+      {
+	/* For AA&N IDCT method, divisors are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * We apply a further scale factor of 8.
+	 */
+#define CONST_BITS 14
+	static const INT16 aanscales[DCTSIZE2] = {
+	  /* precomputed values scaled up by 14 bits */
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
+	  21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
+	  19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
+	   8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
+	   4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
+	};
+	SHIFT_TEMPS
+
+	if (fdct->divisors[qtblno] == NULL) {
+	  fdct->divisors[qtblno] = (DCTELEM *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					DCTSIZE2 * SIZEOF(DCTELEM));
+	}
+	dtbl = fdct->divisors[qtblno];
+	for (i = 0; i < DCTSIZE2; i++) {
+	  dtbl[i] = (DCTELEM)
+	    DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
+				  (INT32) aanscales[i]),
+		    CONST_BITS-3);
+	}
+      }
+      fdct->pub.forward_DCT[ci] = forward_DCT;
+      break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+    case JDCT_FLOAT:
+      {
+	/* For float AA&N IDCT method, divisors are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * We apply a further scale factor of 8.
+	 * What's actually stored is 1/divisor so that the inner loop can
+	 * use a multiplication rather than a division.
+	 */
+	FAST_FLOAT * fdtbl;
+	int row, col;
+	static const double aanscalefactor[DCTSIZE] = {
+	  1.0, 1.387039845, 1.306562965, 1.175875602,
+	  1.0, 0.785694958, 0.541196100, 0.275899379
+	};
+
+	if (fdct->float_divisors[qtblno] == NULL) {
+	  fdct->float_divisors[qtblno] = (FAST_FLOAT *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					DCTSIZE2 * SIZEOF(FAST_FLOAT));
+	}
+	fdtbl = fdct->float_divisors[qtblno];
+	i = 0;
+	for (row = 0; row < DCTSIZE; row++) {
+	  for (col = 0; col < DCTSIZE; col++) {
+	    fdtbl[i] = (FAST_FLOAT)
+	      (1.0 / (((double) qtbl->quantval[i] *
+		       aanscalefactor[row] * aanscalefactor[col] * 8.0)));
+	    i++;
+	  }
+	}
+      }
+      fdct->pub.forward_DCT[ci] = forward_DCT_float;
+      break;
+#endif
+    default:
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+      break;
+    }
+  }
+}
+
+
+/*
+ * Initialize FDCT manager.
+ */
+
+GLOBAL(void)
+jinit_forward_dct (j_compress_ptr cinfo)
+{
+  my_fdct_ptr fdct;
+  int i;
+
+  fdct = (my_fdct_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_fdct_controller));
+  cinfo->fdct = (struct jpeg_forward_dct *) fdct;
+  fdct->pub.start_pass = start_pass_fdctmgr;
+
+  /* Mark divisor tables unallocated */
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    fdct->divisors[i] = NULL;
+#ifdef DCT_FLOAT_SUPPORTED
+    fdct->float_divisors[i] = NULL;
+#endif
+  }
+}

Added: trunk/code/jpeg-8c/jchuff.c
===================================================================
--- trunk/code/jpeg-8c/jchuff.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jchuff.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,1576 @@
+/*
+ * jchuff.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2006-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy encoding routines.
+ * Both sequential and progressive modes are supported in this single module.
+ *
+ * Much of the complexity here has to do with supporting output suspension.
+ * If the data destination module demands suspension, we want to be able to
+ * back up to the start of the current MCU.  To do this, we copy state
+ * variables into local working storage, and update them back to the
+ * permanent JPEG objects only upon successful completion of an MCU.
+ *
+ * We do not support output suspension for the progressive JPEG mode, since
+ * the library currently does not allow multiple-scan files to be written
+ * with output suspension.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* The legal range of a DCT coefficient is
+ *  -1024 .. +1023  for 8-bit data;
+ * -16384 .. +16383 for 12-bit data.
+ * Hence the magnitude should always fit in 10 or 14 bits respectively.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MAX_COEF_BITS 10
+#else
+#define MAX_COEF_BITS 14
+#endif
+
+/* Derived data constructed for each Huffman table */
+
+typedef struct {
+  unsigned int ehufco[256];	/* code for each symbol */
+  char ehufsi[256];		/* length of code for each symbol */
+  /* If no code has been allocated for a symbol S, ehufsi[S] contains 0 */
+} c_derived_tbl;
+
+
+/* Expanded entropy encoder object for Huffman encoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  INT32 put_buffer;		/* current bit-accumulation buffer */
+  int put_bits;			/* # of bits now in it */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+} savable_state;
+
+/* This macro is to work around compilers with missing or broken
+ * structure assignment.  You'll need to fix this code if you have
+ * such a compiler and you change MAX_COMPS_IN_SCAN.
+ */
+
+#ifndef NO_STRUCT_ASSIGN
+#define ASSIGN_STATE(dest,src)  ((dest) = (src))
+#else
+#if MAX_COMPS_IN_SCAN == 4
+#define ASSIGN_STATE(dest,src)  \
+	((dest).put_buffer = (src).put_buffer, \
+	 (dest).put_bits = (src).put_bits, \
+	 (dest).last_dc_val[0] = (src).last_dc_val[0], \
+	 (dest).last_dc_val[1] = (src).last_dc_val[1], \
+	 (dest).last_dc_val[2] = (src).last_dc_val[2], \
+	 (dest).last_dc_val[3] = (src).last_dc_val[3])
+#endif
+#endif
+
+
+typedef struct {
+  struct jpeg_entropy_encoder pub; /* public fields */
+
+  savable_state saved;		/* Bit buffer & DC state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+  int next_restart_num;		/* next restart number to write (0-7) */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  c_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
+  c_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
+
+  /* Statistics tables for optimization */
+  long * dc_count_ptrs[NUM_HUFF_TBLS];
+  long * ac_count_ptrs[NUM_HUFF_TBLS];
+
+  /* Following fields used only in progressive mode */
+
+  /* Mode flag: TRUE for optimization, FALSE for actual data output */
+  boolean gather_statistics;
+
+  /* next_output_byte/free_in_buffer are local copies of cinfo->dest fields.
+   */
+  JOCTET * next_output_byte;	/* => next byte to write in buffer */
+  size_t free_in_buffer;	/* # of byte spaces remaining in buffer */
+  j_compress_ptr cinfo;		/* link to cinfo (needed for dump_buffer) */
+
+  /* Coding status for AC components */
+  int ac_tbl_no;		/* the table number of the single component */
+  unsigned int EOBRUN;		/* run length of EOBs */
+  unsigned int BE;		/* # of buffered correction bits before MCU */
+  char * bit_buffer;		/* buffer for correction bits (1 per char) */
+  /* packing correction bits tightly would save some space but cost time... */
+} huff_entropy_encoder;
+
+typedef huff_entropy_encoder * huff_entropy_ptr;
+
+/* Working state while writing an MCU (sequential mode).
+ * This struct contains all the fields that are needed by subroutines.
+ */
+
+typedef struct {
+  JOCTET * next_output_byte;	/* => next byte to write in buffer */
+  size_t free_in_buffer;	/* # of byte spaces remaining in buffer */
+  savable_state cur;		/* Current bit buffer & DC state */
+  j_compress_ptr cinfo;		/* dump_buffer needs access to this */
+} working_state;
+
+/* MAX_CORR_BITS is the number of bits the AC refinement correction-bit
+ * buffer can hold.  Larger sizes may slightly improve compression, but
+ * 1000 is already well into the realm of overkill.
+ * The minimum safe size is 64 bits.
+ */
+
+#define MAX_CORR_BITS  1000	/* Max # of correction bits I can buffer */
+
+/* IRIGHT_SHIFT is like RIGHT_SHIFT, but works on int rather than INT32.
+ * We assume that int right shift is unsigned if INT32 right shift is,
+ * which should be safe.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS	int ishift_temp;
+#define IRIGHT_SHIFT(x,shft)  \
+	((ishift_temp = (x)) < 0 ? \
+	 (ishift_temp >> (shft)) | ((~0) << (16-(shft))) : \
+	 (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+
+/*
+ * Compute the derived values for a Huffman table.
+ * This routine also performs some validation checks on the table.
+ */
+
+LOCAL(void)
+jpeg_make_c_derived_tbl (j_compress_ptr cinfo, boolean isDC, int tblno,
+			 c_derived_tbl ** pdtbl)
+{
+  JHUFF_TBL *htbl;
+  c_derived_tbl *dtbl;
+  int p, i, l, lastp, si, maxsymbol;
+  char huffsize[257];
+  unsigned int huffcode[257];
+  unsigned int code;
+
+  /* Note that huffsize[] and huffcode[] are filled in code-length order,
+   * paralleling the order of the symbols themselves in htbl->huffval[].
+   */
+
+  /* Find the input Huffman table */
+  if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+  htbl =
+    isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+
+  /* Allocate a workspace if we haven't already done so. */
+  if (*pdtbl == NULL)
+    *pdtbl = (c_derived_tbl *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(c_derived_tbl));
+  dtbl = *pdtbl;
+  
+  /* Figure C.1: make table of Huffman code length for each symbol */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    i = (int) htbl->bits[l];
+    if (i < 0 || p + i > 256)	/* protect against table overrun */
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    while (i--)
+      huffsize[p++] = (char) l;
+  }
+  huffsize[p] = 0;
+  lastp = p;
+  
+  /* Figure C.2: generate the codes themselves */
+  /* We also validate that the counts represent a legal Huffman code tree. */
+
+  code = 0;
+  si = huffsize[0];
+  p = 0;
+  while (huffsize[p]) {
+    while (((int) huffsize[p]) == si) {
+      huffcode[p++] = code;
+      code++;
+    }
+    /* code is now 1 more than the last code used for codelength si; but
+     * it must still fit in si bits, since no code is allowed to be all ones.
+     */
+    if (((INT32) code) >= (((INT32) 1) << si))
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    code <<= 1;
+    si++;
+  }
+  
+  /* Figure C.3: generate encoding tables */
+  /* These are code and size indexed by symbol value */
+
+  /* Set all codeless symbols to have code length 0;
+   * this lets us detect duplicate VAL entries here, and later
+   * allows emit_bits to detect any attempt to emit such symbols.
+   */
+  MEMZERO(dtbl->ehufsi, SIZEOF(dtbl->ehufsi));
+
+  /* This is also a convenient place to check for out-of-range
+   * and duplicated VAL entries.  We allow 0..255 for AC symbols
+   * but only 0..15 for DC.  (We could constrain them further
+   * based on data depth and mode, but this seems enough.)
+   */
+  maxsymbol = isDC ? 15 : 255;
+
+  for (p = 0; p < lastp; p++) {
+    i = htbl->huffval[p];
+    if (i < 0 || i > maxsymbol || dtbl->ehufsi[i])
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    dtbl->ehufco[i] = huffcode[p];
+    dtbl->ehufsi[i] = huffsize[p];
+  }
+}
+
+
+/* Outputting bytes to the file.
+ * NB: these must be called only when actually outputting,
+ * that is, entropy->gather_statistics == FALSE.
+ */
+
+/* Emit a byte, taking 'action' if must suspend. */
+#define emit_byte_s(state,val,action)  \
+	{ *(state)->next_output_byte++ = (JOCTET) (val);  \
+	  if (--(state)->free_in_buffer == 0)  \
+	    if (! dump_buffer_s(state))  \
+	      { action; } }
+
+/* Emit a byte */
+#define emit_byte_e(entropy,val)  \
+	{ *(entropy)->next_output_byte++ = (JOCTET) (val);  \
+	  if (--(entropy)->free_in_buffer == 0)  \
+	    dump_buffer_e(entropy); }
+
+
+LOCAL(boolean)
+dump_buffer_s (working_state * state)
+/* Empty the output buffer; return TRUE if successful, FALSE if must suspend */
+{
+  struct jpeg_destination_mgr * dest = state->cinfo->dest;
+
+  if (! (*dest->empty_output_buffer) (state->cinfo))
+    return FALSE;
+  /* After a successful buffer dump, must reset buffer pointers */
+  state->next_output_byte = dest->next_output_byte;
+  state->free_in_buffer = dest->free_in_buffer;
+  return TRUE;
+}
+
+
+LOCAL(void)
+dump_buffer_e (huff_entropy_ptr entropy)
+/* Empty the output buffer; we do not support suspension in this case. */
+{
+  struct jpeg_destination_mgr * dest = entropy->cinfo->dest;
+
+  if (! (*dest->empty_output_buffer) (entropy->cinfo))
+    ERREXIT(entropy->cinfo, JERR_CANT_SUSPEND);
+  /* After a successful buffer dump, must reset buffer pointers */
+  entropy->next_output_byte = dest->next_output_byte;
+  entropy->free_in_buffer = dest->free_in_buffer;
+}
+
+
+/* Outputting bits to the file */
+
+/* Only the right 24 bits of put_buffer are used; the valid bits are
+ * left-justified in this part.  At most 16 bits can be passed to emit_bits
+ * in one call, and we never retain more than 7 bits in put_buffer
+ * between calls, so 24 bits are sufficient.
+ */
+
+INLINE
+LOCAL(boolean)
+emit_bits_s (working_state * state, unsigned int code, int size)
+/* Emit some bits; return TRUE if successful, FALSE if must suspend */
+{
+  /* This routine is heavily used, so it's worth coding tightly. */
+  register INT32 put_buffer = (INT32) code;
+  register int put_bits = state->cur.put_bits;
+
+  /* if size is 0, caller used an invalid Huffman table entry */
+  if (size == 0)
+    ERREXIT(state->cinfo, JERR_HUFF_MISSING_CODE);
+
+  put_buffer &= (((INT32) 1)<<size) - 1; /* mask off any extra bits in code */
+  
+  put_bits += size;		/* new number of bits in buffer */
+  
+  put_buffer <<= 24 - put_bits; /* align incoming bits */
+
+  put_buffer |= state->cur.put_buffer; /* and merge with old buffer contents */
+  
+  while (put_bits >= 8) {
+    int c = (int) ((put_buffer >> 16) & 0xFF);
+    
+    emit_byte_s(state, c, return FALSE);
+    if (c == 0xFF) {		/* need to stuff a zero byte? */
+      emit_byte_s(state, 0, return FALSE);
+    }
+    put_buffer <<= 8;
+    put_bits -= 8;
+  }
+
+  state->cur.put_buffer = put_buffer; /* update state variables */
+  state->cur.put_bits = put_bits;
+
+  return TRUE;
+}
+
+
+INLINE
+LOCAL(void)
+emit_bits_e (huff_entropy_ptr entropy, unsigned int code, int size)
+/* Emit some bits, unless we are in gather mode */
+{
+  /* This routine is heavily used, so it's worth coding tightly. */
+  register INT32 put_buffer = (INT32) code;
+  register int put_bits = entropy->saved.put_bits;
+
+  /* if size is 0, caller used an invalid Huffman table entry */
+  if (size == 0)
+    ERREXIT(entropy->cinfo, JERR_HUFF_MISSING_CODE);
+
+  if (entropy->gather_statistics)
+    return;			/* do nothing if we're only getting stats */
+
+  put_buffer &= (((INT32) 1)<<size) - 1; /* mask off any extra bits in code */
+  
+  put_bits += size;		/* new number of bits in buffer */
+
+  put_buffer <<= 24 - put_bits; /* align incoming bits */
+
+  /* and merge with old buffer contents */
+  put_buffer |= entropy->saved.put_buffer;
+
+  while (put_bits >= 8) {
+    int c = (int) ((put_buffer >> 16) & 0xFF);
+
+    emit_byte_e(entropy, c);
+    if (c == 0xFF) {		/* need to stuff a zero byte? */
+      emit_byte_e(entropy, 0);
+    }
+    put_buffer <<= 8;
+    put_bits -= 8;
+  }
+
+  entropy->saved.put_buffer = put_buffer; /* update variables */
+  entropy->saved.put_bits = put_bits;
+}
+
+
+LOCAL(boolean)
+flush_bits_s (working_state * state)
+{
+  if (! emit_bits_s(state, 0x7F, 7)) /* fill any partial byte with ones */
+    return FALSE;
+  state->cur.put_buffer = 0;	     /* and reset bit-buffer to empty */
+  state->cur.put_bits = 0;
+  return TRUE;
+}
+
+
+LOCAL(void)
+flush_bits_e (huff_entropy_ptr entropy)
+{
+  emit_bits_e(entropy, 0x7F, 7); /* fill any partial byte with ones */
+  entropy->saved.put_buffer = 0; /* and reset bit-buffer to empty */
+  entropy->saved.put_bits = 0;
+}
+
+
+/*
+ * Emit (or just count) a Huffman symbol.
+ */
+
+INLINE
+LOCAL(void)
+emit_dc_symbol (huff_entropy_ptr entropy, int tbl_no, int symbol)
+{
+  if (entropy->gather_statistics)
+    entropy->dc_count_ptrs[tbl_no][symbol]++;
+  else {
+    c_derived_tbl * tbl = entropy->dc_derived_tbls[tbl_no];
+    emit_bits_e(entropy, tbl->ehufco[symbol], tbl->ehufsi[symbol]);
+  }
+}
+
+
+INLINE
+LOCAL(void)
+emit_ac_symbol (huff_entropy_ptr entropy, int tbl_no, int symbol)
+{
+  if (entropy->gather_statistics)
+    entropy->ac_count_ptrs[tbl_no][symbol]++;
+  else {
+    c_derived_tbl * tbl = entropy->ac_derived_tbls[tbl_no];
+    emit_bits_e(entropy, tbl->ehufco[symbol], tbl->ehufsi[symbol]);
+  }
+}
+
+
+/*
+ * Emit bits from a correction bit buffer.
+ */
+
+LOCAL(void)
+emit_buffered_bits (huff_entropy_ptr entropy, char * bufstart,
+		    unsigned int nbits)
+{
+  if (entropy->gather_statistics)
+    return;			/* no real work */
+
+  while (nbits > 0) {
+    emit_bits_e(entropy, (unsigned int) (*bufstart), 1);
+    bufstart++;
+    nbits--;
+  }
+}
+
+
+/*
+ * Emit any pending EOBRUN symbol.
+ */
+
+LOCAL(void)
+emit_eobrun (huff_entropy_ptr entropy)
+{
+  register int temp, nbits;
+
+  if (entropy->EOBRUN > 0) {	/* if there is any pending EOBRUN */
+    temp = entropy->EOBRUN;
+    nbits = 0;
+    while ((temp >>= 1))
+      nbits++;
+    /* safety check: shouldn't happen given limited correction-bit buffer */
+    if (nbits > 14)
+      ERREXIT(entropy->cinfo, JERR_HUFF_MISSING_CODE);
+
+    emit_ac_symbol(entropy, entropy->ac_tbl_no, nbits << 4);
+    if (nbits)
+      emit_bits_e(entropy, entropy->EOBRUN, nbits);
+
+    entropy->EOBRUN = 0;
+
+    /* Emit any buffered correction bits */
+    emit_buffered_bits(entropy, entropy->bit_buffer, entropy->BE);
+    entropy->BE = 0;
+  }
+}
+
+
+/*
+ * Emit a restart marker & resynchronize predictions.
+ */
+
+LOCAL(boolean)
+emit_restart_s (working_state * state, int restart_num)
+{
+  int ci;
+
+  if (! flush_bits_s(state))
+    return FALSE;
+
+  emit_byte_s(state, 0xFF, return FALSE);
+  emit_byte_s(state, JPEG_RST0 + restart_num, return FALSE);
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < state->cinfo->comps_in_scan; ci++)
+    state->cur.last_dc_val[ci] = 0;
+
+  /* The restart counter is not updated until we successfully write the MCU. */
+
+  return TRUE;
+}
+
+
+LOCAL(void)
+emit_restart_e (huff_entropy_ptr entropy, int restart_num)
+{
+  int ci;
+
+  emit_eobrun(entropy);
+
+  if (! entropy->gather_statistics) {
+    flush_bits_e(entropy);
+    emit_byte_e(entropy, 0xFF);
+    emit_byte_e(entropy, JPEG_RST0 + restart_num);
+  }
+
+  if (entropy->cinfo->Ss == 0) {
+    /* Re-initialize DC predictions to 0 */
+    for (ci = 0; ci < entropy->cinfo->comps_in_scan; ci++)
+      entropy->saved.last_dc_val[ci] = 0;
+  } else {
+    /* Re-initialize all AC-related fields to 0 */
+    entropy->EOBRUN = 0;
+    entropy->BE = 0;
+  }
+}
+
+
+/*
+ * MCU encoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  register int temp, temp2;
+  register int nbits;
+  int blkn, ci;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+  jpeg_component_info * compptr;
+  ISHIFT_TEMPS
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart_e(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Compute the DC value after the required point transform by Al.
+     * This is simply an arithmetic right shift.
+     */
+    temp2 = IRIGHT_SHIFT((int) ((*block)[0]), Al);
+
+    /* DC differences are figured on the point-transformed values. */
+    temp = temp2 - entropy->saved.last_dc_val[ci];
+    entropy->saved.last_dc_val[ci] = temp2;
+
+    /* Encode the DC coefficient difference per section G.1.2.1 */
+    temp2 = temp;
+    if (temp < 0) {
+      temp = -temp;		/* temp is abs value of input */
+      /* For a negative input, want temp2 = bitwise complement of abs(input) */
+      /* This code assumes we are on a two's complement machine */
+      temp2--;
+    }
+    
+    /* Find the number of bits needed for the magnitude of the coefficient */
+    nbits = 0;
+    while (temp) {
+      nbits++;
+      temp >>= 1;
+    }
+    /* Check for out-of-range coefficient values.
+     * Since we're encoding a difference, the range limit is twice as much.
+     */
+    if (nbits > MAX_COEF_BITS+1)
+      ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+    
+    /* Count/emit the Huffman-coded symbol for the number of bits */
+    emit_dc_symbol(entropy, compptr->dc_tbl_no, nbits);
+    
+    /* Emit that number of bits of the value, if positive, */
+    /* or the complement of its magnitude, if negative. */
+    if (nbits)			/* emit_bits rejects calls with size 0 */
+      emit_bits_e(entropy, (unsigned int) temp2, nbits);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_first (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  register int temp, temp2;
+  register int nbits;
+  register int r, k;
+  int Se, Al;
+  const int * natural_order;
+  JBLOCKROW block;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart_e(entropy, entropy->next_restart_num);
+
+  Se = cinfo->Se;
+  Al = cinfo->Al;
+  natural_order = cinfo->natural_order;
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+
+  /* Encode the AC coefficients per section G.1.2.2, fig. G.3 */
+  
+  r = 0;			/* r = run length of zeros */
+   
+  for (k = cinfo->Ss; k <= Se; k++) {
+    if ((temp = (*block)[natural_order[k]]) == 0) {
+      r++;
+      continue;
+    }
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value; so the code is
+     * interwoven with finding the abs value (temp) and output bits (temp2).
+     */
+    if (temp < 0) {
+      temp = -temp;		/* temp is abs value of input */
+      temp >>= Al;		/* apply the point transform */
+      /* For a negative coef, want temp2 = bitwise complement of abs(coef) */
+      temp2 = ~temp;
+    } else {
+      temp >>= Al;		/* apply the point transform */
+      temp2 = temp;
+    }
+    /* Watch out for case that nonzero coef is zero after point transform */
+    if (temp == 0) {
+      r++;
+      continue;
+    }
+
+    /* Emit any pending EOBRUN */
+    if (entropy->EOBRUN > 0)
+      emit_eobrun(entropy);
+    /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+    while (r > 15) {
+      emit_ac_symbol(entropy, entropy->ac_tbl_no, 0xF0);
+      r -= 16;
+    }
+
+    /* Find the number of bits needed for the magnitude of the coefficient */
+    nbits = 1;			/* there must be at least one 1 bit */
+    while ((temp >>= 1))
+      nbits++;
+    /* Check for out-of-range coefficient values */
+    if (nbits > MAX_COEF_BITS)
+      ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+    /* Count/emit Huffman symbol for run length / number of bits */
+    emit_ac_symbol(entropy, entropy->ac_tbl_no, (r << 4) + nbits);
+
+    /* Emit that number of bits of the value, if positive, */
+    /* or the complement of its magnitude, if negative. */
+    emit_bits_e(entropy, (unsigned int) temp2, nbits);
+
+    r = 0;			/* reset zero run length */
+  }
+
+  if (r > 0) {			/* If there are trailing zeroes, */
+    entropy->EOBRUN++;		/* count an EOB */
+    if (entropy->EOBRUN == 0x7FFF)
+      emit_eobrun(entropy);	/* force it out to avoid overflow */
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for DC successive approximation refinement scan.
+ * Note: we assume such scans can be multi-component, although the spec
+ * is not very clear on the point.
+ */
+
+METHODDEF(boolean)
+encode_mcu_DC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  register int temp;
+  int blkn;
+  int Al = cinfo->Al;
+  JBLOCKROW block;
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart_e(entropy, entropy->next_restart_num);
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+
+    /* We simply emit the Al'th bit of the DC coefficient value. */
+    temp = (*block)[0];
+    emit_bits_e(entropy, (unsigned int) (temp >> Al), 1);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU encoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+encode_mcu_AC_refine (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  register int temp;
+  register int r, k;
+  int EOB;
+  char *BR_buffer;
+  unsigned int BR;
+  int Se, Al;
+  const int * natural_order;
+  JBLOCKROW block;
+  int absvalues[DCTSIZE2];
+
+  entropy->next_output_byte = cinfo->dest->next_output_byte;
+  entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval)
+    if (entropy->restarts_to_go == 0)
+      emit_restart_e(entropy, entropy->next_restart_num);
+
+  Se = cinfo->Se;
+  Al = cinfo->Al;
+  natural_order = cinfo->natural_order;
+
+  /* Encode the MCU data block */
+  block = MCU_data[0];
+
+  /* It is convenient to make a pre-pass to determine the transformed
+   * coefficients' absolute values and the EOB position.
+   */
+  EOB = 0;
+  for (k = cinfo->Ss; k <= Se; k++) {
+    temp = (*block)[natural_order[k]];
+    /* We must apply the point transform by Al.  For AC coefficients this
+     * is an integer division with rounding towards 0.  To do this portably
+     * in C, we shift after obtaining the absolute value.
+     */
+    if (temp < 0)
+      temp = -temp;		/* temp is abs value of input */
+    temp >>= Al;		/* apply the point transform */
+    absvalues[k] = temp;	/* save abs value for main pass */
+    if (temp == 1)
+      EOB = k;			/* EOB = index of last newly-nonzero coef */
+  }
+
+  /* Encode the AC coefficients per section G.1.2.3, fig. G.7 */
+  
+  r = 0;			/* r = run length of zeros */
+  BR = 0;			/* BR = count of buffered bits added now */
+  BR_buffer = entropy->bit_buffer + entropy->BE; /* Append bits to buffer */
+
+  for (k = cinfo->Ss; k <= Se; k++) {
+    if ((temp = absvalues[k]) == 0) {
+      r++;
+      continue;
+    }
+
+    /* Emit any required ZRLs, but not if they can be folded into EOB */
+    while (r > 15 && k <= EOB) {
+      /* emit any pending EOBRUN and the BE correction bits */
+      emit_eobrun(entropy);
+      /* Emit ZRL */
+      emit_ac_symbol(entropy, entropy->ac_tbl_no, 0xF0);
+      r -= 16;
+      /* Emit buffered correction bits that must be associated with ZRL */
+      emit_buffered_bits(entropy, BR_buffer, BR);
+      BR_buffer = entropy->bit_buffer; /* BE bits are gone now */
+      BR = 0;
+    }
+
+    /* If the coef was previously nonzero, it only needs a correction bit.
+     * NOTE: a straight translation of the spec's figure G.7 would suggest
+     * that we also need to test r > 15.  But if r > 15, we can only get here
+     * if k > EOB, which implies that this coefficient is not 1.
+     */
+    if (temp > 1) {
+      /* The correction bit is the next bit of the absolute value. */
+      BR_buffer[BR++] = (char) (temp & 1);
+      continue;
+    }
+
+    /* Emit any pending EOBRUN and the BE correction bits */
+    emit_eobrun(entropy);
+
+    /* Count/emit Huffman symbol for run length / number of bits */
+    emit_ac_symbol(entropy, entropy->ac_tbl_no, (r << 4) + 1);
+
+    /* Emit output bit for newly-nonzero coef */
+    temp = ((*block)[natural_order[k]] < 0) ? 0 : 1;
+    emit_bits_e(entropy, (unsigned int) temp, 1);
+
+    /* Emit buffered correction bits that must be associated with this code */
+    emit_buffered_bits(entropy, BR_buffer, BR);
+    BR_buffer = entropy->bit_buffer; /* BE bits are gone now */
+    BR = 0;
+    r = 0;			/* reset zero run length */
+  }
+
+  if (r > 0 || BR > 0) {	/* If there are trailing zeroes, */
+    entropy->EOBRUN++;		/* count an EOB */
+    entropy->BE += BR;		/* concat my correction bits to older ones */
+    /* We force out the EOB if we risk either:
+     * 1. overflow of the EOB counter;
+     * 2. overflow of the correction bit buffer during the next MCU.
+     */
+    if (entropy->EOBRUN == 0x7FFF || entropy->BE > (MAX_CORR_BITS-DCTSIZE2+1))
+      emit_eobrun(entropy);
+  }
+
+  cinfo->dest->next_output_byte = entropy->next_output_byte;
+  cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/* Encode a single block's worth of coefficients */
+
+LOCAL(boolean)
+encode_one_block (working_state * state, JCOEFPTR block, int last_dc_val,
+		  c_derived_tbl *dctbl, c_derived_tbl *actbl)
+{
+  register int temp, temp2;
+  register int nbits;
+  register int k, r, i;
+  int Se = state->cinfo->lim_Se;
+  const int * natural_order = state->cinfo->natural_order;
+
+  /* Encode the DC coefficient difference per section F.1.2.1 */
+
+  temp = temp2 = block[0] - last_dc_val;
+
+  if (temp < 0) {
+    temp = -temp;		/* temp is abs value of input */
+    /* For a negative input, want temp2 = bitwise complement of abs(input) */
+    /* This code assumes we are on a two's complement machine */
+    temp2--;
+  }
+
+  /* Find the number of bits needed for the magnitude of the coefficient */
+  nbits = 0;
+  while (temp) {
+    nbits++;
+    temp >>= 1;
+  }
+  /* Check for out-of-range coefficient values.
+   * Since we're encoding a difference, the range limit is twice as much.
+   */
+  if (nbits > MAX_COEF_BITS+1)
+    ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
+
+  /* Emit the Huffman-coded symbol for the number of bits */
+  if (! emit_bits_s(state, dctbl->ehufco[nbits], dctbl->ehufsi[nbits]))
+    return FALSE;
+
+  /* Emit that number of bits of the value, if positive, */
+  /* or the complement of its magnitude, if negative. */
+  if (nbits)			/* emit_bits rejects calls with size 0 */
+    if (! emit_bits_s(state, (unsigned int) temp2, nbits))
+      return FALSE;
+
+  /* Encode the AC coefficients per section F.1.2.2 */
+
+  r = 0;			/* r = run length of zeros */
+
+  for (k = 1; k <= Se; k++) {
+    if ((temp = block[natural_order[k]]) == 0) {
+      r++;
+    } else {
+      /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+      while (r > 15) {
+	if (! emit_bits_s(state, actbl->ehufco[0xF0], actbl->ehufsi[0xF0]))
+	  return FALSE;
+	r -= 16;
+      }
+
+      temp2 = temp;
+      if (temp < 0) {
+	temp = -temp;		/* temp is abs value of input */
+	/* This code assumes we are on a two's complement machine */
+	temp2--;
+      }
+
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      nbits = 1;		/* there must be at least one 1 bit */
+      while ((temp >>= 1))
+	nbits++;
+      /* Check for out-of-range coefficient values */
+      if (nbits > MAX_COEF_BITS)
+	ERREXIT(state->cinfo, JERR_BAD_DCT_COEF);
+
+      /* Emit Huffman symbol for run length / number of bits */
+      i = (r << 4) + nbits;
+      if (! emit_bits_s(state, actbl->ehufco[i], actbl->ehufsi[i]))
+	return FALSE;
+
+      /* Emit that number of bits of the value, if positive, */
+      /* or the complement of its magnitude, if negative. */
+      if (! emit_bits_s(state, (unsigned int) temp2, nbits))
+	return FALSE;
+
+      r = 0;
+    }
+  }
+
+  /* If the last coef(s) were zero, emit an end-of-block code */
+  if (r > 0)
+    if (! emit_bits_s(state, actbl->ehufco[0], actbl->ehufsi[0]))
+      return FALSE;
+
+  return TRUE;
+}
+
+
+/*
+ * Encode and output one MCU's worth of Huffman-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+encode_mcu_huff (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  working_state state;
+  int blkn, ci;
+  jpeg_component_info * compptr;
+
+  /* Load up working state */
+  state.next_output_byte = cinfo->dest->next_output_byte;
+  state.free_in_buffer = cinfo->dest->free_in_buffer;
+  ASSIGN_STATE(state.cur, entropy->saved);
+  state.cinfo = cinfo;
+
+  /* Emit restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! emit_restart_s(&state, entropy->next_restart_num))
+	return FALSE;
+  }
+
+  /* Encode the MCU data blocks */
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    if (! encode_one_block(&state,
+			   MCU_data[blkn][0], state.cur.last_dc_val[ci],
+			   entropy->dc_derived_tbls[compptr->dc_tbl_no],
+			   entropy->ac_derived_tbls[compptr->ac_tbl_no]))
+      return FALSE;
+    /* Update last_dc_val */
+    state.cur.last_dc_val[ci] = MCU_data[blkn][0][0];
+  }
+
+  /* Completed MCU, so update state */
+  cinfo->dest->next_output_byte = state.next_output_byte;
+  cinfo->dest->free_in_buffer = state.free_in_buffer;
+  ASSIGN_STATE(entropy->saved, state.cur);
+
+  /* Update restart-interval state too */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      entropy->restarts_to_go = cinfo->restart_interval;
+      entropy->next_restart_num++;
+      entropy->next_restart_num &= 7;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Finish up at the end of a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+finish_pass_huff (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  working_state state;
+
+  if (cinfo->progressive_mode) {
+    entropy->next_output_byte = cinfo->dest->next_output_byte;
+    entropy->free_in_buffer = cinfo->dest->free_in_buffer;
+
+    /* Flush out any buffered data */
+    emit_eobrun(entropy);
+    flush_bits_e(entropy);
+
+    cinfo->dest->next_output_byte = entropy->next_output_byte;
+    cinfo->dest->free_in_buffer = entropy->free_in_buffer;
+  } else {
+    /* Load up working state ... flush_bits needs it */
+    state.next_output_byte = cinfo->dest->next_output_byte;
+    state.free_in_buffer = cinfo->dest->free_in_buffer;
+    ASSIGN_STATE(state.cur, entropy->saved);
+    state.cinfo = cinfo;
+
+    /* Flush out the last data */
+    if (! flush_bits_s(&state))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+
+    /* Update state */
+    cinfo->dest->next_output_byte = state.next_output_byte;
+    cinfo->dest->free_in_buffer = state.free_in_buffer;
+    ASSIGN_STATE(entropy->saved, state.cur);
+  }
+}
+
+
+/*
+ * Huffman coding optimization.
+ *
+ * We first scan the supplied data and count the number of uses of each symbol
+ * that is to be Huffman-coded. (This process MUST agree with the code above.)
+ * Then we build a Huffman coding tree for the observed counts.
+ * Symbols which are not needed at all for the particular image are not
+ * assigned any code, which saves space in the DHT marker as well as in
+ * the compressed data.
+ */
+
+
+/* Process a single block's worth of coefficients */
+
+LOCAL(void)
+htest_one_block (j_compress_ptr cinfo, JCOEFPTR block, int last_dc_val,
+		 long dc_counts[], long ac_counts[])
+{
+  register int temp;
+  register int nbits;
+  register int k, r;
+  int Se = cinfo->lim_Se;
+  const int * natural_order = cinfo->natural_order;
+  
+  /* Encode the DC coefficient difference per section F.1.2.1 */
+  
+  temp = block[0] - last_dc_val;
+  if (temp < 0)
+    temp = -temp;
+  
+  /* Find the number of bits needed for the magnitude of the coefficient */
+  nbits = 0;
+  while (temp) {
+    nbits++;
+    temp >>= 1;
+  }
+  /* Check for out-of-range coefficient values.
+   * Since we're encoding a difference, the range limit is twice as much.
+   */
+  if (nbits > MAX_COEF_BITS+1)
+    ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+
+  /* Count the Huffman symbol for the number of bits */
+  dc_counts[nbits]++;
+  
+  /* Encode the AC coefficients per section F.1.2.2 */
+  
+  r = 0;			/* r = run length of zeros */
+  
+  for (k = 1; k <= Se; k++) {
+    if ((temp = block[natural_order[k]]) == 0) {
+      r++;
+    } else {
+      /* if run length > 15, must emit special run-length-16 codes (0xF0) */
+      while (r > 15) {
+	ac_counts[0xF0]++;
+	r -= 16;
+      }
+      
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      if (temp < 0)
+	temp = -temp;
+      
+      /* Find the number of bits needed for the magnitude of the coefficient */
+      nbits = 1;		/* there must be at least one 1 bit */
+      while ((temp >>= 1))
+	nbits++;
+      /* Check for out-of-range coefficient values */
+      if (nbits > MAX_COEF_BITS)
+	ERREXIT(cinfo, JERR_BAD_DCT_COEF);
+      
+      /* Count Huffman symbol for run length / number of bits */
+      ac_counts[(r << 4) + nbits]++;
+      
+      r = 0;
+    }
+  }
+
+  /* If the last coef(s) were zero, emit an end-of-block code */
+  if (r > 0)
+    ac_counts[0]++;
+}
+
+
+/*
+ * Trial-encode one MCU's worth of Huffman-compressed coefficients.
+ * No data is actually output, so no suspension return is possible.
+ */
+
+METHODDEF(boolean)
+encode_mcu_gather (j_compress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int blkn, ci;
+  jpeg_component_info * compptr;
+
+  /* Take care of restart intervals if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0) {
+      /* Re-initialize DC predictions to 0 */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+	entropy->saved.last_dc_val[ci] = 0;
+      /* Update restart state */
+      entropy->restarts_to_go = cinfo->restart_interval;
+    }
+    entropy->restarts_to_go--;
+  }
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+    htest_one_block(cinfo, MCU_data[blkn][0], entropy->saved.last_dc_val[ci],
+		    entropy->dc_count_ptrs[compptr->dc_tbl_no],
+		    entropy->ac_count_ptrs[compptr->ac_tbl_no]);
+    entropy->saved.last_dc_val[ci] = MCU_data[blkn][0][0];
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Generate the best Huffman code table for the given counts, fill htbl.
+ *
+ * The JPEG standard requires that no symbol be assigned a codeword of all
+ * one bits (so that padding bits added at the end of a compressed segment
+ * can't look like a valid code).  Because of the canonical ordering of
+ * codewords, this just means that there must be an unused slot in the
+ * longest codeword length category.  Section K.2 of the JPEG spec suggests
+ * reserving such a slot by pretending that symbol 256 is a valid symbol
+ * with count 1.  In theory that's not optimal; giving it count zero but
+ * including it in the symbol set anyway should give a better Huffman code.
+ * But the theoretically better code actually seems to come out worse in
+ * practice, because it produces more all-ones bytes (which incur stuffed
+ * zero bytes in the final file).  In any case the difference is tiny.
+ *
+ * The JPEG standard requires Huffman codes to be no more than 16 bits long.
+ * If some symbols have a very small but nonzero probability, the Huffman tree
+ * must be adjusted to meet the code length restriction.  We currently use
+ * the adjustment method suggested in JPEG section K.2.  This method is *not*
+ * optimal; it may not choose the best possible limited-length code.  But
+ * typically only very-low-frequency symbols will be given less-than-optimal
+ * lengths, so the code is almost optimal.  Experimental comparisons against
+ * an optimal limited-length-code algorithm indicate that the difference is
+ * microscopic --- usually less than a hundredth of a percent of total size.
+ * So the extra complexity of an optimal algorithm doesn't seem worthwhile.
+ */
+
+LOCAL(void)
+jpeg_gen_optimal_table (j_compress_ptr cinfo, JHUFF_TBL * htbl, long freq[])
+{
+#define MAX_CLEN 32		/* assumed maximum initial code length */
+  UINT8 bits[MAX_CLEN+1];	/* bits[k] = # of symbols with code length k */
+  int codesize[257];		/* codesize[k] = code length of symbol k */
+  int others[257];		/* next symbol in current branch of tree */
+  int c1, c2;
+  int p, i, j;
+  long v;
+
+  /* This algorithm is explained in section K.2 of the JPEG standard */
+
+  MEMZERO(bits, SIZEOF(bits));
+  MEMZERO(codesize, SIZEOF(codesize));
+  for (i = 0; i < 257; i++)
+    others[i] = -1;		/* init links to empty */
+  
+  freq[256] = 1;		/* make sure 256 has a nonzero count */
+  /* Including the pseudo-symbol 256 in the Huffman procedure guarantees
+   * that no real symbol is given code-value of all ones, because 256
+   * will be placed last in the largest codeword category.
+   */
+
+  /* Huffman's basic algorithm to assign optimal code lengths to symbols */
+
+  for (;;) {
+    /* Find the smallest nonzero frequency, set c1 = its symbol */
+    /* In case of ties, take the larger symbol number */
+    c1 = -1;
+    v = 1000000000L;
+    for (i = 0; i <= 256; i++) {
+      if (freq[i] && freq[i] <= v) {
+	v = freq[i];
+	c1 = i;
+      }
+    }
+
+    /* Find the next smallest nonzero frequency, set c2 = its symbol */
+    /* In case of ties, take the larger symbol number */
+    c2 = -1;
+    v = 1000000000L;
+    for (i = 0; i <= 256; i++) {
+      if (freq[i] && freq[i] <= v && i != c1) {
+	v = freq[i];
+	c2 = i;
+      }
+    }
+
+    /* Done if we've merged everything into one frequency */
+    if (c2 < 0)
+      break;
+    
+    /* Else merge the two counts/trees */
+    freq[c1] += freq[c2];
+    freq[c2] = 0;
+
+    /* Increment the codesize of everything in c1's tree branch */
+    codesize[c1]++;
+    while (others[c1] >= 0) {
+      c1 = others[c1];
+      codesize[c1]++;
+    }
+    
+    others[c1] = c2;		/* chain c2 onto c1's tree branch */
+    
+    /* Increment the codesize of everything in c2's tree branch */
+    codesize[c2]++;
+    while (others[c2] >= 0) {
+      c2 = others[c2];
+      codesize[c2]++;
+    }
+  }
+
+  /* Now count the number of symbols of each code length */
+  for (i = 0; i <= 256; i++) {
+    if (codesize[i]) {
+      /* The JPEG standard seems to think that this can't happen, */
+      /* but I'm paranoid... */
+      if (codesize[i] > MAX_CLEN)
+	ERREXIT(cinfo, JERR_HUFF_CLEN_OVERFLOW);
+
+      bits[codesize[i]]++;
+    }
+  }
+
+  /* JPEG doesn't allow symbols with code lengths over 16 bits, so if the pure
+   * Huffman procedure assigned any such lengths, we must adjust the coding.
+   * Here is what the JPEG spec says about how this next bit works:
+   * Since symbols are paired for the longest Huffman code, the symbols are
+   * removed from this length category two at a time.  The prefix for the pair
+   * (which is one bit shorter) is allocated to one of the pair; then,
+   * skipping the BITS entry for that prefix length, a code word from the next
+   * shortest nonzero BITS entry is converted into a prefix for two code words
+   * one bit longer.
+   */
+  
+  for (i = MAX_CLEN; i > 16; i--) {
+    while (bits[i] > 0) {
+      j = i - 2;		/* find length of new prefix to be used */
+      while (bits[j] == 0)
+	j--;
+      
+      bits[i] -= 2;		/* remove two symbols */
+      bits[i-1]++;		/* one goes in this length */
+      bits[j+1] += 2;		/* two new symbols in this length */
+      bits[j]--;		/* symbol of this length is now a prefix */
+    }
+  }
+
+  /* Remove the count for the pseudo-symbol 256 from the largest codelength */
+  while (bits[i] == 0)		/* find largest codelength still in use */
+    i--;
+  bits[i]--;
+  
+  /* Return final symbol counts (only for lengths 0..16) */
+  MEMCOPY(htbl->bits, bits, SIZEOF(htbl->bits));
+  
+  /* Return a list of the symbols sorted by code length */
+  /* It's not real clear to me why we don't need to consider the codelength
+   * changes made above, but the JPEG spec seems to think this works.
+   */
+  p = 0;
+  for (i = 1; i <= MAX_CLEN; i++) {
+    for (j = 0; j <= 255; j++) {
+      if (codesize[j] == i) {
+	htbl->huffval[p] = (UINT8) j;
+	p++;
+      }
+    }
+  }
+
+  /* Set sent_table FALSE so updated table will be written to JPEG file. */
+  htbl->sent_table = FALSE;
+}
+
+
+/*
+ * Finish up a statistics-gathering pass and create the new Huffman tables.
+ */
+
+METHODDEF(void)
+finish_pass_gather (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+  JHUFF_TBL **htblptr;
+  boolean did_dc[NUM_HUFF_TBLS];
+  boolean did_ac[NUM_HUFF_TBLS];
+
+  /* It's important not to apply jpeg_gen_optimal_table more than once
+   * per table, because it clobbers the input frequency counts!
+   */
+  if (cinfo->progressive_mode)
+    /* Flush out buffered data (all we care about is counting the EOB symbol) */
+    emit_eobrun(entropy);
+
+  MEMZERO(did_dc, SIZEOF(did_dc));
+  MEMZERO(did_ac, SIZEOF(did_ac));
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* DC needs no table for refinement scan */
+    if (cinfo->Ss == 0 && cinfo->Ah == 0) {
+      tbl = compptr->dc_tbl_no;
+      if (! did_dc[tbl]) {
+	htblptr = & cinfo->dc_huff_tbl_ptrs[tbl];
+	if (*htblptr == NULL)
+	  *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+	jpeg_gen_optimal_table(cinfo, *htblptr, entropy->dc_count_ptrs[tbl]);
+	did_dc[tbl] = TRUE;
+      }
+    }
+    /* AC needs no table when not present */
+    if (cinfo->Se) {
+      tbl = compptr->ac_tbl_no;
+      if (! did_ac[tbl]) {
+	htblptr = & cinfo->ac_huff_tbl_ptrs[tbl];
+	if (*htblptr == NULL)
+	  *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+	jpeg_gen_optimal_table(cinfo, *htblptr, entropy->ac_count_ptrs[tbl]);
+	did_ac[tbl] = TRUE;
+      }
+    }
+  }
+}
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ * If gather_statistics is TRUE, we do not output anything during the scan,
+ * just count the Huffman symbols used and generate Huffman code tables.
+ */
+
+METHODDEF(void)
+start_pass_huff (j_compress_ptr cinfo, boolean gather_statistics)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+
+  if (gather_statistics)
+    entropy->pub.finish_pass = finish_pass_gather;
+  else
+    entropy->pub.finish_pass = finish_pass_huff;
+
+  if (cinfo->progressive_mode) {
+    entropy->cinfo = cinfo;
+    entropy->gather_statistics = gather_statistics;
+
+    /* We assume jcmaster.c already validated the scan parameters. */
+
+    /* Select execution routine */
+    if (cinfo->Ah == 0) {
+      if (cinfo->Ss == 0)
+	entropy->pub.encode_mcu = encode_mcu_DC_first;
+      else
+	entropy->pub.encode_mcu = encode_mcu_AC_first;
+    } else {
+      if (cinfo->Ss == 0)
+	entropy->pub.encode_mcu = encode_mcu_DC_refine;
+      else {
+	entropy->pub.encode_mcu = encode_mcu_AC_refine;
+	/* AC refinement needs a correction bit buffer */
+	if (entropy->bit_buffer == NULL)
+	  entropy->bit_buffer = (char *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					MAX_CORR_BITS * SIZEOF(char));
+      }
+    }
+
+    /* Initialize AC stuff */
+    entropy->ac_tbl_no = cinfo->cur_comp_info[0]->ac_tbl_no;
+    entropy->EOBRUN = 0;
+    entropy->BE = 0;
+  } else {
+    if (gather_statistics)
+      entropy->pub.encode_mcu = encode_mcu_gather;
+    else
+      entropy->pub.encode_mcu = encode_mcu_huff;
+  }
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* DC needs no table for refinement scan */
+    if (cinfo->Ss == 0 && cinfo->Ah == 0) {
+      tbl = compptr->dc_tbl_no;
+      if (gather_statistics) {
+	/* Check for invalid table index */
+	/* (make_c_derived_tbl does this in the other path) */
+	if (tbl < 0 || tbl >= NUM_HUFF_TBLS)
+	  ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tbl);
+	/* Allocate and zero the statistics tables */
+	/* Note that jpeg_gen_optimal_table expects 257 entries in each table! */
+	if (entropy->dc_count_ptrs[tbl] == NULL)
+	  entropy->dc_count_ptrs[tbl] = (long *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					257 * SIZEOF(long));
+	MEMZERO(entropy->dc_count_ptrs[tbl], 257 * SIZEOF(long));
+      } else {
+	/* Compute derived values for Huffman tables */
+	/* We may do this more than once for a table, but it's not expensive */
+	jpeg_make_c_derived_tbl(cinfo, TRUE, tbl,
+				& entropy->dc_derived_tbls[tbl]);
+      }
+      /* Initialize DC predictions to 0 */
+      entropy->saved.last_dc_val[ci] = 0;
+    }
+    /* AC needs no table when not present */
+    if (cinfo->Se) {
+      tbl = compptr->ac_tbl_no;
+      if (gather_statistics) {
+	if (tbl < 0 || tbl >= NUM_HUFF_TBLS)
+	  ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tbl);
+	if (entropy->ac_count_ptrs[tbl] == NULL)
+	  entropy->ac_count_ptrs[tbl] = (long *)
+	    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+					257 * SIZEOF(long));
+	MEMZERO(entropy->ac_count_ptrs[tbl], 257 * SIZEOF(long));
+      } else {
+	jpeg_make_c_derived_tbl(cinfo, FALSE, tbl,
+				& entropy->ac_derived_tbls[tbl]);
+      }
+    }
+  }
+
+  /* Initialize bit buffer to empty */
+  entropy->saved.put_buffer = 0;
+  entropy->saved.put_bits = 0;
+
+  /* Initialize restart stuff */
+  entropy->restarts_to_go = cinfo->restart_interval;
+  entropy->next_restart_num = 0;
+}
+
+
+/*
+ * Module initialization routine for Huffman entropy encoding.
+ */
+
+GLOBAL(void)
+jinit_huff_encoder (j_compress_ptr cinfo)
+{
+  huff_entropy_ptr entropy;
+  int i;
+
+  entropy = (huff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(huff_entropy_encoder));
+  cinfo->entropy = (struct jpeg_entropy_encoder *) entropy;
+  entropy->pub.start_pass = start_pass_huff;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
+    entropy->dc_count_ptrs[i] = entropy->ac_count_ptrs[i] = NULL;
+  }
+
+  if (cinfo->progressive_mode)
+    entropy->bit_buffer = NULL;	/* needed only in AC refinement scan */
+}

Added: trunk/code/jpeg-8c/jcinit.c
===================================================================
--- trunk/code/jpeg-8c/jcinit.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcinit.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,65 @@
+/*
+ * jcinit.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains initialization logic for the JPEG compressor.
+ * This routine is in charge of selecting the modules to be executed and
+ * making an initialization call to each one.
+ *
+ * Logically, this code belongs in jcmaster.c.  It's split out because
+ * linking this routine implies linking the entire compression library.
+ * For a transcoding-only application, we want to be able to use jcmaster.c
+ * without linking in the whole library.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Master selection of compression modules.
+ * This is done once at the start of processing an image.  We determine
+ * which modules will be used and give them appropriate initialization calls.
+ */
+
+GLOBAL(void)
+jinit_compress_master (j_compress_ptr cinfo)
+{
+  /* Initialize master control (includes parameter checking/processing) */
+  jinit_c_master_control(cinfo, FALSE /* full compression */);
+
+  /* Preprocessing */
+  if (! cinfo->raw_data_in) {
+    jinit_color_converter(cinfo);
+    jinit_downsampler(cinfo);
+    jinit_c_prep_controller(cinfo, FALSE /* never need full buffer here */);
+  }
+  /* Forward DCT */
+  jinit_forward_dct(cinfo);
+  /* Entropy encoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code)
+    jinit_arith_encoder(cinfo);
+  else {
+    jinit_huff_encoder(cinfo);
+  }
+
+  /* Need a full-image coefficient buffer in any multi-pass mode. */
+  jinit_c_coef_controller(cinfo,
+		(boolean) (cinfo->num_scans > 1 || cinfo->optimize_coding));
+  jinit_c_main_controller(cinfo, FALSE /* never need full buffer here */);
+
+  jinit_marker_writer(cinfo);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Write the datastream header (SOI) immediately.
+   * Frame and scan headers are postponed till later.
+   * This lets application insert special markers after the SOI.
+   */
+  (*cinfo->marker->write_file_header) (cinfo);
+}

Added: trunk/code/jpeg-8c/jcmainct.c
===================================================================
--- trunk/code/jpeg-8c/jcmainct.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcmainct.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,293 @@
+/*
+ * jcmainct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the main buffer controller for compression.
+ * The main buffer lies between the pre-processor and the JPEG
+ * compressor proper; it holds downsampled data in the JPEG colorspace.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Note: currently, there is no operating mode in which a full-image buffer
+ * is needed at this step.  If there were, that mode could not be used with
+ * "raw data" input, since this module is bypassed in that case.  However,
+ * we've left the code here for possible use in special applications.
+ */
+#undef FULL_MAIN_BUFFER_SUPPORTED
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_main_controller pub; /* public fields */
+
+  JDIMENSION cur_iMCU_row;	/* number of current iMCU row */
+  JDIMENSION rowgroup_ctr;	/* counts row groups received in iMCU row */
+  boolean suspended;		/* remember if we suspended output */
+  J_BUF_MODE pass_mode;		/* current operating mode */
+
+  /* If using just a strip buffer, this points to the entire set of buffers
+   * (we allocate one for each component).  In the full-image case, this
+   * points to the currently accessible strips of the virtual arrays.
+   */
+  JSAMPARRAY buffer[MAX_COMPONENTS];
+
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+  /* If using full-image storage, this array holds pointers to virtual-array
+   * control blocks for each component.  Unused if not full-image storage.
+   */
+  jvirt_sarray_ptr whole_image[MAX_COMPONENTS];
+#endif
+} my_main_controller;
+
+typedef my_main_controller * my_main_ptr;
+
+
+/* Forward declarations */
+METHODDEF(void) process_data_simple_main
+	JPP((j_compress_ptr cinfo, JSAMPARRAY input_buf,
+	     JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail));
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+METHODDEF(void) process_data_buffer_main
+	JPP((j_compress_ptr cinfo, JSAMPARRAY input_buf,
+	     JDIMENSION *in_row_ctr, JDIMENSION in_rows_avail));
+#endif
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_main (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  /* Do nothing in raw-data mode. */
+  if (cinfo->raw_data_in)
+    return;
+
+  main->cur_iMCU_row = 0;	/* initialize counters */
+  main->rowgroup_ctr = 0;
+  main->suspended = FALSE;
+  main->pass_mode = pass_mode;	/* save mode for use by process_data */
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    if (main->whole_image[0] != NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+    main->pub.process_data = process_data_simple_main;
+    break;
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+  case JBUF_SAVE_SOURCE:
+  case JBUF_CRANK_DEST:
+  case JBUF_SAVE_AND_PASS:
+    if (main->whole_image[0] == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    main->pub.process_data = process_data_buffer_main;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This routine handles the simple pass-through mode,
+ * where we have only a strip buffer.
+ */
+
+METHODDEF(void)
+process_data_simple_main (j_compress_ptr cinfo,
+			  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+			  JDIMENSION in_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  while (main->cur_iMCU_row < cinfo->total_iMCU_rows) {
+    /* Read input data if we haven't filled the main buffer yet */
+    if (main->rowgroup_ctr < (JDIMENSION) cinfo->min_DCT_v_scaled_size)
+      (*cinfo->prep->pre_process_data) (cinfo,
+					input_buf, in_row_ctr, in_rows_avail,
+					main->buffer, &main->rowgroup_ctr,
+					(JDIMENSION) cinfo->min_DCT_v_scaled_size);
+
+    /* If we don't have a full iMCU row buffered, return to application for
+     * more data.  Note that preprocessor will always pad to fill the iMCU row
+     * at the bottom of the image.
+     */
+    if (main->rowgroup_ctr != (JDIMENSION) cinfo->min_DCT_v_scaled_size)
+      return;
+
+    /* Send the completed row to the compressor */
+    if (! (*cinfo->coef->compress_data) (cinfo, main->buffer)) {
+      /* If compressor did not consume the whole row, then we must need to
+       * suspend processing and return to the application.  In this situation
+       * we pretend we didn't yet consume the last input row; otherwise, if
+       * it happened to be the last row of the image, the application would
+       * think we were done.
+       */
+      if (! main->suspended) {
+	(*in_row_ctr)--;
+	main->suspended = TRUE;
+      }
+      return;
+    }
+    /* We did finish the row.  Undo our little suspension hack if a previous
+     * call suspended; then mark the main buffer empty.
+     */
+    if (main->suspended) {
+      (*in_row_ctr)++;
+      main->suspended = FALSE;
+    }
+    main->rowgroup_ctr = 0;
+    main->cur_iMCU_row++;
+  }
+}
+
+
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+
+/*
+ * Process some data.
+ * This routine handles all of the modes that use a full-size buffer.
+ */
+
+METHODDEF(void)
+process_data_buffer_main (j_compress_ptr cinfo,
+			  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+			  JDIMENSION in_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci;
+  jpeg_component_info *compptr;
+  boolean writing = (main->pass_mode != JBUF_CRANK_DEST);
+
+  while (main->cur_iMCU_row < cinfo->total_iMCU_rows) {
+    /* Realign the virtual buffers if at the start of an iMCU row. */
+    if (main->rowgroup_ctr == 0) {
+      for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	   ci++, compptr++) {
+	main->buffer[ci] = (*cinfo->mem->access_virt_sarray)
+	  ((j_common_ptr) cinfo, main->whole_image[ci],
+	   main->cur_iMCU_row * (compptr->v_samp_factor * DCTSIZE),
+	   (JDIMENSION) (compptr->v_samp_factor * DCTSIZE), writing);
+      }
+      /* In a read pass, pretend we just read some source data. */
+      if (! writing) {
+	*in_row_ctr += cinfo->max_v_samp_factor * DCTSIZE;
+	main->rowgroup_ctr = DCTSIZE;
+      }
+    }
+
+    /* If a write pass, read input data until the current iMCU row is full. */
+    /* Note: preprocessor will pad if necessary to fill the last iMCU row. */
+    if (writing) {
+      (*cinfo->prep->pre_process_data) (cinfo,
+					input_buf, in_row_ctr, in_rows_avail,
+					main->buffer, &main->rowgroup_ctr,
+					(JDIMENSION) DCTSIZE);
+      /* Return to application if we need more data to fill the iMCU row. */
+      if (main->rowgroup_ctr < DCTSIZE)
+	return;
+    }
+
+    /* Emit data, unless this is a sink-only pass. */
+    if (main->pass_mode != JBUF_SAVE_SOURCE) {
+      if (! (*cinfo->coef->compress_data) (cinfo, main->buffer)) {
+	/* If compressor did not consume the whole row, then we must need to
+	 * suspend processing and return to the application.  In this situation
+	 * we pretend we didn't yet consume the last input row; otherwise, if
+	 * it happened to be the last row of the image, the application would
+	 * think we were done.
+	 */
+	if (! main->suspended) {
+	  (*in_row_ctr)--;
+	  main->suspended = TRUE;
+	}
+	return;
+      }
+      /* We did finish the row.  Undo our little suspension hack if a previous
+       * call suspended; then mark the main buffer empty.
+       */
+      if (main->suspended) {
+	(*in_row_ctr)++;
+	main->suspended = FALSE;
+      }
+    }
+
+    /* If get here, we are done with this iMCU row.  Mark buffer empty. */
+    main->rowgroup_ctr = 0;
+    main->cur_iMCU_row++;
+  }
+}
+
+#endif /* FULL_MAIN_BUFFER_SUPPORTED */
+
+
+/*
+ * Initialize main buffer controller.
+ */
+
+GLOBAL(void)
+jinit_c_main_controller (j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_main_ptr main;
+  int ci;
+  jpeg_component_info *compptr;
+
+  main = (my_main_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_main_controller));
+  cinfo->main = (struct jpeg_c_main_controller *) main;
+  main->pub.start_pass = start_pass_main;
+
+  /* We don't need to create a buffer in raw-data mode. */
+  if (cinfo->raw_data_in)
+    return;
+
+  /* Create the buffer.  It holds downsampled data, so each component
+   * may be of a different size.
+   */
+  if (need_full_buffer) {
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    /* Allocate a full-image virtual array for each component */
+    /* Note we pad the bottom to a multiple of the iMCU height */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      main->whole_image[ci] = (*cinfo->mem->request_virt_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 compptr->width_in_blocks * compptr->DCT_h_scaled_size,
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor) * DCTSIZE,
+	 (JDIMENSION) (compptr->v_samp_factor * compptr->DCT_v_scaled_size));
+    }
+#else
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif
+  } else {
+#ifdef FULL_MAIN_BUFFER_SUPPORTED
+    main->whole_image[0] = NULL; /* flag for no virtual arrays */
+#endif
+    /* Allocate a strip buffer for each component */
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      main->buffer[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 compptr->width_in_blocks * compptr->DCT_h_scaled_size,
+	 (JDIMENSION) (compptr->v_samp_factor * compptr->DCT_v_scaled_size));
+    }
+  }
+}

Added: trunk/code/jpeg-8c/jcmarker.c
===================================================================
--- trunk/code/jpeg-8c/jcmarker.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcmarker.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,682 @@
+/*
+ * jcmarker.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modified 2003-2010 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to write JPEG datastream markers.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+typedef enum {			/* JPEG marker codes */
+  M_SOF0  = 0xc0,
+  M_SOF1  = 0xc1,
+  M_SOF2  = 0xc2,
+  M_SOF3  = 0xc3,
+  
+  M_SOF5  = 0xc5,
+  M_SOF6  = 0xc6,
+  M_SOF7  = 0xc7,
+  
+  M_JPG   = 0xc8,
+  M_SOF9  = 0xc9,
+  M_SOF10 = 0xca,
+  M_SOF11 = 0xcb,
+  
+  M_SOF13 = 0xcd,
+  M_SOF14 = 0xce,
+  M_SOF15 = 0xcf,
+  
+  M_DHT   = 0xc4,
+  
+  M_DAC   = 0xcc,
+  
+  M_RST0  = 0xd0,
+  M_RST1  = 0xd1,
+  M_RST2  = 0xd2,
+  M_RST3  = 0xd3,
+  M_RST4  = 0xd4,
+  M_RST5  = 0xd5,
+  M_RST6  = 0xd6,
+  M_RST7  = 0xd7,
+  
+  M_SOI   = 0xd8,
+  M_EOI   = 0xd9,
+  M_SOS   = 0xda,
+  M_DQT   = 0xdb,
+  M_DNL   = 0xdc,
+  M_DRI   = 0xdd,
+  M_DHP   = 0xde,
+  M_EXP   = 0xdf,
+  
+  M_APP0  = 0xe0,
+  M_APP1  = 0xe1,
+  M_APP2  = 0xe2,
+  M_APP3  = 0xe3,
+  M_APP4  = 0xe4,
+  M_APP5  = 0xe5,
+  M_APP6  = 0xe6,
+  M_APP7  = 0xe7,
+  M_APP8  = 0xe8,
+  M_APP9  = 0xe9,
+  M_APP10 = 0xea,
+  M_APP11 = 0xeb,
+  M_APP12 = 0xec,
+  M_APP13 = 0xed,
+  M_APP14 = 0xee,
+  M_APP15 = 0xef,
+  
+  M_JPG0  = 0xf0,
+  M_JPG13 = 0xfd,
+  M_COM   = 0xfe,
+  
+  M_TEM   = 0x01,
+  
+  M_ERROR = 0x100
+} JPEG_MARKER;
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_marker_writer pub; /* public fields */
+
+  unsigned int last_restart_interval; /* last DRI value emitted; 0 after SOI */
+} my_marker_writer;
+
+typedef my_marker_writer * my_marker_ptr;
+
+
+/*
+ * Basic output routines.
+ *
+ * Note that we do not support suspension while writing a marker.
+ * Therefore, an application using suspension must ensure that there is
+ * enough buffer space for the initial markers (typ. 600-700 bytes) before
+ * calling jpeg_start_compress, and enough space to write the trailing EOI
+ * (a few bytes) before calling jpeg_finish_compress.  Multipass compression
+ * modes are not supported at all with suspension, so those two are the only
+ * points where markers will be written.
+ */
+
+LOCAL(void)
+emit_byte (j_compress_ptr cinfo, int val)
+/* Emit a byte */
+{
+  struct jpeg_destination_mgr * dest = cinfo->dest;
+
+  *(dest->next_output_byte)++ = (JOCTET) val;
+  if (--dest->free_in_buffer == 0) {
+    if (! (*dest->empty_output_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+  }
+}
+
+
+LOCAL(void)
+emit_marker (j_compress_ptr cinfo, JPEG_MARKER mark)
+/* Emit a marker code */
+{
+  emit_byte(cinfo, 0xFF);
+  emit_byte(cinfo, (int) mark);
+}
+
+
+LOCAL(void)
+emit_2bytes (j_compress_ptr cinfo, int value)
+/* Emit a 2-byte integer; these are always MSB first in JPEG files */
+{
+  emit_byte(cinfo, (value >> 8) & 0xFF);
+  emit_byte(cinfo, value & 0xFF);
+}
+
+
+/*
+ * Routines to write specific marker types.
+ */
+
+LOCAL(int)
+emit_dqt (j_compress_ptr cinfo, int index)
+/* Emit a DQT marker */
+/* Returns the precision used (0 = 8bits, 1 = 16bits) for baseline checking */
+{
+  JQUANT_TBL * qtbl = cinfo->quant_tbl_ptrs[index];
+  int prec;
+  int i;
+
+  if (qtbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, index);
+
+  prec = 0;
+  for (i = 0; i <= cinfo->lim_Se; i++) {
+    if (qtbl->quantval[cinfo->natural_order[i]] > 255)
+      prec = 1;
+  }
+
+  if (! qtbl->sent_table) {
+    emit_marker(cinfo, M_DQT);
+
+    emit_2bytes(cinfo,
+      prec ? cinfo->lim_Se * 2 + 2 + 1 + 2 : cinfo->lim_Se + 1 + 1 + 2);
+
+    emit_byte(cinfo, index + (prec<<4));
+
+    for (i = 0; i <= cinfo->lim_Se; i++) {
+      /* The table entries must be emitted in zigzag order. */
+      unsigned int qval = qtbl->quantval[cinfo->natural_order[i]];
+      if (prec)
+	emit_byte(cinfo, (int) (qval >> 8));
+      emit_byte(cinfo, (int) (qval & 0xFF));
+    }
+
+    qtbl->sent_table = TRUE;
+  }
+
+  return prec;
+}
+
+
+LOCAL(void)
+emit_dht (j_compress_ptr cinfo, int index, boolean is_ac)
+/* Emit a DHT marker */
+{
+  JHUFF_TBL * htbl;
+  int length, i;
+  
+  if (is_ac) {
+    htbl = cinfo->ac_huff_tbl_ptrs[index];
+    index += 0x10;		/* output index has AC bit set */
+  } else {
+    htbl = cinfo->dc_huff_tbl_ptrs[index];
+  }
+
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, index);
+  
+  if (! htbl->sent_table) {
+    emit_marker(cinfo, M_DHT);
+    
+    length = 0;
+    for (i = 1; i <= 16; i++)
+      length += htbl->bits[i];
+    
+    emit_2bytes(cinfo, length + 2 + 1 + 16);
+    emit_byte(cinfo, index);
+    
+    for (i = 1; i <= 16; i++)
+      emit_byte(cinfo, htbl->bits[i]);
+    
+    for (i = 0; i < length; i++)
+      emit_byte(cinfo, htbl->huffval[i]);
+    
+    htbl->sent_table = TRUE;
+  }
+}
+
+
+LOCAL(void)
+emit_dac (j_compress_ptr cinfo)
+/* Emit a DAC marker */
+/* Since the useful info is so small, we want to emit all the tables in */
+/* one DAC marker.  Therefore this routine does its own scan of the table. */
+{
+#ifdef C_ARITH_CODING_SUPPORTED
+  char dc_in_use[NUM_ARITH_TBLS];
+  char ac_in_use[NUM_ARITH_TBLS];
+  int length, i;
+  jpeg_component_info *compptr;
+
+  for (i = 0; i < NUM_ARITH_TBLS; i++)
+    dc_in_use[i] = ac_in_use[i] = 0;
+
+  for (i = 0; i < cinfo->comps_in_scan; i++) {
+    compptr = cinfo->cur_comp_info[i];
+    /* DC needs no table for refinement scan */
+    if (cinfo->Ss == 0 && cinfo->Ah == 0)
+      dc_in_use[compptr->dc_tbl_no] = 1;
+    /* AC needs no table when not present */
+    if (cinfo->Se)
+      ac_in_use[compptr->ac_tbl_no] = 1;
+  }
+
+  length = 0;
+  for (i = 0; i < NUM_ARITH_TBLS; i++)
+    length += dc_in_use[i] + ac_in_use[i];
+
+  if (length) {
+    emit_marker(cinfo, M_DAC);
+
+    emit_2bytes(cinfo, length*2 + 2);
+
+    for (i = 0; i < NUM_ARITH_TBLS; i++) {
+      if (dc_in_use[i]) {
+	emit_byte(cinfo, i);
+	emit_byte(cinfo, cinfo->arith_dc_L[i] + (cinfo->arith_dc_U[i]<<4));
+      }
+      if (ac_in_use[i]) {
+	emit_byte(cinfo, i + 0x10);
+	emit_byte(cinfo, cinfo->arith_ac_K[i]);
+      }
+    }
+  }
+#endif /* C_ARITH_CODING_SUPPORTED */
+}
+
+
+LOCAL(void)
+emit_dri (j_compress_ptr cinfo)
+/* Emit a DRI marker */
+{
+  emit_marker(cinfo, M_DRI);
+  
+  emit_2bytes(cinfo, 4);	/* fixed length */
+
+  emit_2bytes(cinfo, (int) cinfo->restart_interval);
+}
+
+
+LOCAL(void)
+emit_sof (j_compress_ptr cinfo, JPEG_MARKER code)
+/* Emit a SOF marker */
+{
+  int ci;
+  jpeg_component_info *compptr;
+  
+  emit_marker(cinfo, code);
+  
+  emit_2bytes(cinfo, 3 * cinfo->num_components + 2 + 5 + 1); /* length */
+
+  /* Make sure image isn't bigger than SOF field can handle */
+  if ((long) cinfo->jpeg_height > 65535L ||
+      (long) cinfo->jpeg_width > 65535L)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) 65535);
+
+  emit_byte(cinfo, cinfo->data_precision);
+  emit_2bytes(cinfo, (int) cinfo->jpeg_height);
+  emit_2bytes(cinfo, (int) cinfo->jpeg_width);
+
+  emit_byte(cinfo, cinfo->num_components);
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    emit_byte(cinfo, compptr->component_id);
+    emit_byte(cinfo, (compptr->h_samp_factor << 4) + compptr->v_samp_factor);
+    emit_byte(cinfo, compptr->quant_tbl_no);
+  }
+}
+
+
+LOCAL(void)
+emit_sos (j_compress_ptr cinfo)
+/* Emit a SOS marker */
+{
+  int i, td, ta;
+  jpeg_component_info *compptr;
+  
+  emit_marker(cinfo, M_SOS);
+  
+  emit_2bytes(cinfo, 2 * cinfo->comps_in_scan + 2 + 1 + 3); /* length */
+  
+  emit_byte(cinfo, cinfo->comps_in_scan);
+  
+  for (i = 0; i < cinfo->comps_in_scan; i++) {
+    compptr = cinfo->cur_comp_info[i];
+    emit_byte(cinfo, compptr->component_id);
+
+    /* We emit 0 for unused field(s); this is recommended by the P&M text
+     * but does not seem to be specified in the standard.
+     */
+
+    /* DC needs no table for refinement scan */
+    td = cinfo->Ss == 0 && cinfo->Ah == 0 ? compptr->dc_tbl_no : 0;
+    /* AC needs no table when not present */
+    ta = cinfo->Se ? compptr->ac_tbl_no : 0;
+
+    emit_byte(cinfo, (td << 4) + ta);
+  }
+
+  emit_byte(cinfo, cinfo->Ss);
+  emit_byte(cinfo, cinfo->Se);
+  emit_byte(cinfo, (cinfo->Ah << 4) + cinfo->Al);
+}
+
+
+LOCAL(void)
+emit_pseudo_sos (j_compress_ptr cinfo)
+/* Emit a pseudo SOS marker */
+{
+  emit_marker(cinfo, M_SOS);
+  
+  emit_2bytes(cinfo, 2 + 1 + 3); /* length */
+  
+  emit_byte(cinfo, 0); /* Ns */
+
+  emit_byte(cinfo, 0); /* Ss */
+  emit_byte(cinfo, cinfo->block_size * cinfo->block_size - 1); /* Se */
+  emit_byte(cinfo, 0); /* Ah/Al */
+}
+
+
+LOCAL(void)
+emit_jfif_app0 (j_compress_ptr cinfo)
+/* Emit a JFIF-compliant APP0 marker */
+{
+  /*
+   * Length of APP0 block	(2 bytes)
+   * Block ID			(4 bytes - ASCII "JFIF")
+   * Zero byte			(1 byte to terminate the ID string)
+   * Version Major, Minor	(2 bytes - major first)
+   * Units			(1 byte - 0x00 = none, 0x01 = inch, 0x02 = cm)
+   * Xdpu			(2 bytes - dots per unit horizontal)
+   * Ydpu			(2 bytes - dots per unit vertical)
+   * Thumbnail X size		(1 byte)
+   * Thumbnail Y size		(1 byte)
+   */
+  
+  emit_marker(cinfo, M_APP0);
+  
+  emit_2bytes(cinfo, 2 + 4 + 1 + 2 + 1 + 2 + 2 + 1 + 1); /* length */
+
+  emit_byte(cinfo, 0x4A);	/* Identifier: ASCII "JFIF" */
+  emit_byte(cinfo, 0x46);
+  emit_byte(cinfo, 0x49);
+  emit_byte(cinfo, 0x46);
+  emit_byte(cinfo, 0);
+  emit_byte(cinfo, cinfo->JFIF_major_version); /* Version fields */
+  emit_byte(cinfo, cinfo->JFIF_minor_version);
+  emit_byte(cinfo, cinfo->density_unit); /* Pixel size information */
+  emit_2bytes(cinfo, (int) cinfo->X_density);
+  emit_2bytes(cinfo, (int) cinfo->Y_density);
+  emit_byte(cinfo, 0);		/* No thumbnail image */
+  emit_byte(cinfo, 0);
+}
+
+
+LOCAL(void)
+emit_adobe_app14 (j_compress_ptr cinfo)
+/* Emit an Adobe APP14 marker */
+{
+  /*
+   * Length of APP14 block	(2 bytes)
+   * Block ID			(5 bytes - ASCII "Adobe")
+   * Version Number		(2 bytes - currently 100)
+   * Flags0			(2 bytes - currently 0)
+   * Flags1			(2 bytes - currently 0)
+   * Color transform		(1 byte)
+   *
+   * Although Adobe TN 5116 mentions Version = 101, all the Adobe files
+   * now in circulation seem to use Version = 100, so that's what we write.
+   *
+   * We write the color transform byte as 1 if the JPEG color space is
+   * YCbCr, 2 if it's YCCK, 0 otherwise.  Adobe's definition has to do with
+   * whether the encoder performed a transformation, which is pretty useless.
+   */
+  
+  emit_marker(cinfo, M_APP14);
+  
+  emit_2bytes(cinfo, 2 + 5 + 2 + 2 + 2 + 1); /* length */
+
+  emit_byte(cinfo, 0x41);	/* Identifier: ASCII "Adobe" */
+  emit_byte(cinfo, 0x64);
+  emit_byte(cinfo, 0x6F);
+  emit_byte(cinfo, 0x62);
+  emit_byte(cinfo, 0x65);
+  emit_2bytes(cinfo, 100);	/* Version */
+  emit_2bytes(cinfo, 0);	/* Flags0 */
+  emit_2bytes(cinfo, 0);	/* Flags1 */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_YCbCr:
+    emit_byte(cinfo, 1);	/* Color transform = 1 */
+    break;
+  case JCS_YCCK:
+    emit_byte(cinfo, 2);	/* Color transform = 2 */
+    break;
+  default:
+    emit_byte(cinfo, 0);	/* Color transform = 0 */
+    break;
+  }
+}
+
+
+/*
+ * These routines allow writing an arbitrary marker with parameters.
+ * The only intended use is to emit COM or APPn markers after calling
+ * write_file_header and before calling write_frame_header.
+ * Other uses are not guaranteed to produce desirable results.
+ * Counting the parameter bytes properly is the caller's responsibility.
+ */
+
+METHODDEF(void)
+write_marker_header (j_compress_ptr cinfo, int marker, unsigned int datalen)
+/* Emit an arbitrary marker header */
+{
+  if (datalen > (unsigned int) 65533)		/* safety check */
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  emit_marker(cinfo, (JPEG_MARKER) marker);
+
+  emit_2bytes(cinfo, (int) (datalen + 2));	/* total length */
+}
+
+METHODDEF(void)
+write_marker_byte (j_compress_ptr cinfo, int val)
+/* Emit one byte of marker parameters following write_marker_header */
+{
+  emit_byte(cinfo, val);
+}
+
+
+/*
+ * Write datastream header.
+ * This consists of an SOI and optional APPn markers.
+ * We recommend use of the JFIF marker, but not the Adobe marker,
+ * when using YCbCr or grayscale data.  The JFIF marker should NOT
+ * be used for any other JPEG colorspace.  The Adobe marker is helpful
+ * to distinguish RGB, CMYK, and YCCK colorspaces.
+ * Note that an application can write additional header markers after
+ * jpeg_start_compress returns.
+ */
+
+METHODDEF(void)
+write_file_header (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  emit_marker(cinfo, M_SOI);	/* first the SOI */
+
+  /* SOI is defined to reset restart interval to 0 */
+  marker->last_restart_interval = 0;
+
+  if (cinfo->write_JFIF_header)	/* next an optional JFIF APP0 */
+    emit_jfif_app0(cinfo);
+  if (cinfo->write_Adobe_marker) /* next an optional Adobe APP14 */
+    emit_adobe_app14(cinfo);
+}
+
+
+/*
+ * Write frame header.
+ * This consists of DQT and SOFn markers, and a conditional pseudo SOS marker.
+ * Note that we do not emit the SOF until we have emitted the DQT(s).
+ * This avoids compatibility problems with incorrect implementations that
+ * try to error-check the quant table numbers as soon as they see the SOF.
+ */
+
+METHODDEF(void)
+write_frame_header (j_compress_ptr cinfo)
+{
+  int ci, prec;
+  boolean is_baseline;
+  jpeg_component_info *compptr;
+  
+  /* Emit DQT for each quantization table.
+   * Note that emit_dqt() suppresses any duplicate tables.
+   */
+  prec = 0;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    prec += emit_dqt(cinfo, compptr->quant_tbl_no);
+  }
+  /* now prec is nonzero iff there are any 16-bit quant tables. */
+
+  /* Check for a non-baseline specification.
+   * Note we assume that Huffman table numbers won't be changed later.
+   */
+  if (cinfo->arith_code || cinfo->progressive_mode ||
+      cinfo->data_precision != 8 || cinfo->block_size != DCTSIZE) {
+    is_baseline = FALSE;
+  } else {
+    is_baseline = TRUE;
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      if (compptr->dc_tbl_no > 1 || compptr->ac_tbl_no > 1)
+	is_baseline = FALSE;
+    }
+    if (prec && is_baseline) {
+      is_baseline = FALSE;
+      /* If it's baseline except for quantizer size, warn the user */
+      TRACEMS(cinfo, 0, JTRC_16BIT_TABLES);
+    }
+  }
+
+  /* Emit the proper SOF marker */
+  if (cinfo->arith_code) {
+    if (cinfo->progressive_mode)
+      emit_sof(cinfo, M_SOF10); /* SOF code for progressive arithmetic */
+    else
+      emit_sof(cinfo, M_SOF9);  /* SOF code for sequential arithmetic */
+  } else {
+    if (cinfo->progressive_mode)
+      emit_sof(cinfo, M_SOF2);	/* SOF code for progressive Huffman */
+    else if (is_baseline)
+      emit_sof(cinfo, M_SOF0);	/* SOF code for baseline implementation */
+    else
+      emit_sof(cinfo, M_SOF1);	/* SOF code for non-baseline Huffman file */
+  }
+
+  /* Check to emit pseudo SOS marker */
+  if (cinfo->progressive_mode && cinfo->block_size != DCTSIZE)
+    emit_pseudo_sos(cinfo);
+}
+
+
+/*
+ * Write scan header.
+ * This consists of DHT or DAC markers, optional DRI, and SOS.
+ * Compressed data will be written following the SOS.
+ */
+
+METHODDEF(void)
+write_scan_header (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  int i;
+  jpeg_component_info *compptr;
+
+  if (cinfo->arith_code) {
+    /* Emit arith conditioning info.  We may have some duplication
+     * if the file has multiple scans, but it's so small it's hardly
+     * worth worrying about.
+     */
+    emit_dac(cinfo);
+  } else {
+    /* Emit Huffman tables.
+     * Note that emit_dht() suppresses any duplicate tables.
+     */
+    for (i = 0; i < cinfo->comps_in_scan; i++) {
+      compptr = cinfo->cur_comp_info[i];
+      /* DC needs no table for refinement scan */
+      if (cinfo->Ss == 0 && cinfo->Ah == 0)
+	emit_dht(cinfo, compptr->dc_tbl_no, FALSE);
+      /* AC needs no table when not present */
+      if (cinfo->Se)
+	emit_dht(cinfo, compptr->ac_tbl_no, TRUE);
+    }
+  }
+
+  /* Emit DRI if required --- note that DRI value could change for each scan.
+   * We avoid wasting space with unnecessary DRIs, however.
+   */
+  if (cinfo->restart_interval != marker->last_restart_interval) {
+    emit_dri(cinfo);
+    marker->last_restart_interval = cinfo->restart_interval;
+  }
+
+  emit_sos(cinfo);
+}
+
+
+/*
+ * Write datastream trailer.
+ */
+
+METHODDEF(void)
+write_file_trailer (j_compress_ptr cinfo)
+{
+  emit_marker(cinfo, M_EOI);
+}
+
+
+/*
+ * Write an abbreviated table-specification datastream.
+ * This consists of SOI, DQT and DHT tables, and EOI.
+ * Any table that is defined and not marked sent_table = TRUE will be
+ * emitted.  Note that all tables will be marked sent_table = TRUE at exit.
+ */
+
+METHODDEF(void)
+write_tables_only (j_compress_ptr cinfo)
+{
+  int i;
+
+  emit_marker(cinfo, M_SOI);
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++) {
+    if (cinfo->quant_tbl_ptrs[i] != NULL)
+      (void) emit_dqt(cinfo, i);
+  }
+
+  if (! cinfo->arith_code) {
+    for (i = 0; i < NUM_HUFF_TBLS; i++) {
+      if (cinfo->dc_huff_tbl_ptrs[i] != NULL)
+	emit_dht(cinfo, i, FALSE);
+      if (cinfo->ac_huff_tbl_ptrs[i] != NULL)
+	emit_dht(cinfo, i, TRUE);
+    }
+  }
+
+  emit_marker(cinfo, M_EOI);
+}
+
+
+/*
+ * Initialize the marker writer module.
+ */
+
+GLOBAL(void)
+jinit_marker_writer (j_compress_ptr cinfo)
+{
+  my_marker_ptr marker;
+
+  /* Create the subobject */
+  marker = (my_marker_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_marker_writer));
+  cinfo->marker = (struct jpeg_marker_writer *) marker;
+  /* Initialize method pointers */
+  marker->pub.write_file_header = write_file_header;
+  marker->pub.write_frame_header = write_frame_header;
+  marker->pub.write_scan_header = write_scan_header;
+  marker->pub.write_file_trailer = write_file_trailer;
+  marker->pub.write_tables_only = write_tables_only;
+  marker->pub.write_marker_header = write_marker_header;
+  marker->pub.write_marker_byte = write_marker_byte;
+  /* Initialize private state */
+  marker->last_restart_interval = 0;
+}

Added: trunk/code/jpeg-8c/jcmaster.c
===================================================================
--- trunk/code/jpeg-8c/jcmaster.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcmaster.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,858 @@
+/*
+ * jcmaster.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2003-2011 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains master control logic for the JPEG compressor.
+ * These routines are concerned with parameter validation, initial setup,
+ * and inter-pass control (determining the number of passes and the work 
+ * to be done in each pass).
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private state */
+
+typedef enum {
+	main_pass,		/* input data, also do first output step */
+	huff_opt_pass,		/* Huffman code optimization pass */
+	output_pass		/* data output pass */
+} c_pass_type;
+
+typedef struct {
+  struct jpeg_comp_master pub;	/* public fields */
+
+  c_pass_type pass_type;	/* the type of the current pass */
+
+  int pass_number;		/* # of passes completed */
+  int total_passes;		/* total # of passes needed */
+
+  int scan_number;		/* current index in scan_info[] */
+} my_comp_master;
+
+typedef my_comp_master * my_master_ptr;
+
+
+/*
+ * Support routines that do various essential calculations.
+ */
+
+/*
+ * Compute JPEG image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ */
+
+GLOBAL(void)
+jpeg_calc_jpeg_dimensions (j_compress_ptr cinfo)
+/* Do computations that are needed before master selection phase */
+{
+#ifdef DCT_SCALING_SUPPORTED
+
+  /* Sanity check on input image dimensions to prevent overflow in
+   * following calculation.
+   * We do check jpeg_width and jpeg_height in initial_setup below,
+   * but image_width and image_height can come from arbitrary data,
+   * and we need some space for multiplication by block_size.
+   */
+  if (((long) cinfo->image_width >> 24) || ((long) cinfo->image_height >> 24))
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) JPEG_MAX_DIMENSION);
+
+  /* Compute actual JPEG image dimensions and DCT scaling choices. */
+  if (cinfo->scale_num >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/1 scaling */
+    cinfo->jpeg_width = cinfo->image_width * cinfo->block_size;
+    cinfo->jpeg_height = cinfo->image_height * cinfo->block_size;
+    cinfo->min_DCT_h_scaled_size = 1;
+    cinfo->min_DCT_v_scaled_size = 1;
+  } else if (cinfo->scale_num * 2 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/2 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 2L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 2L);
+    cinfo->min_DCT_h_scaled_size = 2;
+    cinfo->min_DCT_v_scaled_size = 2;
+  } else if (cinfo->scale_num * 3 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/3 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 3L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 3L);
+    cinfo->min_DCT_h_scaled_size = 3;
+    cinfo->min_DCT_v_scaled_size = 3;
+  } else if (cinfo->scale_num * 4 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/4 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 4L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 4L);
+    cinfo->min_DCT_h_scaled_size = 4;
+    cinfo->min_DCT_v_scaled_size = 4;
+  } else if (cinfo->scale_num * 5 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/5 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 5L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 5L);
+    cinfo->min_DCT_h_scaled_size = 5;
+    cinfo->min_DCT_v_scaled_size = 5;
+  } else if (cinfo->scale_num * 6 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/6 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 6L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 6L);
+    cinfo->min_DCT_h_scaled_size = 6;
+    cinfo->min_DCT_v_scaled_size = 6;
+  } else if (cinfo->scale_num * 7 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/7 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 7L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 7L);
+    cinfo->min_DCT_h_scaled_size = 7;
+    cinfo->min_DCT_v_scaled_size = 7;
+  } else if (cinfo->scale_num * 8 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/8 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 8L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 8L);
+    cinfo->min_DCT_h_scaled_size = 8;
+    cinfo->min_DCT_v_scaled_size = 8;
+  } else if (cinfo->scale_num * 9 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/9 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 9L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 9L);
+    cinfo->min_DCT_h_scaled_size = 9;
+    cinfo->min_DCT_v_scaled_size = 9;
+  } else if (cinfo->scale_num * 10 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/10 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 10L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 10L);
+    cinfo->min_DCT_h_scaled_size = 10;
+    cinfo->min_DCT_v_scaled_size = 10;
+  } else if (cinfo->scale_num * 11 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/11 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 11L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 11L);
+    cinfo->min_DCT_h_scaled_size = 11;
+    cinfo->min_DCT_v_scaled_size = 11;
+  } else if (cinfo->scale_num * 12 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/12 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 12L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 12L);
+    cinfo->min_DCT_h_scaled_size = 12;
+    cinfo->min_DCT_v_scaled_size = 12;
+  } else if (cinfo->scale_num * 13 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/13 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 13L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 13L);
+    cinfo->min_DCT_h_scaled_size = 13;
+    cinfo->min_DCT_v_scaled_size = 13;
+  } else if (cinfo->scale_num * 14 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/14 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 14L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 14L);
+    cinfo->min_DCT_h_scaled_size = 14;
+    cinfo->min_DCT_v_scaled_size = 14;
+  } else if (cinfo->scale_num * 15 >= cinfo->scale_denom * cinfo->block_size) {
+    /* Provide block_size/15 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 15L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 15L);
+    cinfo->min_DCT_h_scaled_size = 15;
+    cinfo->min_DCT_v_scaled_size = 15;
+  } else {
+    /* Provide block_size/16 scaling */
+    cinfo->jpeg_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * cinfo->block_size, 16L);
+    cinfo->jpeg_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * cinfo->block_size, 16L);
+    cinfo->min_DCT_h_scaled_size = 16;
+    cinfo->min_DCT_v_scaled_size = 16;
+  }
+
+#else /* !DCT_SCALING_SUPPORTED */
+
+  /* Hardwire it to "no scaling" */
+  cinfo->jpeg_width = cinfo->image_width;
+  cinfo->jpeg_height = cinfo->image_height;
+  cinfo->min_DCT_h_scaled_size = DCTSIZE;
+  cinfo->min_DCT_v_scaled_size = DCTSIZE;
+
+#endif /* DCT_SCALING_SUPPORTED */
+}
+
+
+LOCAL(void)
+jpeg_calc_trans_dimensions (j_compress_ptr cinfo)
+{
+  if (cinfo->min_DCT_h_scaled_size != cinfo->min_DCT_v_scaled_size)
+    ERREXIT2(cinfo, JERR_BAD_DCTSIZE,
+	     cinfo->min_DCT_h_scaled_size, cinfo->min_DCT_v_scaled_size);
+
+  cinfo->block_size = cinfo->min_DCT_h_scaled_size;
+}
+
+
+LOCAL(void)
+initial_setup (j_compress_ptr cinfo, boolean transcode_only)
+/* Do computations that are needed before master selection phase */
+{
+  int ci, ssize;
+  jpeg_component_info *compptr;
+  long samplesperrow;
+  JDIMENSION jd_samplesperrow;
+
+  if (transcode_only)
+    jpeg_calc_trans_dimensions(cinfo);
+  else
+    jpeg_calc_jpeg_dimensions(cinfo);
+
+  /* Sanity check on block_size */
+  if (cinfo->block_size < 1 || cinfo->block_size > 16)
+    ERREXIT2(cinfo, JERR_BAD_DCTSIZE, cinfo->block_size, cinfo->block_size);
+
+  /* Derive natural_order from block_size */
+  switch (cinfo->block_size) {
+  case 2: cinfo->natural_order = jpeg_natural_order2; break;
+  case 3: cinfo->natural_order = jpeg_natural_order3; break;
+  case 4: cinfo->natural_order = jpeg_natural_order4; break;
+  case 5: cinfo->natural_order = jpeg_natural_order5; break;
+  case 6: cinfo->natural_order = jpeg_natural_order6; break;
+  case 7: cinfo->natural_order = jpeg_natural_order7; break;
+  default: cinfo->natural_order = jpeg_natural_order; break;
+  }
+
+  /* Derive lim_Se from block_size */
+  cinfo->lim_Se = cinfo->block_size < DCTSIZE ?
+    cinfo->block_size * cinfo->block_size - 1 : DCTSIZE2-1;
+
+  /* Sanity check on image dimensions */
+  if (cinfo->jpeg_height <= 0 || cinfo->jpeg_width <= 0 ||
+      cinfo->num_components <= 0 || cinfo->input_components <= 0)
+    ERREXIT(cinfo, JERR_EMPTY_IMAGE);
+
+  /* Make sure image isn't bigger than I can handle */
+  if ((long) cinfo->jpeg_height > (long) JPEG_MAX_DIMENSION ||
+      (long) cinfo->jpeg_width > (long) JPEG_MAX_DIMENSION)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) JPEG_MAX_DIMENSION);
+
+  /* Width of an input scanline must be representable as JDIMENSION. */
+  samplesperrow = (long) cinfo->image_width * (long) cinfo->input_components;
+  jd_samplesperrow = (JDIMENSION) samplesperrow;
+  if ((long) jd_samplesperrow != samplesperrow)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* For now, precision must match compiled-in value... */
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  /* Check that number of components won't exceed internal array sizes */
+  if (cinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	     MAX_COMPONENTS);
+
+  /* Compute maximum sampling factors; check factor validity */
+  cinfo->max_h_samp_factor = 1;
+  cinfo->max_v_samp_factor = 1;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor<=0 || compptr->h_samp_factor>MAX_SAMP_FACTOR ||
+	compptr->v_samp_factor<=0 || compptr->v_samp_factor>MAX_SAMP_FACTOR)
+      ERREXIT(cinfo, JERR_BAD_SAMPLING);
+    cinfo->max_h_samp_factor = MAX(cinfo->max_h_samp_factor,
+				   compptr->h_samp_factor);
+    cinfo->max_v_samp_factor = MAX(cinfo->max_v_samp_factor,
+				   compptr->v_samp_factor);
+  }
+
+  /* Compute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Fill in the correct component_index value; don't rely on application */
+    compptr->component_index = ci;
+    /* In selecting the actual DCT scaling for each component, we try to
+     * scale down the chroma components via DCT scaling rather than downsampling.
+     * This saves time if the downsampler gets to use 1:1 scaling.
+     * Note this code adapts subsampling ratios which are powers of 2.
+     */
+    ssize = 1;
+#ifdef DCT_SCALING_SUPPORTED
+    while (cinfo->min_DCT_h_scaled_size * ssize <=
+	   (cinfo->do_fancy_downsampling ? DCTSIZE : DCTSIZE / 2) &&
+	   (cinfo->max_h_samp_factor % (compptr->h_samp_factor * ssize * 2)) == 0) {
+      ssize = ssize * 2;
+    }
+#endif
+    compptr->DCT_h_scaled_size = cinfo->min_DCT_h_scaled_size * ssize;
+    ssize = 1;
+#ifdef DCT_SCALING_SUPPORTED
+    while (cinfo->min_DCT_v_scaled_size * ssize <=
+	   (cinfo->do_fancy_downsampling ? DCTSIZE : DCTSIZE / 2) &&
+	   (cinfo->max_v_samp_factor % (compptr->v_samp_factor * ssize * 2)) == 0) {
+      ssize = ssize * 2;
+    }
+#endif
+    compptr->DCT_v_scaled_size = cinfo->min_DCT_v_scaled_size * ssize;
+
+    /* We don't support DCT ratios larger than 2. */
+    if (compptr->DCT_h_scaled_size > compptr->DCT_v_scaled_size * 2)
+	compptr->DCT_h_scaled_size = compptr->DCT_v_scaled_size * 2;
+    else if (compptr->DCT_v_scaled_size > compptr->DCT_h_scaled_size * 2)
+	compptr->DCT_v_scaled_size = compptr->DCT_h_scaled_size * 2;
+
+    /* Size in DCT blocks */
+    compptr->width_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->jpeg_width * (long) compptr->h_samp_factor,
+		    (long) (cinfo->max_h_samp_factor * cinfo->block_size));
+    compptr->height_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->jpeg_height * (long) compptr->v_samp_factor,
+		    (long) (cinfo->max_v_samp_factor * cinfo->block_size));
+    /* Size in samples */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->jpeg_width *
+		    (long) (compptr->h_samp_factor * compptr->DCT_h_scaled_size),
+		    (long) (cinfo->max_h_samp_factor * cinfo->block_size));
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->jpeg_height *
+		    (long) (compptr->v_samp_factor * compptr->DCT_v_scaled_size),
+		    (long) (cinfo->max_v_samp_factor * cinfo->block_size));
+    /* Mark component needed (this flag isn't actually used for compression) */
+    compptr->component_needed = TRUE;
+  }
+
+  /* Compute number of fully interleaved MCU rows (number of times that
+   * main controller will call coefficient controller).
+   */
+  cinfo->total_iMCU_rows = (JDIMENSION)
+    jdiv_round_up((long) cinfo->jpeg_height,
+		  (long) (cinfo->max_v_samp_factor * cinfo->block_size));
+}
+
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+
+LOCAL(void)
+validate_script (j_compress_ptr cinfo)
+/* Verify that the scan script in cinfo->scan_info[] is valid; also
+ * determine whether it uses progressive JPEG, and set cinfo->progressive_mode.
+ */
+{
+  const jpeg_scan_info * scanptr;
+  int scanno, ncomps, ci, coefi, thisi;
+  int Ss, Se, Ah, Al;
+  boolean component_sent[MAX_COMPONENTS];
+#ifdef C_PROGRESSIVE_SUPPORTED
+  int * last_bitpos_ptr;
+  int last_bitpos[MAX_COMPONENTS][DCTSIZE2];
+  /* -1 until that coefficient has been seen; then last Al for it */
+#endif
+
+  if (cinfo->num_scans <= 0)
+    ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, 0);
+
+  /* For sequential JPEG, all scans must have Ss=0, Se=DCTSIZE2-1;
+   * for progressive JPEG, no scan can have this.
+   */
+  scanptr = cinfo->scan_info;
+  if (scanptr->Ss != 0 || scanptr->Se != DCTSIZE2-1) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+    cinfo->progressive_mode = TRUE;
+    last_bitpos_ptr = & last_bitpos[0][0];
+    for (ci = 0; ci < cinfo->num_components; ci++) 
+      for (coefi = 0; coefi < DCTSIZE2; coefi++)
+	*last_bitpos_ptr++ = -1;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    cinfo->progressive_mode = FALSE;
+    for (ci = 0; ci < cinfo->num_components; ci++) 
+      component_sent[ci] = FALSE;
+  }
+
+  for (scanno = 1; scanno <= cinfo->num_scans; scanptr++, scanno++) {
+    /* Validate component indexes */
+    ncomps = scanptr->comps_in_scan;
+    if (ncomps <= 0 || ncomps > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, ncomps, MAX_COMPS_IN_SCAN);
+    for (ci = 0; ci < ncomps; ci++) {
+      thisi = scanptr->component_index[ci];
+      if (thisi < 0 || thisi >= cinfo->num_components)
+	ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+      /* Components must appear in SOF order within each scan */
+      if (ci > 0 && thisi <= scanptr->component_index[ci-1])
+	ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+    }
+    /* Validate progression parameters */
+    Ss = scanptr->Ss;
+    Se = scanptr->Se;
+    Ah = scanptr->Ah;
+    Al = scanptr->Al;
+    if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+      /* The JPEG spec simply gives the ranges 0..13 for Ah and Al, but that
+       * seems wrong: the upper bound ought to depend on data precision.
+       * Perhaps they really meant 0..N+1 for N-bit precision.
+       * Here we allow 0..10 for 8-bit data; Al larger than 10 results in
+       * out-of-range reconstructed DC values during the first DC scan,
+       * which might cause problems for some decoders.
+       */
+#if BITS_IN_JSAMPLE == 8
+#define MAX_AH_AL 10
+#else
+#define MAX_AH_AL 13
+#endif
+      if (Ss < 0 || Ss >= DCTSIZE2 || Se < Ss || Se >= DCTSIZE2 ||
+	  Ah < 0 || Ah > MAX_AH_AL || Al < 0 || Al > MAX_AH_AL)
+	ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      if (Ss == 0) {
+	if (Se != 0)		/* DC and AC together not OK */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      } else {
+	if (ncomps != 1)	/* AC scans must be for only one component */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      }
+      for (ci = 0; ci < ncomps; ci++) {
+	last_bitpos_ptr = & last_bitpos[scanptr->component_index[ci]][0];
+	if (Ss != 0 && last_bitpos_ptr[0] < 0) /* AC without prior DC scan */
+	  ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	for (coefi = Ss; coefi <= Se; coefi++) {
+	  if (last_bitpos_ptr[coefi] < 0) {
+	    /* first scan of this coefficient */
+	    if (Ah != 0)
+	      ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	  } else {
+	    /* not first scan */
+	    if (Ah != last_bitpos_ptr[coefi] || Al != Ah-1)
+	      ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+	  }
+	  last_bitpos_ptr[coefi] = Al;
+	}
+      }
+#endif
+    } else {
+      /* For sequential JPEG, all progression parameters must be these: */
+      if (Ss != 0 || Se != DCTSIZE2-1 || Ah != 0 || Al != 0)
+	ERREXIT1(cinfo, JERR_BAD_PROG_SCRIPT, scanno);
+      /* Make sure components are not sent twice */
+      for (ci = 0; ci < ncomps; ci++) {
+	thisi = scanptr->component_index[ci];
+	if (component_sent[thisi])
+	  ERREXIT1(cinfo, JERR_BAD_SCAN_SCRIPT, scanno);
+	component_sent[thisi] = TRUE;
+      }
+    }
+  }
+
+  /* Now verify that everything got sent. */
+  if (cinfo->progressive_mode) {
+#ifdef C_PROGRESSIVE_SUPPORTED
+    /* For progressive mode, we only check that at least some DC data
+     * got sent for each component; the spec does not require that all bits
+     * of all coefficients be transmitted.  Would it be wiser to enforce
+     * transmission of all coefficient bits??
+     */
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      if (last_bitpos[ci][0] < 0)
+	ERREXIT(cinfo, JERR_MISSING_DATA);
+    }
+#endif
+  } else {
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      if (! component_sent[ci])
+	ERREXIT(cinfo, JERR_MISSING_DATA);
+    }
+  }
+}
+
+
+LOCAL(void)
+reduce_script (j_compress_ptr cinfo)
+/* Adapt scan script for use with reduced block size;
+ * assume that script has been validated before.
+ */
+{
+  jpeg_scan_info * scanptr;
+  int idxout, idxin;
+
+  /* Circumvent const declaration for this function */
+  scanptr = (jpeg_scan_info *) cinfo->scan_info;
+  idxout = 0;
+
+  for (idxin = 0; idxin < cinfo->num_scans; idxin++) {
+    /* After skipping, idxout becomes smaller than idxin */
+    if (idxin != idxout)
+      /* Copy rest of data;
+       * note we stay in given chunk of allocated memory.
+       */
+      scanptr[idxout] = scanptr[idxin];
+    if (scanptr[idxout].Ss > cinfo->lim_Se)
+      /* Entire scan out of range - skip this entry */
+      continue;
+    if (scanptr[idxout].Se > cinfo->lim_Se)
+      /* Limit scan to end of block */
+      scanptr[idxout].Se = cinfo->lim_Se;
+    idxout++;
+  }
+
+  cinfo->num_scans = idxout;
+}
+
+#endif /* C_MULTISCAN_FILES_SUPPORTED */
+
+
+LOCAL(void)
+select_scan_parameters (j_compress_ptr cinfo)
+/* Set up the scan parameters for the current scan */
+{
+  int ci;
+
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+  if (cinfo->scan_info != NULL) {
+    /* Prepare for current scan --- the script is already validated */
+    my_master_ptr master = (my_master_ptr) cinfo->master;
+    const jpeg_scan_info * scanptr = cinfo->scan_info + master->scan_number;
+
+    cinfo->comps_in_scan = scanptr->comps_in_scan;
+    for (ci = 0; ci < scanptr->comps_in_scan; ci++) {
+      cinfo->cur_comp_info[ci] =
+	&cinfo->comp_info[scanptr->component_index[ci]];
+    }
+    if (cinfo->progressive_mode) {
+      cinfo->Ss = scanptr->Ss;
+      cinfo->Se = scanptr->Se;
+      cinfo->Ah = scanptr->Ah;
+      cinfo->Al = scanptr->Al;
+      return;
+    }
+  }
+  else
+#endif
+  {
+    /* Prepare for single sequential-JPEG scan containing all components */
+    if (cinfo->num_components > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	       MAX_COMPS_IN_SCAN);
+    cinfo->comps_in_scan = cinfo->num_components;
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      cinfo->cur_comp_info[ci] = &cinfo->comp_info[ci];
+    }
+  }
+  cinfo->Ss = 0;
+  cinfo->Se = cinfo->block_size * cinfo->block_size - 1;
+  cinfo->Ah = 0;
+  cinfo->Al = 0;
+}
+
+
+LOCAL(void)
+per_scan_setup (j_compress_ptr cinfo)
+/* Do computations that are needed before processing a JPEG scan */
+/* cinfo->comps_in_scan and cinfo->cur_comp_info[] are already set */
+{
+  int ci, mcublks, tmp;
+  jpeg_component_info *compptr;
+  
+  if (cinfo->comps_in_scan == 1) {
+    
+    /* Noninterleaved (single-component) scan */
+    compptr = cinfo->cur_comp_info[0];
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = compptr->width_in_blocks;
+    cinfo->MCU_rows_in_scan = compptr->height_in_blocks;
+    
+    /* For noninterleaved scan, always one block per MCU */
+    compptr->MCU_width = 1;
+    compptr->MCU_height = 1;
+    compptr->MCU_blocks = 1;
+    compptr->MCU_sample_width = compptr->DCT_h_scaled_size;
+    compptr->last_col_width = 1;
+    /* For noninterleaved scans, it is convenient to define last_row_height
+     * as the number of block rows present in the last iMCU row.
+     */
+    tmp = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+    if (tmp == 0) tmp = compptr->v_samp_factor;
+    compptr->last_row_height = tmp;
+    
+    /* Prepare array describing MCU composition */
+    cinfo->blocks_in_MCU = 1;
+    cinfo->MCU_membership[0] = 0;
+    
+  } else {
+    
+    /* Interleaved (multi-component) scan */
+    if (cinfo->comps_in_scan <= 0 || cinfo->comps_in_scan > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->comps_in_scan,
+	       MAX_COMPS_IN_SCAN);
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = (JDIMENSION)
+      jdiv_round_up((long) cinfo->jpeg_width,
+		    (long) (cinfo->max_h_samp_factor * cinfo->block_size));
+    cinfo->MCU_rows_in_scan = (JDIMENSION)
+      jdiv_round_up((long) cinfo->jpeg_height,
+		    (long) (cinfo->max_v_samp_factor * cinfo->block_size));
+    
+    cinfo->blocks_in_MCU = 0;
+    
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Sampling factors give # of blocks of component in each MCU */
+      compptr->MCU_width = compptr->h_samp_factor;
+      compptr->MCU_height = compptr->v_samp_factor;
+      compptr->MCU_blocks = compptr->MCU_width * compptr->MCU_height;
+      compptr->MCU_sample_width = compptr->MCU_width * compptr->DCT_h_scaled_size;
+      /* Figure number of non-dummy blocks in last MCU column & row */
+      tmp = (int) (compptr->width_in_blocks % compptr->MCU_width);
+      if (tmp == 0) tmp = compptr->MCU_width;
+      compptr->last_col_width = tmp;
+      tmp = (int) (compptr->height_in_blocks % compptr->MCU_height);
+      if (tmp == 0) tmp = compptr->MCU_height;
+      compptr->last_row_height = tmp;
+      /* Prepare array describing MCU composition */
+      mcublks = compptr->MCU_blocks;
+      if (cinfo->blocks_in_MCU + mcublks > C_MAX_BLOCKS_IN_MCU)
+	ERREXIT(cinfo, JERR_BAD_MCU_SIZE);
+      while (mcublks-- > 0) {
+	cinfo->MCU_membership[cinfo->blocks_in_MCU++] = ci;
+      }
+    }
+    
+  }
+
+  /* Convert restart specified in rows to actual MCU count. */
+  /* Note that count must fit in 16 bits, so we provide limiting. */
+  if (cinfo->restart_in_rows > 0) {
+    long nominal = (long) cinfo->restart_in_rows * (long) cinfo->MCUs_per_row;
+    cinfo->restart_interval = (unsigned int) MIN(nominal, 65535L);
+  }
+}
+
+
+/*
+ * Per-pass setup.
+ * This is called at the beginning of each pass.  We determine which modules
+ * will be active during this pass and give them appropriate start_pass calls.
+ * We also set is_last_pass to indicate whether any more passes will be
+ * required.
+ */
+
+METHODDEF(void)
+prepare_for_pass (j_compress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  switch (master->pass_type) {
+  case main_pass:
+    /* Initial pass: will collect input data, and do either Huffman
+     * optimization or data output for the first scan.
+     */
+    select_scan_parameters(cinfo);
+    per_scan_setup(cinfo);
+    if (! cinfo->raw_data_in) {
+      (*cinfo->cconvert->start_pass) (cinfo);
+      (*cinfo->downsample->start_pass) (cinfo);
+      (*cinfo->prep->start_pass) (cinfo, JBUF_PASS_THRU);
+    }
+    (*cinfo->fdct->start_pass) (cinfo);
+    (*cinfo->entropy->start_pass) (cinfo, cinfo->optimize_coding);
+    (*cinfo->coef->start_pass) (cinfo,
+				(master->total_passes > 1 ?
+				 JBUF_SAVE_AND_PASS : JBUF_PASS_THRU));
+    (*cinfo->main->start_pass) (cinfo, JBUF_PASS_THRU);
+    if (cinfo->optimize_coding) {
+      /* No immediate data output; postpone writing frame/scan headers */
+      master->pub.call_pass_startup = FALSE;
+    } else {
+      /* Will write frame/scan headers at first jpeg_write_scanlines call */
+      master->pub.call_pass_startup = TRUE;
+    }
+    break;
+#ifdef ENTROPY_OPT_SUPPORTED
+  case huff_opt_pass:
+    /* Do Huffman optimization for a scan after the first one. */
+    select_scan_parameters(cinfo);
+    per_scan_setup(cinfo);
+    if (cinfo->Ss != 0 || cinfo->Ah == 0) {
+      (*cinfo->entropy->start_pass) (cinfo, TRUE);
+      (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
+      master->pub.call_pass_startup = FALSE;
+      break;
+    }
+    /* Special case: Huffman DC refinement scans need no Huffman table
+     * and therefore we can skip the optimization pass for them.
+     */
+    master->pass_type = output_pass;
+    master->pass_number++;
+    /*FALLTHROUGH*/
+#endif
+  case output_pass:
+    /* Do a data-output pass. */
+    /* We need not repeat per-scan setup if prior optimization pass did it. */
+    if (! cinfo->optimize_coding) {
+      select_scan_parameters(cinfo);
+      per_scan_setup(cinfo);
+    }
+    (*cinfo->entropy->start_pass) (cinfo, FALSE);
+    (*cinfo->coef->start_pass) (cinfo, JBUF_CRANK_DEST);
+    /* We emit frame/scan headers now */
+    if (master->scan_number == 0)
+      (*cinfo->marker->write_frame_header) (cinfo);
+    (*cinfo->marker->write_scan_header) (cinfo);
+    master->pub.call_pass_startup = FALSE;
+    break;
+  default:
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+  }
+
+  master->pub.is_last_pass = (master->pass_number == master->total_passes-1);
+
+  /* Set up progress monitor's pass info if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->completed_passes = master->pass_number;
+    cinfo->progress->total_passes = master->total_passes;
+  }
+}
+
+
+/*
+ * Special start-of-pass hook.
+ * This is called by jpeg_write_scanlines if call_pass_startup is TRUE.
+ * In single-pass processing, we need this hook because we don't want to
+ * write frame/scan headers during jpeg_start_compress; we want to let the
+ * application write COM markers etc. between jpeg_start_compress and the
+ * jpeg_write_scanlines loop.
+ * In multi-pass processing, this routine is not used.
+ */
+
+METHODDEF(void)
+pass_startup (j_compress_ptr cinfo)
+{
+  cinfo->master->call_pass_startup = FALSE; /* reset flag so call only once */
+
+  (*cinfo->marker->write_frame_header) (cinfo);
+  (*cinfo->marker->write_scan_header) (cinfo);
+}
+
+
+/*
+ * Finish up at end of pass.
+ */
+
+METHODDEF(void)
+finish_pass_master (j_compress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  /* The entropy coder always needs an end-of-pass call,
+   * either to analyze statistics or to flush its output buffer.
+   */
+  (*cinfo->entropy->finish_pass) (cinfo);
+
+  /* Update state for next pass */
+  switch (master->pass_type) {
+  case main_pass:
+    /* next pass is either output of scan 0 (after optimization)
+     * or output of scan 1 (if no optimization).
+     */
+    master->pass_type = output_pass;
+    if (! cinfo->optimize_coding)
+      master->scan_number++;
+    break;
+  case huff_opt_pass:
+    /* next pass is always output of current scan */
+    master->pass_type = output_pass;
+    break;
+  case output_pass:
+    /* next pass is either optimization or output of next scan */
+    if (cinfo->optimize_coding)
+      master->pass_type = huff_opt_pass;
+    master->scan_number++;
+    break;
+  }
+
+  master->pass_number++;
+}
+
+
+/*
+ * Initialize master compression control.
+ */
+
+GLOBAL(void)
+jinit_c_master_control (j_compress_ptr cinfo, boolean transcode_only)
+{
+  my_master_ptr master;
+
+  master = (my_master_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(my_comp_master));
+  cinfo->master = (struct jpeg_comp_master *) master;
+  master->pub.prepare_for_pass = prepare_for_pass;
+  master->pub.pass_startup = pass_startup;
+  master->pub.finish_pass = finish_pass_master;
+  master->pub.is_last_pass = FALSE;
+
+  /* Validate parameters, determine derived values */
+  initial_setup(cinfo, transcode_only);
+
+  if (cinfo->scan_info != NULL) {
+#ifdef C_MULTISCAN_FILES_SUPPORTED
+    validate_script(cinfo);
+    if (cinfo->block_size < DCTSIZE)
+      reduce_script(cinfo);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    cinfo->progressive_mode = FALSE;
+    cinfo->num_scans = 1;
+  }
+
+  if ((cinfo->progressive_mode || cinfo->block_size < DCTSIZE) &&
+      !cinfo->arith_code)			/*  TEMPORARY HACK ??? */
+    /* assume default tables no good for progressive or downscale mode */
+    cinfo->optimize_coding = TRUE;
+
+  /* Initialize my private state */
+  if (transcode_only) {
+    /* no main pass in transcoding */
+    if (cinfo->optimize_coding)
+      master->pass_type = huff_opt_pass;
+    else
+      master->pass_type = output_pass;
+  } else {
+    /* for normal compression, first pass is always this type: */
+    master->pass_type = main_pass;
+  }
+  master->scan_number = 0;
+  master->pass_number = 0;
+  if (cinfo->optimize_coding)
+    master->total_passes = cinfo->num_scans * 2;
+  else
+    master->total_passes = cinfo->num_scans;
+}

Added: trunk/code/jpeg-8c/jcomapi.c
===================================================================
--- trunk/code/jpeg-8c/jcomapi.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcomapi.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,106 @@
+/*
+ * jcomapi.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface routines that are used for both
+ * compression and decompression.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Abort processing of a JPEG compression or decompression operation,
+ * but don't destroy the object itself.
+ *
+ * For this, we merely clean up all the nonpermanent memory pools.
+ * Note that temp files (virtual arrays) are not allowed to belong to
+ * the permanent pool, so we will be able to close all temp files here.
+ * Closing a data source or destination, if necessary, is the application's
+ * responsibility.
+ */
+
+GLOBAL(void)
+jpeg_abort (j_common_ptr cinfo)
+{
+  int pool;
+
+  /* Do nothing if called on a not-initialized or destroyed JPEG object. */
+  if (cinfo->mem == NULL)
+    return;
+
+  /* Releasing pools in reverse order might help avoid fragmentation
+   * with some (brain-damaged) malloc libraries.
+   */
+  for (pool = JPOOL_NUMPOOLS-1; pool > JPOOL_PERMANENT; pool--) {
+    (*cinfo->mem->free_pool) (cinfo, pool);
+  }
+
+  /* Reset overall state for possible reuse of object */
+  if (cinfo->is_decompressor) {
+    cinfo->global_state = DSTATE_START;
+    /* Try to keep application from accessing now-deleted marker list.
+     * A bit kludgy to do it here, but this is the most central place.
+     */
+    ((j_decompress_ptr) cinfo)->marker_list = NULL;
+  } else {
+    cinfo->global_state = CSTATE_START;
+  }
+}
+
+
+/*
+ * Destruction of a JPEG object.
+ *
+ * Everything gets deallocated except the master jpeg_compress_struct itself
+ * and the error manager struct.  Both of these are supplied by the application
+ * and must be freed, if necessary, by the application.  (Often they are on
+ * the stack and so don't need to be freed anyway.)
+ * Closing a data source or destination, if necessary, is the application's
+ * responsibility.
+ */
+
+GLOBAL(void)
+jpeg_destroy (j_common_ptr cinfo)
+{
+  /* We need only tell the memory manager to release everything. */
+  /* NB: mem pointer is NULL if memory mgr failed to initialize. */
+  if (cinfo->mem != NULL)
+    (*cinfo->mem->self_destruct) (cinfo);
+  cinfo->mem = NULL;		/* be safe if jpeg_destroy is called twice */
+  cinfo->global_state = 0;	/* mark it destroyed */
+}
+
+
+/*
+ * Convenience routines for allocating quantization and Huffman tables.
+ * (Would jutils.c be a more reasonable place to put these?)
+ */
+
+GLOBAL(JQUANT_TBL *)
+jpeg_alloc_quant_table (j_common_ptr cinfo)
+{
+  JQUANT_TBL *tbl;
+
+  tbl = (JQUANT_TBL *)
+    (*cinfo->mem->alloc_small) (cinfo, JPOOL_PERMANENT, SIZEOF(JQUANT_TBL));
+  tbl->sent_table = FALSE;	/* make sure this is false in any new table */
+  return tbl;
+}
+
+
+GLOBAL(JHUFF_TBL *)
+jpeg_alloc_huff_table (j_common_ptr cinfo)
+{
+  JHUFF_TBL *tbl;
+
+  tbl = (JHUFF_TBL *)
+    (*cinfo->mem->alloc_small) (cinfo, JPOOL_PERMANENT, SIZEOF(JHUFF_TBL));
+  tbl->sent_table = FALSE;	/* make sure this is false in any new table */
+  return tbl;
+}

Added: trunk/code/jpeg-8c/jcparam.c
===================================================================
--- trunk/code/jpeg-8c/jcparam.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcparam.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,632 @@
+/*
+ * jcparam.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modified 2003-2008 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains optional default-setting code for the JPEG compressor.
+ * Applications do not have to use this file, but those that don't use it
+ * must know a lot more about the innards of the JPEG code.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Quantization table setup routines
+ */
+
+GLOBAL(void)
+jpeg_add_quant_table (j_compress_ptr cinfo, int which_tbl,
+		      const unsigned int *basic_table,
+		      int scale_factor, boolean force_baseline)
+/* Define a quantization table equal to the basic_table times
+ * a scale factor (given as a percentage).
+ * If force_baseline is TRUE, the computed quantization table entries
+ * are limited to 1..255 for JPEG baseline compatibility.
+ */
+{
+  JQUANT_TBL ** qtblptr;
+  int i;
+  long temp;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (which_tbl < 0 || which_tbl >= NUM_QUANT_TBLS)
+    ERREXIT1(cinfo, JERR_DQT_INDEX, which_tbl);
+
+  qtblptr = & cinfo->quant_tbl_ptrs[which_tbl];
+
+  if (*qtblptr == NULL)
+    *qtblptr = jpeg_alloc_quant_table((j_common_ptr) cinfo);
+
+  for (i = 0; i < DCTSIZE2; i++) {
+    temp = ((long) basic_table[i] * scale_factor + 50L) / 100L;
+    /* limit the values to the valid range */
+    if (temp <= 0L) temp = 1L;
+    if (temp > 32767L) temp = 32767L; /* max quantizer needed for 12 bits */
+    if (force_baseline && temp > 255L)
+      temp = 255L;		/* limit to baseline range if requested */
+    (*qtblptr)->quantval[i] = (UINT16) temp;
+  }
+
+  /* Initialize sent_table FALSE so table will be written to JPEG file. */
+  (*qtblptr)->sent_table = FALSE;
+}
+
+
+/* These are the sample quantization tables given in JPEG spec section K.1.
+ * The spec says that the values given produce "good" quality, and
+ * when divided by 2, "very good" quality.
+ */
+static const unsigned int std_luminance_quant_tbl[DCTSIZE2] = {
+  16,  11,  10,  16,  24,  40,  51,  61,
+  12,  12,  14,  19,  26,  58,  60,  55,
+  14,  13,  16,  24,  40,  57,  69,  56,
+  14,  17,  22,  29,  51,  87,  80,  62,
+  18,  22,  37,  56,  68, 109, 103,  77,
+  24,  35,  55,  64,  81, 104, 113,  92,
+  49,  64,  78,  87, 103, 121, 120, 101,
+  72,  92,  95,  98, 112, 100, 103,  99
+};
+static const unsigned int std_chrominance_quant_tbl[DCTSIZE2] = {
+  17,  18,  24,  47,  99,  99,  99,  99,
+  18,  21,  26,  66,  99,  99,  99,  99,
+  24,  26,  56,  99,  99,  99,  99,  99,
+  47,  66,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99,
+  99,  99,  99,  99,  99,  99,  99,  99
+};
+
+
+GLOBAL(void)
+jpeg_default_qtables (j_compress_ptr cinfo, boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables
+ * and straight percentage-scaling quality scales.
+ * This entry point allows different scalings for luminance and chrominance.
+ */
+{
+  /* Set up two quantization tables using the specified scaling */
+  jpeg_add_quant_table(cinfo, 0, std_luminance_quant_tbl,
+		       cinfo->q_scale_factor[0], force_baseline);
+  jpeg_add_quant_table(cinfo, 1, std_chrominance_quant_tbl,
+		       cinfo->q_scale_factor[1], force_baseline);
+}
+
+
+GLOBAL(void)
+jpeg_set_linear_quality (j_compress_ptr cinfo, int scale_factor,
+			 boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables
+ * and a straight percentage-scaling quality scale.  In most cases it's better
+ * to use jpeg_set_quality (below); this entry point is provided for
+ * applications that insist on a linear percentage scaling.
+ */
+{
+  /* Set up two quantization tables using the specified scaling */
+  jpeg_add_quant_table(cinfo, 0, std_luminance_quant_tbl,
+		       scale_factor, force_baseline);
+  jpeg_add_quant_table(cinfo, 1, std_chrominance_quant_tbl,
+		       scale_factor, force_baseline);
+}
+
+
+GLOBAL(int)
+jpeg_quality_scaling (int quality)
+/* Convert a user-specified quality rating to a percentage scaling factor
+ * for an underlying quantization table, using our recommended scaling curve.
+ * The input 'quality' factor should be 0 (terrible) to 100 (very good).
+ */
+{
+  /* Safety limit on quality factor.  Convert 0 to 1 to avoid zero divide. */
+  if (quality <= 0) quality = 1;
+  if (quality > 100) quality = 100;
+
+  /* The basic table is used as-is (scaling 100) for a quality of 50.
+   * Qualities 50..100 are converted to scaling percentage 200 - 2*Q;
+   * note that at Q=100 the scaling is 0, which will cause jpeg_add_quant_table
+   * to make all the table entries 1 (hence, minimum quantization loss).
+   * Qualities 1..50 are converted to scaling percentage 5000/Q.
+   */
+  if (quality < 50)
+    quality = 5000 / quality;
+  else
+    quality = 200 - quality*2;
+
+  return quality;
+}
+
+
+GLOBAL(void)
+jpeg_set_quality (j_compress_ptr cinfo, int quality, boolean force_baseline)
+/* Set or change the 'quality' (quantization) setting, using default tables.
+ * This is the standard quality-adjusting entry point for typical user
+ * interfaces; only those who want detailed control over quantization tables
+ * would use the preceding three routines directly.
+ */
+{
+  /* Convert user 0-100 rating to percentage scaling */
+  quality = jpeg_quality_scaling(quality);
+
+  /* Set up standard quality tables */
+  jpeg_set_linear_quality(cinfo, quality, force_baseline);
+}
+
+
+/*
+ * Huffman table setup routines
+ */
+
+LOCAL(void)
+add_huff_table (j_compress_ptr cinfo,
+		JHUFF_TBL **htblptr, const UINT8 *bits, const UINT8 *val)
+/* Define a Huffman table */
+{
+  int nsymbols, len;
+
+  if (*htblptr == NULL)
+    *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+
+  /* Copy the number-of-symbols-of-each-code-length counts */
+  MEMCOPY((*htblptr)->bits, bits, SIZEOF((*htblptr)->bits));
+
+  /* Validate the counts.  We do this here mainly so we can copy the right
+   * number of symbols from the val[] array, without risking marching off
+   * the end of memory.  jchuff.c will do a more thorough test later.
+   */
+  nsymbols = 0;
+  for (len = 1; len <= 16; len++)
+    nsymbols += bits[len];
+  if (nsymbols < 1 || nsymbols > 256)
+    ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+
+  MEMCOPY((*htblptr)->huffval, val, nsymbols * SIZEOF(UINT8));
+
+  /* Initialize sent_table FALSE so table will be written to JPEG file. */
+  (*htblptr)->sent_table = FALSE;
+}
+
+
+LOCAL(void)
+std_huff_tables (j_compress_ptr cinfo)
+/* Set up the standard Huffman tables (cf. JPEG standard section K.3) */
+/* IMPORTANT: these are only valid for 8-bit data precision! */
+{
+  static const UINT8 bits_dc_luminance[17] =
+    { /* 0-base */ 0, 0, 1, 5, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0 };
+  static const UINT8 val_dc_luminance[] =
+    { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 };
+  
+  static const UINT8 bits_dc_chrominance[17] =
+    { /* 0-base */ 0, 0, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0 };
+  static const UINT8 val_dc_chrominance[] =
+    { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 };
+  
+  static const UINT8 bits_ac_luminance[17] =
+    { /* 0-base */ 0, 0, 2, 1, 3, 3, 2, 4, 3, 5, 5, 4, 4, 0, 0, 1, 0x7d };
+  static const UINT8 val_ac_luminance[] =
+    { 0x01, 0x02, 0x03, 0x00, 0x04, 0x11, 0x05, 0x12,
+      0x21, 0x31, 0x41, 0x06, 0x13, 0x51, 0x61, 0x07,
+      0x22, 0x71, 0x14, 0x32, 0x81, 0x91, 0xa1, 0x08,
+      0x23, 0x42, 0xb1, 0xc1, 0x15, 0x52, 0xd1, 0xf0,
+      0x24, 0x33, 0x62, 0x72, 0x82, 0x09, 0x0a, 0x16,
+      0x17, 0x18, 0x19, 0x1a, 0x25, 0x26, 0x27, 0x28,
+      0x29, 0x2a, 0x34, 0x35, 0x36, 0x37, 0x38, 0x39,
+      0x3a, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48, 0x49,
+      0x4a, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59,
+      0x5a, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69,
+      0x6a, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78, 0x79,
+      0x7a, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89,
+      0x8a, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98,
+      0x99, 0x9a, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7,
+      0xa8, 0xa9, 0xaa, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6,
+      0xb7, 0xb8, 0xb9, 0xba, 0xc2, 0xc3, 0xc4, 0xc5,
+      0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xd2, 0xd3, 0xd4,
+      0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda, 0xe1, 0xe2,
+      0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea,
+      0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+      0xf9, 0xfa };
+  
+  static const UINT8 bits_ac_chrominance[17] =
+    { /* 0-base */ 0, 0, 2, 1, 2, 4, 4, 3, 4, 7, 5, 4, 4, 0, 1, 2, 0x77 };
+  static const UINT8 val_ac_chrominance[] =
+    { 0x00, 0x01, 0x02, 0x03, 0x11, 0x04, 0x05, 0x21,
+      0x31, 0x06, 0x12, 0x41, 0x51, 0x07, 0x61, 0x71,
+      0x13, 0x22, 0x32, 0x81, 0x08, 0x14, 0x42, 0x91,
+      0xa1, 0xb1, 0xc1, 0x09, 0x23, 0x33, 0x52, 0xf0,
+      0x15, 0x62, 0x72, 0xd1, 0x0a, 0x16, 0x24, 0x34,
+      0xe1, 0x25, 0xf1, 0x17, 0x18, 0x19, 0x1a, 0x26,
+      0x27, 0x28, 0x29, 0x2a, 0x35, 0x36, 0x37, 0x38,
+      0x39, 0x3a, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48,
+      0x49, 0x4a, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58,
+      0x59, 0x5a, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68,
+      0x69, 0x6a, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78,
+      0x79, 0x7a, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
+      0x88, 0x89, 0x8a, 0x92, 0x93, 0x94, 0x95, 0x96,
+      0x97, 0x98, 0x99, 0x9a, 0xa2, 0xa3, 0xa4, 0xa5,
+      0xa6, 0xa7, 0xa8, 0xa9, 0xaa, 0xb2, 0xb3, 0xb4,
+      0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xc2, 0xc3,
+      0xc4, 0xc5, 0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xd2,
+      0xd3, 0xd4, 0xd5, 0xd6, 0xd7, 0xd8, 0xd9, 0xda,
+      0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9,
+      0xea, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8,
+      0xf9, 0xfa };
+  
+  add_huff_table(cinfo, &cinfo->dc_huff_tbl_ptrs[0],
+		 bits_dc_luminance, val_dc_luminance);
+  add_huff_table(cinfo, &cinfo->ac_huff_tbl_ptrs[0],
+		 bits_ac_luminance, val_ac_luminance);
+  add_huff_table(cinfo, &cinfo->dc_huff_tbl_ptrs[1],
+		 bits_dc_chrominance, val_dc_chrominance);
+  add_huff_table(cinfo, &cinfo->ac_huff_tbl_ptrs[1],
+		 bits_ac_chrominance, val_ac_chrominance);
+}
+
+
+/*
+ * Default parameter setup for compression.
+ *
+ * Applications that don't choose to use this routine must do their
+ * own setup of all these parameters.  Alternately, you can call this
+ * to establish defaults and then alter parameters selectively.  This
+ * is the recommended approach since, if we add any new parameters,
+ * your code will still work (they'll be set to reasonable defaults).
+ */
+
+GLOBAL(void)
+jpeg_set_defaults (j_compress_ptr cinfo)
+{
+  int i;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Allocate comp_info array large enough for maximum component count.
+   * Array is made permanent in case application wants to compress
+   * multiple images at same param settings.
+   */
+  if (cinfo->comp_info == NULL)
+    cinfo->comp_info = (jpeg_component_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  MAX_COMPONENTS * SIZEOF(jpeg_component_info));
+
+  /* Initialize everything not dependent on the color space */
+
+  cinfo->scale_num = 1;		/* 1:1 scaling */
+  cinfo->scale_denom = 1;
+  cinfo->data_precision = BITS_IN_JSAMPLE;
+  /* Set up two quantization tables using default quality of 75 */
+  jpeg_set_quality(cinfo, 75, TRUE);
+  /* Set up two Huffman tables */
+  std_huff_tables(cinfo);
+
+  /* Initialize default arithmetic coding conditioning */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    cinfo->arith_dc_L[i] = 0;
+    cinfo->arith_dc_U[i] = 1;
+    cinfo->arith_ac_K[i] = 5;
+  }
+
+  /* Default is no multiple-scan output */
+  cinfo->scan_info = NULL;
+  cinfo->num_scans = 0;
+
+  /* Expect normal source image, not raw downsampled data */
+  cinfo->raw_data_in = FALSE;
+
+  /* Use Huffman coding, not arithmetic coding, by default */
+  cinfo->arith_code = FALSE;
+
+  /* By default, don't do extra passes to optimize entropy coding */
+  cinfo->optimize_coding = FALSE;
+  /* The standard Huffman tables are only valid for 8-bit data precision.
+   * If the precision is higher, force optimization on so that usable
+   * tables will be computed.  This test can be removed if default tables
+   * are supplied that are valid for the desired precision.
+   */
+  if (cinfo->data_precision > 8)
+    cinfo->optimize_coding = TRUE;
+
+  /* By default, use the simpler non-cosited sampling alignment */
+  cinfo->CCIR601_sampling = FALSE;
+
+  /* By default, apply fancy downsampling */
+  cinfo->do_fancy_downsampling = TRUE;
+
+  /* No input smoothing */
+  cinfo->smoothing_factor = 0;
+
+  /* DCT algorithm preference */
+  cinfo->dct_method = JDCT_DEFAULT;
+
+  /* No restart markers */
+  cinfo->restart_interval = 0;
+  cinfo->restart_in_rows = 0;
+
+  /* Fill in default JFIF marker parameters.  Note that whether the marker
+   * will actually be written is determined by jpeg_set_colorspace.
+   *
+   * By default, the library emits JFIF version code 1.01.
+   * An application that wants to emit JFIF 1.02 extension markers should set
+   * JFIF_minor_version to 2.  We could probably get away with just defaulting
+   * to 1.02, but there may still be some decoders in use that will complain
+   * about that; saying 1.01 should minimize compatibility problems.
+   */
+  cinfo->JFIF_major_version = 1; /* Default JFIF version = 1.01 */
+  cinfo->JFIF_minor_version = 1;
+  cinfo->density_unit = 0;	/* Pixel size is unknown by default */
+  cinfo->X_density = 1;		/* Pixel aspect ratio is square by default */
+  cinfo->Y_density = 1;
+
+  /* Choose JPEG colorspace based on input space, set defaults accordingly */
+
+  jpeg_default_colorspace(cinfo);
+}
+
+
+/*
+ * Select an appropriate JPEG colorspace for in_color_space.
+ */
+
+GLOBAL(void)
+jpeg_default_colorspace (j_compress_ptr cinfo)
+{
+  switch (cinfo->in_color_space) {
+  case JCS_GRAYSCALE:
+    jpeg_set_colorspace(cinfo, JCS_GRAYSCALE);
+    break;
+  case JCS_RGB:
+    jpeg_set_colorspace(cinfo, JCS_YCbCr);
+    break;
+  case JCS_YCbCr:
+    jpeg_set_colorspace(cinfo, JCS_YCbCr);
+    break;
+  case JCS_CMYK:
+    jpeg_set_colorspace(cinfo, JCS_CMYK); /* By default, no translation */
+    break;
+  case JCS_YCCK:
+    jpeg_set_colorspace(cinfo, JCS_YCCK);
+    break;
+  case JCS_UNKNOWN:
+    jpeg_set_colorspace(cinfo, JCS_UNKNOWN);
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_IN_COLORSPACE);
+  }
+}
+
+
+/*
+ * Set the JPEG colorspace, and choose colorspace-dependent default values.
+ */
+
+GLOBAL(void)
+jpeg_set_colorspace (j_compress_ptr cinfo, J_COLOR_SPACE colorspace)
+{
+  jpeg_component_info * compptr;
+  int ci;
+
+#define SET_COMP(index,id,hsamp,vsamp,quant,dctbl,actbl)  \
+  (compptr = &cinfo->comp_info[index], \
+   compptr->component_id = (id), \
+   compptr->h_samp_factor = (hsamp), \
+   compptr->v_samp_factor = (vsamp), \
+   compptr->quant_tbl_no = (quant), \
+   compptr->dc_tbl_no = (dctbl), \
+   compptr->ac_tbl_no = (actbl) )
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* For all colorspaces, we use Q and Huff tables 0 for luminance components,
+   * tables 1 for chrominance components.
+   */
+
+  cinfo->jpeg_color_space = colorspace;
+
+  cinfo->write_JFIF_header = FALSE; /* No marker for non-JFIF colorspaces */
+  cinfo->write_Adobe_marker = FALSE; /* write no Adobe marker by default */
+
+  switch (colorspace) {
+  case JCS_GRAYSCALE:
+    cinfo->write_JFIF_header = TRUE; /* Write a JFIF marker */
+    cinfo->num_components = 1;
+    /* JFIF specifies component ID 1 */
+    SET_COMP(0, 1, 1,1, 0, 0,0);
+    break;
+  case JCS_RGB:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag RGB */
+    cinfo->num_components = 3;
+    SET_COMP(0, 0x52 /* 'R' */, 1,1, 0, 0,0);
+    SET_COMP(1, 0x47 /* 'G' */, 1,1, 0, 0,0);
+    SET_COMP(2, 0x42 /* 'B' */, 1,1, 0, 0,0);
+    break;
+  case JCS_YCbCr:
+    cinfo->write_JFIF_header = TRUE; /* Write a JFIF marker */
+    cinfo->num_components = 3;
+    /* JFIF specifies component IDs 1,2,3 */
+    /* We default to 2x2 subsamples of chrominance */
+    SET_COMP(0, 1, 2,2, 0, 0,0);
+    SET_COMP(1, 2, 1,1, 1, 1,1);
+    SET_COMP(2, 3, 1,1, 1, 1,1);
+    break;
+  case JCS_CMYK:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag CMYK */
+    cinfo->num_components = 4;
+    SET_COMP(0, 0x43 /* 'C' */, 1,1, 0, 0,0);
+    SET_COMP(1, 0x4D /* 'M' */, 1,1, 0, 0,0);
+    SET_COMP(2, 0x59 /* 'Y' */, 1,1, 0, 0,0);
+    SET_COMP(3, 0x4B /* 'K' */, 1,1, 0, 0,0);
+    break;
+  case JCS_YCCK:
+    cinfo->write_Adobe_marker = TRUE; /* write Adobe marker to flag YCCK */
+    cinfo->num_components = 4;
+    SET_COMP(0, 1, 2,2, 0, 0,0);
+    SET_COMP(1, 2, 1,1, 1, 1,1);
+    SET_COMP(2, 3, 1,1, 1, 1,1);
+    SET_COMP(3, 4, 2,2, 0, 0,0);
+    break;
+  case JCS_UNKNOWN:
+    cinfo->num_components = cinfo->input_components;
+    if (cinfo->num_components < 1 || cinfo->num_components > MAX_COMPONENTS)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	       MAX_COMPONENTS);
+    for (ci = 0; ci < cinfo->num_components; ci++) {
+      SET_COMP(ci, ci, 1,1, 0, 0,0);
+    }
+    break;
+  default:
+    ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+  }
+}
+
+
+#ifdef C_PROGRESSIVE_SUPPORTED
+
+LOCAL(jpeg_scan_info *)
+fill_a_scan (jpeg_scan_info * scanptr, int ci,
+	     int Ss, int Se, int Ah, int Al)
+/* Support routine: generate one scan for specified component */
+{
+  scanptr->comps_in_scan = 1;
+  scanptr->component_index[0] = ci;
+  scanptr->Ss = Ss;
+  scanptr->Se = Se;
+  scanptr->Ah = Ah;
+  scanptr->Al = Al;
+  scanptr++;
+  return scanptr;
+}
+
+LOCAL(jpeg_scan_info *)
+fill_scans (jpeg_scan_info * scanptr, int ncomps,
+	    int Ss, int Se, int Ah, int Al)
+/* Support routine: generate one scan for each component */
+{
+  int ci;
+
+  for (ci = 0; ci < ncomps; ci++) {
+    scanptr->comps_in_scan = 1;
+    scanptr->component_index[0] = ci;
+    scanptr->Ss = Ss;
+    scanptr->Se = Se;
+    scanptr->Ah = Ah;
+    scanptr->Al = Al;
+    scanptr++;
+  }
+  return scanptr;
+}
+
+LOCAL(jpeg_scan_info *)
+fill_dc_scans (jpeg_scan_info * scanptr, int ncomps, int Ah, int Al)
+/* Support routine: generate interleaved DC scan if possible, else N scans */
+{
+  int ci;
+
+  if (ncomps <= MAX_COMPS_IN_SCAN) {
+    /* Single interleaved DC scan */
+    scanptr->comps_in_scan = ncomps;
+    for (ci = 0; ci < ncomps; ci++)
+      scanptr->component_index[ci] = ci;
+    scanptr->Ss = scanptr->Se = 0;
+    scanptr->Ah = Ah;
+    scanptr->Al = Al;
+    scanptr++;
+  } else {
+    /* Noninterleaved DC scan for each component */
+    scanptr = fill_scans(scanptr, ncomps, 0, 0, Ah, Al);
+  }
+  return scanptr;
+}
+
+
+/*
+ * Create a recommended progressive-JPEG script.
+ * cinfo->num_components and cinfo->jpeg_color_space must be correct.
+ */
+
+GLOBAL(void)
+jpeg_simple_progression (j_compress_ptr cinfo)
+{
+  int ncomps = cinfo->num_components;
+  int nscans;
+  jpeg_scan_info * scanptr;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Figure space needed for script.  Calculation must match code below! */
+  if (ncomps == 3 && cinfo->jpeg_color_space == JCS_YCbCr) {
+    /* Custom script for YCbCr color images. */
+    nscans = 10;
+  } else {
+    /* All-purpose script for other color spaces. */
+    if (ncomps > MAX_COMPS_IN_SCAN)
+      nscans = 6 * ncomps;	/* 2 DC + 4 AC scans per component */
+    else
+      nscans = 2 + 4 * ncomps;	/* 2 DC scans; 4 AC scans per component */
+  }
+
+  /* Allocate space for script.
+   * We need to put it in the permanent pool in case the application performs
+   * multiple compressions without changing the settings.  To avoid a memory
+   * leak if jpeg_simple_progression is called repeatedly for the same JPEG
+   * object, we try to re-use previously allocated space, and we allocate
+   * enough space to handle YCbCr even if initially asked for grayscale.
+   */
+  if (cinfo->script_space == NULL || cinfo->script_space_size < nscans) {
+    cinfo->script_space_size = MAX(nscans, 10);
+    cinfo->script_space = (jpeg_scan_info *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+			cinfo->script_space_size * SIZEOF(jpeg_scan_info));
+  }
+  scanptr = cinfo->script_space;
+  cinfo->scan_info = scanptr;
+  cinfo->num_scans = nscans;
+
+  if (ncomps == 3 && cinfo->jpeg_color_space == JCS_YCbCr) {
+    /* Custom script for YCbCr color images. */
+    /* Initial DC scan */
+    scanptr = fill_dc_scans(scanptr, ncomps, 0, 1);
+    /* Initial AC scan: get some luma data out in a hurry */
+    scanptr = fill_a_scan(scanptr, 0, 1, 5, 0, 2);
+    /* Chroma data is too small to be worth expending many scans on */
+    scanptr = fill_a_scan(scanptr, 2, 1, 63, 0, 1);
+    scanptr = fill_a_scan(scanptr, 1, 1, 63, 0, 1);
+    /* Complete spectral selection for luma AC */
+    scanptr = fill_a_scan(scanptr, 0, 6, 63, 0, 2);
+    /* Refine next bit of luma AC */
+    scanptr = fill_a_scan(scanptr, 0, 1, 63, 2, 1);
+    /* Finish DC successive approximation */
+    scanptr = fill_dc_scans(scanptr, ncomps, 1, 0);
+    /* Finish AC successive approximation */
+    scanptr = fill_a_scan(scanptr, 2, 1, 63, 1, 0);
+    scanptr = fill_a_scan(scanptr, 1, 1, 63, 1, 0);
+    /* Luma bottom bit comes last since it's usually largest scan */
+    scanptr = fill_a_scan(scanptr, 0, 1, 63, 1, 0);
+  } else {
+    /* All-purpose script for other color spaces. */
+    /* Successive approximation first pass */
+    scanptr = fill_dc_scans(scanptr, ncomps, 0, 1);
+    scanptr = fill_scans(scanptr, ncomps, 1, 5, 0, 2);
+    scanptr = fill_scans(scanptr, ncomps, 6, 63, 0, 2);
+    /* Successive approximation second pass */
+    scanptr = fill_scans(scanptr, ncomps, 1, 63, 2, 1);
+    /* Successive approximation final pass */
+    scanptr = fill_dc_scans(scanptr, ncomps, 1, 0);
+    scanptr = fill_scans(scanptr, ncomps, 1, 63, 1, 0);
+  }
+}
+
+#endif /* C_PROGRESSIVE_SUPPORTED */

Added: trunk/code/jpeg-8c/jcprepct.c
===================================================================
--- trunk/code/jpeg-8c/jcprepct.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcprepct.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,358 @@
+/*
+ * jcprepct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the compression preprocessing controller.
+ * This controller manages the color conversion, downsampling,
+ * and edge expansion steps.
+ *
+ * Most of the complexity here is associated with buffering input rows
+ * as required by the downsampler.  See the comments at the head of
+ * jcsample.c for the downsampler's needs.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* At present, jcsample.c can request context rows only for smoothing.
+ * In the future, we might also need context rows for CCIR601 sampling
+ * or other more-complex downsampling procedures.  The code to support
+ * context rows should be compiled only if needed.
+ */
+#ifdef INPUT_SMOOTHING_SUPPORTED
+#define CONTEXT_ROWS_SUPPORTED
+#endif
+
+
+/*
+ * For the simple (no-context-row) case, we just need to buffer one
+ * row group's worth of pixels for the downsampling step.  At the bottom of
+ * the image, we pad to a full row group by replicating the last pixel row.
+ * The downsampler's last output row is then replicated if needed to pad
+ * out to a full iMCU row.
+ *
+ * When providing context rows, we must buffer three row groups' worth of
+ * pixels.  Three row groups are physically allocated, but the row pointer
+ * arrays are made five row groups high, with the extra pointers above and
+ * below "wrapping around" to point to the last and first real row groups.
+ * This allows the downsampler to access the proper context rows.
+ * At the top and bottom of the image, we create dummy context rows by
+ * copying the first or last real pixel row.  This copying could be avoided
+ * by pointer hacking as is done in jdmainct.c, but it doesn't seem worth the
+ * trouble on the compression side.
+ */
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_prep_controller pub; /* public fields */
+
+  /* Downsampling input buffer.  This buffer holds color-converted data
+   * until we have enough to do a downsample step.
+   */
+  JSAMPARRAY color_buf[MAX_COMPONENTS];
+
+  JDIMENSION rows_to_go;	/* counts rows remaining in source image */
+  int next_buf_row;		/* index of next row to store in color_buf */
+
+#ifdef CONTEXT_ROWS_SUPPORTED	/* only needed for context case */
+  int this_row_group;		/* starting row index of group to process */
+  int next_buf_stop;		/* downsample when we reach this index */
+#endif
+} my_prep_controller;
+
+typedef my_prep_controller * my_prep_ptr;
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_prep (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+
+  if (pass_mode != JBUF_PASS_THRU)
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  /* Initialize total-height counter for detecting bottom of image */
+  prep->rows_to_go = cinfo->image_height;
+  /* Mark the conversion buffer empty */
+  prep->next_buf_row = 0;
+#ifdef CONTEXT_ROWS_SUPPORTED
+  /* Preset additional state variables for context mode.
+   * These aren't used in non-context mode, so we needn't test which mode.
+   */
+  prep->this_row_group = 0;
+  /* Set next_buf_stop to stop after two row groups have been read in. */
+  prep->next_buf_stop = 2 * cinfo->max_v_samp_factor;
+#endif
+}
+
+
+/*
+ * Expand an image vertically from height input_rows to height output_rows,
+ * by duplicating the bottom row.
+ */
+
+LOCAL(void)
+expand_bottom_edge (JSAMPARRAY image_data, JDIMENSION num_cols,
+		    int input_rows, int output_rows)
+{
+  register int row;
+
+  for (row = input_rows; row < output_rows; row++) {
+    jcopy_sample_rows(image_data, input_rows-1, image_data, row,
+		      1, num_cols);
+  }
+}
+
+
+/*
+ * Process some data in the simple no-context case.
+ *
+ * Preprocessor output data is counted in "row groups".  A row group
+ * is defined to be v_samp_factor sample rows of each component.
+ * Downsampling will produce this much data from each max_v_samp_factor
+ * input rows.
+ */
+
+METHODDEF(void)
+pre_process_data (j_compress_ptr cinfo,
+		  JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+		  JDIMENSION in_rows_avail,
+		  JSAMPIMAGE output_buf, JDIMENSION *out_row_group_ctr,
+		  JDIMENSION out_row_groups_avail)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int numrows, ci;
+  JDIMENSION inrows;
+  jpeg_component_info * compptr;
+
+  while (*in_row_ctr < in_rows_avail &&
+	 *out_row_group_ctr < out_row_groups_avail) {
+    /* Do color conversion to fill the conversion buffer. */
+    inrows = in_rows_avail - *in_row_ctr;
+    numrows = cinfo->max_v_samp_factor - prep->next_buf_row;
+    numrows = (int) MIN((JDIMENSION) numrows, inrows);
+    (*cinfo->cconvert->color_convert) (cinfo, input_buf + *in_row_ctr,
+				       prep->color_buf,
+				       (JDIMENSION) prep->next_buf_row,
+				       numrows);
+    *in_row_ctr += numrows;
+    prep->next_buf_row += numrows;
+    prep->rows_to_go -= numrows;
+    /* If at bottom of image, pad to fill the conversion buffer. */
+    if (prep->rows_to_go == 0 &&
+	prep->next_buf_row < cinfo->max_v_samp_factor) {
+      for (ci = 0; ci < cinfo->num_components; ci++) {
+	expand_bottom_edge(prep->color_buf[ci], cinfo->image_width,
+			   prep->next_buf_row, cinfo->max_v_samp_factor);
+      }
+      prep->next_buf_row = cinfo->max_v_samp_factor;
+    }
+    /* If we've filled the conversion buffer, empty it. */
+    if (prep->next_buf_row == cinfo->max_v_samp_factor) {
+      (*cinfo->downsample->downsample) (cinfo,
+					prep->color_buf, (JDIMENSION) 0,
+					output_buf, *out_row_group_ctr);
+      prep->next_buf_row = 0;
+      (*out_row_group_ctr)++;
+    }
+    /* If at bottom of image, pad the output to a full iMCU height.
+     * Note we assume the caller is providing a one-iMCU-height output buffer!
+     */
+    if (prep->rows_to_go == 0 &&
+	*out_row_group_ctr < out_row_groups_avail) {
+      for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	   ci++, compptr++) {
+	numrows = (compptr->v_samp_factor * compptr->DCT_v_scaled_size) /
+		  cinfo->min_DCT_v_scaled_size;
+	expand_bottom_edge(output_buf[ci],
+			   compptr->width_in_blocks * compptr->DCT_h_scaled_size,
+			   (int) (*out_row_group_ctr * numrows),
+			   (int) (out_row_groups_avail * numrows));
+      }
+      *out_row_group_ctr = out_row_groups_avail;
+      break;			/* can exit outer loop without test */
+    }
+  }
+}
+
+
+#ifdef CONTEXT_ROWS_SUPPORTED
+
+/*
+ * Process some data in the context case.
+ */
+
+METHODDEF(void)
+pre_process_context (j_compress_ptr cinfo,
+		     JSAMPARRAY input_buf, JDIMENSION *in_row_ctr,
+		     JDIMENSION in_rows_avail,
+		     JSAMPIMAGE output_buf, JDIMENSION *out_row_group_ctr,
+		     JDIMENSION out_row_groups_avail)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int numrows, ci;
+  int buf_height = cinfo->max_v_samp_factor * 3;
+  JDIMENSION inrows;
+
+  while (*out_row_group_ctr < out_row_groups_avail) {
+    if (*in_row_ctr < in_rows_avail) {
+      /* Do color conversion to fill the conversion buffer. */
+      inrows = in_rows_avail - *in_row_ctr;
+      numrows = prep->next_buf_stop - prep->next_buf_row;
+      numrows = (int) MIN((JDIMENSION) numrows, inrows);
+      (*cinfo->cconvert->color_convert) (cinfo, input_buf + *in_row_ctr,
+					 prep->color_buf,
+					 (JDIMENSION) prep->next_buf_row,
+					 numrows);
+      /* Pad at top of image, if first time through */
+      if (prep->rows_to_go == cinfo->image_height) {
+	for (ci = 0; ci < cinfo->num_components; ci++) {
+	  int row;
+	  for (row = 1; row <= cinfo->max_v_samp_factor; row++) {
+	    jcopy_sample_rows(prep->color_buf[ci], 0,
+			      prep->color_buf[ci], -row,
+			      1, cinfo->image_width);
+	  }
+	}
+      }
+      *in_row_ctr += numrows;
+      prep->next_buf_row += numrows;
+      prep->rows_to_go -= numrows;
+    } else {
+      /* Return for more data, unless we are at the bottom of the image. */
+      if (prep->rows_to_go != 0)
+	break;
+      /* When at bottom of image, pad to fill the conversion buffer. */
+      if (prep->next_buf_row < prep->next_buf_stop) {
+	for (ci = 0; ci < cinfo->num_components; ci++) {
+	  expand_bottom_edge(prep->color_buf[ci], cinfo->image_width,
+			     prep->next_buf_row, prep->next_buf_stop);
+	}
+	prep->next_buf_row = prep->next_buf_stop;
+      }
+    }
+    /* If we've gotten enough data, downsample a row group. */
+    if (prep->next_buf_row == prep->next_buf_stop) {
+      (*cinfo->downsample->downsample) (cinfo,
+					prep->color_buf,
+					(JDIMENSION) prep->this_row_group,
+					output_buf, *out_row_group_ctr);
+      (*out_row_group_ctr)++;
+      /* Advance pointers with wraparound as necessary. */
+      prep->this_row_group += cinfo->max_v_samp_factor;
+      if (prep->this_row_group >= buf_height)
+	prep->this_row_group = 0;
+      if (prep->next_buf_row >= buf_height)
+	prep->next_buf_row = 0;
+      prep->next_buf_stop = prep->next_buf_row + cinfo->max_v_samp_factor;
+    }
+  }
+}
+
+
+/*
+ * Create the wrapped-around downsampling input buffer needed for context mode.
+ */
+
+LOCAL(void)
+create_context_buffer (j_compress_ptr cinfo)
+{
+  my_prep_ptr prep = (my_prep_ptr) cinfo->prep;
+  int rgroup_height = cinfo->max_v_samp_factor;
+  int ci, i;
+  jpeg_component_info * compptr;
+  JSAMPARRAY true_buffer, fake_buffer;
+
+  /* Grab enough space for fake row pointers for all the components;
+   * we need five row groups' worth of pointers for each component.
+   */
+  fake_buffer = (JSAMPARRAY)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(cinfo->num_components * 5 * rgroup_height) *
+				SIZEOF(JSAMPROW));
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Allocate the actual buffer space (3 row groups) for this component.
+     * We make the buffer wide enough to allow the downsampler to edge-expand
+     * horizontally within the buffer, if it so chooses.
+     */
+    true_buffer = (*cinfo->mem->alloc_sarray)
+      ((j_common_ptr) cinfo, JPOOL_IMAGE,
+       (JDIMENSION) (((long) compptr->width_in_blocks *
+		      cinfo->min_DCT_h_scaled_size *
+		      cinfo->max_h_samp_factor) / compptr->h_samp_factor),
+       (JDIMENSION) (3 * rgroup_height));
+    /* Copy true buffer row pointers into the middle of the fake row array */
+    MEMCOPY(fake_buffer + rgroup_height, true_buffer,
+	    3 * rgroup_height * SIZEOF(JSAMPROW));
+    /* Fill in the above and below wraparound pointers */
+    for (i = 0; i < rgroup_height; i++) {
+      fake_buffer[i] = true_buffer[2 * rgroup_height + i];
+      fake_buffer[4 * rgroup_height + i] = true_buffer[i];
+    }
+    prep->color_buf[ci] = fake_buffer + rgroup_height;
+    fake_buffer += 5 * rgroup_height; /* point to space for next component */
+  }
+}
+
+#endif /* CONTEXT_ROWS_SUPPORTED */
+
+
+/*
+ * Initialize preprocessing controller.
+ */
+
+GLOBAL(void)
+jinit_c_prep_controller (j_compress_ptr cinfo, boolean need_full_buffer)
+{
+  my_prep_ptr prep;
+  int ci;
+  jpeg_component_info * compptr;
+
+  if (need_full_buffer)		/* safety check */
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  prep = (my_prep_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_prep_controller));
+  cinfo->prep = (struct jpeg_c_prep_controller *) prep;
+  prep->pub.start_pass = start_pass_prep;
+
+  /* Allocate the color conversion buffer.
+   * We make the buffer wide enough to allow the downsampler to edge-expand
+   * horizontally within the buffer, if it so chooses.
+   */
+  if (cinfo->downsample->need_context_rows) {
+    /* Set up to provide context rows */
+#ifdef CONTEXT_ROWS_SUPPORTED
+    prep->pub.pre_process_data = pre_process_context;
+    create_context_buffer(cinfo);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* No context, just make it tall enough for one row group */
+    prep->pub.pre_process_data = pre_process_data;
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      prep->color_buf[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 (JDIMENSION) (((long) compptr->width_in_blocks *
+			cinfo->min_DCT_h_scaled_size *
+			cinfo->max_h_samp_factor) / compptr->h_samp_factor),
+	 (JDIMENSION) cinfo->max_v_samp_factor);
+    }
+  }
+}

Added: trunk/code/jpeg-8c/jcsample.c
===================================================================
--- trunk/code/jpeg-8c/jcsample.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jcsample.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,545 @@
+/*
+ * jcsample.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains downsampling routines.
+ *
+ * Downsampling input data is counted in "row groups".  A row group
+ * is defined to be max_v_samp_factor pixel rows of each component,
+ * from which the downsampler produces v_samp_factor sample rows.
+ * A single row group is processed in each call to the downsampler module.
+ *
+ * The downsampler is responsible for edge-expansion of its output data
+ * to fill an integral number of DCT blocks horizontally.  The source buffer
+ * may be modified if it is helpful for this purpose (the source buffer is
+ * allocated wide enough to correspond to the desired output width).
+ * The caller (the prep controller) is responsible for vertical padding.
+ *
+ * The downsampler may request "context rows" by setting need_context_rows
+ * during startup.  In this case, the input arrays will contain at least
+ * one row group's worth of pixels above and below the passed-in data;
+ * the caller will create dummy rows at image top and bottom by replicating
+ * the first or last real pixel row.
+ *
+ * An excellent reference for image resampling is
+ *   Digital Image Warping, George Wolberg, 1990.
+ *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
+ *
+ * The downsampling algorithm used here is a simple average of the source
+ * pixels covered by the output pixel.  The hi-falutin sampling literature
+ * refers to this as a "box filter".  In general the characteristics of a box
+ * filter are not very good, but for the specific cases we normally use (1:1
+ * and 2:1 ratios) the box is equivalent to a "triangle filter" which is not
+ * nearly so bad.  If you intend to use other sampling ratios, you'd be well
+ * advised to improve this code.
+ *
+ * A simple input-smoothing capability is provided.  This is mainly intended
+ * for cleaning up color-dithered GIF input files (if you find it inadequate,
+ * we suggest using an external filtering program such as pnmconvol).  When
+ * enabled, each input pixel P is replaced by a weighted sum of itself and its
+ * eight neighbors.  P's weight is 1-8*SF and each neighbor's weight is SF,
+ * where SF = (smoothing_factor / 1024).
+ * Currently, smoothing is only supported for 2h2v sampling factors.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Pointer to routine to downsample a single component */
+typedef JMETHOD(void, downsample1_ptr,
+		(j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data));
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_downsampler pub;	/* public fields */
+
+  /* Downsampling method pointers, one per component */
+  downsample1_ptr methods[MAX_COMPONENTS];
+
+  /* Height of an output row group for each component. */
+  int rowgroup_height[MAX_COMPONENTS];
+
+  /* These arrays save pixel expansion factors so that int_downsample need not
+   * recompute them each time.  They are unused for other downsampling methods.
+   */
+  UINT8 h_expand[MAX_COMPONENTS];
+  UINT8 v_expand[MAX_COMPONENTS];
+} my_downsampler;
+
+typedef my_downsampler * my_downsample_ptr;
+
+
+/*
+ * Initialize for a downsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_downsample (j_compress_ptr cinfo)
+{
+  /* no work for now */
+}
+
+
+/*
+ * Expand a component horizontally from width input_cols to width output_cols,
+ * by duplicating the rightmost samples.
+ */
+
+LOCAL(void)
+expand_right_edge (JSAMPARRAY image_data, int num_rows,
+		   JDIMENSION input_cols, JDIMENSION output_cols)
+{
+  register JSAMPROW ptr;
+  register JSAMPLE pixval;
+  register int count;
+  int row;
+  int numcols = (int) (output_cols - input_cols);
+
+  if (numcols > 0) {
+    for (row = 0; row < num_rows; row++) {
+      ptr = image_data[row] + input_cols;
+      pixval = ptr[-1];		/* don't need GETJSAMPLE() here */
+      for (count = numcols; count > 0; count--)
+	*ptr++ = pixval;
+    }
+  }
+}
+
+
+/*
+ * Do downsampling for a whole row group (all components).
+ *
+ * In this version we simply downsample each component independently.
+ */
+
+METHODDEF(void)
+sep_downsample (j_compress_ptr cinfo,
+		JSAMPIMAGE input_buf, JDIMENSION in_row_index,
+		JSAMPIMAGE output_buf, JDIMENSION out_row_group_index)
+{
+  my_downsample_ptr downsample = (my_downsample_ptr) cinfo->downsample;
+  int ci;
+  jpeg_component_info * compptr;
+  JSAMPARRAY in_ptr, out_ptr;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    in_ptr = input_buf[ci] + in_row_index;
+    out_ptr = output_buf[ci] +
+	      (out_row_group_index * downsample->rowgroup_height[ci]);
+    (*downsample->methods[ci]) (cinfo, compptr, in_ptr, out_ptr);
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * One row group is processed per call.
+ * This version handles arbitrary integral sampling ratios, without smoothing.
+ * Note that this version is not actually used for customary sampling ratios.
+ */
+
+METHODDEF(void)
+int_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  my_downsample_ptr downsample = (my_downsample_ptr) cinfo->downsample;
+  int inrow, outrow, h_expand, v_expand, numpix, numpix2, h, v;
+  JDIMENSION outcol, outcol_h;	/* outcol_h == outcol*h_expand */
+  JDIMENSION output_cols = compptr->width_in_blocks * compptr->DCT_h_scaled_size;
+  JSAMPROW inptr, outptr;
+  INT32 outvalue;
+
+  h_expand = downsample->h_expand[compptr->component_index];
+  v_expand = downsample->v_expand[compptr->component_index];
+  numpix = h_expand * v_expand;
+  numpix2 = numpix/2;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * h_expand);
+
+  inrow = outrow = 0;
+  while (inrow < cinfo->max_v_samp_factor) {
+    outptr = output_data[outrow];
+    for (outcol = 0, outcol_h = 0; outcol < output_cols;
+	 outcol++, outcol_h += h_expand) {
+      outvalue = 0;
+      for (v = 0; v < v_expand; v++) {
+	inptr = input_data[inrow+v] + outcol_h;
+	for (h = 0; h < h_expand; h++) {
+	  outvalue += (INT32) GETJSAMPLE(*inptr++);
+	}
+      }
+      *outptr++ = (JSAMPLE) ((outvalue + numpix2) / numpix);
+    }
+    inrow += v_expand;
+    outrow++;
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the special case of a full-size component,
+ * without smoothing.
+ */
+
+METHODDEF(void)
+fullsize_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		     JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  /* Copy the data */
+  jcopy_sample_rows(input_data, 0, output_data, 0,
+		    cinfo->max_v_samp_factor, cinfo->image_width);
+  /* Edge-expand */
+  expand_right_edge(output_data, cinfo->max_v_samp_factor, cinfo->image_width,
+		    compptr->width_in_blocks * compptr->DCT_h_scaled_size);
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the common case of 2:1 horizontal and 1:1 vertical,
+ * without smoothing.
+ *
+ * A note about the "bias" calculations: when rounding fractional values to
+ * integer, we do not want to always round 0.5 up to the next integer.
+ * If we did that, we'd introduce a noticeable bias towards larger values.
+ * Instead, this code is arranged so that 0.5 will be rounded up or down at
+ * alternate pixel locations (a simple ordered dither pattern).
+ */
+
+METHODDEF(void)
+h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow;
+  JDIMENSION outcol;
+  JDIMENSION output_cols = compptr->width_in_blocks * compptr->DCT_h_scaled_size;
+  register JSAMPROW inptr, outptr;
+  register int bias;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * 2);
+
+  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
+    outptr = output_data[inrow];
+    inptr = input_data[inrow];
+    bias = 0;			/* bias = 0,1,0,1,... for successive samples */
+    for (outcol = 0; outcol < output_cols; outcol++) {
+      *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr) + GETJSAMPLE(inptr[1])
+			      + bias) >> 1);
+      bias ^= 1;		/* 0=>1, 1=>0 */
+      inptr += 2;
+    }
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * without smoothing.
+ */
+
+METHODDEF(void)
+h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow;
+  JDIMENSION outcol;
+  JDIMENSION output_cols = compptr->width_in_blocks * compptr->DCT_h_scaled_size;
+  register JSAMPROW inptr0, inptr1, outptr;
+  register int bias;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data, cinfo->max_v_samp_factor,
+		    cinfo->image_width, output_cols * 2);
+
+  inrow = outrow = 0;
+  while (inrow < cinfo->max_v_samp_factor) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow+1];
+    bias = 1;			/* bias = 1,2,1,2,... for successive samples */
+    for (outcol = 0; outcol < output_cols; outcol++) {
+      *outptr++ = (JSAMPLE) ((GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+			      GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1])
+			      + bias) >> 2);
+      bias ^= 3;		/* 1=>2, 2=>1 */
+      inptr0 += 2; inptr1 += 2;
+    }
+    inrow += 2;
+    outrow++;
+  }
+}
+
+
+#ifdef INPUT_SMOOTHING_SUPPORTED
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
+ * with smoothing.  One row of context is required.
+ */
+
+METHODDEF(void)
+h2v2_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
+			JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow, outrow;
+  JDIMENSION colctr;
+  JDIMENSION output_cols = compptr->width_in_blocks * compptr->DCT_h_scaled_size;
+  register JSAMPROW inptr0, inptr1, above_ptr, below_ptr, outptr;
+  INT32 membersum, neighsum, memberscale, neighscale;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
+		    cinfo->image_width, output_cols * 2);
+
+  /* We don't bother to form the individual "smoothed" input pixel values;
+   * we can directly compute the output which is the average of the four
+   * smoothed values.  Each of the four member pixels contributes a fraction
+   * (1-8*SF) to its own smoothed image and a fraction SF to each of the three
+   * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final
+   * output.  The four corner-adjacent neighbor pixels contribute a fraction
+   * SF to just one smoothed pixel, or SF/4 to the final output; while the
+   * eight edge-adjacent neighbors contribute SF to each of two smoothed
+   * pixels, or SF/2 overall.  In order to use integer arithmetic, these
+   * factors are scaled by 2^16 = 65536.
+   * Also recall that SF = smoothing_factor / 1024.
+   */
+
+  memberscale = 16384 - cinfo->smoothing_factor * 80; /* scaled (1-5*SF)/4 */
+  neighscale = cinfo->smoothing_factor * 16; /* scaled SF/4 */
+
+  inrow = outrow = 0;
+  while (inrow < cinfo->max_v_samp_factor) {
+    outptr = output_data[outrow];
+    inptr0 = input_data[inrow];
+    inptr1 = input_data[inrow+1];
+    above_ptr = input_data[inrow-1];
+    below_ptr = input_data[inrow+2];
+
+    /* Special case for first column: pretend column -1 is same as column 0 */
+    membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+    neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+	       GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[2]) +
+	       GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[2]);
+    neighsum += neighsum;
+    neighsum += GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[2]) +
+		GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[2]);
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+    inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
+
+    for (colctr = output_cols - 2; colctr > 0; colctr--) {
+      /* sum of pixels directly mapped to this output element */
+      membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		  GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+      /* sum of edge-neighbor pixels */
+      neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+		 GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+		 GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[2]) +
+		 GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[2]);
+      /* The edge-neighbors count twice as much as corner-neighbors */
+      neighsum += neighsum;
+      /* Add in the corner-neighbors */
+      neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[2]) +
+		  GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[2]);
+      /* form final output scaled up by 2^16 */
+      membersum = membersum * memberscale + neighsum * neighscale;
+      /* round, descale and output it */
+      *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+      inptr0 += 2; inptr1 += 2; above_ptr += 2; below_ptr += 2;
+    }
+
+    /* Special case for last column */
+    membersum = GETJSAMPLE(*inptr0) + GETJSAMPLE(inptr0[1]) +
+		GETJSAMPLE(*inptr1) + GETJSAMPLE(inptr1[1]);
+    neighsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(above_ptr[1]) +
+	       GETJSAMPLE(*below_ptr) + GETJSAMPLE(below_ptr[1]) +
+	       GETJSAMPLE(inptr0[-1]) + GETJSAMPLE(inptr0[1]) +
+	       GETJSAMPLE(inptr1[-1]) + GETJSAMPLE(inptr1[1]);
+    neighsum += neighsum;
+    neighsum += GETJSAMPLE(above_ptr[-1]) + GETJSAMPLE(above_ptr[1]) +
+		GETJSAMPLE(below_ptr[-1]) + GETJSAMPLE(below_ptr[1]);
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr = (JSAMPLE) ((membersum + 32768) >> 16);
+
+    inrow += 2;
+    outrow++;
+  }
+}
+
+
+/*
+ * Downsample pixel values of a single component.
+ * This version handles the special case of a full-size component,
+ * with smoothing.  One row of context is required.
+ */
+
+METHODDEF(void)
+fullsize_smooth_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr,
+			    JSAMPARRAY input_data, JSAMPARRAY output_data)
+{
+  int inrow;
+  JDIMENSION colctr;
+  JDIMENSION output_cols = compptr->width_in_blocks * compptr->DCT_h_scaled_size;
+  register JSAMPROW inptr, above_ptr, below_ptr, outptr;
+  INT32 membersum, neighsum, memberscale, neighscale;
+  int colsum, lastcolsum, nextcolsum;
+
+  /* Expand input data enough to let all the output samples be generated
+   * by the standard loop.  Special-casing padded output would be more
+   * efficient.
+   */
+  expand_right_edge(input_data - 1, cinfo->max_v_samp_factor + 2,
+		    cinfo->image_width, output_cols);
+
+  /* Each of the eight neighbor pixels contributes a fraction SF to the
+   * smoothed pixel, while the main pixel contributes (1-8*SF).  In order
+   * to use integer arithmetic, these factors are multiplied by 2^16 = 65536.
+   * Also recall that SF = smoothing_factor / 1024.
+   */
+
+  memberscale = 65536L - cinfo->smoothing_factor * 512L; /* scaled 1-8*SF */
+  neighscale = cinfo->smoothing_factor * 64; /* scaled SF */
+
+  for (inrow = 0; inrow < cinfo->max_v_samp_factor; inrow++) {
+    outptr = output_data[inrow];
+    inptr = input_data[inrow];
+    above_ptr = input_data[inrow-1];
+    below_ptr = input_data[inrow+1];
+
+    /* Special case for first column */
+    colsum = GETJSAMPLE(*above_ptr++) + GETJSAMPLE(*below_ptr++) +
+	     GETJSAMPLE(*inptr);
+    membersum = GETJSAMPLE(*inptr++);
+    nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
+		 GETJSAMPLE(*inptr);
+    neighsum = colsum + (colsum - membersum) + nextcolsum;
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+    lastcolsum = colsum; colsum = nextcolsum;
+
+    for (colctr = output_cols - 2; colctr > 0; colctr--) {
+      membersum = GETJSAMPLE(*inptr++);
+      above_ptr++; below_ptr++;
+      nextcolsum = GETJSAMPLE(*above_ptr) + GETJSAMPLE(*below_ptr) +
+		   GETJSAMPLE(*inptr);
+      neighsum = lastcolsum + (colsum - membersum) + nextcolsum;
+      membersum = membersum * memberscale + neighsum * neighscale;
+      *outptr++ = (JSAMPLE) ((membersum + 32768) >> 16);
+      lastcolsum = colsum; colsum = nextcolsum;
+    }
+
+    /* Special case for last column */
+    membersum = GETJSAMPLE(*inptr);
+    neighsum = lastcolsum + (colsum - membersum) + colsum;
+    membersum = membersum * memberscale + neighsum * neighscale;
+    *outptr = (JSAMPLE) ((membersum + 32768) >> 16);
+
+  }
+}
+
+#endif /* INPUT_SMOOTHING_SUPPORTED */
+
+
+/*
+ * Module initialization routine for downsampling.
+ * Note that we must select a routine for each component.
+ */
+
+GLOBAL(void)
+jinit_downsampler (j_compress_ptr cinfo)
+{
+  my_downsample_ptr downsample;
+  int ci;
+  jpeg_component_info * compptr;
+  boolean smoothok = TRUE;
+  int h_in_group, v_in_group, h_out_group, v_out_group;
+
+  downsample = (my_downsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_downsampler));
+  cinfo->downsample = (struct jpeg_downsampler *) downsample;
+  downsample->pub.start_pass = start_pass_downsample;
+  downsample->pub.downsample = sep_downsample;
+  downsample->pub.need_context_rows = FALSE;
+
+  if (cinfo->CCIR601_sampling)
+    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
+
+  /* Verify we can handle the sampling factors, and set up method pointers */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Compute size of an "output group" for DCT scaling.  This many samples
+     * are to be converted from max_h_samp_factor * max_v_samp_factor pixels.
+     */
+    h_out_group = (compptr->h_samp_factor * compptr->DCT_h_scaled_size) /
+		  cinfo->min_DCT_h_scaled_size;
+    v_out_group = (compptr->v_samp_factor * compptr->DCT_v_scaled_size) /
+		  cinfo->min_DCT_v_scaled_size;
+    h_in_group = cinfo->max_h_samp_factor;
+    v_in_group = cinfo->max_v_samp_factor;
+    downsample->rowgroup_height[ci] = v_out_group; /* save for use later */
+    if (h_in_group == h_out_group && v_in_group == v_out_group) {
+#ifdef INPUT_SMOOTHING_SUPPORTED
+      if (cinfo->smoothing_factor) {
+	downsample->methods[ci] = fullsize_smooth_downsample;
+	downsample->pub.need_context_rows = TRUE;
+      } else
+#endif
+	downsample->methods[ci] = fullsize_downsample;
+    } else if (h_in_group == h_out_group * 2 &&
+	       v_in_group == v_out_group) {
+      smoothok = FALSE;
+      downsample->methods[ci] = h2v1_downsample;
+    } else if (h_in_group == h_out_group * 2 &&
+	       v_in_group == v_out_group * 2) {
+#ifdef INPUT_SMOOTHING_SUPPORTED
+      if (cinfo->smoothing_factor) {
+	downsample->methods[ci] = h2v2_smooth_downsample;
+	downsample->pub.need_context_rows = TRUE;
+      } else
+#endif
+	downsample->methods[ci] = h2v2_downsample;
+    } else if ((h_in_group % h_out_group) == 0 &&
+	       (v_in_group % v_out_group) == 0) {
+      smoothok = FALSE;
+      downsample->methods[ci] = int_downsample;
+      downsample->h_expand[ci] = (UINT8) (h_in_group / h_out_group);
+      downsample->v_expand[ci] = (UINT8) (v_in_group / v_out_group);
+    } else
+      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
+  }
+
+#ifdef INPUT_SMOOTHING_SUPPORTED
+  if (cinfo->smoothing_factor && !smoothok)
+    TRACEMS(cinfo, 0, JTRC_SMOOTH_NOTIMPL);
+#endif
+}

Added: trunk/code/jpeg-8c/jctrans.c
===================================================================
--- trunk/code/jpeg-8c/jctrans.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jctrans.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,382 @@
+/*
+ * jctrans.c
+ *
+ * Copyright (C) 1995-1998, Thomas G. Lane.
+ * Modified 2000-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains library routines for transcoding compression,
+ * that is, writing raw DCT coefficient arrays to an output JPEG file.
+ * The routines in jcapimin.c will also be needed by a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Forward declarations */
+LOCAL(void) transencode_master_selection
+	JPP((j_compress_ptr cinfo, jvirt_barray_ptr * coef_arrays));
+LOCAL(void) transencode_coef_controller
+	JPP((j_compress_ptr cinfo, jvirt_barray_ptr * coef_arrays));
+
+
+/*
+ * Compression initialization for writing raw-coefficient data.
+ * Before calling this, all parameters and a data destination must be set up.
+ * Call jpeg_finish_compress() to actually write the data.
+ *
+ * The number of passed virtual arrays must match cinfo->num_components.
+ * Note that the virtual arrays need not be filled or even realized at
+ * the time write_coefficients is called; indeed, if the virtual arrays
+ * were requested from this compression object's memory manager, they
+ * typically will be realized during this routine and filled afterwards.
+ */
+
+GLOBAL(void)
+jpeg_write_coefficients (j_compress_ptr cinfo, jvirt_barray_ptr * coef_arrays)
+{
+  if (cinfo->global_state != CSTATE_START)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Mark all tables to be written */
+  jpeg_suppress_tables(cinfo, FALSE);
+  /* (Re)initialize error mgr and destination modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->dest->init_destination) (cinfo);
+  /* Perform master selection of active modules */
+  transencode_master_selection(cinfo, coef_arrays);
+  /* Wait for jpeg_finish_compress() call */
+  cinfo->next_scanline = 0;	/* so jpeg_write_marker works */
+  cinfo->global_state = CSTATE_WRCOEFS;
+}
+
+
+/*
+ * Initialize the compression object with default parameters,
+ * then copy from the source object all parameters needed for lossless
+ * transcoding.  Parameters that can be varied without loss (such as
+ * scan script and Huffman optimization) are left in their default states.
+ */
+
+GLOBAL(void)
+jpeg_copy_critical_parameters (j_decompress_ptr srcinfo,
+			       j_compress_ptr dstinfo)
+{
+  JQUANT_TBL ** qtblptr;
+  jpeg_component_info *incomp, *outcomp;
+  JQUANT_TBL *c_quant, *slot_quant;
+  int tblno, ci, coefi;
+
+  /* Safety check to ensure start_compress not called yet. */
+  if (dstinfo->global_state != CSTATE_START)
+    ERREXIT1(dstinfo, JERR_BAD_STATE, dstinfo->global_state);
+  /* Copy fundamental image dimensions */
+  dstinfo->image_width = srcinfo->image_width;
+  dstinfo->image_height = srcinfo->image_height;
+  dstinfo->input_components = srcinfo->num_components;
+  dstinfo->in_color_space = srcinfo->jpeg_color_space;
+  dstinfo->jpeg_width = srcinfo->output_width;
+  dstinfo->jpeg_height = srcinfo->output_height;
+  dstinfo->min_DCT_h_scaled_size = srcinfo->min_DCT_h_scaled_size;
+  dstinfo->min_DCT_v_scaled_size = srcinfo->min_DCT_v_scaled_size;
+  /* Initialize all parameters to default values */
+  jpeg_set_defaults(dstinfo);
+  /* jpeg_set_defaults may choose wrong colorspace, eg YCbCr if input is RGB.
+   * Fix it to get the right header markers for the image colorspace.
+   */
+  jpeg_set_colorspace(dstinfo, srcinfo->jpeg_color_space);
+  dstinfo->data_precision = srcinfo->data_precision;
+  dstinfo->CCIR601_sampling = srcinfo->CCIR601_sampling;
+  /* Copy the source's quantization tables. */
+  for (tblno = 0; tblno < NUM_QUANT_TBLS; tblno++) {
+    if (srcinfo->quant_tbl_ptrs[tblno] != NULL) {
+      qtblptr = & dstinfo->quant_tbl_ptrs[tblno];
+      if (*qtblptr == NULL)
+	*qtblptr = jpeg_alloc_quant_table((j_common_ptr) dstinfo);
+      MEMCOPY((*qtblptr)->quantval,
+	      srcinfo->quant_tbl_ptrs[tblno]->quantval,
+	      SIZEOF((*qtblptr)->quantval));
+      (*qtblptr)->sent_table = FALSE;
+    }
+  }
+  /* Copy the source's per-component info.
+   * Note we assume jpeg_set_defaults has allocated the dest comp_info array.
+   */
+  dstinfo->num_components = srcinfo->num_components;
+  if (dstinfo->num_components < 1 || dstinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(dstinfo, JERR_COMPONENT_COUNT, dstinfo->num_components,
+	     MAX_COMPONENTS);
+  for (ci = 0, incomp = srcinfo->comp_info, outcomp = dstinfo->comp_info;
+       ci < dstinfo->num_components; ci++, incomp++, outcomp++) {
+    outcomp->component_id = incomp->component_id;
+    outcomp->h_samp_factor = incomp->h_samp_factor;
+    outcomp->v_samp_factor = incomp->v_samp_factor;
+    outcomp->quant_tbl_no = incomp->quant_tbl_no;
+    /* Make sure saved quantization table for component matches the qtable
+     * slot.  If not, the input file re-used this qtable slot.
+     * IJG encoder currently cannot duplicate this.
+     */
+    tblno = outcomp->quant_tbl_no;
+    if (tblno < 0 || tblno >= NUM_QUANT_TBLS ||
+	srcinfo->quant_tbl_ptrs[tblno] == NULL)
+      ERREXIT1(dstinfo, JERR_NO_QUANT_TABLE, tblno);
+    slot_quant = srcinfo->quant_tbl_ptrs[tblno];
+    c_quant = incomp->quant_table;
+    if (c_quant != NULL) {
+      for (coefi = 0; coefi < DCTSIZE2; coefi++) {
+	if (c_quant->quantval[coefi] != slot_quant->quantval[coefi])
+	  ERREXIT1(dstinfo, JERR_MISMATCHED_QUANT_TABLE, tblno);
+      }
+    }
+    /* Note: we do not copy the source's Huffman table assignments;
+     * instead we rely on jpeg_set_colorspace to have made a suitable choice.
+     */
+  }
+  /* Also copy JFIF version and resolution information, if available.
+   * Strictly speaking this isn't "critical" info, but it's nearly
+   * always appropriate to copy it if available.  In particular,
+   * if the application chooses to copy JFIF 1.02 extension markers from
+   * the source file, we need to copy the version to make sure we don't
+   * emit a file that has 1.02 extensions but a claimed version of 1.01.
+   * We will *not*, however, copy version info from mislabeled "2.01" files.
+   */
+  if (srcinfo->saw_JFIF_marker) {
+    if (srcinfo->JFIF_major_version == 1) {
+      dstinfo->JFIF_major_version = srcinfo->JFIF_major_version;
+      dstinfo->JFIF_minor_version = srcinfo->JFIF_minor_version;
+    }
+    dstinfo->density_unit = srcinfo->density_unit;
+    dstinfo->X_density = srcinfo->X_density;
+    dstinfo->Y_density = srcinfo->Y_density;
+  }
+}
+
+
+/*
+ * Master selection of compression modules for transcoding.
+ * This substitutes for jcinit.c's initialization of the full compressor.
+ */
+
+LOCAL(void)
+transencode_master_selection (j_compress_ptr cinfo,
+			      jvirt_barray_ptr * coef_arrays)
+{
+  /* Initialize master control (includes parameter checking/processing) */
+  jinit_c_master_control(cinfo, TRUE /* transcode only */);
+
+  /* Entropy encoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code)
+    jinit_arith_encoder(cinfo);
+  else {
+    jinit_huff_encoder(cinfo);
+  }
+
+  /* We need a special coefficient buffer controller. */
+  transencode_coef_controller(cinfo, coef_arrays);
+
+  jinit_marker_writer(cinfo);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Write the datastream header (SOI, JFIF) immediately.
+   * Frame and scan headers are postponed till later.
+   * This lets application insert special markers after the SOI.
+   */
+  (*cinfo->marker->write_file_header) (cinfo);
+}
+
+
+/*
+ * The rest of this file is a special implementation of the coefficient
+ * buffer controller.  This is similar to jccoefct.c, but it handles only
+ * output from presupplied virtual arrays.  Furthermore, we generate any
+ * dummy padding blocks on-the-fly rather than expecting them to be present
+ * in the arrays.
+ */
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_c_coef_controller pub; /* public fields */
+
+  JDIMENSION iMCU_row_num;	/* iMCU row # within image */
+  JDIMENSION mcu_ctr;		/* counts MCUs processed in current row */
+  int MCU_vert_offset;		/* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;	/* number of such rows needed */
+
+  /* Virtual block array for each component. */
+  jvirt_barray_ptr * whole_image;
+
+  /* Workspace for constructing dummy blocks at right/bottom edges. */
+  JBLOCKROW dummy_buffer[C_MAX_BLOCKS_IN_MCU];
+} my_coef_controller;
+
+typedef my_coef_controller * my_coef_ptr;
+
+
+LOCAL(void)
+start_iMCU_row (j_compress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row */
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (coef->iMCU_row_num < (cinfo->total_iMCU_rows-1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->mcu_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_coef (j_compress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  if (pass_mode != JBUF_CRANK_DEST)
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  coef->iMCU_row_num = 0;
+  start_iMCU_row(cinfo);
+}
+
+
+/*
+ * Process some data.
+ * We process the equivalent of one fully interleaved MCU row ("iMCU" row)
+ * per call, ie, v_samp_factor block rows for each component in the scan.
+ * The data is obtained from the virtual arrays and fed to the entropy coder.
+ * Returns TRUE if the iMCU row is completed, FALSE if suspended.
+ *
+ * NB: input_buf is ignored; it is likely to be a NULL pointer.
+ */
+
+METHODDEF(boolean)
+compress_output (j_compress_ptr cinfo, JSAMPIMAGE input_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, ci, xindex, yindex, yoffset, blockcnt;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW MCU_buffer[C_MAX_BLOCKS_IN_MCU];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan. */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
+       coef->iMCU_row_num * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, FALSE);
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->mcu_ctr; MCU_col_num < cinfo->MCUs_per_row;
+	 MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	start_col = MCU_col_num * compptr->MCU_width;
+	blockcnt = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
+						: compptr->last_col_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  if (coef->iMCU_row_num < last_iMCU_row ||
+	      yindex+yoffset < compptr->last_row_height) {
+	    /* Fill in pointers to real blocks in this row */
+	    buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
+	    for (xindex = 0; xindex < blockcnt; xindex++)
+	      MCU_buffer[blkn++] = buffer_ptr++;
+	  } else {
+	    /* At bottom of image, need a whole row of dummy blocks */
+	    xindex = 0;
+	  }
+	  /* Fill in any dummy blocks needed in this row.
+	   * Dummy blocks are filled in the same way as in jccoefct.c:
+	   * all zeroes in the AC entries, DC entries equal to previous
+	   * block's DC value.  The init routine has already zeroed the
+	   * AC entries, so we need only set the DC entries correctly.
+	   */
+	  for (; xindex < compptr->MCU_width; xindex++) {
+	    MCU_buffer[blkn] = coef->dummy_buffer[blkn];
+	    MCU_buffer[blkn][0][0] = MCU_buffer[blkn-1][0][0];
+	    blkn++;
+	  }
+	}
+      }
+      /* Try to write the MCU. */
+      if (! (*cinfo->entropy->encode_mcu) (cinfo, MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->mcu_ctr = MCU_col_num;
+	return FALSE;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->mcu_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  coef->iMCU_row_num++;
+  start_iMCU_row(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Initialize coefficient buffer controller.
+ *
+ * Each passed coefficient array must be the right size for that
+ * coefficient: width_in_blocks wide and height_in_blocks high,
+ * with unitheight at least v_samp_factor.
+ */
+
+LOCAL(void)
+transencode_coef_controller (j_compress_ptr cinfo,
+			     jvirt_barray_ptr * coef_arrays)
+{
+  my_coef_ptr coef;
+  JBLOCKROW buffer;
+  int i;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_coef_controller));
+  cinfo->coef = (struct jpeg_c_coef_controller *) coef;
+  coef->pub.start_pass = start_pass_coef;
+  coef->pub.compress_data = compress_output;
+
+  /* Save pointer to virtual arrays */
+  coef->whole_image = coef_arrays;
+
+  /* Allocate and pre-zero space for dummy DCT blocks. */
+  buffer = (JBLOCKROW)
+    (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				C_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+  jzero_far((void FAR *) buffer, C_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+  for (i = 0; i < C_MAX_BLOCKS_IN_MCU; i++) {
+    coef->dummy_buffer[i] = buffer + i;
+  }
+}

Added: trunk/code/jpeg-8c/jdapimin.c
===================================================================
--- trunk/code/jpeg-8c/jdapimin.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdapimin.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,396 @@
+/*
+ * jdapimin.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * Modified 2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the decompression half
+ * of the JPEG library.  These are the "minimum" API routines that may be
+ * needed in either the normal full-decompression case or the
+ * transcoding-only case.
+ *
+ * Most of the routines intended to be called directly by an application
+ * are in this file or in jdapistd.c.  But also see jcomapi.c for routines
+ * shared by compression and decompression, and jdtrans.c for the transcoding
+ * case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * Initialization of a JPEG decompression object.
+ * The error manager must already be set up (in case memory manager fails).
+ */
+
+GLOBAL(void)
+jpeg_CreateDecompress (j_decompress_ptr cinfo, int version, size_t structsize)
+{
+  int i;
+
+  /* Guard against version mismatches between library and caller. */
+  cinfo->mem = NULL;		/* so jpeg_destroy knows mem mgr not called */
+  if (version != JPEG_LIB_VERSION)
+    ERREXIT2(cinfo, JERR_BAD_LIB_VERSION, JPEG_LIB_VERSION, version);
+  if (structsize != SIZEOF(struct jpeg_decompress_struct))
+    ERREXIT2(cinfo, JERR_BAD_STRUCT_SIZE, 
+	     (int) SIZEOF(struct jpeg_decompress_struct), (int) structsize);
+
+  /* For debugging purposes, we zero the whole master structure.
+   * But the application has already set the err pointer, and may have set
+   * client_data, so we have to save and restore those fields.
+   * Note: if application hasn't set client_data, tools like Purify may
+   * complain here.
+   */
+  {
+    struct jpeg_error_mgr * err = cinfo->err;
+    void * client_data = cinfo->client_data; /* ignore Purify complaint here */
+    MEMZERO(cinfo, SIZEOF(struct jpeg_decompress_struct));
+    cinfo->err = err;
+    cinfo->client_data = client_data;
+  }
+  cinfo->is_decompressor = TRUE;
+
+  /* Initialize a memory manager instance for this object */
+  jinit_memory_mgr((j_common_ptr) cinfo);
+
+  /* Zero out pointers to permanent structures. */
+  cinfo->progress = NULL;
+  cinfo->src = NULL;
+
+  for (i = 0; i < NUM_QUANT_TBLS; i++)
+    cinfo->quant_tbl_ptrs[i] = NULL;
+
+  for (i = 0; i < NUM_HUFF_TBLS; i++) {
+    cinfo->dc_huff_tbl_ptrs[i] = NULL;
+    cinfo->ac_huff_tbl_ptrs[i] = NULL;
+  }
+
+  /* Initialize marker processor so application can override methods
+   * for COM, APPn markers before calling jpeg_read_header.
+   */
+  cinfo->marker_list = NULL;
+  jinit_marker_reader(cinfo);
+
+  /* And initialize the overall input controller. */
+  jinit_input_controller(cinfo);
+
+  /* OK, I'm ready */
+  cinfo->global_state = DSTATE_START;
+}
+
+
+/*
+ * Destruction of a JPEG decompression object
+ */
+
+GLOBAL(void)
+jpeg_destroy_decompress (j_decompress_ptr cinfo)
+{
+  jpeg_destroy((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Abort processing of a JPEG decompression operation,
+ * but don't destroy the object itself.
+ */
+
+GLOBAL(void)
+jpeg_abort_decompress (j_decompress_ptr cinfo)
+{
+  jpeg_abort((j_common_ptr) cinfo); /* use common routine */
+}
+
+
+/*
+ * Set default decompression parameters.
+ */
+
+LOCAL(void)
+default_decompress_parms (j_decompress_ptr cinfo)
+{
+  /* Guess the input colorspace, and set output colorspace accordingly. */
+  /* (Wish JPEG committee had provided a real way to specify this...) */
+  /* Note application may override our guesses. */
+  switch (cinfo->num_components) {
+  case 1:
+    cinfo->jpeg_color_space = JCS_GRAYSCALE;
+    cinfo->out_color_space = JCS_GRAYSCALE;
+    break;
+    
+  case 3:
+    if (cinfo->saw_JFIF_marker) {
+      cinfo->jpeg_color_space = JCS_YCbCr; /* JFIF implies YCbCr */
+    } else if (cinfo->saw_Adobe_marker) {
+      switch (cinfo->Adobe_transform) {
+      case 0:
+	cinfo->jpeg_color_space = JCS_RGB;
+	break;
+      case 1:
+	cinfo->jpeg_color_space = JCS_YCbCr;
+	break;
+      default:
+	WARNMS1(cinfo, JWRN_ADOBE_XFORM, cinfo->Adobe_transform);
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume it's YCbCr */
+	break;
+      }
+    } else {
+      /* Saw no special markers, try to guess from the component IDs */
+      int cid0 = cinfo->comp_info[0].component_id;
+      int cid1 = cinfo->comp_info[1].component_id;
+      int cid2 = cinfo->comp_info[2].component_id;
+
+      if (cid0 == 1 && cid1 == 2 && cid2 == 3)
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume JFIF w/out marker */
+      else if (cid0 == 82 && cid1 == 71 && cid2 == 66)
+	cinfo->jpeg_color_space = JCS_RGB; /* ASCII 'R', 'G', 'B' */
+      else {
+	TRACEMS3(cinfo, 1, JTRC_UNKNOWN_IDS, cid0, cid1, cid2);
+	cinfo->jpeg_color_space = JCS_YCbCr; /* assume it's YCbCr */
+      }
+    }
+    /* Always guess RGB is proper output colorspace. */
+    cinfo->out_color_space = JCS_RGB;
+    break;
+    
+  case 4:
+    if (cinfo->saw_Adobe_marker) {
+      switch (cinfo->Adobe_transform) {
+      case 0:
+	cinfo->jpeg_color_space = JCS_CMYK;
+	break;
+      case 2:
+	cinfo->jpeg_color_space = JCS_YCCK;
+	break;
+      default:
+	WARNMS1(cinfo, JWRN_ADOBE_XFORM, cinfo->Adobe_transform);
+	cinfo->jpeg_color_space = JCS_YCCK; /* assume it's YCCK */
+	break;
+      }
+    } else {
+      /* No special markers, assume straight CMYK. */
+      cinfo->jpeg_color_space = JCS_CMYK;
+    }
+    cinfo->out_color_space = JCS_CMYK;
+    break;
+    
+  default:
+    cinfo->jpeg_color_space = JCS_UNKNOWN;
+    cinfo->out_color_space = JCS_UNKNOWN;
+    break;
+  }
+
+  /* Set defaults for other decompression parameters. */
+  cinfo->scale_num = cinfo->block_size;		/* 1:1 scaling */
+  cinfo->scale_denom = cinfo->block_size;
+  cinfo->output_gamma = 1.0;
+  cinfo->buffered_image = FALSE;
+  cinfo->raw_data_out = FALSE;
+  cinfo->dct_method = JDCT_DEFAULT;
+  cinfo->do_fancy_upsampling = TRUE;
+  cinfo->do_block_smoothing = TRUE;
+  cinfo->quantize_colors = FALSE;
+  /* We set these in case application only sets quantize_colors. */
+  cinfo->dither_mode = JDITHER_FS;
+#ifdef QUANT_2PASS_SUPPORTED
+  cinfo->two_pass_quantize = TRUE;
+#else
+  cinfo->two_pass_quantize = FALSE;
+#endif
+  cinfo->desired_number_of_colors = 256;
+  cinfo->colormap = NULL;
+  /* Initialize for no mode change in buffered-image mode. */
+  cinfo->enable_1pass_quant = FALSE;
+  cinfo->enable_external_quant = FALSE;
+  cinfo->enable_2pass_quant = FALSE;
+}
+
+
+/*
+ * Decompression startup: read start of JPEG datastream to see what's there.
+ * Need only initialize JPEG object and supply a data source before calling.
+ *
+ * This routine will read as far as the first SOS marker (ie, actual start of
+ * compressed data), and will save all tables and parameters in the JPEG
+ * object.  It will also initialize the decompression parameters to default
+ * values, and finally return JPEG_HEADER_OK.  On return, the application may
+ * adjust the decompression parameters and then call jpeg_start_decompress.
+ * (Or, if the application only wanted to determine the image parameters,
+ * the data need not be decompressed.  In that case, call jpeg_abort or
+ * jpeg_destroy to release any temporary space.)
+ * If an abbreviated (tables only) datastream is presented, the routine will
+ * return JPEG_HEADER_TABLES_ONLY upon reaching EOI.  The application may then
+ * re-use the JPEG object to read the abbreviated image datastream(s).
+ * It is unnecessary (but OK) to call jpeg_abort in this case.
+ * The JPEG_SUSPENDED return code only occurs if the data source module
+ * requests suspension of the decompressor.  In this case the application
+ * should load more source data and then re-call jpeg_read_header to resume
+ * processing.
+ * If a non-suspending data source is used and require_image is TRUE, then the
+ * return code need not be inspected since only JPEG_HEADER_OK is possible.
+ *
+ * This routine is now just a front end to jpeg_consume_input, with some
+ * extra error checking.
+ */
+
+GLOBAL(int)
+jpeg_read_header (j_decompress_ptr cinfo, boolean require_image)
+{
+  int retcode;
+
+  if (cinfo->global_state != DSTATE_START &&
+      cinfo->global_state != DSTATE_INHEADER)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  retcode = jpeg_consume_input(cinfo);
+
+  switch (retcode) {
+  case JPEG_REACHED_SOS:
+    retcode = JPEG_HEADER_OK;
+    break;
+  case JPEG_REACHED_EOI:
+    if (require_image)		/* Complain if application wanted an image */
+      ERREXIT(cinfo, JERR_NO_IMAGE);
+    /* Reset to start state; it would be safer to require the application to
+     * call jpeg_abort, but we can't change it now for compatibility reasons.
+     * A side effect is to free any temporary memory (there shouldn't be any).
+     */
+    jpeg_abort((j_common_ptr) cinfo); /* sets state = DSTATE_START */
+    retcode = JPEG_HEADER_TABLES_ONLY;
+    break;
+  case JPEG_SUSPENDED:
+    /* no work */
+    break;
+  }
+
+  return retcode;
+}
+
+
+/*
+ * Consume data in advance of what the decompressor requires.
+ * This can be called at any time once the decompressor object has
+ * been created and a data source has been set up.
+ *
+ * This routine is essentially a state machine that handles a couple
+ * of critical state-transition actions, namely initial setup and
+ * transition from header scanning to ready-for-start_decompress.
+ * All the actual input is done via the input controller's consume_input
+ * method.
+ */
+
+GLOBAL(int)
+jpeg_consume_input (j_decompress_ptr cinfo)
+{
+  int retcode = JPEG_SUSPENDED;
+
+  /* NB: every possible DSTATE value should be listed in this switch */
+  switch (cinfo->global_state) {
+  case DSTATE_START:
+    /* Start-of-datastream actions: reset appropriate modules */
+    (*cinfo->inputctl->reset_input_controller) (cinfo);
+    /* Initialize application's data source module */
+    (*cinfo->src->init_source) (cinfo);
+    cinfo->global_state = DSTATE_INHEADER;
+    /*FALLTHROUGH*/
+  case DSTATE_INHEADER:
+    retcode = (*cinfo->inputctl->consume_input) (cinfo);
+    if (retcode == JPEG_REACHED_SOS) { /* Found SOS, prepare to decompress */
+      /* Set up default parameters based on header data */
+      default_decompress_parms(cinfo);
+      /* Set global state: ready for start_decompress */
+      cinfo->global_state = DSTATE_READY;
+    }
+    break;
+  case DSTATE_READY:
+    /* Can't advance past first SOS until start_decompress is called */
+    retcode = JPEG_REACHED_SOS;
+    break;
+  case DSTATE_PRELOAD:
+  case DSTATE_PRESCAN:
+  case DSTATE_SCANNING:
+  case DSTATE_RAW_OK:
+  case DSTATE_BUFIMAGE:
+  case DSTATE_BUFPOST:
+  case DSTATE_STOPPING:
+    retcode = (*cinfo->inputctl->consume_input) (cinfo);
+    break;
+  default:
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  return retcode;
+}
+
+
+/*
+ * Have we finished reading the input file?
+ */
+
+GLOBAL(boolean)
+jpeg_input_complete (j_decompress_ptr cinfo)
+{
+  /* Check for valid jpeg object */
+  if (cinfo->global_state < DSTATE_START ||
+      cinfo->global_state > DSTATE_STOPPING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return cinfo->inputctl->eoi_reached;
+}
+
+
+/*
+ * Is there more than one scan?
+ */
+
+GLOBAL(boolean)
+jpeg_has_multiple_scans (j_decompress_ptr cinfo)
+{
+  /* Only valid after jpeg_read_header completes */
+  if (cinfo->global_state < DSTATE_READY ||
+      cinfo->global_state > DSTATE_STOPPING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return cinfo->inputctl->has_multiple_scans;
+}
+
+
+/*
+ * Finish JPEG decompression.
+ *
+ * This will normally just verify the file trailer and release temp storage.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_finish_decompress (j_decompress_ptr cinfo)
+{
+  if ((cinfo->global_state == DSTATE_SCANNING ||
+       cinfo->global_state == DSTATE_RAW_OK) && ! cinfo->buffered_image) {
+    /* Terminate final pass of non-buffered mode */
+    if (cinfo->output_scanline < cinfo->output_height)
+      ERREXIT(cinfo, JERR_TOO_LITTLE_DATA);
+    (*cinfo->master->finish_output_pass) (cinfo);
+    cinfo->global_state = DSTATE_STOPPING;
+  } else if (cinfo->global_state == DSTATE_BUFIMAGE) {
+    /* Finishing after a buffered-image operation */
+    cinfo->global_state = DSTATE_STOPPING;
+  } else if (cinfo->global_state != DSTATE_STOPPING) {
+    /* STOPPING = repeat call after a suspension, anything else is error */
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  /* Read until EOI */
+  while (! cinfo->inputctl->eoi_reached) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return FALSE;		/* Suspend, come back later */
+  }
+  /* Do final cleanup */
+  (*cinfo->src->term_source) (cinfo);
+  /* We can use jpeg_abort to release memory and reset global_state */
+  jpeg_abort((j_common_ptr) cinfo);
+  return TRUE;
+}

Added: trunk/code/jpeg-8c/jdapistd.c
===================================================================
--- trunk/code/jpeg-8c/jdapistd.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdapistd.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,275 @@
+/*
+ * jdapistd.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains application interface code for the decompression half
+ * of the JPEG library.  These are the "standard" API routines that are
+ * used in the normal full-decompression case.  They are not used by a
+ * transcoding-only application.  Note that if an application links in
+ * jpeg_start_decompress, it will end up linking in the entire decompressor.
+ * We thus must separate this file from jdapimin.c to avoid linking the
+ * whole decompression library into a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Forward declarations */
+LOCAL(boolean) output_pass_setup JPP((j_decompress_ptr cinfo));
+
+
+/*
+ * Decompression initialization.
+ * jpeg_read_header must be completed before calling this.
+ *
+ * If a multipass operating mode was selected, this will do all but the
+ * last pass, and thus may take a great deal of time.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_start_decompress (j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state == DSTATE_READY) {
+    /* First call: initialize master control, select active modules */
+    jinit_master_decompress(cinfo);
+    if (cinfo->buffered_image) {
+      /* No more work here; expecting jpeg_start_output next */
+      cinfo->global_state = DSTATE_BUFIMAGE;
+      return TRUE;
+    }
+    cinfo->global_state = DSTATE_PRELOAD;
+  }
+  if (cinfo->global_state == DSTATE_PRELOAD) {
+    /* If file has multiple scans, absorb them all into the coef buffer */
+    if (cinfo->inputctl->has_multiple_scans) {
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+      for (;;) {
+	int retcode;
+	/* Call progress monitor hook if present */
+	if (cinfo->progress != NULL)
+	  (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+	/* Absorb some more input */
+	retcode = (*cinfo->inputctl->consume_input) (cinfo);
+	if (retcode == JPEG_SUSPENDED)
+	  return FALSE;
+	if (retcode == JPEG_REACHED_EOI)
+	  break;
+	/* Advance progress counter if appropriate */
+	if (cinfo->progress != NULL &&
+	    (retcode == JPEG_ROW_COMPLETED || retcode == JPEG_REACHED_SOS)) {
+	  if (++cinfo->progress->pass_counter >= cinfo->progress->pass_limit) {
+	    /* jdmaster underestimated number of scans; ratchet up one scan */
+	    cinfo->progress->pass_limit += (long) cinfo->total_iMCU_rows;
+	  }
+	}
+      }
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+    }
+    cinfo->output_scan_number = cinfo->input_scan_number;
+  } else if (cinfo->global_state != DSTATE_PRESCAN)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Perform any dummy output passes, and set up for the final pass */
+  return output_pass_setup(cinfo);
+}
+
+
+/*
+ * Set up for an output pass, and perform any dummy pass(es) needed.
+ * Common subroutine for jpeg_start_decompress and jpeg_start_output.
+ * Entry: global_state = DSTATE_PRESCAN only if previously suspended.
+ * Exit: If done, returns TRUE and sets global_state for proper output mode.
+ *       If suspended, returns FALSE and sets global_state = DSTATE_PRESCAN.
+ */
+
+LOCAL(boolean)
+output_pass_setup (j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state != DSTATE_PRESCAN) {
+    /* First call: do pass setup */
+    (*cinfo->master->prepare_for_output_pass) (cinfo);
+    cinfo->output_scanline = 0;
+    cinfo->global_state = DSTATE_PRESCAN;
+  }
+  /* Loop over any required dummy passes */
+  while (cinfo->master->is_dummy_pass) {
+#ifdef QUANT_2PASS_SUPPORTED
+    /* Crank through the dummy pass */
+    while (cinfo->output_scanline < cinfo->output_height) {
+      JDIMENSION last_scanline;
+      /* Call progress monitor hook if present */
+      if (cinfo->progress != NULL) {
+	cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+	cinfo->progress->pass_limit = (long) cinfo->output_height;
+	(*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+      }
+      /* Process some data */
+      last_scanline = cinfo->output_scanline;
+      (*cinfo->main->process_data) (cinfo, (JSAMPARRAY) NULL,
+				    &cinfo->output_scanline, (JDIMENSION) 0);
+      if (cinfo->output_scanline == last_scanline)
+	return FALSE;		/* No progress made, must suspend */
+    }
+    /* Finish up dummy pass, and set up for another one */
+    (*cinfo->master->finish_output_pass) (cinfo);
+    (*cinfo->master->prepare_for_output_pass) (cinfo);
+    cinfo->output_scanline = 0;
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* QUANT_2PASS_SUPPORTED */
+  }
+  /* Ready for application to drive output pass through
+   * jpeg_read_scanlines or jpeg_read_raw_data.
+   */
+  cinfo->global_state = cinfo->raw_data_out ? DSTATE_RAW_OK : DSTATE_SCANNING;
+  return TRUE;
+}
+
+
+/*
+ * Read some scanlines of data from the JPEG decompressor.
+ *
+ * The return value will be the number of lines actually read.
+ * This may be less than the number requested in several cases,
+ * including bottom of image, data source suspension, and operating
+ * modes that emit multiple scanlines at a time.
+ *
+ * Note: we warn about excess calls to jpeg_read_scanlines() since
+ * this likely signals an application programmer error.  However,
+ * an oversize buffer (max_lines > scanlines remaining) is not an error.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_read_scanlines (j_decompress_ptr cinfo, JSAMPARRAY scanlines,
+		     JDIMENSION max_lines)
+{
+  JDIMENSION row_ctr;
+
+  if (cinfo->global_state != DSTATE_SCANNING)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->output_scanline >= cinfo->output_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->output_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Process some data */
+  row_ctr = 0;
+  (*cinfo->main->process_data) (cinfo, scanlines, &row_ctr, max_lines);
+  cinfo->output_scanline += row_ctr;
+  return row_ctr;
+}
+
+
+/*
+ * Alternate entry point to read raw data.
+ * Processes exactly one iMCU row per call, unless suspended.
+ */
+
+GLOBAL(JDIMENSION)
+jpeg_read_raw_data (j_decompress_ptr cinfo, JSAMPIMAGE data,
+		    JDIMENSION max_lines)
+{
+  JDIMENSION lines_per_iMCU_row;
+
+  if (cinfo->global_state != DSTATE_RAW_OK)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  if (cinfo->output_scanline >= cinfo->output_height) {
+    WARNMS(cinfo, JWRN_TOO_MUCH_DATA);
+    return 0;
+  }
+
+  /* Call progress monitor hook if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->pass_counter = (long) cinfo->output_scanline;
+    cinfo->progress->pass_limit = (long) cinfo->output_height;
+    (*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+  }
+
+  /* Verify that at least one iMCU row can be returned. */
+  lines_per_iMCU_row = cinfo->max_v_samp_factor * cinfo->min_DCT_v_scaled_size;
+  if (max_lines < lines_per_iMCU_row)
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* Decompress directly into user's buffer. */
+  if (! (*cinfo->coef->decompress_data) (cinfo, data))
+    return 0;			/* suspension forced, can do nothing more */
+
+  /* OK, we processed one iMCU row. */
+  cinfo->output_scanline += lines_per_iMCU_row;
+  return lines_per_iMCU_row;
+}
+
+
+/* Additional entry points for buffered-image mode. */
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Initialize for an output pass in buffered-image mode.
+ */
+
+GLOBAL(boolean)
+jpeg_start_output (j_decompress_ptr cinfo, int scan_number)
+{
+  if (cinfo->global_state != DSTATE_BUFIMAGE &&
+      cinfo->global_state != DSTATE_PRESCAN)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  /* Limit scan number to valid range */
+  if (scan_number <= 0)
+    scan_number = 1;
+  if (cinfo->inputctl->eoi_reached &&
+      scan_number > cinfo->input_scan_number)
+    scan_number = cinfo->input_scan_number;
+  cinfo->output_scan_number = scan_number;
+  /* Perform any dummy output passes, and set up for the real pass */
+  return output_pass_setup(cinfo);
+}
+
+
+/*
+ * Finish up after an output pass in buffered-image mode.
+ *
+ * Returns FALSE if suspended.  The return value need be inspected only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(boolean)
+jpeg_finish_output (j_decompress_ptr cinfo)
+{
+  if ((cinfo->global_state == DSTATE_SCANNING ||
+       cinfo->global_state == DSTATE_RAW_OK) && cinfo->buffered_image) {
+    /* Terminate this pass. */
+    /* We do not require the whole pass to have been completed. */
+    (*cinfo->master->finish_output_pass) (cinfo);
+    cinfo->global_state = DSTATE_BUFPOST;
+  } else if (cinfo->global_state != DSTATE_BUFPOST) {
+    /* BUFPOST = repeat call after a suspension, anything else is error */
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  }
+  /* Read markers looking for SOS or EOI */
+  while (cinfo->input_scan_number <= cinfo->output_scan_number &&
+	 ! cinfo->inputctl->eoi_reached) {
+    if ((*cinfo->inputctl->consume_input) (cinfo) == JPEG_SUSPENDED)
+      return FALSE;		/* Suspend, come back later */
+  }
+  cinfo->global_state = DSTATE_BUFIMAGE;
+  return TRUE;
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */

Added: trunk/code/jpeg-8c/jdarith.c
===================================================================
--- trunk/code/jpeg-8c/jdarith.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdarith.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,772 @@
+/*
+ * jdarith.c
+ *
+ * Developed 1997-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains portable arithmetic entropy decoding routines for JPEG
+ * (implementing the ISO/IEC IS 10918-1 and CCITT Recommendation ITU-T T.81).
+ *
+ * Both sequential and progressive modes are supported in this single module.
+ *
+ * Suspension is not currently supported in this module.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Expanded entropy decoder object for arithmetic decoding. */
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  INT32 c;       /* C register, base of coding interval + input bit buffer */
+  INT32 a;               /* A register, normalized size of coding interval */
+  int ct;     /* bit shift counter, # of bits left in bit buffer part of C */
+                                                         /* init: ct = -16 */
+                                                         /* run: ct = 0..7 */
+                                                         /* error: ct = -1 */
+  int last_dc_val[MAX_COMPS_IN_SCAN]; /* last DC coef for each component */
+  int dc_context[MAX_COMPS_IN_SCAN]; /* context index for DC conditioning */
+
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+
+  /* Pointers to statistics areas (these workspaces have image lifespan) */
+  unsigned char * dc_stats[NUM_ARITH_TBLS];
+  unsigned char * ac_stats[NUM_ARITH_TBLS];
+
+  /* Statistics bin for coding with fixed probability 0.5 */
+  unsigned char fixed_bin[4];
+} arith_entropy_decoder;
+
+typedef arith_entropy_decoder * arith_entropy_ptr;
+
+/* The following two definitions specify the allocation chunk size
+ * for the statistics area.
+ * According to sections F.1.4.4.1.3 and F.1.4.4.2, we need at least
+ * 49 statistics bins for DC, and 245 statistics bins for AC coding.
+ *
+ * We use a compact representation with 1 byte per statistics bin,
+ * thus the numbers directly represent byte sizes.
+ * This 1 byte per statistics bin contains the meaning of the MPS
+ * (more probable symbol) in the highest bit (mask 0x80), and the
+ * index into the probability estimation state machine table
+ * in the lower bits (mask 0x7F).
+ */
+
+#define DC_STAT_BINS 64
+#define AC_STAT_BINS 256
+
+
+LOCAL(int)
+get_byte (j_decompress_ptr cinfo)
+/* Read next input byte; we do not support suspension in this module. */
+{
+  struct jpeg_source_mgr * src = cinfo->src;
+
+  if (src->bytes_in_buffer == 0)
+    if (! (*src->fill_input_buffer) (cinfo))
+      ERREXIT(cinfo, JERR_CANT_SUSPEND);
+  src->bytes_in_buffer--;
+  return GETJOCTET(*src->next_input_byte++);
+}
+
+
+/*
+ * The core arithmetic decoding routine (common in JPEG and JBIG).
+ * This needs to go as fast as possible.
+ * Machine-dependent optimization facilities
+ * are not utilized in this portable implementation.
+ * However, this code should be fairly efficient and
+ * may be a good base for further optimizations anyway.
+ *
+ * Return value is 0 or 1 (binary decision).
+ *
+ * Note: I've changed the handling of the code base & bit
+ * buffer register C compared to other implementations
+ * based on the standards layout & procedures.
+ * While it also contains both the actual base of the
+ * coding interval (16 bits) and the next-bits buffer,
+ * the cut-point between these two parts is floating
+ * (instead of fixed) with the bit shift counter CT.
+ * Thus, we also need only one (variable instead of
+ * fixed size) shift for the LPS/MPS decision, and
+ * we can get away with any renormalization update
+ * of C (except for new data insertion, of course).
+ *
+ * I've also introduced a new scheme for accessing
+ * the probability estimation state machine table,
+ * derived from Markus Kuhn's JBIG implementation.
+ */
+
+LOCAL(int)
+arith_decode (j_decompress_ptr cinfo, unsigned char *st)
+{
+  register arith_entropy_ptr e = (arith_entropy_ptr) cinfo->entropy;
+  register unsigned char nl, nm;
+  register INT32 qe, temp;
+  register int sv, data;
+
+  /* Renormalization & data input per section D.2.6 */
+  while (e->a < 0x8000L) {
+    if (--e->ct < 0) {
+      /* Need to fetch next data byte */
+      if (cinfo->unread_marker)
+	data = 0;		/* stuff zero data */
+      else {
+	data = get_byte(cinfo);	/* read next input byte */
+	if (data == 0xFF) {	/* zero stuff or marker code */
+	  do data = get_byte(cinfo);
+	  while (data == 0xFF);	/* swallow extra 0xFF bytes */
+	  if (data == 0)
+	    data = 0xFF;	/* discard stuffed zero byte */
+	  else {
+	    /* Note: Different from the Huffman decoder, hitting
+	     * a marker while processing the compressed data
+	     * segment is legal in arithmetic coding.
+	     * The convention is to supply zero data
+	     * then until decoding is complete.
+	     */
+	    cinfo->unread_marker = data;
+	    data = 0;
+	  }
+	}
+      }
+      e->c = (e->c << 8) | data; /* insert data into C register */
+      if ((e->ct += 8) < 0)	 /* update bit shift counter */
+	/* Need more initial bytes */
+	if (++e->ct == 0)
+	  /* Got 2 initial bytes -> re-init A and exit loop */
+	  e->a = 0x8000L; /* => e->a = 0x10000L after loop exit */
+    }
+    e->a <<= 1;
+  }
+
+  /* Fetch values from our compact representation of Table D.2:
+   * Qe values and probability estimation state machine
+   */
+  sv = *st;
+  qe = jpeg_aritab[sv & 0x7F];	/* => Qe_Value */
+  nl = qe & 0xFF; qe >>= 8;	/* Next_Index_LPS + Switch_MPS */
+  nm = qe & 0xFF; qe >>= 8;	/* Next_Index_MPS */
+
+  /* Decode & estimation procedures per sections D.2.4 & D.2.5 */
+  temp = e->a - qe;
+  e->a = temp;
+  temp <<= e->ct;
+  if (e->c >= temp) {
+    e->c -= temp;
+    /* Conditional LPS (less probable symbol) exchange */
+    if (e->a < qe) {
+      e->a = qe;
+      *st = (sv & 0x80) ^ nm;	/* Estimate_after_MPS */
+    } else {
+      e->a = qe;
+      *st = (sv & 0x80) ^ nl;	/* Estimate_after_LPS */
+      sv ^= 0x80;		/* Exchange LPS/MPS */
+    }
+  } else if (e->a < 0x8000L) {
+    /* Conditional MPS (more probable symbol) exchange */
+    if (e->a < qe) {
+      *st = (sv & 0x80) ^ nl;	/* Estimate_after_LPS */
+      sv ^= 0x80;		/* Exchange LPS/MPS */
+    } else {
+      *st = (sv & 0x80) ^ nm;	/* Estimate_after_MPS */
+    }
+  }
+
+  return sv >> 7;
+}
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ */
+
+LOCAL(void)
+process_restart (j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  int ci;
+  jpeg_component_info * compptr;
+
+  /* Advance past the RSTn marker */
+  if (! (*cinfo->marker->read_restart_marker) (cinfo))
+    ERREXIT(cinfo, JERR_CANT_SUSPEND);
+
+  /* Re-initialize statistics areas */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    if (! cinfo->progressive_mode || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      MEMZERO(entropy->dc_stats[compptr->dc_tbl_no], DC_STAT_BINS);
+      /* Reset DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    if ((! cinfo->progressive_mode && cinfo->lim_Se) ||
+	(cinfo->progressive_mode && cinfo->Ss)) {
+      MEMZERO(entropy->ac_stats[compptr->ac_tbl_no], AC_STAT_BINS);
+    }
+  }
+
+  /* Reset arithmetic decoding variables */
+  entropy->c = 0;
+  entropy->a = 0;
+  entropy->ct = -16;	/* force reading 2 initial bytes to fill C */
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Arithmetic MCU decoding.
+ * Each of these routines decodes and returns one MCU's worth of
+ * arithmetic-compressed coefficients.
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA IS INITIALLY ZEROED BY THE CALLER.
+ */
+
+/*
+ * MCU decoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, sign;
+  int v, m;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;	/* if error do nothing */
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    tbl = cinfo->cur_comp_info[ci]->dc_tbl_no;
+
+    /* Sections F.2.4.1 & F.1.4.4.1: Decoding of DC coefficients */
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.19: Decode_DC_DIFF */
+    if (arith_decode(cinfo, st) == 0)
+      entropy->dc_context[ci] = 0;
+    else {
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      sign = arith_decode(cinfo, st + 1);
+      st += 2; st += sign;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+	st = entropy->dc_stats[tbl] + 20;	/* Table F.4: X1 = 20 */
+	while (arith_decode(cinfo, st)) {
+	  if ((m <<= 1) == 0x8000) {
+	    WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	    entropy->ct = -1;			/* magnitude overflow */
+	    return TRUE;
+	  }
+	  st += 1;
+	}
+      }
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int) ((1L << cinfo->arith_dc_L[tbl]) >> 1))
+	entropy->dc_context[ci] = 0;		   /* zero diff category */
+      else if (m > (int) ((1L << cinfo->arith_dc_U[tbl]) >> 1))
+	entropy->dc_context[ci] = 12 + (sign * 4); /* large diff category */
+      else
+	entropy->dc_context[ci] = 4 + (sign * 4);  /* small diff category */
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	if (arith_decode(cinfo, st)) v |= m;
+      v += 1; if (sign) v = -v;
+      entropy->last_dc_val[ci] += v;
+    }
+
+    /* Scale and output the DC coefficient (assumes jpeg_natural_order[0]=0) */
+    (*block)[0] = (JCOEF) (entropy->last_dc_val[ci] << cinfo->Al);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  unsigned char *st;
+  int tbl, sign, k;
+  int v, m;
+  const int * natural_order;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;	/* if error do nothing */
+
+  natural_order = cinfo->natural_order;
+
+  /* There is always only one block per MCU */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  /* Sections F.2.4.2 & F.1.4.4.2: Decoding of AC coefficients */
+
+  /* Figure F.20: Decode_AC_coefficients */
+  for (k = cinfo->Ss; k <= cinfo->Se; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (arith_decode(cinfo, st)) break;		/* EOB flag */
+    while (arith_decode(cinfo, st + 1) == 0) {
+      st += 3; k++;
+      if (k > cinfo->Se) {
+	WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	entropy->ct = -1;			/* spectral overflow */
+	return TRUE;
+      }
+    }
+    /* Figure F.21: Decoding nonzero value v */
+    /* Figure F.22: Decoding the sign of v */
+    sign = arith_decode(cinfo, entropy->fixed_bin);
+    st += 2;
+    /* Figure F.23: Decoding the magnitude category of v */
+    if ((m = arith_decode(cinfo, st)) != 0) {
+      if (arith_decode(cinfo, st)) {
+	m <<= 1;
+	st = entropy->ac_stats[tbl] +
+	     (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+	while (arith_decode(cinfo, st)) {
+	  if ((m <<= 1) == 0x8000) {
+	    WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	    entropy->ct = -1;			/* magnitude overflow */
+	    return TRUE;
+	  }
+	  st += 1;
+	}
+      }
+    }
+    v = m;
+    /* Figure F.24: Decoding the magnitude bit pattern of v */
+    st += 14;
+    while (m >>= 1)
+      if (arith_decode(cinfo, st)) v |= m;
+    v += 1; if (sign) v = -v;
+    /* Scale and output coefficient in natural (dezigzagged) order */
+    (*block)[natural_order[k]] = (JCOEF) (v << cinfo->Al);
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for DC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  unsigned char *st;
+  int p1, blkn;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  st = entropy->fixed_bin;	/* use fixed probability estimation */
+  p1 = 1 << cinfo->Al;		/* 1 in the bit position being coded */
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    /* Encoded data is simply the next bit of the two's-complement DC value */
+    if (arith_decode(cinfo, st))
+      MCU_data[blkn][0][0] |= p1;
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  JBLOCKROW block;
+  JCOEFPTR thiscoef;
+  unsigned char *st;
+  int tbl, k, kex;
+  int p1, m1;
+  const int * natural_order;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;	/* if error do nothing */
+
+  natural_order = cinfo->natural_order;
+
+  /* There is always only one block per MCU */
+  block = MCU_data[0];
+  tbl = cinfo->cur_comp_info[0]->ac_tbl_no;
+
+  p1 = 1 << cinfo->Al;		/* 1 in the bit position being coded */
+  m1 = (-1) << cinfo->Al;	/* -1 in the bit position being coded */
+
+  /* Establish EOBx (previous stage end-of-block) index */
+  for (kex = cinfo->Se; kex > 0; kex--)
+    if ((*block)[natural_order[kex]]) break;
+
+  for (k = cinfo->Ss; k <= cinfo->Se; k++) {
+    st = entropy->ac_stats[tbl] + 3 * (k - 1);
+    if (k > kex)
+      if (arith_decode(cinfo, st)) break;	/* EOB flag */
+    for (;;) {
+      thiscoef = *block + natural_order[k];
+      if (*thiscoef) {				/* previously nonzero coef */
+	if (arith_decode(cinfo, st + 2)) {
+	  if (*thiscoef < 0)
+	    *thiscoef += m1;
+	  else
+	    *thiscoef += p1;
+	}
+	break;
+      }
+      if (arith_decode(cinfo, st + 1)) {	/* newly nonzero coef */
+	if (arith_decode(cinfo, entropy->fixed_bin))
+	  *thiscoef = m1;
+	else
+	  *thiscoef = p1;
+	break;
+      }
+      st += 3; k++;
+      if (k > cinfo->Se) {
+	WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	entropy->ct = -1;			/* spectral overflow */
+	return TRUE;
+      }
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Decode one MCU's worth of arithmetic-compressed coefficients.
+ */
+
+METHODDEF(boolean)
+decode_mcu (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  jpeg_component_info * compptr;
+  JBLOCKROW block;
+  unsigned char *st;
+  int blkn, ci, tbl, sign, k;
+  int v, m;
+  const int * natural_order;
+
+  /* Process restart marker if needed */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      process_restart(cinfo);
+    entropy->restarts_to_go--;
+  }
+
+  if (entropy->ct == -1) return TRUE;	/* if error do nothing */
+
+  natural_order = cinfo->natural_order;
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+    ci = cinfo->MCU_membership[blkn];
+    compptr = cinfo->cur_comp_info[ci];
+
+    /* Sections F.2.4.1 & F.1.4.4.1: Decoding of DC coefficients */
+
+    tbl = compptr->dc_tbl_no;
+
+    /* Table F.4: Point to statistics bin S0 for DC coefficient coding */
+    st = entropy->dc_stats[tbl] + entropy->dc_context[ci];
+
+    /* Figure F.19: Decode_DC_DIFF */
+    if (arith_decode(cinfo, st) == 0)
+      entropy->dc_context[ci] = 0;
+    else {
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      sign = arith_decode(cinfo, st + 1);
+      st += 2; st += sign;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+	st = entropy->dc_stats[tbl] + 20;	/* Table F.4: X1 = 20 */
+	while (arith_decode(cinfo, st)) {
+	  if ((m <<= 1) == 0x8000) {
+	    WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	    entropy->ct = -1;			/* magnitude overflow */
+	    return TRUE;
+	  }
+	  st += 1;
+	}
+      }
+      /* Section F.1.4.4.1.2: Establish dc_context conditioning category */
+      if (m < (int) ((1L << cinfo->arith_dc_L[tbl]) >> 1))
+	entropy->dc_context[ci] = 0;		   /* zero diff category */
+      else if (m > (int) ((1L << cinfo->arith_dc_U[tbl]) >> 1))
+	entropy->dc_context[ci] = 12 + (sign * 4); /* large diff category */
+      else
+	entropy->dc_context[ci] = 4 + (sign * 4);  /* small diff category */
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	if (arith_decode(cinfo, st)) v |= m;
+      v += 1; if (sign) v = -v;
+      entropy->last_dc_val[ci] += v;
+    }
+
+    (*block)[0] = (JCOEF) entropy->last_dc_val[ci];
+
+    /* Sections F.2.4.2 & F.1.4.4.2: Decoding of AC coefficients */
+
+    tbl = compptr->ac_tbl_no;
+
+    /* Figure F.20: Decode_AC_coefficients */
+    for (k = 1; k <= cinfo->lim_Se; k++) {
+      st = entropy->ac_stats[tbl] + 3 * (k - 1);
+      if (arith_decode(cinfo, st)) break;	/* EOB flag */
+      while (arith_decode(cinfo, st + 1) == 0) {
+	st += 3; k++;
+	if (k > cinfo->lim_Se) {
+	  WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	  entropy->ct = -1;			/* spectral overflow */
+	  return TRUE;
+	}
+      }
+      /* Figure F.21: Decoding nonzero value v */
+      /* Figure F.22: Decoding the sign of v */
+      sign = arith_decode(cinfo, entropy->fixed_bin);
+      st += 2;
+      /* Figure F.23: Decoding the magnitude category of v */
+      if ((m = arith_decode(cinfo, st)) != 0) {
+	if (arith_decode(cinfo, st)) {
+	  m <<= 1;
+	  st = entropy->ac_stats[tbl] +
+	       (k <= cinfo->arith_ac_K[tbl] ? 189 : 217);
+	  while (arith_decode(cinfo, st)) {
+	    if ((m <<= 1) == 0x8000) {
+	      WARNMS(cinfo, JWRN_ARITH_BAD_CODE);
+	      entropy->ct = -1;			/* magnitude overflow */
+	      return TRUE;
+	    }
+	    st += 1;
+	  }
+	}
+      }
+      v = m;
+      /* Figure F.24: Decoding the magnitude bit pattern of v */
+      st += 14;
+      while (m >>= 1)
+	if (arith_decode(cinfo, st)) v |= m;
+      v += 1; if (sign) v = -v;
+      (*block)[natural_order[k]] = (JCOEF) v;
+    }
+  }
+
+  return TRUE;
+}
+
+
+/*
+ * Initialize for an arithmetic-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass (j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy = (arith_entropy_ptr) cinfo->entropy;
+  int ci, tbl;
+  jpeg_component_info * compptr;
+
+  if (cinfo->progressive_mode) {
+    /* Validate progressive scan parameters */
+    if (cinfo->Ss == 0) {
+      if (cinfo->Se != 0)
+	goto bad;
+    } else {
+      /* need not check Ss/Se < 0 since they came from unsigned bytes */
+      if (cinfo->Se < cinfo->Ss || cinfo->Se > cinfo->lim_Se)
+	goto bad;
+      /* AC scans may have only one component */
+      if (cinfo->comps_in_scan != 1)
+	goto bad;
+    }
+    if (cinfo->Ah != 0) {
+      /* Successive approximation refinement scan: must have Al = Ah-1. */
+      if (cinfo->Ah-1 != cinfo->Al)
+	goto bad;
+    }
+    if (cinfo->Al > 13) {	/* need not check for < 0 */
+      bad:
+      ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+	       cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+    }
+    /* Update progression status, and verify that scan order is legal.
+     * Note that inter-scan inconsistencies are treated as warnings
+     * not fatal errors ... not clear if this is right way to behave.
+     */
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      int coefi, cindex = cinfo->cur_comp_info[ci]->component_index;
+      int *coef_bit_ptr = & cinfo->coef_bits[cindex][0];
+      if (cinfo->Ss && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
+	WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+      for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
+	int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
+	if (cinfo->Ah != expected)
+	  WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, coefi);
+	coef_bit_ptr[coefi] = cinfo->Al;
+      }
+    }
+    /* Select MCU decoding routine */
+    if (cinfo->Ah == 0) {
+      if (cinfo->Ss == 0)
+	entropy->pub.decode_mcu = decode_mcu_DC_first;
+      else
+	entropy->pub.decode_mcu = decode_mcu_AC_first;
+    } else {
+      if (cinfo->Ss == 0)
+	entropy->pub.decode_mcu = decode_mcu_DC_refine;
+      else
+	entropy->pub.decode_mcu = decode_mcu_AC_refine;
+    }
+  } else {
+    /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
+     * This ought to be an error condition, but we make it a warning.
+     */
+    if (cinfo->Ss != 0 || cinfo->Ah != 0 || cinfo->Al != 0 ||
+	(cinfo->Se < DCTSIZE2 && cinfo->Se != cinfo->lim_Se))
+      WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
+    /* Select MCU decoding routine */
+    entropy->pub.decode_mcu = decode_mcu;
+  }
+
+  /* Allocate & initialize requested statistics areas */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    if (! cinfo->progressive_mode || (cinfo->Ss == 0 && cinfo->Ah == 0)) {
+      tbl = compptr->dc_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+	ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->dc_stats[tbl] == NULL)
+	entropy->dc_stats[tbl] = (unsigned char *) (*cinfo->mem->alloc_small)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, DC_STAT_BINS);
+      MEMZERO(entropy->dc_stats[tbl], DC_STAT_BINS);
+      /* Initialize DC predictions to 0 */
+      entropy->last_dc_val[ci] = 0;
+      entropy->dc_context[ci] = 0;
+    }
+    if ((! cinfo->progressive_mode && cinfo->lim_Se) ||
+	(cinfo->progressive_mode && cinfo->Ss)) {
+      tbl = compptr->ac_tbl_no;
+      if (tbl < 0 || tbl >= NUM_ARITH_TBLS)
+	ERREXIT1(cinfo, JERR_NO_ARITH_TABLE, tbl);
+      if (entropy->ac_stats[tbl] == NULL)
+	entropy->ac_stats[tbl] = (unsigned char *) (*cinfo->mem->alloc_small)
+	  ((j_common_ptr) cinfo, JPOOL_IMAGE, AC_STAT_BINS);
+      MEMZERO(entropy->ac_stats[tbl], AC_STAT_BINS);
+    }
+  }
+
+  /* Initialize arithmetic decoding variables */
+  entropy->c = 0;
+  entropy->a = 0;
+  entropy->ct = -16;	/* force reading 2 initial bytes to fill C */
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Module initialization routine for arithmetic entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_arith_decoder (j_decompress_ptr cinfo)
+{
+  arith_entropy_ptr entropy;
+  int i;
+
+  entropy = (arith_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(arith_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *) entropy;
+  entropy->pub.start_pass = start_pass;
+
+  /* Mark tables unallocated */
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    entropy->dc_stats[i] = NULL;
+    entropy->ac_stats[i] = NULL;
+  }
+
+  /* Initialize index for fixed probability estimation */
+  entropy->fixed_bin[0] = 113;
+
+  if (cinfo->progressive_mode) {
+    /* Create progression status table */
+    int *coef_bit_ptr, ci;
+    cinfo->coef_bits = (int (*)[DCTSIZE2])
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  cinfo->num_components*DCTSIZE2*SIZEOF(int));
+    coef_bit_ptr = & cinfo->coef_bits[0][0];
+    for (ci = 0; ci < cinfo->num_components; ci++) 
+      for (i = 0; i < DCTSIZE2; i++)
+	*coef_bit_ptr++ = -1;
+  }
+}

Added: trunk/code/jpeg-8c/jdatadst.c
===================================================================
--- trunk/code/jpeg-8c/jdatadst.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdatadst.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,267 @@
+/*
+ * jdatadst.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains compression data destination routines for the case of
+ * emitting JPEG data to memory or to a file (or any stdio stream).
+ * While these routines are sufficient for most applications,
+ * some will want to use a different destination manager.
+ * IMPORTANT: we assume that fwrite() will correctly transcribe an array of
+ * JOCTETs into 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+#ifndef HAVE_STDLIB_H		/* <stdlib.h> should declare malloc(),free() */
+extern void * malloc JPP((size_t size));
+extern void free JPP((void *ptr));
+#endif
+
+
+/* Expanded data destination object for stdio output */
+
+typedef struct {
+  struct jpeg_destination_mgr pub; /* public fields */
+
+  FILE * outfile;		/* target stream */
+  JOCTET * buffer;		/* start of buffer */
+} my_destination_mgr;
+
+typedef my_destination_mgr * my_dest_ptr;
+
+#define OUTPUT_BUF_SIZE  4096	/* choose an efficiently fwrite'able size */
+
+
+/* Expanded data destination object for memory output */
+
+typedef struct {
+  struct jpeg_destination_mgr pub; /* public fields */
+
+  unsigned char ** outbuffer;	/* target buffer */
+  unsigned long * outsize;
+  unsigned char * newbuffer;	/* newly allocated buffer */
+  JOCTET * buffer;		/* start of buffer */
+  size_t bufsize;
+} my_mem_destination_mgr;
+
+typedef my_mem_destination_mgr * my_mem_dest_ptr;
+
+
+/*
+ * Initialize destination --- called by jpeg_start_compress
+ * before any data is actually written.
+ */
+
+METHODDEF(void)
+init_destination (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+
+  /* Allocate the output buffer --- it will be released when done with image */
+  dest->buffer = (JOCTET *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  OUTPUT_BUF_SIZE * SIZEOF(JOCTET));
+
+  dest->pub.next_output_byte = dest->buffer;
+  dest->pub.free_in_buffer = OUTPUT_BUF_SIZE;
+}
+
+METHODDEF(void)
+init_mem_destination (j_compress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Empty the output buffer --- called whenever buffer fills up.
+ *
+ * In typical applications, this should write the entire output buffer
+ * (ignoring the current state of next_output_byte & free_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been dumped.
+ *
+ * In applications that need to be able to suspend compression due to output
+ * overrun, a FALSE return indicates that the buffer cannot be emptied now.
+ * In this situation, the compressor will return to its caller (possibly with
+ * an indication that it has not accepted all the supplied scanlines).  The
+ * application should resume compression after it has made more room in the
+ * output buffer.  Note that there are substantial restrictions on the use of
+ * suspension --- see the documentation.
+ *
+ * When suspending, the compressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_output_byte & free_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point will be regenerated after resumption, so do not
+ * write it out when emptying the buffer externally.
+ */
+
+METHODDEF(boolean)
+empty_output_buffer (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+
+  if (JFWRITE(dest->outfile, dest->buffer, OUTPUT_BUF_SIZE) !=
+      (size_t) OUTPUT_BUF_SIZE)
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+
+  dest->pub.next_output_byte = dest->buffer;
+  dest->pub.free_in_buffer = OUTPUT_BUF_SIZE;
+
+  return TRUE;
+}
+
+METHODDEF(boolean)
+empty_mem_output_buffer (j_compress_ptr cinfo)
+{
+  size_t nextsize;
+  JOCTET * nextbuffer;
+  my_mem_dest_ptr dest = (my_mem_dest_ptr) cinfo->dest;
+
+  /* Try to allocate new buffer with double size */
+  nextsize = dest->bufsize * 2;
+  nextbuffer = malloc(nextsize);
+
+  if (nextbuffer == NULL)
+    ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);
+
+  MEMCOPY(nextbuffer, dest->buffer, dest->bufsize);
+
+  if (dest->newbuffer != NULL)
+    free(dest->newbuffer);
+
+  dest->newbuffer = nextbuffer;
+
+  dest->pub.next_output_byte = nextbuffer + dest->bufsize;
+  dest->pub.free_in_buffer = dest->bufsize;
+
+  dest->buffer = nextbuffer;
+  dest->bufsize = nextsize;
+
+  return TRUE;
+}
+
+
+/*
+ * Terminate destination --- called by jpeg_finish_compress
+ * after all data has been written.  Usually needs to flush buffer.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_destination (j_compress_ptr cinfo)
+{
+  my_dest_ptr dest = (my_dest_ptr) cinfo->dest;
+  size_t datacount = OUTPUT_BUF_SIZE - dest->pub.free_in_buffer;
+
+  /* Write any data remaining in the buffer */
+  if (datacount > 0) {
+    if (JFWRITE(dest->outfile, dest->buffer, datacount) != datacount)
+      ERREXIT(cinfo, JERR_FILE_WRITE);
+  }
+  fflush(dest->outfile);
+  /* Make sure we wrote the output file OK */
+  if (ferror(dest->outfile))
+    ERREXIT(cinfo, JERR_FILE_WRITE);
+}
+
+METHODDEF(void)
+term_mem_destination (j_compress_ptr cinfo)
+{
+  my_mem_dest_ptr dest = (my_mem_dest_ptr) cinfo->dest;
+
+  *dest->outbuffer = dest->buffer;
+  *dest->outsize = dest->bufsize - dest->pub.free_in_buffer;
+}
+
+
+/*
+ * Prepare for output to a stdio stream.
+ * The caller must have already opened the stream, and is responsible
+ * for closing it after finishing compression.
+ */
+
+GLOBAL(void)
+jpeg_stdio_dest (j_compress_ptr cinfo, FILE * outfile)
+{
+  my_dest_ptr dest;
+
+  /* The destination object is made permanent so that multiple JPEG images
+   * can be written to the same file without re-executing jpeg_stdio_dest.
+   * This makes it dangerous to use this manager and a different destination
+   * manager serially with the same JPEG object, because their private object
+   * sizes may be different.  Caveat programmer.
+   */
+  if (cinfo->dest == NULL) {	/* first time for this JPEG object? */
+    cinfo->dest = (struct jpeg_destination_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(my_destination_mgr));
+  }
+
+  dest = (my_dest_ptr) cinfo->dest;
+  dest->pub.init_destination = init_destination;
+  dest->pub.empty_output_buffer = empty_output_buffer;
+  dest->pub.term_destination = term_destination;
+  dest->outfile = outfile;
+}
+
+
+/*
+ * Prepare for output to a memory buffer.
+ * The caller may supply an own initial buffer with appropriate size.
+ * Otherwise, or when the actual data output exceeds the given size,
+ * the library adapts the buffer size as necessary.
+ * The standard library functions malloc/free are used for allocating
+ * larger memory, so the buffer is available to the application after
+ * finishing compression, and then the application is responsible for
+ * freeing the requested memory.
+ */
+
+GLOBAL(void)
+jpeg_mem_dest (j_compress_ptr cinfo,
+	       unsigned char ** outbuffer, unsigned long * outsize)
+{
+  my_mem_dest_ptr dest;
+
+  if (outbuffer == NULL || outsize == NULL)	/* sanity check */
+    ERREXIT(cinfo, JERR_BUFFER_SIZE);
+
+  /* The destination object is made permanent so that multiple JPEG images
+   * can be written to the same buffer without re-executing jpeg_mem_dest.
+   */
+  if (cinfo->dest == NULL) {	/* first time for this JPEG object? */
+    cinfo->dest = (struct jpeg_destination_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(my_mem_destination_mgr));
+  }
+
+  dest = (my_mem_dest_ptr) cinfo->dest;
+  dest->pub.init_destination = init_mem_destination;
+  dest->pub.empty_output_buffer = empty_mem_output_buffer;
+  dest->pub.term_destination = term_mem_destination;
+  dest->outbuffer = outbuffer;
+  dest->outsize = outsize;
+  dest->newbuffer = NULL;
+
+  if (*outbuffer == NULL || *outsize == 0) {
+    /* Allocate initial buffer */
+    dest->newbuffer = *outbuffer = malloc(OUTPUT_BUF_SIZE);
+    if (dest->newbuffer == NULL)
+      ERREXIT1(cinfo, JERR_OUT_OF_MEMORY, 10);
+    *outsize = OUTPUT_BUF_SIZE;
+  }
+
+  dest->pub.next_output_byte = dest->buffer = *outbuffer;
+  dest->pub.free_in_buffer = dest->bufsize = *outsize;
+}

Added: trunk/code/jpeg-8c/jdatasrc.c
===================================================================
--- trunk/code/jpeg-8c/jdatasrc.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdatasrc.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,274 @@
+/*
+ * jdatasrc.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2009-2010 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains decompression data source routines for the case of
+ * reading JPEG data from memory or from a file (or any stdio stream).
+ * While these routines are sufficient for most applications,
+ * some will want to use a different source manager.
+ * IMPORTANT: we assume that fread() will correctly transcribe an array of
+ * JOCTETs from 8-bit-wide elements on external storage.  If char is wider
+ * than 8 bits on your machine, you may need to do some tweaking.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jerror.h"
+
+
+/* Expanded data source object for stdio input */
+
+typedef struct {
+  struct jpeg_source_mgr pub;	/* public fields */
+
+  FILE * infile;		/* source stream */
+  JOCTET * buffer;		/* start of buffer */
+  boolean start_of_file;	/* have we gotten any data yet? */
+} my_source_mgr;
+
+typedef my_source_mgr * my_src_ptr;
+
+#define INPUT_BUF_SIZE  4096	/* choose an efficiently fread'able size */
+
+
+/*
+ * Initialize source --- called by jpeg_read_header
+ * before any data is actually read.
+ */
+
+METHODDEF(void)
+init_source (j_decompress_ptr cinfo)
+{
+  my_src_ptr src = (my_src_ptr) cinfo->src;
+
+  /* We reset the empty-input-file flag for each image,
+   * but we don't clear the input buffer.
+   * This is correct behavior for reading a series of images from one source.
+   */
+  src->start_of_file = TRUE;
+}
+
+METHODDEF(void)
+init_mem_source (j_decompress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Fill the input buffer --- called whenever buffer is emptied.
+ *
+ * In typical applications, this should read fresh data into the buffer
+ * (ignoring the current state of next_input_byte & bytes_in_buffer),
+ * reset the pointer & count to the start of the buffer, and return TRUE
+ * indicating that the buffer has been reloaded.  It is not necessary to
+ * fill the buffer entirely, only to obtain at least one more byte.
+ *
+ * There is no such thing as an EOF return.  If the end of the file has been
+ * reached, the routine has a choice of ERREXIT() or inserting fake data into
+ * the buffer.  In most cases, generating a warning message and inserting a
+ * fake EOI marker is the best course of action --- this will allow the
+ * decompressor to output however much of the image is there.  However,
+ * the resulting error message is misleading if the real problem is an empty
+ * input file, so we handle that case specially.
+ *
+ * In applications that need to be able to suspend compression due to input
+ * not being available yet, a FALSE return indicates that no more data can be
+ * obtained right now, but more may be forthcoming later.  In this situation,
+ * the decompressor will return to its caller (with an indication of the
+ * number of scanlines it has read, if any).  The application should resume
+ * decompression after it has loaded more data into the input buffer.  Note
+ * that there are substantial restrictions on the use of suspension --- see
+ * the documentation.
+ *
+ * When suspending, the decompressor will back up to a convenient restart point
+ * (typically the start of the current MCU). next_input_byte & bytes_in_buffer
+ * indicate where the restart point will be if the current call returns FALSE.
+ * Data beyond this point must be rescanned after resumption, so move it to
+ * the front of the buffer rather than discarding it.
+ */
+
+METHODDEF(boolean)
+fill_input_buffer (j_decompress_ptr cinfo)
+{
+  my_src_ptr src = (my_src_ptr) cinfo->src;
+  size_t nbytes;
+
+  nbytes = JFREAD(src->infile, src->buffer, INPUT_BUF_SIZE);
+
+  if (nbytes <= 0) {
+    if (src->start_of_file)	/* Treat empty input file as fatal error */
+      ERREXIT(cinfo, JERR_INPUT_EMPTY);
+    WARNMS(cinfo, JWRN_JPEG_EOF);
+    /* Insert a fake EOI marker */
+    src->buffer[0] = (JOCTET) 0xFF;
+    src->buffer[1] = (JOCTET) JPEG_EOI;
+    nbytes = 2;
+  }
+
+  src->pub.next_input_byte = src->buffer;
+  src->pub.bytes_in_buffer = nbytes;
+  src->start_of_file = FALSE;
+
+  return TRUE;
+}
+
+METHODDEF(boolean)
+fill_mem_input_buffer (j_decompress_ptr cinfo)
+{
+  static JOCTET mybuffer[4];
+
+  /* The whole JPEG data is expected to reside in the supplied memory
+   * buffer, so any request for more data beyond the given buffer size
+   * is treated as an error.
+   */
+  WARNMS(cinfo, JWRN_JPEG_EOF);
+  /* Insert a fake EOI marker */
+  mybuffer[0] = (JOCTET) 0xFF;
+  mybuffer[1] = (JOCTET) JPEG_EOI;
+
+  cinfo->src->next_input_byte = mybuffer;
+  cinfo->src->bytes_in_buffer = 2;
+
+  return TRUE;
+}
+
+
+/*
+ * Skip data --- used to skip over a potentially large amount of
+ * uninteresting data (such as an APPn marker).
+ *
+ * Writers of suspendable-input applications must note that skip_input_data
+ * is not granted the right to give a suspension return.  If the skip extends
+ * beyond the data currently in the buffer, the buffer can be marked empty so
+ * that the next read will cause a fill_input_buffer call that can suspend.
+ * Arranging for additional bytes to be discarded before reloading the input
+ * buffer is the application writer's problem.
+ */
+
+METHODDEF(void)
+skip_input_data (j_decompress_ptr cinfo, long num_bytes)
+{
+  struct jpeg_source_mgr * src = cinfo->src;
+
+  /* Just a dumb implementation for now.  Could use fseek() except
+   * it doesn't work on pipes.  Not clear that being smart is worth
+   * any trouble anyway --- large skips are infrequent.
+   */
+  if (num_bytes > 0) {
+    while (num_bytes > (long) src->bytes_in_buffer) {
+      num_bytes -= (long) src->bytes_in_buffer;
+      (void) (*src->fill_input_buffer) (cinfo);
+      /* note we assume that fill_input_buffer will never return FALSE,
+       * so suspension need not be handled.
+       */
+    }
+    src->next_input_byte += (size_t) num_bytes;
+    src->bytes_in_buffer -= (size_t) num_bytes;
+  }
+}
+
+
+/*
+ * An additional method that can be provided by data source modules is the
+ * resync_to_restart method for error recovery in the presence of RST markers.
+ * For the moment, this source module just uses the default resync method
+ * provided by the JPEG library.  That method assumes that no backtracking
+ * is possible.
+ */
+
+
+/*
+ * Terminate source --- called by jpeg_finish_decompress
+ * after all data has been read.  Often a no-op.
+ *
+ * NB: *not* called by jpeg_abort or jpeg_destroy; surrounding
+ * application must deal with any cleanup that should happen even
+ * for error exit.
+ */
+
+METHODDEF(void)
+term_source (j_decompress_ptr cinfo)
+{
+  /* no work necessary here */
+}
+
+
+/*
+ * Prepare for input from a stdio stream.
+ * The caller must have already opened the stream, and is responsible
+ * for closing it after finishing decompression.
+ */
+
+GLOBAL(void)
+jpeg_stdio_src (j_decompress_ptr cinfo, FILE * infile)
+{
+  my_src_ptr src;
+
+  /* The source object and input buffer are made permanent so that a series
+   * of JPEG images can be read from the same file by calling jpeg_stdio_src
+   * only before the first one.  (If we discarded the buffer at the end of
+   * one image, we'd likely lose the start of the next one.)
+   * This makes it unsafe to use this manager and a different source
+   * manager serially with the same JPEG object.  Caveat programmer.
+   */
+  if (cinfo->src == NULL) {	/* first time for this JPEG object? */
+    cinfo->src = (struct jpeg_source_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(my_source_mgr));
+    src = (my_src_ptr) cinfo->src;
+    src->buffer = (JOCTET *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  INPUT_BUF_SIZE * SIZEOF(JOCTET));
+  }
+
+  src = (my_src_ptr) cinfo->src;
+  src->pub.init_source = init_source;
+  src->pub.fill_input_buffer = fill_input_buffer;
+  src->pub.skip_input_data = skip_input_data;
+  src->pub.resync_to_restart = jpeg_resync_to_restart; /* use default method */
+  src->pub.term_source = term_source;
+  src->infile = infile;
+  src->pub.bytes_in_buffer = 0; /* forces fill_input_buffer on first read */
+  src->pub.next_input_byte = NULL; /* until buffer loaded */
+}
+
+
+/*
+ * Prepare for input from a supplied memory buffer.
+ * The buffer must contain the whole JPEG data.
+ */
+
+GLOBAL(void)
+jpeg_mem_src (j_decompress_ptr cinfo,
+	      unsigned char * inbuffer, unsigned long insize)
+{
+  struct jpeg_source_mgr * src;
+
+  if (inbuffer == NULL || insize == 0)	/* Treat empty input as fatal error */
+    ERREXIT(cinfo, JERR_INPUT_EMPTY);
+
+  /* The source object is made permanent so that a series of JPEG images
+   * can be read from the same buffer by calling jpeg_mem_src only before
+   * the first one.
+   */
+  if (cinfo->src == NULL) {	/* first time for this JPEG object? */
+    cinfo->src = (struct jpeg_source_mgr *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				  SIZEOF(struct jpeg_source_mgr));
+  }
+
+  src = cinfo->src;
+  src->init_source = init_mem_source;
+  src->fill_input_buffer = fill_mem_input_buffer;
+  src->skip_input_data = skip_input_data;
+  src->resync_to_restart = jpeg_resync_to_restart; /* use default method */
+  src->term_source = term_source;
+  src->bytes_in_buffer = (size_t) insize;
+  src->next_input_byte = (JOCTET *) inbuffer;
+}

Added: trunk/code/jpeg-8c/jdcoefct.c
===================================================================
--- trunk/code/jpeg-8c/jdcoefct.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdcoefct.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,736 @@
+/*
+ * jdcoefct.c
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the coefficient buffer controller for decompression.
+ * This controller is the top level of the JPEG decompressor proper.
+ * The coefficient buffer lies between entropy decoding and inverse-DCT steps.
+ *
+ * In buffered-image mode, this controller is the interface between
+ * input-oriented processing and output-oriented processing.
+ * Also, the input side (only) is used when reading a file for transcoding.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+/* Block smoothing is only applicable for progressive JPEG, so: */
+#ifndef D_PROGRESSIVE_SUPPORTED
+#undef BLOCK_SMOOTHING_SUPPORTED
+#endif
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_coef_controller pub; /* public fields */
+
+  /* These variables keep track of the current location of the input side. */
+  /* cinfo->input_iMCU_row is also used for this. */
+  JDIMENSION MCU_ctr;		/* counts MCUs processed in current row */
+  int MCU_vert_offset;		/* counts MCU rows within iMCU row */
+  int MCU_rows_per_iMCU_row;	/* number of such rows needed */
+
+  /* The output side's location is represented by cinfo->output_iMCU_row. */
+
+  /* In single-pass modes, it's sufficient to buffer just one MCU.
+   * We allocate a workspace of D_MAX_BLOCKS_IN_MCU coefficient blocks,
+   * and let the entropy decoder write into that workspace each time.
+   * (On 80x86, the workspace is FAR even though it's not really very big;
+   * this is to keep the module interfaces unchanged when a large coefficient
+   * buffer is necessary.)
+   * In multi-pass modes, this array points to the current MCU's blocks
+   * within the virtual arrays; it is used only by the input side.
+   */
+  JBLOCKROW MCU_buffer[D_MAX_BLOCKS_IN_MCU];
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* In multi-pass modes, we need a virtual block array for each component. */
+  jvirt_barray_ptr whole_image[MAX_COMPONENTS];
+#endif
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  /* When doing block smoothing, we latch coefficient Al values here */
+  int * coef_bits_latch;
+#define SAVED_COEFS  6		/* we save coef_bits[0..5] */
+#endif
+} my_coef_controller;
+
+typedef my_coef_controller * my_coef_ptr;
+
+/* Forward declarations */
+METHODDEF(int) decompress_onepass
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+METHODDEF(int) decompress_data
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#endif
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+LOCAL(boolean) smoothing_ok JPP((j_decompress_ptr cinfo));
+METHODDEF(int) decompress_smooth_data
+	JPP((j_decompress_ptr cinfo, JSAMPIMAGE output_buf));
+#endif
+
+
+LOCAL(void)
+start_iMCU_row (j_decompress_ptr cinfo)
+/* Reset within-iMCU-row counters for a new row (input side) */
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* In an interleaved scan, an MCU row is the same as an iMCU row.
+   * In a noninterleaved scan, an iMCU row has v_samp_factor MCU rows.
+   * But at the bottom of the image, process only what's left.
+   */
+  if (cinfo->comps_in_scan > 1) {
+    coef->MCU_rows_per_iMCU_row = 1;
+  } else {
+    if (cinfo->input_iMCU_row < (cinfo->total_iMCU_rows-1))
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->v_samp_factor;
+    else
+      coef->MCU_rows_per_iMCU_row = cinfo->cur_comp_info[0]->last_row_height;
+  }
+
+  coef->MCU_ctr = 0;
+  coef->MCU_vert_offset = 0;
+}
+
+
+/*
+ * Initialize for an input processing pass.
+ */
+
+METHODDEF(void)
+start_input_pass (j_decompress_ptr cinfo)
+{
+  cinfo->input_iMCU_row = 0;
+  start_iMCU_row(cinfo);
+}
+
+
+/*
+ * Initialize for an output processing pass.
+ */
+
+METHODDEF(void)
+start_output_pass (j_decompress_ptr cinfo)
+{
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+
+  /* If multipass, check to see whether to use block smoothing on this pass */
+  if (coef->pub.coef_arrays != NULL) {
+    if (cinfo->do_block_smoothing && smoothing_ok(cinfo))
+      coef->pub.decompress_data = decompress_smooth_data;
+    else
+      coef->pub.decompress_data = decompress_data;
+  }
+#endif
+  cinfo->output_iMCU_row = 0;
+}
+
+
+/*
+ * Decompress and return some data in the single-pass case.
+ * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
+ * Input and output must run in lockstep since we have only a one-MCU buffer.
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ *
+ * NB: output_buf contains a plane for each component in image,
+ * which we index according to the component's SOF position.
+ */
+
+METHODDEF(int)
+decompress_onepass (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  JDIMENSION last_MCU_col = cinfo->MCUs_per_row - 1;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  int blkn, ci, xindex, yindex, yoffset, useful_width;
+  JSAMPARRAY output_ptr;
+  JDIMENSION start_col, output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+
+  /* Loop to process as much as one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->MCU_ctr; MCU_col_num <= last_MCU_col;
+	 MCU_col_num++) {
+      /* Try to fetch an MCU.  Entropy decoder expects buffer to be zeroed. */
+      jzero_far((void FAR *) coef->MCU_buffer[0],
+		(size_t) (cinfo->blocks_in_MCU * SIZEOF(JBLOCK)));
+      if (! (*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->MCU_ctr = MCU_col_num;
+	return JPEG_SUSPENDED;
+      }
+      /* Determine where data should go in output_buf and do the IDCT thing.
+       * We skip dummy blocks at the right and bottom edges (but blkn gets
+       * incremented past them!).  Note the inner loop relies on having
+       * allocated the MCU_buffer[] blocks sequentially.
+       */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	/* Don't bother to IDCT an uninteresting component. */
+	if (! compptr->component_needed) {
+	  blkn += compptr->MCU_blocks;
+	  continue;
+	}
+	inverse_DCT = cinfo->idct->inverse_DCT[compptr->component_index];
+	useful_width = (MCU_col_num < last_MCU_col) ? compptr->MCU_width
+						    : compptr->last_col_width;
+	output_ptr = output_buf[compptr->component_index] +
+	  yoffset * compptr->DCT_v_scaled_size;
+	start_col = MCU_col_num * compptr->MCU_sample_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  if (cinfo->input_iMCU_row < last_iMCU_row ||
+	      yoffset+yindex < compptr->last_row_height) {
+	    output_col = start_col;
+	    for (xindex = 0; xindex < useful_width; xindex++) {
+	      (*inverse_DCT) (cinfo, compptr,
+			      (JCOEFPTR) coef->MCU_buffer[blkn+xindex],
+			      output_ptr, output_col);
+	      output_col += compptr->DCT_h_scaled_size;
+	    }
+	  }
+	  blkn += compptr->MCU_width;
+	  output_ptr += compptr->DCT_v_scaled_size;
+	}
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->MCU_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  cinfo->output_iMCU_row++;
+  if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
+    start_iMCU_row(cinfo);
+    return JPEG_ROW_COMPLETED;
+  }
+  /* Completed the scan */
+  (*cinfo->inputctl->finish_input_pass) (cinfo);
+  return JPEG_SCAN_COMPLETED;
+}
+
+
+/*
+ * Dummy consume-input routine for single-pass operation.
+ */
+
+METHODDEF(int)
+dummy_consume_data (j_decompress_ptr cinfo)
+{
+  return JPEG_SUSPENDED;	/* Always indicate nothing was done */
+}
+
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Consume input data and store it in the full-image coefficient buffer.
+ * We read as much as one fully interleaved MCU row ("iMCU" row) per call,
+ * ie, v_samp_factor block rows for each component in the scan.
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ */
+
+METHODDEF(int)
+consume_data (j_decompress_ptr cinfo)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION MCU_col_num;	/* index of current MCU within row */
+  int blkn, ci, xindex, yindex, yoffset;
+  JDIMENSION start_col;
+  JBLOCKARRAY buffer[MAX_COMPS_IN_SCAN];
+  JBLOCKROW buffer_ptr;
+  jpeg_component_info *compptr;
+
+  /* Align the virtual buffers for the components used in this scan. */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    buffer[ci] = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[compptr->component_index],
+       cinfo->input_iMCU_row * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, TRUE);
+    /* Note: entropy decoder expects buffer to be zeroed,
+     * but this is handled automatically by the memory manager
+     * because we requested a pre-zeroed array.
+     */
+  }
+
+  /* Loop to process one whole iMCU row */
+  for (yoffset = coef->MCU_vert_offset; yoffset < coef->MCU_rows_per_iMCU_row;
+       yoffset++) {
+    for (MCU_col_num = coef->MCU_ctr; MCU_col_num < cinfo->MCUs_per_row;
+	 MCU_col_num++) {
+      /* Construct list of pointers to DCT blocks belonging to this MCU */
+      blkn = 0;			/* index of current DCT block within MCU */
+      for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+	compptr = cinfo->cur_comp_info[ci];
+	start_col = MCU_col_num * compptr->MCU_width;
+	for (yindex = 0; yindex < compptr->MCU_height; yindex++) {
+	  buffer_ptr = buffer[ci][yindex+yoffset] + start_col;
+	  for (xindex = 0; xindex < compptr->MCU_width; xindex++) {
+	    coef->MCU_buffer[blkn++] = buffer_ptr++;
+	  }
+	}
+      }
+      /* Try to fetch the MCU. */
+      if (! (*cinfo->entropy->decode_mcu) (cinfo, coef->MCU_buffer)) {
+	/* Suspension forced; update state counters and exit */
+	coef->MCU_vert_offset = yoffset;
+	coef->MCU_ctr = MCU_col_num;
+	return JPEG_SUSPENDED;
+      }
+    }
+    /* Completed an MCU row, but perhaps not an iMCU row */
+    coef->MCU_ctr = 0;
+  }
+  /* Completed the iMCU row, advance counters for next one */
+  if (++(cinfo->input_iMCU_row) < cinfo->total_iMCU_rows) {
+    start_iMCU_row(cinfo);
+    return JPEG_ROW_COMPLETED;
+  }
+  /* Completed the scan */
+  (*cinfo->inputctl->finish_input_pass) (cinfo);
+  return JPEG_SCAN_COMPLETED;
+}
+
+
+/*
+ * Decompress and return some data in the multi-pass case.
+ * Always attempts to emit one fully interleaved MCU row ("iMCU" row).
+ * Return value is JPEG_ROW_COMPLETED, JPEG_SCAN_COMPLETED, or JPEG_SUSPENDED.
+ *
+ * NB: output_buf contains a plane for each component in image.
+ */
+
+METHODDEF(int)
+decompress_data (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION block_num;
+  int ci, block_row, block_rows;
+  JBLOCKARRAY buffer;
+  JBLOCKROW buffer_ptr;
+  JSAMPARRAY output_ptr;
+  JDIMENSION output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+
+  /* Force some input to be done if we are getting ahead of the input. */
+  while (cinfo->input_scan_number < cinfo->output_scan_number ||
+	 (cinfo->input_scan_number == cinfo->output_scan_number &&
+	  cinfo->input_iMCU_row <= cinfo->output_iMCU_row)) {
+    if ((*cinfo->inputctl->consume_input)(cinfo) == JPEG_SUSPENDED)
+      return JPEG_SUSPENDED;
+  }
+
+  /* OK, output from the virtual arrays. */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Don't bother to IDCT an uninteresting component. */
+    if (! compptr->component_needed)
+      continue;
+    /* Align the virtual buffer for this component. */
+    buffer = (*cinfo->mem->access_virt_barray)
+      ((j_common_ptr) cinfo, coef->whole_image[ci],
+       cinfo->output_iMCU_row * compptr->v_samp_factor,
+       (JDIMENSION) compptr->v_samp_factor, FALSE);
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (cinfo->output_iMCU_row < last_iMCU_row)
+      block_rows = compptr->v_samp_factor;
+    else {
+      /* NB: can't use last_row_height here; it is input-side-dependent! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+    }
+    inverse_DCT = cinfo->idct->inverse_DCT[ci];
+    output_ptr = output_buf[ci];
+    /* Loop over all DCT blocks to be processed. */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      buffer_ptr = buffer[block_row];
+      output_col = 0;
+      for (block_num = 0; block_num < compptr->width_in_blocks; block_num++) {
+	(*inverse_DCT) (cinfo, compptr, (JCOEFPTR) buffer_ptr,
+			output_ptr, output_col);
+	buffer_ptr++;
+	output_col += compptr->DCT_h_scaled_size;
+      }
+      output_ptr += compptr->DCT_v_scaled_size;
+    }
+  }
+
+  if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
+    return JPEG_ROW_COMPLETED;
+  return JPEG_SCAN_COMPLETED;
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+
+/*
+ * This code applies interblock smoothing as described by section K.8
+ * of the JPEG standard: the first 5 AC coefficients are estimated from
+ * the DC values of a DCT block and its 8 neighboring blocks.
+ * We apply smoothing only for progressive JPEG decoding, and only if
+ * the coefficients it can estimate are not yet known to full precision.
+ */
+
+/* Natural-order array positions of the first 5 zigzag-order coefficients */
+#define Q01_POS  1
+#define Q10_POS  8
+#define Q20_POS  16
+#define Q11_POS  9
+#define Q02_POS  2
+
+/*
+ * Determine whether block smoothing is applicable and safe.
+ * We also latch the current states of the coef_bits[] entries for the
+ * AC coefficients; otherwise, if the input side of the decompressor
+ * advances into a new scan, we might think the coefficients are known
+ * more accurately than they really are.
+ */
+
+LOCAL(boolean)
+smoothing_ok (j_decompress_ptr cinfo)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  boolean smoothing_useful = FALSE;
+  int ci, coefi;
+  jpeg_component_info *compptr;
+  JQUANT_TBL * qtable;
+  int * coef_bits;
+  int * coef_bits_latch;
+
+  if (! cinfo->progressive_mode || cinfo->coef_bits == NULL)
+    return FALSE;
+
+  /* Allocate latch area if not already done */
+  if (coef->coef_bits_latch == NULL)
+    coef->coef_bits_latch = (int *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  cinfo->num_components *
+				  (SAVED_COEFS * SIZEOF(int)));
+  coef_bits_latch = coef->coef_bits_latch;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* All components' quantization values must already be latched. */
+    if ((qtable = compptr->quant_table) == NULL)
+      return FALSE;
+    /* Verify DC & first 5 AC quantizers are nonzero to avoid zero-divide. */
+    if (qtable->quantval[0] == 0 ||
+	qtable->quantval[Q01_POS] == 0 ||
+	qtable->quantval[Q10_POS] == 0 ||
+	qtable->quantval[Q20_POS] == 0 ||
+	qtable->quantval[Q11_POS] == 0 ||
+	qtable->quantval[Q02_POS] == 0)
+      return FALSE;
+    /* DC values must be at least partly known for all components. */
+    coef_bits = cinfo->coef_bits[ci];
+    if (coef_bits[0] < 0)
+      return FALSE;
+    /* Block smoothing is helpful if some AC coefficients remain inaccurate. */
+    for (coefi = 1; coefi <= 5; coefi++) {
+      coef_bits_latch[coefi] = coef_bits[coefi];
+      if (coef_bits[coefi] != 0)
+	smoothing_useful = TRUE;
+    }
+    coef_bits_latch += SAVED_COEFS;
+  }
+
+  return smoothing_useful;
+}
+
+
+/*
+ * Variant of decompress_data for use when doing block smoothing.
+ */
+
+METHODDEF(int)
+decompress_smooth_data (j_decompress_ptr cinfo, JSAMPIMAGE output_buf)
+{
+  my_coef_ptr coef = (my_coef_ptr) cinfo->coef;
+  JDIMENSION last_iMCU_row = cinfo->total_iMCU_rows - 1;
+  JDIMENSION block_num, last_block_column;
+  int ci, block_row, block_rows, access_rows;
+  JBLOCKARRAY buffer;
+  JBLOCKROW buffer_ptr, prev_block_row, next_block_row;
+  JSAMPARRAY output_ptr;
+  JDIMENSION output_col;
+  jpeg_component_info *compptr;
+  inverse_DCT_method_ptr inverse_DCT;
+  boolean first_row, last_row;
+  JBLOCK workspace;
+  int *coef_bits;
+  JQUANT_TBL *quanttbl;
+  INT32 Q00,Q01,Q02,Q10,Q11,Q20, num;
+  int DC1,DC2,DC3,DC4,DC5,DC6,DC7,DC8,DC9;
+  int Al, pred;
+
+  /* Force some input to be done if we are getting ahead of the input. */
+  while (cinfo->input_scan_number <= cinfo->output_scan_number &&
+	 ! cinfo->inputctl->eoi_reached) {
+    if (cinfo->input_scan_number == cinfo->output_scan_number) {
+      /* If input is working on current scan, we ordinarily want it to
+       * have completed the current row.  But if input scan is DC,
+       * we want it to keep one row ahead so that next block row's DC
+       * values are up to date.
+       */
+      JDIMENSION delta = (cinfo->Ss == 0) ? 1 : 0;
+      if (cinfo->input_iMCU_row > cinfo->output_iMCU_row+delta)
+	break;
+    }
+    if ((*cinfo->inputctl->consume_input)(cinfo) == JPEG_SUSPENDED)
+      return JPEG_SUSPENDED;
+  }
+
+  /* OK, output from the virtual arrays. */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Don't bother to IDCT an uninteresting component. */
+    if (! compptr->component_needed)
+      continue;
+    /* Count non-dummy DCT block rows in this iMCU row. */
+    if (cinfo->output_iMCU_row < last_iMCU_row) {
+      block_rows = compptr->v_samp_factor;
+      access_rows = block_rows * 2; /* this and next iMCU row */
+      last_row = FALSE;
+    } else {
+      /* NB: can't use last_row_height here; it is input-side-dependent! */
+      block_rows = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+      if (block_rows == 0) block_rows = compptr->v_samp_factor;
+      access_rows = block_rows; /* this iMCU row only */
+      last_row = TRUE;
+    }
+    /* Align the virtual buffer for this component. */
+    if (cinfo->output_iMCU_row > 0) {
+      access_rows += compptr->v_samp_factor; /* prior iMCU row too */
+      buffer = (*cinfo->mem->access_virt_barray)
+	((j_common_ptr) cinfo, coef->whole_image[ci],
+	 (cinfo->output_iMCU_row - 1) * compptr->v_samp_factor,
+	 (JDIMENSION) access_rows, FALSE);
+      buffer += compptr->v_samp_factor;	/* point to current iMCU row */
+      first_row = FALSE;
+    } else {
+      buffer = (*cinfo->mem->access_virt_barray)
+	((j_common_ptr) cinfo, coef->whole_image[ci],
+	 (JDIMENSION) 0, (JDIMENSION) access_rows, FALSE);
+      first_row = TRUE;
+    }
+    /* Fetch component-dependent info */
+    coef_bits = coef->coef_bits_latch + (ci * SAVED_COEFS);
+    quanttbl = compptr->quant_table;
+    Q00 = quanttbl->quantval[0];
+    Q01 = quanttbl->quantval[Q01_POS];
+    Q10 = quanttbl->quantval[Q10_POS];
+    Q20 = quanttbl->quantval[Q20_POS];
+    Q11 = quanttbl->quantval[Q11_POS];
+    Q02 = quanttbl->quantval[Q02_POS];
+    inverse_DCT = cinfo->idct->inverse_DCT[ci];
+    output_ptr = output_buf[ci];
+    /* Loop over all DCT blocks to be processed. */
+    for (block_row = 0; block_row < block_rows; block_row++) {
+      buffer_ptr = buffer[block_row];
+      if (first_row && block_row == 0)
+	prev_block_row = buffer_ptr;
+      else
+	prev_block_row = buffer[block_row-1];
+      if (last_row && block_row == block_rows-1)
+	next_block_row = buffer_ptr;
+      else
+	next_block_row = buffer[block_row+1];
+      /* We fetch the surrounding DC values using a sliding-register approach.
+       * Initialize all nine here so as to do the right thing on narrow pics.
+       */
+      DC1 = DC2 = DC3 = (int) prev_block_row[0][0];
+      DC4 = DC5 = DC6 = (int) buffer_ptr[0][0];
+      DC7 = DC8 = DC9 = (int) next_block_row[0][0];
+      output_col = 0;
+      last_block_column = compptr->width_in_blocks - 1;
+      for (block_num = 0; block_num <= last_block_column; block_num++) {
+	/* Fetch current DCT block into workspace so we can modify it. */
+	jcopy_block_row(buffer_ptr, (JBLOCKROW) workspace, (JDIMENSION) 1);
+	/* Update DC values */
+	if (block_num < last_block_column) {
+	  DC3 = (int) prev_block_row[1][0];
+	  DC6 = (int) buffer_ptr[1][0];
+	  DC9 = (int) next_block_row[1][0];
+	}
+	/* Compute coefficient estimates per K.8.
+	 * An estimate is applied only if coefficient is still zero,
+	 * and is not known to be fully accurate.
+	 */
+	/* AC01 */
+	if ((Al=coef_bits[1]) != 0 && workspace[1] == 0) {
+	  num = 36 * Q00 * (DC4 - DC6);
+	  if (num >= 0) {
+	    pred = (int) (((Q01<<7) + num) / (Q01<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q01<<7) - num) / (Q01<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[1] = (JCOEF) pred;
+	}
+	/* AC10 */
+	if ((Al=coef_bits[2]) != 0 && workspace[8] == 0) {
+	  num = 36 * Q00 * (DC2 - DC8);
+	  if (num >= 0) {
+	    pred = (int) (((Q10<<7) + num) / (Q10<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q10<<7) - num) / (Q10<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[8] = (JCOEF) pred;
+	}
+	/* AC20 */
+	if ((Al=coef_bits[3]) != 0 && workspace[16] == 0) {
+	  num = 9 * Q00 * (DC2 + DC8 - 2*DC5);
+	  if (num >= 0) {
+	    pred = (int) (((Q20<<7) + num) / (Q20<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q20<<7) - num) / (Q20<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[16] = (JCOEF) pred;
+	}
+	/* AC11 */
+	if ((Al=coef_bits[4]) != 0 && workspace[9] == 0) {
+	  num = 5 * Q00 * (DC1 - DC3 - DC7 + DC9);
+	  if (num >= 0) {
+	    pred = (int) (((Q11<<7) + num) / (Q11<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q11<<7) - num) / (Q11<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[9] = (JCOEF) pred;
+	}
+	/* AC02 */
+	if ((Al=coef_bits[5]) != 0 && workspace[2] == 0) {
+	  num = 9 * Q00 * (DC4 + DC6 - 2*DC5);
+	  if (num >= 0) {
+	    pred = (int) (((Q02<<7) + num) / (Q02<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	  } else {
+	    pred = (int) (((Q02<<7) - num) / (Q02<<8));
+	    if (Al > 0 && pred >= (1<<Al))
+	      pred = (1<<Al)-1;
+	    pred = -pred;
+	  }
+	  workspace[2] = (JCOEF) pred;
+	}
+	/* OK, do the IDCT */
+	(*inverse_DCT) (cinfo, compptr, (JCOEFPTR) workspace,
+			output_ptr, output_col);
+	/* Advance for next column */
+	DC1 = DC2; DC2 = DC3;
+	DC4 = DC5; DC5 = DC6;
+	DC7 = DC8; DC8 = DC9;
+	buffer_ptr++, prev_block_row++, next_block_row++;
+	output_col += compptr->DCT_h_scaled_size;
+      }
+      output_ptr += compptr->DCT_v_scaled_size;
+    }
+  }
+
+  if (++(cinfo->output_iMCU_row) < cinfo->total_iMCU_rows)
+    return JPEG_ROW_COMPLETED;
+  return JPEG_SCAN_COMPLETED;
+}
+
+#endif /* BLOCK_SMOOTHING_SUPPORTED */
+
+
+/*
+ * Initialize coefficient buffer controller.
+ */
+
+GLOBAL(void)
+jinit_d_coef_controller (j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_coef_ptr coef;
+
+  coef = (my_coef_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_coef_controller));
+  cinfo->coef = (struct jpeg_d_coef_controller *) coef;
+  coef->pub.start_input_pass = start_input_pass;
+  coef->pub.start_output_pass = start_output_pass;
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+  coef->coef_bits_latch = NULL;
+#endif
+
+  /* Create the coefficient buffer. */
+  if (need_full_buffer) {
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+    /* Allocate a full-image virtual array for each component, */
+    /* padded to a multiple of samp_factor DCT blocks in each direction. */
+    /* Note we ask for a pre-zeroed array. */
+    int ci, access_rows;
+    jpeg_component_info *compptr;
+
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      access_rows = compptr->v_samp_factor;
+#ifdef BLOCK_SMOOTHING_SUPPORTED
+      /* If block smoothing could be used, need a bigger window */
+      if (cinfo->progressive_mode)
+	access_rows *= 3;
+#endif
+      coef->whole_image[ci] = (*cinfo->mem->request_virt_barray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, TRUE,
+	 (JDIMENSION) jround_up((long) compptr->width_in_blocks,
+				(long) compptr->h_samp_factor),
+	 (JDIMENSION) jround_up((long) compptr->height_in_blocks,
+				(long) compptr->v_samp_factor),
+	 (JDIMENSION) access_rows);
+    }
+    coef->pub.consume_data = consume_data;
+    coef->pub.decompress_data = decompress_data;
+    coef->pub.coef_arrays = coef->whole_image; /* link to virtual arrays */
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+  } else {
+    /* We only need a single-MCU buffer. */
+    JBLOCKROW buffer;
+    int i;
+
+    buffer = (JBLOCKROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  D_MAX_BLOCKS_IN_MCU * SIZEOF(JBLOCK));
+    for (i = 0; i < D_MAX_BLOCKS_IN_MCU; i++) {
+      coef->MCU_buffer[i] = buffer + i;
+    }
+    coef->pub.consume_data = dummy_consume_data;
+    coef->pub.decompress_data = decompress_onepass;
+    coef->pub.coef_arrays = NULL; /* flag for no virtual arrays */
+  }
+}

Added: trunk/code/jpeg-8c/jdcolor.c
===================================================================
--- trunk/code/jpeg-8c/jdcolor.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdcolor.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,396 @@
+/*
+ * jdcolor.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains output colorspace conversion routines.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_color_deconverter pub; /* public fields */
+
+  /* Private state for YCC->RGB conversion */
+  int * Cr_r_tab;		/* => table for Cr to R conversion */
+  int * Cb_b_tab;		/* => table for Cb to B conversion */
+  INT32 * Cr_g_tab;		/* => table for Cr to G conversion */
+  INT32 * Cb_g_tab;		/* => table for Cb to G conversion */
+} my_color_deconverter;
+
+typedef my_color_deconverter * my_cconvert_ptr;
+
+
+/**************** YCbCr -> RGB conversion: most common case **************/
+
+/*
+ * YCbCr is defined per CCIR 601-1, except that Cb and Cr are
+ * normalized to the range 0..MAXJSAMPLE rather than -0.5 .. 0.5.
+ * The conversion equations to be implemented are therefore
+ *	R = Y                + 1.40200 * Cr
+ *	G = Y - 0.34414 * Cb - 0.71414 * Cr
+ *	B = Y + 1.77200 * Cb
+ * where Cb and Cr represent the incoming values less CENTERJSAMPLE.
+ * (These numbers are derived from TIFF 6.0 section 21, dated 3-June-92.)
+ *
+ * To avoid floating-point arithmetic, we represent the fractional constants
+ * as integers scaled up by 2^16 (about 4 digits precision); we have to divide
+ * the products by 2^16, with appropriate rounding, to get the correct answer.
+ * Notice that Y, being an integral input, does not contribute any fraction
+ * so it need not participate in the rounding.
+ *
+ * For even more speed, we avoid doing any multiplications in the inner loop
+ * by precalculating the constants times Cb and Cr for all possible values.
+ * For 8-bit JSAMPLEs this is very reasonable (only 256 entries per table);
+ * for 12-bit samples it is still acceptable.  It's not very reasonable for
+ * 16-bit samples, but if you want lossless storage you shouldn't be changing
+ * colorspace anyway.
+ * The Cr=>R and Cb=>B values can be rounded to integers in advance; the
+ * values for the G calculation are left scaled up, since we must add them
+ * together before rounding.
+ */
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define ONE_HALF	((INT32) 1 << (SCALEBITS-1))
+#define FIX(x)		((INT32) ((x) * (1L<<SCALEBITS) + 0.5))
+
+
+/*
+ * Initialize tables for YCC->RGB colorspace conversion.
+ */
+
+LOCAL(void)
+build_ycc_rgb_table (j_decompress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  int i;
+  INT32 x;
+  SHIFT_TEMPS
+
+  cconvert->Cr_r_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  cconvert->Cb_b_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  cconvert->Cr_g_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(INT32));
+  cconvert->Cb_g_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(INT32));
+
+  for (i = 0, x = -CENTERJSAMPLE; i <= MAXJSAMPLE; i++, x++) {
+    /* i is the actual input pixel value, in the range 0..MAXJSAMPLE */
+    /* The Cb or Cr value we are thinking of is x = i - CENTERJSAMPLE */
+    /* Cr=>R value is nearest int to 1.40200 * x */
+    cconvert->Cr_r_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.40200) * x + ONE_HALF, SCALEBITS);
+    /* Cb=>B value is nearest int to 1.77200 * x */
+    cconvert->Cb_b_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.77200) * x + ONE_HALF, SCALEBITS);
+    /* Cr=>G value is scaled-up -0.71414 * x */
+    cconvert->Cr_g_tab[i] = (- FIX(0.71414)) * x;
+    /* Cb=>G value is scaled-up -0.34414 * x */
+    /* We also add in ONE_HALF so that need not do it in inner loop */
+    cconvert->Cb_g_tab[i] = (- FIX(0.34414)) * x + ONE_HALF;
+  }
+}
+
+
+/*
+ * Convert some rows of samples to the output colorspace.
+ *
+ * Note that we change from noninterleaved, one-plane-per-component format
+ * to interleaved-pixel format.  The output buffer is therefore three times
+ * as wide as the input buffer.
+ * A starting row offset is provided only for the input buffer.  The caller
+ * can easily adjust the passed output_buf value to accommodate any row
+ * offset required on that side.
+ */
+
+METHODDEF(void)
+ycc_rgb_convert (j_decompress_ptr cinfo,
+		 JSAMPIMAGE input_buf, JDIMENSION input_row,
+		 JSAMPARRAY output_buf, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int y, cb, cr;
+  register JSAMPROW outptr;
+  register JSAMPROW inptr0, inptr1, inptr2;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  register JSAMPLE * range_limit = cinfo->sample_range_limit;
+  register int * Crrtab = cconvert->Cr_r_tab;
+  register int * Cbbtab = cconvert->Cb_b_tab;
+  register INT32 * Crgtab = cconvert->Cr_g_tab;
+  register INT32 * Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      y  = GETJSAMPLE(inptr0[col]);
+      cb = GETJSAMPLE(inptr1[col]);
+      cr = GETJSAMPLE(inptr2[col]);
+      /* Range-limiting is essential due to noise introduced by DCT losses. */
+      outptr[RGB_RED] =   range_limit[y + Crrtab[cr]];
+      outptr[RGB_GREEN] = range_limit[y +
+			      ((int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+						 SCALEBITS))];
+      outptr[RGB_BLUE] =  range_limit[y + Cbbtab[cb]];
+      outptr += RGB_PIXELSIZE;
+    }
+  }
+}
+
+
+/**************** Cases other than YCbCr -> RGB **************/
+
+
+/*
+ * Color conversion for no colorspace change: just copy the data,
+ * converting from separate-planes to interleaved representation.
+ */
+
+METHODDEF(void)
+null_convert (j_decompress_ptr cinfo,
+	      JSAMPIMAGE input_buf, JDIMENSION input_row,
+	      JSAMPARRAY output_buf, int num_rows)
+{
+  register JSAMPROW inptr, outptr;
+  register JDIMENSION count;
+  register int num_components = cinfo->num_components;
+  JDIMENSION num_cols = cinfo->output_width;
+  int ci;
+
+  while (--num_rows >= 0) {
+    for (ci = 0; ci < num_components; ci++) {
+      inptr = input_buf[ci][input_row];
+      outptr = output_buf[0] + ci;
+      for (count = num_cols; count > 0; count--) {
+	*outptr = *inptr++;	/* needn't bother with GETJSAMPLE() here */
+	outptr += num_components;
+      }
+    }
+    input_row++;
+    output_buf++;
+  }
+}
+
+
+/*
+ * Color conversion for grayscale: just copy the data.
+ * This also works for YCbCr -> grayscale conversion, in which
+ * we just copy the Y (luminance) component and ignore chrominance.
+ */
+
+METHODDEF(void)
+grayscale_convert (j_decompress_ptr cinfo,
+		   JSAMPIMAGE input_buf, JDIMENSION input_row,
+		   JSAMPARRAY output_buf, int num_rows)
+{
+  jcopy_sample_rows(input_buf[0], (int) input_row, output_buf, 0,
+		    num_rows, cinfo->output_width);
+}
+
+
+/*
+ * Convert grayscale to RGB: just duplicate the graylevel three times.
+ * This is provided to support applications that don't want to cope
+ * with grayscale as a separate case.
+ */
+
+METHODDEF(void)
+gray_rgb_convert (j_decompress_ptr cinfo,
+		  JSAMPIMAGE input_buf, JDIMENSION input_row,
+		  JSAMPARRAY output_buf, int num_rows)
+{
+  register JSAMPROW inptr, outptr;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+
+  while (--num_rows >= 0) {
+    inptr = input_buf[0][input_row++];
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      /* We can dispense with GETJSAMPLE() here */
+      outptr[RGB_RED] = outptr[RGB_GREEN] = outptr[RGB_BLUE] = inptr[col];
+      outptr += RGB_PIXELSIZE;
+    }
+  }
+}
+
+
+/*
+ * Adobe-style YCCK->CMYK conversion.
+ * We convert YCbCr to R=1-C, G=1-M, and B=1-Y using the same
+ * conversion as above, while passing K (black) unchanged.
+ * We assume build_ycc_rgb_table has been called.
+ */
+
+METHODDEF(void)
+ycck_cmyk_convert (j_decompress_ptr cinfo,
+		   JSAMPIMAGE input_buf, JDIMENSION input_row,
+		   JSAMPARRAY output_buf, int num_rows)
+{
+  my_cconvert_ptr cconvert = (my_cconvert_ptr) cinfo->cconvert;
+  register int y, cb, cr;
+  register JSAMPROW outptr;
+  register JSAMPROW inptr0, inptr1, inptr2, inptr3;
+  register JDIMENSION col;
+  JDIMENSION num_cols = cinfo->output_width;
+  /* copy these pointers into registers if possible */
+  register JSAMPLE * range_limit = cinfo->sample_range_limit;
+  register int * Crrtab = cconvert->Cr_r_tab;
+  register int * Cbbtab = cconvert->Cb_b_tab;
+  register INT32 * Crgtab = cconvert->Cr_g_tab;
+  register INT32 * Cbgtab = cconvert->Cb_g_tab;
+  SHIFT_TEMPS
+
+  while (--num_rows >= 0) {
+    inptr0 = input_buf[0][input_row];
+    inptr1 = input_buf[1][input_row];
+    inptr2 = input_buf[2][input_row];
+    inptr3 = input_buf[3][input_row];
+    input_row++;
+    outptr = *output_buf++;
+    for (col = 0; col < num_cols; col++) {
+      y  = GETJSAMPLE(inptr0[col]);
+      cb = GETJSAMPLE(inptr1[col]);
+      cr = GETJSAMPLE(inptr2[col]);
+      /* Range-limiting is essential due to noise introduced by DCT losses. */
+      outptr[0] = range_limit[MAXJSAMPLE - (y + Crrtab[cr])];	/* red */
+      outptr[1] = range_limit[MAXJSAMPLE - (y +			/* green */
+			      ((int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr],
+						 SCALEBITS)))];
+      outptr[2] = range_limit[MAXJSAMPLE - (y + Cbbtab[cb])];	/* blue */
+      /* K passes through unchanged */
+      outptr[3] = inptr3[col];	/* don't need GETJSAMPLE here */
+      outptr += 4;
+    }
+  }
+}
+
+
+/*
+ * Empty method for start_pass.
+ */
+
+METHODDEF(void)
+start_pass_dcolor (j_decompress_ptr cinfo)
+{
+  /* no work needed */
+}
+
+
+/*
+ * Module initialization routine for output colorspace conversion.
+ */
+
+GLOBAL(void)
+jinit_color_deconverter (j_decompress_ptr cinfo)
+{
+  my_cconvert_ptr cconvert;
+  int ci;
+
+  cconvert = (my_cconvert_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_color_deconverter));
+  cinfo->cconvert = (struct jpeg_color_deconverter *) cconvert;
+  cconvert->pub.start_pass = start_pass_dcolor;
+
+  /* Make sure num_components agrees with jpeg_color_space */
+  switch (cinfo->jpeg_color_space) {
+  case JCS_GRAYSCALE:
+    if (cinfo->num_components != 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  case JCS_RGB:
+  case JCS_YCbCr:
+    if (cinfo->num_components != 3)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  case JCS_CMYK:
+  case JCS_YCCK:
+    if (cinfo->num_components != 4)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+
+  default:			/* JCS_UNKNOWN can be anything */
+    if (cinfo->num_components < 1)
+      ERREXIT(cinfo, JERR_BAD_J_COLORSPACE);
+    break;
+  }
+
+  /* Set out_color_components and conversion method based on requested space.
+   * Also clear the component_needed flags for any unused components,
+   * so that earlier pipeline stages can avoid useless computation.
+   */
+
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+    cinfo->out_color_components = 1;
+    if (cinfo->jpeg_color_space == JCS_GRAYSCALE ||
+	cinfo->jpeg_color_space == JCS_YCbCr) {
+      cconvert->pub.color_convert = grayscale_convert;
+      /* For color->grayscale conversion, only the Y (0) component is needed */
+      for (ci = 1; ci < cinfo->num_components; ci++)
+	cinfo->comp_info[ci].component_needed = FALSE;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_RGB:
+    cinfo->out_color_components = RGB_PIXELSIZE;
+    if (cinfo->jpeg_color_space == JCS_YCbCr) {
+      cconvert->pub.color_convert = ycc_rgb_convert;
+      build_ycc_rgb_table(cinfo);
+    } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
+      cconvert->pub.color_convert = gray_rgb_convert;
+    } else if (cinfo->jpeg_color_space == JCS_RGB && RGB_PIXELSIZE == 3) {
+      cconvert->pub.color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  case JCS_CMYK:
+    cinfo->out_color_components = 4;
+    if (cinfo->jpeg_color_space == JCS_YCCK) {
+      cconvert->pub.color_convert = ycck_cmyk_convert;
+      build_ycc_rgb_table(cinfo);
+    } else if (cinfo->jpeg_color_space == JCS_CMYK) {
+      cconvert->pub.color_convert = null_convert;
+    } else
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+
+  default:
+    /* Permit null conversion to same output space */
+    if (cinfo->out_color_space == cinfo->jpeg_color_space) {
+      cinfo->out_color_components = cinfo->num_components;
+      cconvert->pub.color_convert = null_convert;
+    } else			/* unsupported non-null conversion */
+      ERREXIT(cinfo, JERR_CONVERSION_NOTIMPL);
+    break;
+  }
+
+  if (cinfo->quantize_colors)
+    cinfo->output_components = 1; /* single colormapped output component */
+  else
+    cinfo->output_components = cinfo->out_color_components;
+}

Added: trunk/code/jpeg-8c/jdct.h
===================================================================
--- trunk/code/jpeg-8c/jdct.h	                        (rev 0)
+++ trunk/code/jpeg-8c/jdct.h	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,393 @@
+/*
+ * jdct.h
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This include file contains common declarations for the forward and
+ * inverse DCT modules.  These declarations are private to the DCT managers
+ * (jcdctmgr.c, jddctmgr.c) and the individual DCT algorithms.
+ * The individual DCT algorithms are kept in separate files to ease 
+ * machine-dependent tuning (e.g., assembly coding).
+ */
+
+
+/*
+ * A forward DCT routine is given a pointer to an input sample array and
+ * a pointer to a work area of type DCTELEM[]; the DCT is to be performed
+ * in-place in that buffer.  Type DCTELEM is int for 8-bit samples, INT32
+ * for 12-bit samples.  (NOTE: Floating-point DCT implementations use an
+ * array of type FAST_FLOAT, instead.)
+ * The input data is to be fetched from the sample array starting at a
+ * specified column.  (Any row offset needed will be applied to the array
+ * pointer before it is passed to the FDCT code.)
+ * Note that the number of samples fetched by the FDCT routine is
+ * DCT_h_scaled_size * DCT_v_scaled_size.
+ * The DCT outputs are returned scaled up by a factor of 8; they therefore
+ * have a range of +-8K for 8-bit data, +-128K for 12-bit data.  This
+ * convention improves accuracy in integer implementations and saves some
+ * work in floating-point ones.
+ * Quantization of the output coefficients is done by jcdctmgr.c.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+typedef int DCTELEM;		/* 16 or 32 bits is fine */
+#else
+typedef INT32 DCTELEM;		/* must have 32 bits */
+#endif
+
+typedef JMETHOD(void, forward_DCT_method_ptr, (DCTELEM * data,
+					       JSAMPARRAY sample_data,
+					       JDIMENSION start_col));
+typedef JMETHOD(void, float_DCT_method_ptr, (FAST_FLOAT * data,
+					     JSAMPARRAY sample_data,
+					     JDIMENSION start_col));
+
+
+/*
+ * An inverse DCT routine is given a pointer to the input JBLOCK and a pointer
+ * to an output sample array.  The routine must dequantize the input data as
+ * well as perform the IDCT; for dequantization, it uses the multiplier table
+ * pointed to by compptr->dct_table.  The output data is to be placed into the
+ * sample array starting at a specified column.  (Any row offset needed will
+ * be applied to the array pointer before it is passed to the IDCT code.)
+ * Note that the number of samples emitted by the IDCT routine is
+ * DCT_h_scaled_size * DCT_v_scaled_size.
+ */
+
+/* typedef inverse_DCT_method_ptr is declared in jpegint.h */
+
+/*
+ * Each IDCT routine has its own ideas about the best dct_table element type.
+ */
+
+typedef MULTIPLIER ISLOW_MULT_TYPE; /* short or int, whichever is faster */
+#if BITS_IN_JSAMPLE == 8
+typedef MULTIPLIER IFAST_MULT_TYPE; /* 16 bits is OK, use short if faster */
+#define IFAST_SCALE_BITS  2	/* fractional bits in scale factors */
+#else
+typedef INT32 IFAST_MULT_TYPE;	/* need 32 bits for scaled quantizers */
+#define IFAST_SCALE_BITS  13	/* fractional bits in scale factors */
+#endif
+typedef FAST_FLOAT FLOAT_MULT_TYPE; /* preferred floating type */
+
+
+/*
+ * Each IDCT routine is responsible for range-limiting its results and
+ * converting them to unsigned form (0..MAXJSAMPLE).  The raw outputs could
+ * be quite far out of range if the input data is corrupt, so a bulletproof
+ * range-limiting step is required.  We use a mask-and-table-lookup method
+ * to do the combined operations quickly.  See the comments with
+ * prepare_range_limit_table (in jdmaster.c) for more info.
+ */
+
+#define IDCT_range_limit(cinfo)  ((cinfo)->sample_range_limit + CENTERJSAMPLE)
+
+#define RANGE_MASK  (MAXJSAMPLE * 4 + 3) /* 2 bits wider than legal samples */
+
+
+/* Short forms of external names for systems with brain-damaged linkers. */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_fdct_islow		jFDislow
+#define jpeg_fdct_ifast		jFDifast
+#define jpeg_fdct_float		jFDfloat
+#define jpeg_fdct_7x7		jFD7x7
+#define jpeg_fdct_6x6		jFD6x6
+#define jpeg_fdct_5x5		jFD5x5
+#define jpeg_fdct_4x4		jFD4x4
+#define jpeg_fdct_3x3		jFD3x3
+#define jpeg_fdct_2x2		jFD2x2
+#define jpeg_fdct_1x1		jFD1x1
+#define jpeg_fdct_9x9		jFD9x9
+#define jpeg_fdct_10x10		jFD10x10
+#define jpeg_fdct_11x11		jFD11x11
+#define jpeg_fdct_12x12		jFD12x12
+#define jpeg_fdct_13x13		jFD13x13
+#define jpeg_fdct_14x14		jFD14x14
+#define jpeg_fdct_15x15		jFD15x15
+#define jpeg_fdct_16x16		jFD16x16
+#define jpeg_fdct_16x8		jFD16x8
+#define jpeg_fdct_14x7		jFD14x7
+#define jpeg_fdct_12x6		jFD12x6
+#define jpeg_fdct_10x5		jFD10x5
+#define jpeg_fdct_8x4		jFD8x4
+#define jpeg_fdct_6x3		jFD6x3
+#define jpeg_fdct_4x2		jFD4x2
+#define jpeg_fdct_2x1		jFD2x1
+#define jpeg_fdct_8x16		jFD8x16
+#define jpeg_fdct_7x14		jFD7x14
+#define jpeg_fdct_6x12		jFD6x12
+#define jpeg_fdct_5x10		jFD5x10
+#define jpeg_fdct_4x8		jFD4x8
+#define jpeg_fdct_3x6		jFD3x6
+#define jpeg_fdct_2x4		jFD2x4
+#define jpeg_fdct_1x2		jFD1x2
+#define jpeg_idct_islow		jRDislow
+#define jpeg_idct_ifast		jRDifast
+#define jpeg_idct_float		jRDfloat
+#define jpeg_idct_7x7		jRD7x7
+#define jpeg_idct_6x6		jRD6x6
+#define jpeg_idct_5x5		jRD5x5
+#define jpeg_idct_4x4		jRD4x4
+#define jpeg_idct_3x3		jRD3x3
+#define jpeg_idct_2x2		jRD2x2
+#define jpeg_idct_1x1		jRD1x1
+#define jpeg_idct_9x9		jRD9x9
+#define jpeg_idct_10x10		jRD10x10
+#define jpeg_idct_11x11		jRD11x11
+#define jpeg_idct_12x12		jRD12x12
+#define jpeg_idct_13x13		jRD13x13
+#define jpeg_idct_14x14		jRD14x14
+#define jpeg_idct_15x15		jRD15x15
+#define jpeg_idct_16x16		jRD16x16
+#define jpeg_idct_16x8		jRD16x8
+#define jpeg_idct_14x7		jRD14x7
+#define jpeg_idct_12x6		jRD12x6
+#define jpeg_idct_10x5		jRD10x5
+#define jpeg_idct_8x4		jRD8x4
+#define jpeg_idct_6x3		jRD6x3
+#define jpeg_idct_4x2		jRD4x2
+#define jpeg_idct_2x1		jRD2x1
+#define jpeg_idct_8x16		jRD8x16
+#define jpeg_idct_7x14		jRD7x14
+#define jpeg_idct_6x12		jRD6x12
+#define jpeg_idct_5x10		jRD5x10
+#define jpeg_idct_4x8		jRD4x8
+#define jpeg_idct_3x6		jRD3x8
+#define jpeg_idct_2x4		jRD2x4
+#define jpeg_idct_1x2		jRD1x2
+#endif /* NEED_SHORT_EXTERNAL_NAMES */
+
+/* Extern declarations for the forward and inverse DCT routines. */
+
+EXTERN(void) jpeg_fdct_islow
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_ifast
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_float
+    JPP((FAST_FLOAT * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_7x7
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_6x6
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_5x5
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_4x4
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_3x3
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_2x2
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_1x1
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_9x9
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_10x10
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_11x11
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_12x12
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_13x13
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_14x14
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_15x15
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_16x16
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_16x8
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_14x7
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_12x6
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_10x5
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_8x4
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_6x3
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_4x2
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_2x1
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_8x16
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_7x14
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_6x12
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_5x10
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_4x8
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_3x6
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_2x4
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+EXTERN(void) jpeg_fdct_1x2
+    JPP((DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col));
+
+EXTERN(void) jpeg_idct_islow
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_ifast
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_float
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_7x7
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_6x6
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_5x5
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_4x4
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_3x3
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_2x2
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_1x1
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_9x9
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_10x10
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_11x11
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_12x12
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_13x13
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_14x14
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_15x15
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_16x16
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_16x8
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_14x7
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_12x6
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_10x5
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_8x4
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_6x3
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_4x2
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_2x1
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_8x16
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_7x14
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_6x12
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_5x10
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_4x8
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_3x6
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_2x4
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+EXTERN(void) jpeg_idct_1x2
+    JPP((j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	 JCOEFPTR coef_block, JSAMPARRAY output_buf, JDIMENSION output_col));
+
+
+/*
+ * Macros for handling fixed-point arithmetic; these are used by many
+ * but not all of the DCT/IDCT modules.
+ *
+ * All values are expected to be of type INT32.
+ * Fractional constants are scaled left by CONST_BITS bits.
+ * CONST_BITS is defined within each module using these macros,
+ * and may differ from one module to the next.
+ */
+
+#define ONE	((INT32) 1)
+#define CONST_SCALE (ONE << CONST_BITS)
+
+/* Convert a positive real constant to an integer scaled by CONST_SCALE.
+ * Caution: some C compilers fail to reduce "FIX(constant)" at compile time,
+ * thus causing a lot of useless floating-point operations at run time.
+ */
+
+#define FIX(x)	((INT32) ((x) * CONST_SCALE + 0.5))
+
+/* Descale and correctly round an INT32 value that's scaled by N bits.
+ * We assume RIGHT_SHIFT rounds towards minus infinity, so adding
+ * the fudge factor is correct for either sign of X.
+ */
+
+#define DESCALE(x,n)  RIGHT_SHIFT((x) + (ONE << ((n)-1)), n)
+
+/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
+ * This macro is used only when the two inputs will actually be no more than
+ * 16 bits wide, so that a 16x16->32 bit multiply can be used instead of a
+ * full 32x32 multiply.  This provides a useful speedup on many machines.
+ * Unfortunately there is no way to specify a 16x16->32 multiply portably
+ * in C, but some C compilers will do the right thing if you provide the
+ * correct combination of casts.
+ */
+
+#ifdef SHORTxSHORT_32		/* may work if 'int' is 32 bits */
+#define MULTIPLY16C16(var,const)  (((INT16) (var)) * ((INT16) (const)))
+#endif
+#ifdef SHORTxLCONST_32		/* known to work with Microsoft C 6.0 */
+#define MULTIPLY16C16(var,const)  (((INT16) (var)) * ((INT32) (const)))
+#endif
+
+#ifndef MULTIPLY16C16		/* default definition */
+#define MULTIPLY16C16(var,const)  ((var) * (const))
+#endif
+
+/* Same except both inputs are variables. */
+
+#ifdef SHORTxSHORT_32		/* may work if 'int' is 32 bits */
+#define MULTIPLY16V16(var1,var2)  (((INT16) (var1)) * ((INT16) (var2)))
+#endif
+
+#ifndef MULTIPLY16V16		/* default definition */
+#define MULTIPLY16V16(var1,var2)  ((var1) * (var2))
+#endif

Added: trunk/code/jpeg-8c/jddctmgr.c
===================================================================
--- trunk/code/jpeg-8c/jddctmgr.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jddctmgr.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,384 @@
+/*
+ * jddctmgr.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2002-2010 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the inverse-DCT management logic.
+ * This code selects a particular IDCT implementation to be used,
+ * and it performs related housekeeping chores.  No code in this file
+ * is executed per IDCT step, only during output pass setup.
+ *
+ * Note that the IDCT routines are responsible for performing coefficient
+ * dequantization as well as the IDCT proper.  This module sets up the
+ * dequantization multiplier table needed by the IDCT routine.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+
+/*
+ * The decompressor input side (jdinput.c) saves away the appropriate
+ * quantization table for each component at the start of the first scan
+ * involving that component.  (This is necessary in order to correctly
+ * decode files that reuse Q-table slots.)
+ * When we are ready to make an output pass, the saved Q-table is converted
+ * to a multiplier table that will actually be used by the IDCT routine.
+ * The multiplier table contents are IDCT-method-dependent.  To support
+ * application changes in IDCT method between scans, we can remake the
+ * multiplier tables if necessary.
+ * In buffered-image mode, the first output pass may occur before any data
+ * has been seen for some components, and thus before their Q-tables have
+ * been saved away.  To handle this case, multiplier tables are preset
+ * to zeroes; the result of the IDCT will be a neutral gray level.
+ */
+
+
+/* Private subobject for this module */
+
+typedef struct {
+  struct jpeg_inverse_dct pub;	/* public fields */
+
+  /* This array contains the IDCT method code that each multiplier table
+   * is currently set up for, or -1 if it's not yet set up.
+   * The actual multiplier tables are pointed to by dct_table in the
+   * per-component comp_info structures.
+   */
+  int cur_method[MAX_COMPONENTS];
+} my_idct_controller;
+
+typedef my_idct_controller * my_idct_ptr;
+
+
+/* Allocated multiplier tables: big enough for any supported variant */
+
+typedef union {
+  ISLOW_MULT_TYPE islow_array[DCTSIZE2];
+#ifdef DCT_IFAST_SUPPORTED
+  IFAST_MULT_TYPE ifast_array[DCTSIZE2];
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+  FLOAT_MULT_TYPE float_array[DCTSIZE2];
+#endif
+} multiplier_table;
+
+
+/* The current scaled-IDCT routines require ISLOW-style multiplier tables,
+ * so be sure to compile that code if either ISLOW or SCALING is requested.
+ */
+#ifdef DCT_ISLOW_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#else
+#ifdef IDCT_SCALING_SUPPORTED
+#define PROVIDE_ISLOW_TABLES
+#endif
+#endif
+
+
+/*
+ * Prepare for an output pass.
+ * Here we select the proper IDCT routine for each component and build
+ * a matching multiplier table.
+ */
+
+METHODDEF(void)
+start_pass (j_decompress_ptr cinfo)
+{
+  my_idct_ptr idct = (my_idct_ptr) cinfo->idct;
+  int ci, i;
+  jpeg_component_info *compptr;
+  int method = 0;
+  inverse_DCT_method_ptr method_ptr = NULL;
+  JQUANT_TBL * qtbl;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Select the proper IDCT routine for this component's scaling */
+    switch ((compptr->DCT_h_scaled_size << 8) + compptr->DCT_v_scaled_size) {
+#ifdef IDCT_SCALING_SUPPORTED
+    case ((1 << 8) + 1):
+      method_ptr = jpeg_idct_1x1;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((2 << 8) + 2):
+      method_ptr = jpeg_idct_2x2;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((3 << 8) + 3):
+      method_ptr = jpeg_idct_3x3;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((4 << 8) + 4):
+      method_ptr = jpeg_idct_4x4;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((5 << 8) + 5):
+      method_ptr = jpeg_idct_5x5;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((6 << 8) + 6):
+      method_ptr = jpeg_idct_6x6;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((7 << 8) + 7):
+      method_ptr = jpeg_idct_7x7;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((9 << 8) + 9):
+      method_ptr = jpeg_idct_9x9;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((10 << 8) + 10):
+      method_ptr = jpeg_idct_10x10;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((11 << 8) + 11):
+      method_ptr = jpeg_idct_11x11;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((12 << 8) + 12):
+      method_ptr = jpeg_idct_12x12;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((13 << 8) + 13):
+      method_ptr = jpeg_idct_13x13;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((14 << 8) + 14):
+      method_ptr = jpeg_idct_14x14;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((15 << 8) + 15):
+      method_ptr = jpeg_idct_15x15;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((16 << 8) + 16):
+      method_ptr = jpeg_idct_16x16;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((16 << 8) + 8):
+      method_ptr = jpeg_idct_16x8;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((14 << 8) + 7):
+      method_ptr = jpeg_idct_14x7;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((12 << 8) + 6):
+      method_ptr = jpeg_idct_12x6;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((10 << 8) + 5):
+      method_ptr = jpeg_idct_10x5;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((8 << 8) + 4):
+      method_ptr = jpeg_idct_8x4;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((6 << 8) + 3):
+      method_ptr = jpeg_idct_6x3;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((4 << 8) + 2):
+      method_ptr = jpeg_idct_4x2;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((2 << 8) + 1):
+      method_ptr = jpeg_idct_2x1;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((8 << 8) + 16):
+      method_ptr = jpeg_idct_8x16;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((7 << 8) + 14):
+      method_ptr = jpeg_idct_7x14;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((6 << 8) + 12):
+      method_ptr = jpeg_idct_6x12;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((5 << 8) + 10):
+      method_ptr = jpeg_idct_5x10;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((4 << 8) + 8):
+      method_ptr = jpeg_idct_4x8;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((3 << 8) + 6):
+      method_ptr = jpeg_idct_3x6;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((2 << 8) + 4):
+      method_ptr = jpeg_idct_2x4;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+    case ((1 << 8) + 2):
+      method_ptr = jpeg_idct_1x2;
+      method = JDCT_ISLOW;	/* jidctint uses islow-style table */
+      break;
+#endif
+    case ((DCTSIZE << 8) + DCTSIZE):
+      switch (cinfo->dct_method) {
+#ifdef DCT_ISLOW_SUPPORTED
+      case JDCT_ISLOW:
+	method_ptr = jpeg_idct_islow;
+	method = JDCT_ISLOW;
+	break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+      case JDCT_IFAST:
+	method_ptr = jpeg_idct_ifast;
+	method = JDCT_IFAST;
+	break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+      case JDCT_FLOAT:
+	method_ptr = jpeg_idct_float;
+	method = JDCT_FLOAT;
+	break;
+#endif
+      default:
+	ERREXIT(cinfo, JERR_NOT_COMPILED);
+	break;
+      }
+      break;
+    default:
+      ERREXIT2(cinfo, JERR_BAD_DCTSIZE,
+	       compptr->DCT_h_scaled_size, compptr->DCT_v_scaled_size);
+      break;
+    }
+    idct->pub.inverse_DCT[ci] = method_ptr;
+    /* Create multiplier table from quant table.
+     * However, we can skip this if the component is uninteresting
+     * or if we already built the table.  Also, if no quant table
+     * has yet been saved for the component, we leave the
+     * multiplier table all-zero; we'll be reading zeroes from the
+     * coefficient controller's buffer anyway.
+     */
+    if (! compptr->component_needed || idct->cur_method[ci] == method)
+      continue;
+    qtbl = compptr->quant_table;
+    if (qtbl == NULL)		/* happens if no data yet for component */
+      continue;
+    idct->cur_method[ci] = method;
+    switch (method) {
+#ifdef PROVIDE_ISLOW_TABLES
+    case JDCT_ISLOW:
+      {
+	/* For LL&M IDCT method, multipliers are equal to raw quantization
+	 * coefficients, but are stored as ints to ensure access efficiency.
+	 */
+	ISLOW_MULT_TYPE * ismtbl = (ISLOW_MULT_TYPE *) compptr->dct_table;
+	for (i = 0; i < DCTSIZE2; i++) {
+	  ismtbl[i] = (ISLOW_MULT_TYPE) qtbl->quantval[i];
+	}
+      }
+      break;
+#endif
+#ifdef DCT_IFAST_SUPPORTED
+    case JDCT_IFAST:
+      {
+	/* For AA&N IDCT method, multipliers are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * For integer operation, the multiplier table is to be scaled by
+	 * IFAST_SCALE_BITS.
+	 */
+	IFAST_MULT_TYPE * ifmtbl = (IFAST_MULT_TYPE *) compptr->dct_table;
+#define CONST_BITS 14
+	static const INT16 aanscales[DCTSIZE2] = {
+	  /* precomputed values scaled up by 14 bits */
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  22725, 31521, 29692, 26722, 22725, 17855, 12299,  6270,
+	  21407, 29692, 27969, 25172, 21407, 16819, 11585,  5906,
+	  19266, 26722, 25172, 22654, 19266, 15137, 10426,  5315,
+	  16384, 22725, 21407, 19266, 16384, 12873,  8867,  4520,
+	  12873, 17855, 16819, 15137, 12873, 10114,  6967,  3552,
+	   8867, 12299, 11585, 10426,  8867,  6967,  4799,  2446,
+	   4520,  6270,  5906,  5315,  4520,  3552,  2446,  1247
+	};
+	SHIFT_TEMPS
+
+	for (i = 0; i < DCTSIZE2; i++) {
+	  ifmtbl[i] = (IFAST_MULT_TYPE)
+	    DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
+				  (INT32) aanscales[i]),
+		    CONST_BITS-IFAST_SCALE_BITS);
+	}
+      }
+      break;
+#endif
+#ifdef DCT_FLOAT_SUPPORTED
+    case JDCT_FLOAT:
+      {
+	/* For float AA&N IDCT method, multipliers are equal to quantization
+	 * coefficients scaled by scalefactor[row]*scalefactor[col], where
+	 *   scalefactor[0] = 1
+	 *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
+	 * We apply a further scale factor of 1/8.
+	 */
+	FLOAT_MULT_TYPE * fmtbl = (FLOAT_MULT_TYPE *) compptr->dct_table;
+	int row, col;
+	static const double aanscalefactor[DCTSIZE] = {
+	  1.0, 1.387039845, 1.306562965, 1.175875602,
+	  1.0, 0.785694958, 0.541196100, 0.275899379
+	};
+
+	i = 0;
+	for (row = 0; row < DCTSIZE; row++) {
+	  for (col = 0; col < DCTSIZE; col++) {
+	    fmtbl[i] = (FLOAT_MULT_TYPE)
+	      ((double) qtbl->quantval[i] *
+	       aanscalefactor[row] * aanscalefactor[col] * 0.125);
+	    i++;
+	  }
+	}
+      }
+      break;
+#endif
+    default:
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+      break;
+    }
+  }
+}
+
+
+/*
+ * Initialize IDCT manager.
+ */
+
+GLOBAL(void)
+jinit_inverse_dct (j_decompress_ptr cinfo)
+{
+  my_idct_ptr idct;
+  int ci;
+  jpeg_component_info *compptr;
+
+  idct = (my_idct_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_idct_controller));
+  cinfo->idct = (struct jpeg_inverse_dct *) idct;
+  idct->pub.start_pass = start_pass;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Allocate and pre-zero a multiplier table for each component */
+    compptr->dct_table =
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(multiplier_table));
+    MEMZERO(compptr->dct_table, SIZEOF(multiplier_table));
+    /* Mark multiplier table not yet set up for any method */
+    idct->cur_method[ci] = -1;
+  }
+}

Added: trunk/code/jpeg-8c/jdhuff.c
===================================================================
--- trunk/code/jpeg-8c/jdhuff.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdhuff.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,1541 @@
+/*
+ * jdhuff.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2006-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains Huffman entropy decoding routines.
+ * Both sequential and progressive modes are supported in this single module.
+ *
+ * Much of the complexity here has to do with supporting input suspension.
+ * If the data source module demands suspension, we want to be able to back
+ * up to the start of the current MCU.  To do this, we copy state variables
+ * into local working storage, and update them back to the permanent
+ * storage only upon successful completion of an MCU.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Derived data constructed for each Huffman table */
+
+#define HUFF_LOOKAHEAD	8	/* # of bits of lookahead */
+
+typedef struct {
+  /* Basic tables: (element [0] of each array is unused) */
+  INT32 maxcode[18];		/* largest code of length k (-1 if none) */
+  /* (maxcode[17] is a sentinel to ensure jpeg_huff_decode terminates) */
+  INT32 valoffset[17];		/* huffval[] offset for codes of length k */
+  /* valoffset[k] = huffval[] index of 1st symbol of code length k, less
+   * the smallest code of length k; so given a code of length k, the
+   * corresponding symbol is huffval[code + valoffset[k]]
+   */
+
+  /* Link to public Huffman table (needed only in jpeg_huff_decode) */
+  JHUFF_TBL *pub;
+
+  /* Lookahead tables: indexed by the next HUFF_LOOKAHEAD bits of
+   * the input data stream.  If the next Huffman code is no more
+   * than HUFF_LOOKAHEAD bits long, we can obtain its length and
+   * the corresponding symbol directly from these tables.
+   */
+  int look_nbits[1<<HUFF_LOOKAHEAD]; /* # bits, or 0 if too long */
+  UINT8 look_sym[1<<HUFF_LOOKAHEAD]; /* symbol, or unused */
+} d_derived_tbl;
+
+
+/*
+ * Fetching the next N bits from the input stream is a time-critical operation
+ * for the Huffman decoders.  We implement it with a combination of inline
+ * macros and out-of-line subroutines.  Note that N (the number of bits
+ * demanded at one time) never exceeds 15 for JPEG use.
+ *
+ * We read source bytes into get_buffer and dole out bits as needed.
+ * If get_buffer already contains enough bits, they are fetched in-line
+ * by the macros CHECK_BIT_BUFFER and GET_BITS.  When there aren't enough
+ * bits, jpeg_fill_bit_buffer is called; it will attempt to fill get_buffer
+ * as full as possible (not just to the number of bits needed; this
+ * prefetching reduces the overhead cost of calling jpeg_fill_bit_buffer).
+ * Note that jpeg_fill_bit_buffer may return FALSE to indicate suspension.
+ * On TRUE return, jpeg_fill_bit_buffer guarantees that get_buffer contains
+ * at least the requested number of bits --- dummy zeroes are inserted if
+ * necessary.
+ */
+
+typedef INT32 bit_buf_type;	/* type of bit-extraction buffer */
+#define BIT_BUF_SIZE  32	/* size of buffer in bits */
+
+/* If long is > 32 bits on your machine, and shifting/masking longs is
+ * reasonably fast, making bit_buf_type be long and setting BIT_BUF_SIZE
+ * appropriately should be a win.  Unfortunately we can't define the size
+ * with something like  #define BIT_BUF_SIZE (sizeof(bit_buf_type)*8)
+ * because not all machines measure sizeof in 8-bit bytes.
+ */
+
+typedef struct {		/* Bitreading state saved across MCUs */
+  bit_buf_type get_buffer;	/* current bit-extraction buffer */
+  int bits_left;		/* # of unused bits in it */
+} bitread_perm_state;
+
+typedef struct {		/* Bitreading working state within an MCU */
+  /* Current data source location */
+  /* We need a copy, rather than munging the original, in case of suspension */
+  const JOCTET * next_input_byte; /* => next byte to read from source */
+  size_t bytes_in_buffer;	/* # of bytes remaining in source buffer */
+  /* Bit input buffer --- note these values are kept in register variables,
+   * not in this struct, inside the inner loops.
+   */
+  bit_buf_type get_buffer;	/* current bit-extraction buffer */
+  int bits_left;		/* # of unused bits in it */
+  /* Pointer needed by jpeg_fill_bit_buffer. */
+  j_decompress_ptr cinfo;	/* back link to decompress master record */
+} bitread_working_state;
+
+/* Macros to declare and load/save bitread local variables. */
+#define BITREAD_STATE_VARS  \
+	register bit_buf_type get_buffer;  \
+	register int bits_left;  \
+	bitread_working_state br_state
+
+#define BITREAD_LOAD_STATE(cinfop,permstate)  \
+	br_state.cinfo = cinfop; \
+	br_state.next_input_byte = cinfop->src->next_input_byte; \
+	br_state.bytes_in_buffer = cinfop->src->bytes_in_buffer; \
+	get_buffer = permstate.get_buffer; \
+	bits_left = permstate.bits_left;
+
+#define BITREAD_SAVE_STATE(cinfop,permstate)  \
+	cinfop->src->next_input_byte = br_state.next_input_byte; \
+	cinfop->src->bytes_in_buffer = br_state.bytes_in_buffer; \
+	permstate.get_buffer = get_buffer; \
+	permstate.bits_left = bits_left
+
+/*
+ * These macros provide the in-line portion of bit fetching.
+ * Use CHECK_BIT_BUFFER to ensure there are N bits in get_buffer
+ * before using GET_BITS, PEEK_BITS, or DROP_BITS.
+ * The variables get_buffer and bits_left are assumed to be locals,
+ * but the state struct might not be (jpeg_huff_decode needs this).
+ *	CHECK_BIT_BUFFER(state,n,action);
+ *		Ensure there are N bits in get_buffer; if suspend, take action.
+ *      val = GET_BITS(n);
+ *		Fetch next N bits.
+ *      val = PEEK_BITS(n);
+ *		Fetch next N bits without removing them from the buffer.
+ *	DROP_BITS(n);
+ *		Discard next N bits.
+ * The value N should be a simple variable, not an expression, because it
+ * is evaluated multiple times.
+ */
+
+#define CHECK_BIT_BUFFER(state,nbits,action) \
+	{ if (bits_left < (nbits)) {  \
+	    if (! jpeg_fill_bit_buffer(&(state),get_buffer,bits_left,nbits))  \
+	      { action; }  \
+	    get_buffer = (state).get_buffer; bits_left = (state).bits_left; } }
+
+#define GET_BITS(nbits) \
+	(((int) (get_buffer >> (bits_left -= (nbits)))) & BIT_MASK(nbits))
+
+#define PEEK_BITS(nbits) \
+	(((int) (get_buffer >> (bits_left -  (nbits)))) & BIT_MASK(nbits))
+
+#define DROP_BITS(nbits) \
+	(bits_left -= (nbits))
+
+
+/*
+ * Code for extracting next Huffman-coded symbol from input bit stream.
+ * Again, this is time-critical and we make the main paths be macros.
+ *
+ * We use a lookahead table to process codes of up to HUFF_LOOKAHEAD bits
+ * without looping.  Usually, more than 95% of the Huffman codes will be 8
+ * or fewer bits long.  The few overlength codes are handled with a loop,
+ * which need not be inline code.
+ *
+ * Notes about the HUFF_DECODE macro:
+ * 1. Near the end of the data segment, we may fail to get enough bits
+ *    for a lookahead.  In that case, we do it the hard way.
+ * 2. If the lookahead table contains no entry, the next code must be
+ *    more than HUFF_LOOKAHEAD bits long.
+ * 3. jpeg_huff_decode returns -1 if forced to suspend.
+ */
+
+#define HUFF_DECODE(result,state,htbl,failaction,slowlabel) \
+{ register int nb, look; \
+  if (bits_left < HUFF_LOOKAHEAD) { \
+    if (! jpeg_fill_bit_buffer(&state,get_buffer,bits_left, 0)) {failaction;} \
+    get_buffer = state.get_buffer; bits_left = state.bits_left; \
+    if (bits_left < HUFF_LOOKAHEAD) { \
+      nb = 1; goto slowlabel; \
+    } \
+  } \
+  look = PEEK_BITS(HUFF_LOOKAHEAD); \
+  if ((nb = htbl->look_nbits[look]) != 0) { \
+    DROP_BITS(nb); \
+    result = htbl->look_sym[look]; \
+  } else { \
+    nb = HUFF_LOOKAHEAD+1; \
+slowlabel: \
+    if ((result=jpeg_huff_decode(&state,get_buffer,bits_left,htbl,nb)) < 0) \
+	{ failaction; } \
+    get_buffer = state.get_buffer; bits_left = state.bits_left; \
+  } \
+}
+
+
+/*
+ * Expanded entropy decoder object for Huffman decoding.
+ *
+ * The savable_state subrecord contains fields that change within an MCU,
+ * but must not be updated permanently until we complete the MCU.
+ */
+
+typedef struct {
+  unsigned int EOBRUN;			/* remaining EOBs in EOBRUN */
+  int last_dc_val[MAX_COMPS_IN_SCAN];	/* last DC coef for each component */
+} savable_state;
+
+/* This macro is to work around compilers with missing or broken
+ * structure assignment.  You'll need to fix this code if you have
+ * such a compiler and you change MAX_COMPS_IN_SCAN.
+ */
+
+#ifndef NO_STRUCT_ASSIGN
+#define ASSIGN_STATE(dest,src)  ((dest) = (src))
+#else
+#if MAX_COMPS_IN_SCAN == 4
+#define ASSIGN_STATE(dest,src)  \
+	((dest).EOBRUN = (src).EOBRUN, \
+	 (dest).last_dc_val[0] = (src).last_dc_val[0], \
+	 (dest).last_dc_val[1] = (src).last_dc_val[1], \
+	 (dest).last_dc_val[2] = (src).last_dc_val[2], \
+	 (dest).last_dc_val[3] = (src).last_dc_val[3])
+#endif
+#endif
+
+
+typedef struct {
+  struct jpeg_entropy_decoder pub; /* public fields */
+
+  /* These fields are loaded into local variables at start of each MCU.
+   * In case of suspension, we exit WITHOUT updating them.
+   */
+  bitread_perm_state bitstate;	/* Bit buffer at start of MCU */
+  savable_state saved;		/* Other state at start of MCU */
+
+  /* These fields are NOT loaded into local working state. */
+  boolean insufficient_data;	/* set TRUE after emitting warning */
+  unsigned int restarts_to_go;	/* MCUs left in this restart interval */
+
+  /* Following two fields used only in progressive mode */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  d_derived_tbl * derived_tbls[NUM_HUFF_TBLS];
+
+  d_derived_tbl * ac_derived_tbl; /* active table during an AC scan */
+
+  /* Following fields used only in sequential mode */
+
+  /* Pointers to derived tables (these workspaces have image lifespan) */
+  d_derived_tbl * dc_derived_tbls[NUM_HUFF_TBLS];
+  d_derived_tbl * ac_derived_tbls[NUM_HUFF_TBLS];
+
+  /* Precalculated info set up by start_pass for use in decode_mcu: */
+
+  /* Pointers to derived tables to be used for each block within an MCU */
+  d_derived_tbl * dc_cur_tbls[D_MAX_BLOCKS_IN_MCU];
+  d_derived_tbl * ac_cur_tbls[D_MAX_BLOCKS_IN_MCU];
+  /* Whether we care about the DC and AC coefficient values for each block */
+  int coef_limit[D_MAX_BLOCKS_IN_MCU];
+} huff_entropy_decoder;
+
+typedef huff_entropy_decoder * huff_entropy_ptr;
+
+
+static const int jpeg_zigzag_order[8][8] = {
+  {  0,  1,  5,  6, 14, 15, 27, 28 },
+  {  2,  4,  7, 13, 16, 26, 29, 42 },
+  {  3,  8, 12, 17, 25, 30, 41, 43 },
+  {  9, 11, 18, 24, 31, 40, 44, 53 },
+  { 10, 19, 23, 32, 39, 45, 52, 54 },
+  { 20, 22, 33, 38, 46, 51, 55, 60 },
+  { 21, 34, 37, 47, 50, 56, 59, 61 },
+  { 35, 36, 48, 49, 57, 58, 62, 63 }
+};
+
+static const int jpeg_zigzag_order7[7][7] = {
+  {  0,  1,  5,  6, 14, 15, 27 },
+  {  2,  4,  7, 13, 16, 26, 28 },
+  {  3,  8, 12, 17, 25, 29, 38 },
+  {  9, 11, 18, 24, 30, 37, 39 },
+  { 10, 19, 23, 31, 36, 40, 45 },
+  { 20, 22, 32, 35, 41, 44, 46 },
+  { 21, 33, 34, 42, 43, 47, 48 }
+};
+
+static const int jpeg_zigzag_order6[6][6] = {
+  {  0,  1,  5,  6, 14, 15 },
+  {  2,  4,  7, 13, 16, 25 },
+  {  3,  8, 12, 17, 24, 26 },
+  {  9, 11, 18, 23, 27, 32 },
+  { 10, 19, 22, 28, 31, 33 },
+  { 20, 21, 29, 30, 34, 35 }
+};
+
+static const int jpeg_zigzag_order5[5][5] = {
+  {  0,  1,  5,  6, 14 },
+  {  2,  4,  7, 13, 15 },
+  {  3,  8, 12, 16, 21 },
+  {  9, 11, 17, 20, 22 },
+  { 10, 18, 19, 23, 24 }
+};
+
+static const int jpeg_zigzag_order4[4][4] = {
+  { 0,  1,  5,  6 },
+  { 2,  4,  7, 12 },
+  { 3,  8, 11, 13 },
+  { 9, 10, 14, 15 }
+};
+
+static const int jpeg_zigzag_order3[3][3] = {
+  { 0, 1, 5 },
+  { 2, 4, 6 },
+  { 3, 7, 8 }
+};
+
+static const int jpeg_zigzag_order2[2][2] = {
+  { 0, 1 },
+  { 2, 3 }
+};
+
+
+/*
+ * Compute the derived values for a Huffman table.
+ * This routine also performs some validation checks on the table.
+ */
+
+LOCAL(void)
+jpeg_make_d_derived_tbl (j_decompress_ptr cinfo, boolean isDC, int tblno,
+			 d_derived_tbl ** pdtbl)
+{
+  JHUFF_TBL *htbl;
+  d_derived_tbl *dtbl;
+  int p, i, l, si, numsymbols;
+  int lookbits, ctr;
+  char huffsize[257];
+  unsigned int huffcode[257];
+  unsigned int code;
+
+  /* Note that huffsize[] and huffcode[] are filled in code-length order,
+   * paralleling the order of the symbols themselves in htbl->huffval[].
+   */
+
+  /* Find the input Huffman table */
+  if (tblno < 0 || tblno >= NUM_HUFF_TBLS)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+  htbl =
+    isDC ? cinfo->dc_huff_tbl_ptrs[tblno] : cinfo->ac_huff_tbl_ptrs[tblno];
+  if (htbl == NULL)
+    ERREXIT1(cinfo, JERR_NO_HUFF_TABLE, tblno);
+
+  /* Allocate a workspace if we haven't already done so. */
+  if (*pdtbl == NULL)
+    *pdtbl = (d_derived_tbl *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(d_derived_tbl));
+  dtbl = *pdtbl;
+  dtbl->pub = htbl;		/* fill in back link */
+  
+  /* Figure C.1: make table of Huffman code length for each symbol */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    i = (int) htbl->bits[l];
+    if (i < 0 || p + i > 256)	/* protect against table overrun */
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    while (i--)
+      huffsize[p++] = (char) l;
+  }
+  huffsize[p] = 0;
+  numsymbols = p;
+  
+  /* Figure C.2: generate the codes themselves */
+  /* We also validate that the counts represent a legal Huffman code tree. */
+  
+  code = 0;
+  si = huffsize[0];
+  p = 0;
+  while (huffsize[p]) {
+    while (((int) huffsize[p]) == si) {
+      huffcode[p++] = code;
+      code++;
+    }
+    /* code is now 1 more than the last code used for codelength si; but
+     * it must still fit in si bits, since no code is allowed to be all ones.
+     */
+    if (((INT32) code) >= (((INT32) 1) << si))
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    code <<= 1;
+    si++;
+  }
+
+  /* Figure F.15: generate decoding tables for bit-sequential decoding */
+
+  p = 0;
+  for (l = 1; l <= 16; l++) {
+    if (htbl->bits[l]) {
+      /* valoffset[l] = huffval[] index of 1st symbol of code length l,
+       * minus the minimum code of length l
+       */
+      dtbl->valoffset[l] = (INT32) p - (INT32) huffcode[p];
+      p += htbl->bits[l];
+      dtbl->maxcode[l] = huffcode[p-1]; /* maximum code of length l */
+    } else {
+      dtbl->maxcode[l] = -1;	/* -1 if no codes of this length */
+    }
+  }
+  dtbl->maxcode[17] = 0xFFFFFL; /* ensures jpeg_huff_decode terminates */
+
+  /* Compute lookahead tables to speed up decoding.
+   * First we set all the table entries to 0, indicating "too long";
+   * then we iterate through the Huffman codes that are short enough and
+   * fill in all the entries that correspond to bit sequences starting
+   * with that code.
+   */
+
+  MEMZERO(dtbl->look_nbits, SIZEOF(dtbl->look_nbits));
+
+  p = 0;
+  for (l = 1; l <= HUFF_LOOKAHEAD; l++) {
+    for (i = 1; i <= (int) htbl->bits[l]; i++, p++) {
+      /* l = current code's length, p = its index in huffcode[] & huffval[]. */
+      /* Generate left-justified code followed by all possible bit sequences */
+      lookbits = huffcode[p] << (HUFF_LOOKAHEAD-l);
+      for (ctr = 1 << (HUFF_LOOKAHEAD-l); ctr > 0; ctr--) {
+	dtbl->look_nbits[lookbits] = l;
+	dtbl->look_sym[lookbits] = htbl->huffval[p];
+	lookbits++;
+      }
+    }
+  }
+
+  /* Validate symbols as being reasonable.
+   * For AC tables, we make no check, but accept all byte values 0..255.
+   * For DC tables, we require the symbols to be in range 0..15.
+   * (Tighter bounds could be applied depending on the data depth and mode,
+   * but this is sufficient to ensure safe decoding.)
+   */
+  if (isDC) {
+    for (i = 0; i < numsymbols; i++) {
+      int sym = htbl->huffval[i];
+      if (sym < 0 || sym > 15)
+	ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+    }
+  }
+}
+
+
+/*
+ * Out-of-line code for bit fetching.
+ * Note: current values of get_buffer and bits_left are passed as parameters,
+ * but are returned in the corresponding fields of the state struct.
+ *
+ * On most machines MIN_GET_BITS should be 25 to allow the full 32-bit width
+ * of get_buffer to be used.  (On machines with wider words, an even larger
+ * buffer could be used.)  However, on some machines 32-bit shifts are
+ * quite slow and take time proportional to the number of places shifted.
+ * (This is true with most PC compilers, for instance.)  In this case it may
+ * be a win to set MIN_GET_BITS to the minimum value of 15.  This reduces the
+ * average shift distance at the cost of more calls to jpeg_fill_bit_buffer.
+ */
+
+#ifdef SLOW_SHIFT_32
+#define MIN_GET_BITS  15	/* minimum allowable value */
+#else
+#define MIN_GET_BITS  (BIT_BUF_SIZE-7)
+#endif
+
+
+LOCAL(boolean)
+jpeg_fill_bit_buffer (bitread_working_state * state,
+		      register bit_buf_type get_buffer, register int bits_left,
+		      int nbits)
+/* Load up the bit buffer to a depth of at least nbits */
+{
+  /* Copy heavily used state fields into locals (hopefully registers) */
+  register const JOCTET * next_input_byte = state->next_input_byte;
+  register size_t bytes_in_buffer = state->bytes_in_buffer;
+  j_decompress_ptr cinfo = state->cinfo;
+
+  /* Attempt to load at least MIN_GET_BITS bits into get_buffer. */
+  /* (It is assumed that no request will be for more than that many bits.) */
+  /* We fail to do so only if we hit a marker or are forced to suspend. */
+
+  if (cinfo->unread_marker == 0) {	/* cannot advance past a marker */
+    while (bits_left < MIN_GET_BITS) {
+      register int c;
+
+      /* Attempt to read a byte */
+      if (bytes_in_buffer == 0) {
+	if (! (*cinfo->src->fill_input_buffer) (cinfo))
+	  return FALSE;
+	next_input_byte = cinfo->src->next_input_byte;
+	bytes_in_buffer = cinfo->src->bytes_in_buffer;
+      }
+      bytes_in_buffer--;
+      c = GETJOCTET(*next_input_byte++);
+
+      /* If it's 0xFF, check and discard stuffed zero byte */
+      if (c == 0xFF) {
+	/* Loop here to discard any padding FF's on terminating marker,
+	 * so that we can save a valid unread_marker value.  NOTE: we will
+	 * accept multiple FF's followed by a 0 as meaning a single FF data
+	 * byte.  This data pattern is not valid according to the standard.
+	 */
+	do {
+	  if (bytes_in_buffer == 0) {
+	    if (! (*cinfo->src->fill_input_buffer) (cinfo))
+	      return FALSE;
+	    next_input_byte = cinfo->src->next_input_byte;
+	    bytes_in_buffer = cinfo->src->bytes_in_buffer;
+	  }
+	  bytes_in_buffer--;
+	  c = GETJOCTET(*next_input_byte++);
+	} while (c == 0xFF);
+
+	if (c == 0) {
+	  /* Found FF/00, which represents an FF data byte */
+	  c = 0xFF;
+	} else {
+	  /* Oops, it's actually a marker indicating end of compressed data.
+	   * Save the marker code for later use.
+	   * Fine point: it might appear that we should save the marker into
+	   * bitread working state, not straight into permanent state.  But
+	   * once we have hit a marker, we cannot need to suspend within the
+	   * current MCU, because we will read no more bytes from the data
+	   * source.  So it is OK to update permanent state right away.
+	   */
+	  cinfo->unread_marker = c;
+	  /* See if we need to insert some fake zero bits. */
+	  goto no_more_bytes;
+	}
+      }
+
+      /* OK, load c into get_buffer */
+      get_buffer = (get_buffer << 8) | c;
+      bits_left += 8;
+    } /* end while */
+  } else {
+  no_more_bytes:
+    /* We get here if we've read the marker that terminates the compressed
+     * data segment.  There should be enough bits in the buffer register
+     * to satisfy the request; if so, no problem.
+     */
+    if (nbits > bits_left) {
+      /* Uh-oh.  Report corrupted data to user and stuff zeroes into
+       * the data stream, so that we can produce some kind of image.
+       * We use a nonvolatile flag to ensure that only one warning message
+       * appears per data segment.
+       */
+      if (! ((huff_entropy_ptr) cinfo->entropy)->insufficient_data) {
+	WARNMS(cinfo, JWRN_HIT_MARKER);
+	((huff_entropy_ptr) cinfo->entropy)->insufficient_data = TRUE;
+      }
+      /* Fill the buffer with zero bits */
+      get_buffer <<= MIN_GET_BITS - bits_left;
+      bits_left = MIN_GET_BITS;
+    }
+  }
+
+  /* Unload the local registers */
+  state->next_input_byte = next_input_byte;
+  state->bytes_in_buffer = bytes_in_buffer;
+  state->get_buffer = get_buffer;
+  state->bits_left = bits_left;
+
+  return TRUE;
+}
+
+
+/*
+ * Figure F.12: extend sign bit.
+ * On some machines, a shift and sub will be faster than a table lookup.
+ */
+
+#ifdef AVOID_TABLES
+
+#define BIT_MASK(nbits)   ((1<<(nbits))-1)
+#define HUFF_EXTEND(x,s)  ((x) < (1<<((s)-1)) ? (x) - ((1<<(s))-1) : (x))
+
+#else
+
+#define BIT_MASK(nbits)   bmask[nbits]
+#define HUFF_EXTEND(x,s)  ((x) <= bmask[(s) - 1] ? (x) - bmask[s] : (x))
+
+static const int bmask[16] =	/* bmask[n] is mask for n rightmost bits */
+  { 0, 0x0001, 0x0003, 0x0007, 0x000F, 0x001F, 0x003F, 0x007F, 0x00FF,
+    0x01FF, 0x03FF, 0x07FF, 0x0FFF, 0x1FFF, 0x3FFF, 0x7FFF };
+
+#endif /* AVOID_TABLES */
+
+
+/*
+ * Out-of-line code for Huffman code decoding.
+ */
+
+LOCAL(int)
+jpeg_huff_decode (bitread_working_state * state,
+		  register bit_buf_type get_buffer, register int bits_left,
+		  d_derived_tbl * htbl, int min_bits)
+{
+  register int l = min_bits;
+  register INT32 code;
+
+  /* HUFF_DECODE has determined that the code is at least min_bits */
+  /* bits long, so fetch that many bits in one swoop. */
+
+  CHECK_BIT_BUFFER(*state, l, return -1);
+  code = GET_BITS(l);
+
+  /* Collect the rest of the Huffman code one bit at a time. */
+  /* This is per Figure F.16 in the JPEG spec. */
+
+  while (code > htbl->maxcode[l]) {
+    code <<= 1;
+    CHECK_BIT_BUFFER(*state, 1, return -1);
+    code |= GET_BITS(1);
+    l++;
+  }
+
+  /* Unload the local registers */
+  state->get_buffer = get_buffer;
+  state->bits_left = bits_left;
+
+  /* With garbage input we may reach the sentinel value l = 17. */
+
+  if (l > 16) {
+    WARNMS(state->cinfo, JWRN_HUFF_BAD_CODE);
+    return 0;			/* fake a zero as the safest result */
+  }
+
+  return htbl->pub->huffval[ (int) (code + htbl->valoffset[l]) ];
+}
+
+
+/*
+ * Check for a restart marker & resynchronize decoder.
+ * Returns FALSE if must suspend.
+ */
+
+LOCAL(boolean)
+process_restart (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci;
+
+  /* Throw away any unused bits remaining in bit buffer; */
+  /* include any full bytes in next_marker's count of discarded bytes */
+  cinfo->marker->discarded_bytes += entropy->bitstate.bits_left / 8;
+  entropy->bitstate.bits_left = 0;
+
+  /* Advance past the RSTn marker */
+  if (! (*cinfo->marker->read_restart_marker) (cinfo))
+    return FALSE;
+
+  /* Re-initialize DC predictions to 0 */
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++)
+    entropy->saved.last_dc_val[ci] = 0;
+  /* Re-init EOB run count, too */
+  entropy->saved.EOBRUN = 0;
+
+  /* Reset restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+
+  /* Reset out-of-data flag, unless read_restart_marker left us smack up
+   * against a marker.  In that case we will end up treating the next data
+   * segment as empty, and we can avoid producing bogus output pixels by
+   * leaving the flag set.
+   */
+  if (cinfo->unread_marker == 0)
+    entropy->insufficient_data = FALSE;
+
+  return TRUE;
+}
+
+
+/*
+ * Huffman MCU decoding.
+ * Each of these routines decodes and returns one MCU's worth of
+ * Huffman-compressed coefficients. 
+ * The coefficients are reordered from zigzag order into natural array order,
+ * but are not dequantized.
+ *
+ * The i'th block of the MCU is stored into the block pointed to by
+ * MCU_data[i].  WE ASSUME THIS AREA IS INITIALLY ZEROED BY THE CALLER.
+ * (Wholesale zeroing is usually a little faster than retail...)
+ *
+ * We return FALSE if data source requested suspension.  In that case no
+ * changes have been made to permanent state.  (Exception: some output
+ * coefficients may already have been assigned.  This is harmless for
+ * spectral selection, since we'll just re-assign them on the next call.
+ * Successive approximation AC refinement has to be more careful, however.)
+ */
+
+/*
+ * MCU decoding for DC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int Al = cinfo->Al;
+  register int s, r;
+  int blkn, ci;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+  savable_state state;
+  d_derived_tbl * tbl;
+  jpeg_component_info * compptr;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(state, entropy->saved);
+
+    /* Outer loop handles each block in the MCU */
+
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      block = MCU_data[blkn];
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      tbl = entropy->derived_tbls[compptr->dc_tbl_no];
+
+      /* Decode a single block's worth of coefficients */
+
+      /* Section F.2.2.1: decode the DC coefficient difference */
+      HUFF_DECODE(s, br_state, tbl, return FALSE, label1);
+      if (s) {
+	CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	r = GET_BITS(s);
+	s = HUFF_EXTEND(r, s);
+      }
+
+      /* Convert DC difference to actual value, update last_dc_val */
+      s += state.last_dc_val[ci];
+      state.last_dc_val[ci] = s;
+      /* Scale and output the coefficient (assumes jpeg_natural_order[0]=0) */
+      (*block)[0] = (JCOEF) (s << Al);
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(entropy->saved, state);
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC initial scan (either spectral selection,
+ * or first pass of successive approximation).
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_first (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  register int s, k, r;
+  unsigned int EOBRUN;
+  int Se, Al;
+  const int * natural_order;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+  d_derived_tbl * tbl;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->insufficient_data) {
+
+    Se = cinfo->Se;
+    Al = cinfo->Al;
+    natural_order = cinfo->natural_order;
+
+    /* Load up working state.
+     * We can avoid loading/saving bitread state if in an EOB run.
+     */
+    EOBRUN = entropy->saved.EOBRUN;	/* only part of saved state we need */
+
+    /* There is always only one block per MCU */
+
+    if (EOBRUN > 0)		/* if it's a band of zeroes... */
+      EOBRUN--;			/* ...process it now (we do nothing) */
+    else {
+      BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+      block = MCU_data[0];
+      tbl = entropy->ac_derived_tbl;
+
+      for (k = cinfo->Ss; k <= Se; k++) {
+	HUFF_DECODE(s, br_state, tbl, return FALSE, label2);
+	r = s >> 4;
+	s &= 15;
+	if (s) {
+	  k += r;
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  r = GET_BITS(s);
+	  s = HUFF_EXTEND(r, s);
+	  /* Scale and output coefficient in natural (dezigzagged) order */
+	  (*block)[natural_order[k]] = (JCOEF) (s << Al);
+	} else {
+	  if (r == 15) {	/* ZRL */
+	    k += 15;		/* skip 15 zeroes in band */
+	  } else {		/* EOBr, run length is 2^r + appended bits */
+	    EOBRUN = 1 << r;
+	    if (r) {		/* EOBr, r > 0 */
+	      CHECK_BIT_BUFFER(br_state, r, return FALSE);
+	      r = GET_BITS(r);
+	      EOBRUN += r;
+	    }
+	    EOBRUN--;		/* this band is processed at this moment */
+	    break;		/* force end-of-band */
+	  }
+	}
+      }
+
+      BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    }
+
+    /* Completed MCU, so update state */
+    entropy->saved.EOBRUN = EOBRUN;	/* only part of saved state we need */
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for DC successive approximation refinement scan.
+ * Note: we assume such scans can be multi-component, although the spec
+ * is not very clear on the point.
+ */
+
+METHODDEF(boolean)
+decode_mcu_DC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int p1 = 1 << cinfo->Al;	/* 1 in the bit position being coded */
+  int blkn;
+  JBLOCKROW block;
+  BITREAD_STATE_VARS;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* Not worth the cycles to check insufficient_data here,
+   * since we will not change the data anyway if we read zeroes.
+   */
+
+  /* Load up working state */
+  BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+
+  /* Outer loop handles each block in the MCU */
+
+  for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+    block = MCU_data[blkn];
+
+    /* Encoded data is simply the next bit of the two's-complement DC value */
+    CHECK_BIT_BUFFER(br_state, 1, return FALSE);
+    if (GET_BITS(1))
+      (*block)[0] |= p1;
+    /* Note: since we use |=, repeating the assignment later is safe */
+  }
+
+  /* Completed MCU, so update state */
+  BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * MCU decoding for AC successive approximation refinement scan.
+ */
+
+METHODDEF(boolean)
+decode_mcu_AC_refine (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{   
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  register int s, k, r;
+  unsigned int EOBRUN;
+  int Se, p1, m1;
+  const int * natural_order;
+  JBLOCKROW block;
+  JCOEFPTR thiscoef;
+  BITREAD_STATE_VARS;
+  d_derived_tbl * tbl;
+  int num_newnz;
+  int newnz_pos[DCTSIZE2];
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, don't modify the MCU.
+   */
+  if (! entropy->insufficient_data) {
+
+    Se = cinfo->Se;
+    p1 = 1 << cinfo->Al;	/* 1 in the bit position being coded */
+    m1 = (-1) << cinfo->Al;	/* -1 in the bit position being coded */
+    natural_order = cinfo->natural_order;
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    EOBRUN = entropy->saved.EOBRUN; /* only part of saved state we need */
+
+    /* There is always only one block per MCU */
+    block = MCU_data[0];
+    tbl = entropy->ac_derived_tbl;
+
+    /* If we are forced to suspend, we must undo the assignments to any newly
+     * nonzero coefficients in the block, because otherwise we'd get confused
+     * next time about which coefficients were already nonzero.
+     * But we need not undo addition of bits to already-nonzero coefficients;
+     * instead, we can test the current bit to see if we already did it.
+     */
+    num_newnz = 0;
+
+    /* initialize coefficient loop counter to start of band */
+    k = cinfo->Ss;
+
+    if (EOBRUN == 0) {
+      for (; k <= Se; k++) {
+	HUFF_DECODE(s, br_state, tbl, goto undoit, label3);
+	r = s >> 4;
+	s &= 15;
+	if (s) {
+	  if (s != 1)		/* size of new coef should always be 1 */
+	    WARNMS(cinfo, JWRN_HUFF_BAD_CODE);
+	  CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	  if (GET_BITS(1))
+	    s = p1;		/* newly nonzero coef is positive */
+	  else
+	    s = m1;		/* newly nonzero coef is negative */
+	} else {
+	  if (r != 15) {
+	    EOBRUN = 1 << r;	/* EOBr, run length is 2^r + appended bits */
+	    if (r) {
+	      CHECK_BIT_BUFFER(br_state, r, goto undoit);
+	      r = GET_BITS(r);
+	      EOBRUN += r;
+	    }
+	    break;		/* rest of block is handled by EOB logic */
+	  }
+	  /* note s = 0 for processing ZRL */
+	}
+	/* Advance over already-nonzero coefs and r still-zero coefs,
+	 * appending correction bits to the nonzeroes.  A correction bit is 1
+	 * if the absolute value of the coefficient must be increased.
+	 */
+	do {
+	  thiscoef = *block + natural_order[k];
+	  if (*thiscoef != 0) {
+	    CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	    if (GET_BITS(1)) {
+	      if ((*thiscoef & p1) == 0) { /* do nothing if already set it */
+		if (*thiscoef >= 0)
+		  *thiscoef += p1;
+		else
+		  *thiscoef += m1;
+	      }
+	    }
+	  } else {
+	    if (--r < 0)
+	      break;		/* reached target zero coefficient */
+	  }
+	  k++;
+	} while (k <= Se);
+	if (s) {
+	  int pos = natural_order[k];
+	  /* Output newly nonzero coefficient */
+	  (*block)[pos] = (JCOEF) s;
+	  /* Remember its position in case we have to suspend */
+	  newnz_pos[num_newnz++] = pos;
+	}
+      }
+    }
+
+    if (EOBRUN > 0) {
+      /* Scan any remaining coefficient positions after the end-of-band
+       * (the last newly nonzero coefficient, if any).  Append a correction
+       * bit to each already-nonzero coefficient.  A correction bit is 1
+       * if the absolute value of the coefficient must be increased.
+       */
+      for (; k <= Se; k++) {
+	thiscoef = *block + natural_order[k];
+	if (*thiscoef != 0) {
+	  CHECK_BIT_BUFFER(br_state, 1, goto undoit);
+	  if (GET_BITS(1)) {
+	    if ((*thiscoef & p1) == 0) { /* do nothing if already changed it */
+	      if (*thiscoef >= 0)
+		*thiscoef += p1;
+	      else
+		*thiscoef += m1;
+	    }
+	  }
+	}
+      }
+      /* Count one block completed in EOB run */
+      EOBRUN--;
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    entropy->saved.EOBRUN = EOBRUN; /* only part of saved state we need */
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+
+undoit:
+  /* Re-zero any output coefficients that we made newly nonzero */
+  while (num_newnz > 0)
+    (*block)[newnz_pos[--num_newnz]] = 0;
+
+  return FALSE;
+}
+
+
+/*
+ * Decode one MCU's worth of Huffman-compressed coefficients,
+ * partial blocks.
+ */
+
+METHODDEF(boolean)
+decode_mcu_sub (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  const int * natural_order;
+  int Se, blkn;
+  BITREAD_STATE_VARS;
+  savable_state state;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->insufficient_data) {
+
+    natural_order = cinfo->natural_order;
+    Se = cinfo->lim_Se;
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(state, entropy->saved);
+
+    /* Outer loop handles each block in the MCU */
+
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      JBLOCKROW block = MCU_data[blkn];
+      d_derived_tbl * htbl;
+      register int s, k, r;
+      int coef_limit, ci;
+
+      /* Decode a single block's worth of coefficients */
+
+      /* Section F.2.2.1: decode the DC coefficient difference */
+      htbl = entropy->dc_cur_tbls[blkn];
+      HUFF_DECODE(s, br_state, htbl, return FALSE, label1);
+
+      htbl = entropy->ac_cur_tbls[blkn];
+      k = 1;
+      coef_limit = entropy->coef_limit[blkn];
+      if (coef_limit) {
+	/* Convert DC difference to actual value, update last_dc_val */
+	if (s) {
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  r = GET_BITS(s);
+	  s = HUFF_EXTEND(r, s);
+	}
+	ci = cinfo->MCU_membership[blkn];
+	s += state.last_dc_val[ci];
+	state.last_dc_val[ci] = s;
+	/* Output the DC coefficient */
+	(*block)[0] = (JCOEF) s;
+
+	/* Section F.2.2.2: decode the AC coefficients */
+	/* Since zeroes are skipped, output area must be cleared beforehand */
+	for (; k < coef_limit; k++) {
+	  HUFF_DECODE(s, br_state, htbl, return FALSE, label2);
+
+	  r = s >> 4;
+	  s &= 15;
+
+	  if (s) {
+	    k += r;
+	    CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	    r = GET_BITS(s);
+	    s = HUFF_EXTEND(r, s);
+	    /* Output coefficient in natural (dezigzagged) order.
+	     * Note: the extra entries in natural_order[] will save us
+	     * if k > Se, which could happen if the data is corrupted.
+	     */
+	    (*block)[natural_order[k]] = (JCOEF) s;
+	  } else {
+	    if (r != 15)
+	      goto EndOfBlock;
+	    k += 15;
+	  }
+	}
+      } else {
+	if (s) {
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  DROP_BITS(s);
+	}
+      }
+
+      /* Section F.2.2.2: decode the AC coefficients */
+      /* In this path we just discard the values */
+      for (; k <= Se; k++) {
+	HUFF_DECODE(s, br_state, htbl, return FALSE, label3);
+
+	r = s >> 4;
+	s &= 15;
+
+	if (s) {
+	  k += r;
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  DROP_BITS(s);
+	} else {
+	  if (r != 15)
+	    break;
+	  k += 15;
+	}
+      }
+
+      EndOfBlock: ;
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(entropy->saved, state);
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * Decode one MCU's worth of Huffman-compressed coefficients,
+ * full-size blocks.
+ */
+
+METHODDEF(boolean)
+decode_mcu (j_decompress_ptr cinfo, JBLOCKROW *MCU_data)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int blkn;
+  BITREAD_STATE_VARS;
+  savable_state state;
+
+  /* Process restart marker if needed; may have to suspend */
+  if (cinfo->restart_interval) {
+    if (entropy->restarts_to_go == 0)
+      if (! process_restart(cinfo))
+	return FALSE;
+  }
+
+  /* If we've run out of data, just leave the MCU set to zeroes.
+   * This way, we return uniform gray for the remainder of the segment.
+   */
+  if (! entropy->insufficient_data) {
+
+    /* Load up working state */
+    BITREAD_LOAD_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(state, entropy->saved);
+
+    /* Outer loop handles each block in the MCU */
+
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      JBLOCKROW block = MCU_data[blkn];
+      d_derived_tbl * htbl;
+      register int s, k, r;
+      int coef_limit, ci;
+
+      /* Decode a single block's worth of coefficients */
+
+      /* Section F.2.2.1: decode the DC coefficient difference */
+      htbl = entropy->dc_cur_tbls[blkn];
+      HUFF_DECODE(s, br_state, htbl, return FALSE, label1);
+
+      htbl = entropy->ac_cur_tbls[blkn];
+      k = 1;
+      coef_limit = entropy->coef_limit[blkn];
+      if (coef_limit) {
+	/* Convert DC difference to actual value, update last_dc_val */
+	if (s) {
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  r = GET_BITS(s);
+	  s = HUFF_EXTEND(r, s);
+	}
+	ci = cinfo->MCU_membership[blkn];
+	s += state.last_dc_val[ci];
+	state.last_dc_val[ci] = s;
+	/* Output the DC coefficient */
+	(*block)[0] = (JCOEF) s;
+
+	/* Section F.2.2.2: decode the AC coefficients */
+	/* Since zeroes are skipped, output area must be cleared beforehand */
+	for (; k < coef_limit; k++) {
+	  HUFF_DECODE(s, br_state, htbl, return FALSE, label2);
+
+	  r = s >> 4;
+	  s &= 15;
+
+	  if (s) {
+	    k += r;
+	    CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	    r = GET_BITS(s);
+	    s = HUFF_EXTEND(r, s);
+	    /* Output coefficient in natural (dezigzagged) order.
+	     * Note: the extra entries in jpeg_natural_order[] will save us
+	     * if k >= DCTSIZE2, which could happen if the data is corrupted.
+	     */
+	    (*block)[jpeg_natural_order[k]] = (JCOEF) s;
+	  } else {
+	    if (r != 15)
+	      goto EndOfBlock;
+	    k += 15;
+	  }
+	}
+      } else {
+	if (s) {
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  DROP_BITS(s);
+	}
+      }
+
+      /* Section F.2.2.2: decode the AC coefficients */
+      /* In this path we just discard the values */
+      for (; k < DCTSIZE2; k++) {
+	HUFF_DECODE(s, br_state, htbl, return FALSE, label3);
+
+	r = s >> 4;
+	s &= 15;
+
+	if (s) {
+	  k += r;
+	  CHECK_BIT_BUFFER(br_state, s, return FALSE);
+	  DROP_BITS(s);
+	} else {
+	  if (r != 15)
+	    break;
+	  k += 15;
+	}
+      }
+
+      EndOfBlock: ;
+    }
+
+    /* Completed MCU, so update state */
+    BITREAD_SAVE_STATE(cinfo,entropy->bitstate);
+    ASSIGN_STATE(entropy->saved, state);
+  }
+
+  /* Account for restart interval (no-op if not using restarts) */
+  entropy->restarts_to_go--;
+
+  return TRUE;
+}
+
+
+/*
+ * Initialize for a Huffman-compressed scan.
+ */
+
+METHODDEF(void)
+start_pass_huff_decoder (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy = (huff_entropy_ptr) cinfo->entropy;
+  int ci, blkn, tbl, i;
+  jpeg_component_info * compptr;
+
+  if (cinfo->progressive_mode) {
+    /* Validate progressive scan parameters */
+    if (cinfo->Ss == 0) {
+      if (cinfo->Se != 0)
+	goto bad;
+    } else {
+      /* need not check Ss/Se < 0 since they came from unsigned bytes */
+      if (cinfo->Se < cinfo->Ss || cinfo->Se > cinfo->lim_Se)
+	goto bad;
+      /* AC scans may have only one component */
+      if (cinfo->comps_in_scan != 1)
+	goto bad;
+    }
+    if (cinfo->Ah != 0) {
+      /* Successive approximation refinement scan: must have Al = Ah-1. */
+      if (cinfo->Ah-1 != cinfo->Al)
+	goto bad;
+    }
+    if (cinfo->Al > 13) {	/* need not check for < 0 */
+      /* Arguably the maximum Al value should be less than 13 for 8-bit precision,
+       * but the spec doesn't say so, and we try to be liberal about what we
+       * accept.  Note: large Al values could result in out-of-range DC
+       * coefficients during early scans, leading to bizarre displays due to
+       * overflows in the IDCT math.  But we won't crash.
+       */
+      bad:
+      ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+	       cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+    }
+    /* Update progression status, and verify that scan order is legal.
+     * Note that inter-scan inconsistencies are treated as warnings
+     * not fatal errors ... not clear if this is right way to behave.
+     */
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      int coefi, cindex = cinfo->cur_comp_info[ci]->component_index;
+      int *coef_bit_ptr = & cinfo->coef_bits[cindex][0];
+      if (cinfo->Ss && coef_bit_ptr[0] < 0) /* AC without prior DC scan */
+	WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, 0);
+      for (coefi = cinfo->Ss; coefi <= cinfo->Se; coefi++) {
+	int expected = (coef_bit_ptr[coefi] < 0) ? 0 : coef_bit_ptr[coefi];
+	if (cinfo->Ah != expected)
+	  WARNMS2(cinfo, JWRN_BOGUS_PROGRESSION, cindex, coefi);
+	coef_bit_ptr[coefi] = cinfo->Al;
+      }
+    }
+
+    /* Select MCU decoding routine */
+    if (cinfo->Ah == 0) {
+      if (cinfo->Ss == 0)
+	entropy->pub.decode_mcu = decode_mcu_DC_first;
+      else
+	entropy->pub.decode_mcu = decode_mcu_AC_first;
+    } else {
+      if (cinfo->Ss == 0)
+	entropy->pub.decode_mcu = decode_mcu_DC_refine;
+      else
+	entropy->pub.decode_mcu = decode_mcu_AC_refine;
+    }
+
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Make sure requested tables are present, and compute derived tables.
+       * We may build same derived table more than once, but it's not expensive.
+       */
+      if (cinfo->Ss == 0) {
+	if (cinfo->Ah == 0) {	/* DC refinement needs no table */
+	  tbl = compptr->dc_tbl_no;
+	  jpeg_make_d_derived_tbl(cinfo, TRUE, tbl,
+				  & entropy->derived_tbls[tbl]);
+	}
+      } else {
+	tbl = compptr->ac_tbl_no;
+	jpeg_make_d_derived_tbl(cinfo, FALSE, tbl,
+				& entropy->derived_tbls[tbl]);
+	/* remember the single active table */
+	entropy->ac_derived_tbl = entropy->derived_tbls[tbl];
+      }
+      /* Initialize DC predictions to 0 */
+      entropy->saved.last_dc_val[ci] = 0;
+    }
+
+    /* Initialize private state variables */
+    entropy->saved.EOBRUN = 0;
+  } else {
+    /* Check that the scan parameters Ss, Se, Ah/Al are OK for sequential JPEG.
+     * This ought to be an error condition, but we make it a warning because
+     * there are some baseline files out there with all zeroes in these bytes.
+     */
+    if (cinfo->Ss != 0 || cinfo->Ah != 0 || cinfo->Al != 0 ||
+	((cinfo->is_baseline || cinfo->Se < DCTSIZE2) &&
+	cinfo->Se != cinfo->lim_Se))
+      WARNMS(cinfo, JWRN_NOT_SEQUENTIAL);
+
+    /* Select MCU decoding routine */
+    /* We retain the hard-coded case for full-size blocks.
+     * This is not necessary, but it appears that this version is slightly
+     * more performant in the given implementation.
+     * With an improved implementation we would prefer a single optimized
+     * function.
+     */
+    if (cinfo->lim_Se != DCTSIZE2-1)
+      entropy->pub.decode_mcu = decode_mcu_sub;
+    else
+      entropy->pub.decode_mcu = decode_mcu;
+
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Compute derived values for Huffman tables */
+      /* We may do this more than once for a table, but it's not expensive */
+      tbl = compptr->dc_tbl_no;
+      jpeg_make_d_derived_tbl(cinfo, TRUE, tbl,
+			      & entropy->dc_derived_tbls[tbl]);
+      if (cinfo->lim_Se) {	/* AC needs no table when not present */
+	tbl = compptr->ac_tbl_no;
+	jpeg_make_d_derived_tbl(cinfo, FALSE, tbl,
+				& entropy->ac_derived_tbls[tbl]);
+      }
+      /* Initialize DC predictions to 0 */
+      entropy->saved.last_dc_val[ci] = 0;
+    }
+
+    /* Precalculate decoding info for each block in an MCU of this scan */
+    for (blkn = 0; blkn < cinfo->blocks_in_MCU; blkn++) {
+      ci = cinfo->MCU_membership[blkn];
+      compptr = cinfo->cur_comp_info[ci];
+      /* Precalculate which table to use for each block */
+      entropy->dc_cur_tbls[blkn] = entropy->dc_derived_tbls[compptr->dc_tbl_no];
+      entropy->ac_cur_tbls[blkn] = entropy->ac_derived_tbls[compptr->ac_tbl_no];
+      /* Decide whether we really care about the coefficient values */
+      if (compptr->component_needed) {
+	ci = compptr->DCT_v_scaled_size;
+	i = compptr->DCT_h_scaled_size;
+	switch (cinfo->lim_Se) {
+	case (1*1-1):
+	  entropy->coef_limit[blkn] = 1;
+	  break;
+	case (2*2-1):
+	  if (ci <= 0 || ci > 2) ci = 2;
+	  if (i <= 0 || i > 2) i = 2;
+	  entropy->coef_limit[blkn] = 1 + jpeg_zigzag_order2[ci - 1][i - 1];
+	  break;
+	case (3*3-1):
+	  if (ci <= 0 || ci > 3) ci = 3;
+	  if (i <= 0 || i > 3) i = 3;
+	  entropy->coef_limit[blkn] = 1 + jpeg_zigzag_order3[ci - 1][i - 1];
+	  break;
+	case (4*4-1):
+	  if (ci <= 0 || ci > 4) ci = 4;
+	  if (i <= 0 || i > 4) i = 4;
+	  entropy->coef_limit[blkn] = 1 + jpeg_zigzag_order4[ci - 1][i - 1];
+	  break;
+	case (5*5-1):
+	  if (ci <= 0 || ci > 5) ci = 5;
+	  if (i <= 0 || i > 5) i = 5;
+	  entropy->coef_limit[blkn] = 1 + jpeg_zigzag_order5[ci - 1][i - 1];
+	  break;
+	case (6*6-1):
+	  if (ci <= 0 || ci > 6) ci = 6;
+	  if (i <= 0 || i > 6) i = 6;
+	  entropy->coef_limit[blkn] = 1 + jpeg_zigzag_order6[ci - 1][i - 1];
+	  break;
+	case (7*7-1):
+	  if (ci <= 0 || ci > 7) ci = 7;
+	  if (i <= 0 || i > 7) i = 7;
+	  entropy->coef_limit[blkn] = 1 + jpeg_zigzag_order7[ci - 1][i - 1];
+	  break;
+	default:
+	  if (ci <= 0 || ci > 8) ci = 8;
+	  if (i <= 0 || i > 8) i = 8;
+	  entropy->coef_limit[blkn] = 1 + jpeg_zigzag_order[ci - 1][i - 1];
+	  break;
+	}
+      } else {
+	entropy->coef_limit[blkn] = 0;
+      }
+    }
+  }
+
+  /* Initialize bitread state variables */
+  entropy->bitstate.bits_left = 0;
+  entropy->bitstate.get_buffer = 0; /* unnecessary, but keeps Purify quiet */
+  entropy->insufficient_data = FALSE;
+
+  /* Initialize restart counter */
+  entropy->restarts_to_go = cinfo->restart_interval;
+}
+
+
+/*
+ * Module initialization routine for Huffman entropy decoding.
+ */
+
+GLOBAL(void)
+jinit_huff_decoder (j_decompress_ptr cinfo)
+{
+  huff_entropy_ptr entropy;
+  int i;
+
+  entropy = (huff_entropy_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(huff_entropy_decoder));
+  cinfo->entropy = (struct jpeg_entropy_decoder *) entropy;
+  entropy->pub.start_pass = start_pass_huff_decoder;
+
+  if (cinfo->progressive_mode) {
+    /* Create progression status table */
+    int *coef_bit_ptr, ci;
+    cinfo->coef_bits = (int (*)[DCTSIZE2])
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  cinfo->num_components*DCTSIZE2*SIZEOF(int));
+    coef_bit_ptr = & cinfo->coef_bits[0][0];
+    for (ci = 0; ci < cinfo->num_components; ci++)
+      for (i = 0; i < DCTSIZE2; i++)
+	*coef_bit_ptr++ = -1;
+
+    /* Mark derived tables unallocated */
+    for (i = 0; i < NUM_HUFF_TBLS; i++) {
+      entropy->derived_tbls[i] = NULL;
+    }
+  } else {
+    /* Mark tables unallocated */
+    for (i = 0; i < NUM_HUFF_TBLS; i++) {
+      entropy->dc_derived_tbls[i] = entropy->ac_derived_tbls[i] = NULL;
+    }
+  }
+}

Added: trunk/code/jpeg-8c/jdinput.c
===================================================================
--- trunk/code/jpeg-8c/jdinput.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdinput.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,661 @@
+/*
+ * jdinput.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2002-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains input control logic for the JPEG decompressor.
+ * These routines are concerned with controlling the decompressor's input
+ * processing (marker reading and coefficient decoding).  The actual input
+ * reading is done in jdmarker.c, jdhuff.c, and jdarith.c.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_input_controller pub; /* public fields */
+
+  int inheaders;		/* Nonzero until first SOS is reached */
+} my_input_controller;
+
+typedef my_input_controller * my_inputctl_ptr;
+
+
+/* Forward declarations */
+METHODDEF(int) consume_markers JPP((j_decompress_ptr cinfo));
+
+
+/*
+ * Routines to calculate various quantities related to the size of the image.
+ */
+
+
+/*
+ * Compute output image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ */
+
+GLOBAL(void)
+jpeg_core_output_dimensions (j_decompress_ptr cinfo)
+/* Do computations that are needed before master selection phase.
+ * This function is used for transcoding and full decompression.
+ */
+{
+#ifdef IDCT_SCALING_SUPPORTED
+  int ci;
+  jpeg_component_info *compptr;
+
+  /* Compute actual output image dimensions and DCT scaling choices. */
+  if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom) {
+    /* Provide 1/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 1;
+    cinfo->min_DCT_v_scaled_size = 1;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 2) {
+    /* Provide 2/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 2L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 2L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 2;
+    cinfo->min_DCT_v_scaled_size = 2;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 3) {
+    /* Provide 3/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 3L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 3L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 3;
+    cinfo->min_DCT_v_scaled_size = 3;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 4) {
+    /* Provide 4/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 4L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 4L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 4;
+    cinfo->min_DCT_v_scaled_size = 4;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 5) {
+    /* Provide 5/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 5L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 5L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 5;
+    cinfo->min_DCT_v_scaled_size = 5;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 6) {
+    /* Provide 6/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 6L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 6L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 6;
+    cinfo->min_DCT_v_scaled_size = 6;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 7) {
+    /* Provide 7/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 7L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 7L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 7;
+    cinfo->min_DCT_v_scaled_size = 7;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 8) {
+    /* Provide 8/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 8L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 8L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 8;
+    cinfo->min_DCT_v_scaled_size = 8;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 9) {
+    /* Provide 9/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 9L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 9L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 9;
+    cinfo->min_DCT_v_scaled_size = 9;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 10) {
+    /* Provide 10/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 10L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 10L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 10;
+    cinfo->min_DCT_v_scaled_size = 10;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 11) {
+    /* Provide 11/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 11L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 11L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 11;
+    cinfo->min_DCT_v_scaled_size = 11;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 12) {
+    /* Provide 12/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 12L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 12L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 12;
+    cinfo->min_DCT_v_scaled_size = 12;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 13) {
+    /* Provide 13/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 13L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 13L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 13;
+    cinfo->min_DCT_v_scaled_size = 13;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 14) {
+    /* Provide 14/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 14L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 14L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 14;
+    cinfo->min_DCT_v_scaled_size = 14;
+  } else if (cinfo->scale_num * cinfo->block_size <= cinfo->scale_denom * 15) {
+    /* Provide 15/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 15L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 15L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 15;
+    cinfo->min_DCT_v_scaled_size = 15;
+  } else {
+    /* Provide 16/block_size scaling */
+    cinfo->output_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * 16L, (long) cinfo->block_size);
+    cinfo->output_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * 16L, (long) cinfo->block_size);
+    cinfo->min_DCT_h_scaled_size = 16;
+    cinfo->min_DCT_v_scaled_size = 16;
+  }
+
+  /* Recompute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    compptr->DCT_h_scaled_size = cinfo->min_DCT_h_scaled_size;
+    compptr->DCT_v_scaled_size = cinfo->min_DCT_v_scaled_size;
+  }
+
+#else /* !IDCT_SCALING_SUPPORTED */
+
+  /* Hardwire it to "no scaling" */
+  cinfo->output_width = cinfo->image_width;
+  cinfo->output_height = cinfo->image_height;
+  /* jdinput.c has already initialized DCT_scaled_size,
+   * and has computed unscaled downsampled_width and downsampled_height.
+   */
+
+#endif /* IDCT_SCALING_SUPPORTED */
+}
+
+
+LOCAL(void)
+initial_setup (j_decompress_ptr cinfo)
+/* Called once, when first SOS marker is reached */
+{
+  int ci;
+  jpeg_component_info *compptr;
+
+  /* Make sure image isn't bigger than I can handle */
+  if ((long) cinfo->image_height > (long) JPEG_MAX_DIMENSION ||
+      (long) cinfo->image_width > (long) JPEG_MAX_DIMENSION)
+    ERREXIT1(cinfo, JERR_IMAGE_TOO_BIG, (unsigned int) JPEG_MAX_DIMENSION);
+
+  /* For now, precision must match compiled-in value... */
+  if (cinfo->data_precision != BITS_IN_JSAMPLE)
+    ERREXIT1(cinfo, JERR_BAD_PRECISION, cinfo->data_precision);
+
+  /* Check that number of components won't exceed internal array sizes */
+  if (cinfo->num_components > MAX_COMPONENTS)
+    ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->num_components,
+	     MAX_COMPONENTS);
+
+  /* Compute maximum sampling factors; check factor validity */
+  cinfo->max_h_samp_factor = 1;
+  cinfo->max_v_samp_factor = 1;
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    if (compptr->h_samp_factor<=0 || compptr->h_samp_factor>MAX_SAMP_FACTOR ||
+	compptr->v_samp_factor<=0 || compptr->v_samp_factor>MAX_SAMP_FACTOR)
+      ERREXIT(cinfo, JERR_BAD_SAMPLING);
+    cinfo->max_h_samp_factor = MAX(cinfo->max_h_samp_factor,
+				   compptr->h_samp_factor);
+    cinfo->max_v_samp_factor = MAX(cinfo->max_v_samp_factor,
+				   compptr->v_samp_factor);
+  }
+
+  /* Derive block_size, natural_order, and lim_Se */
+  if (cinfo->is_baseline || (cinfo->progressive_mode &&
+      cinfo->comps_in_scan)) { /* no pseudo SOS marker */
+    cinfo->block_size = DCTSIZE;
+    cinfo->natural_order = jpeg_natural_order;
+    cinfo->lim_Se = DCTSIZE2-1;
+  } else
+    switch (cinfo->Se) {
+    case (1*1-1):
+      cinfo->block_size = 1;
+      cinfo->natural_order = jpeg_natural_order; /* not needed */
+      cinfo->lim_Se = cinfo->Se;
+      break;
+    case (2*2-1):
+      cinfo->block_size = 2;
+      cinfo->natural_order = jpeg_natural_order2;
+      cinfo->lim_Se = cinfo->Se;
+      break;
+    case (3*3-1):
+      cinfo->block_size = 3;
+      cinfo->natural_order = jpeg_natural_order3;
+      cinfo->lim_Se = cinfo->Se;
+      break;
+    case (4*4-1):
+      cinfo->block_size = 4;
+      cinfo->natural_order = jpeg_natural_order4;
+      cinfo->lim_Se = cinfo->Se;
+      break;
+    case (5*5-1):
+      cinfo->block_size = 5;
+      cinfo->natural_order = jpeg_natural_order5;
+      cinfo->lim_Se = cinfo->Se;
+      break;
+    case (6*6-1):
+      cinfo->block_size = 6;
+      cinfo->natural_order = jpeg_natural_order6;
+      cinfo->lim_Se = cinfo->Se;
+      break;
+    case (7*7-1):
+      cinfo->block_size = 7;
+      cinfo->natural_order = jpeg_natural_order7;
+      cinfo->lim_Se = cinfo->Se;
+      break;
+    case (8*8-1):
+      cinfo->block_size = 8;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    case (9*9-1):
+      cinfo->block_size = 9;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    case (10*10-1):
+      cinfo->block_size = 10;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    case (11*11-1):
+      cinfo->block_size = 11;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    case (12*12-1):
+      cinfo->block_size = 12;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    case (13*13-1):
+      cinfo->block_size = 13;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    case (14*14-1):
+      cinfo->block_size = 14;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    case (15*15-1):
+      cinfo->block_size = 15;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    case (16*16-1):
+      cinfo->block_size = 16;
+      cinfo->natural_order = jpeg_natural_order;
+      cinfo->lim_Se = DCTSIZE2-1;
+      break;
+    default:
+      ERREXIT4(cinfo, JERR_BAD_PROGRESSION,
+	       cinfo->Ss, cinfo->Se, cinfo->Ah, cinfo->Al);
+      break;
+    }
+
+  /* We initialize DCT_scaled_size and min_DCT_scaled_size to block_size.
+   * In the full decompressor,
+   * this will be overridden by jpeg_calc_output_dimensions in jdmaster.c;
+   * but in the transcoder,
+   * jpeg_calc_output_dimensions is not used, so we must do it here.
+   */
+  cinfo->min_DCT_h_scaled_size = cinfo->block_size;
+  cinfo->min_DCT_v_scaled_size = cinfo->block_size;
+
+  /* Compute dimensions of components */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    compptr->DCT_h_scaled_size = cinfo->block_size;
+    compptr->DCT_v_scaled_size = cinfo->block_size;
+    /* Size in DCT blocks */
+    compptr->width_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * (long) compptr->h_samp_factor,
+		    (long) (cinfo->max_h_samp_factor * cinfo->block_size));
+    compptr->height_in_blocks = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * (long) compptr->v_samp_factor,
+		    (long) (cinfo->max_v_samp_factor * cinfo->block_size));
+    /* downsampled_width and downsampled_height will also be overridden by
+     * jdmaster.c if we are doing full decompression.  The transcoder library
+     * doesn't use these values, but the calling application might.
+     */
+    /* Size in samples */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width * (long) compptr->h_samp_factor,
+		    (long) cinfo->max_h_samp_factor);
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height * (long) compptr->v_samp_factor,
+		    (long) cinfo->max_v_samp_factor);
+    /* Mark component needed, until color conversion says otherwise */
+    compptr->component_needed = TRUE;
+    /* Mark no quantization table yet saved for component */
+    compptr->quant_table = NULL;
+  }
+
+  /* Compute number of fully interleaved MCU rows. */
+  cinfo->total_iMCU_rows = (JDIMENSION)
+    jdiv_round_up((long) cinfo->image_height,
+	          (long) (cinfo->max_v_samp_factor * cinfo->block_size));
+
+  /* Decide whether file contains multiple scans */
+  if (cinfo->comps_in_scan < cinfo->num_components || cinfo->progressive_mode)
+    cinfo->inputctl->has_multiple_scans = TRUE;
+  else
+    cinfo->inputctl->has_multiple_scans = FALSE;
+}
+
+
+LOCAL(void)
+per_scan_setup (j_decompress_ptr cinfo)
+/* Do computations that are needed before processing a JPEG scan */
+/* cinfo->comps_in_scan and cinfo->cur_comp_info[] were set from SOS marker */
+{
+  int ci, mcublks, tmp;
+  jpeg_component_info *compptr;
+  
+  if (cinfo->comps_in_scan == 1) {
+    
+    /* Noninterleaved (single-component) scan */
+    compptr = cinfo->cur_comp_info[0];
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = compptr->width_in_blocks;
+    cinfo->MCU_rows_in_scan = compptr->height_in_blocks;
+    
+    /* For noninterleaved scan, always one block per MCU */
+    compptr->MCU_width = 1;
+    compptr->MCU_height = 1;
+    compptr->MCU_blocks = 1;
+    compptr->MCU_sample_width = compptr->DCT_h_scaled_size;
+    compptr->last_col_width = 1;
+    /* For noninterleaved scans, it is convenient to define last_row_height
+     * as the number of block rows present in the last iMCU row.
+     */
+    tmp = (int) (compptr->height_in_blocks % compptr->v_samp_factor);
+    if (tmp == 0) tmp = compptr->v_samp_factor;
+    compptr->last_row_height = tmp;
+    
+    /* Prepare array describing MCU composition */
+    cinfo->blocks_in_MCU = 1;
+    cinfo->MCU_membership[0] = 0;
+    
+  } else {
+    
+    /* Interleaved (multi-component) scan */
+    if (cinfo->comps_in_scan <= 0 || cinfo->comps_in_scan > MAX_COMPS_IN_SCAN)
+      ERREXIT2(cinfo, JERR_COMPONENT_COUNT, cinfo->comps_in_scan,
+	       MAX_COMPS_IN_SCAN);
+    
+    /* Overall image size in MCUs */
+    cinfo->MCUs_per_row = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width,
+		    (long) (cinfo->max_h_samp_factor * cinfo->block_size));
+    cinfo->MCU_rows_in_scan = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height,
+		    (long) (cinfo->max_v_samp_factor * cinfo->block_size));
+    
+    cinfo->blocks_in_MCU = 0;
+    
+    for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+      compptr = cinfo->cur_comp_info[ci];
+      /* Sampling factors give # of blocks of component in each MCU */
+      compptr->MCU_width = compptr->h_samp_factor;
+      compptr->MCU_height = compptr->v_samp_factor;
+      compptr->MCU_blocks = compptr->MCU_width * compptr->MCU_height;
+      compptr->MCU_sample_width = compptr->MCU_width * compptr->DCT_h_scaled_size;
+      /* Figure number of non-dummy blocks in last MCU column & row */
+      tmp = (int) (compptr->width_in_blocks % compptr->MCU_width);
+      if (tmp == 0) tmp = compptr->MCU_width;
+      compptr->last_col_width = tmp;
+      tmp = (int) (compptr->height_in_blocks % compptr->MCU_height);
+      if (tmp == 0) tmp = compptr->MCU_height;
+      compptr->last_row_height = tmp;
+      /* Prepare array describing MCU composition */
+      mcublks = compptr->MCU_blocks;
+      if (cinfo->blocks_in_MCU + mcublks > D_MAX_BLOCKS_IN_MCU)
+	ERREXIT(cinfo, JERR_BAD_MCU_SIZE);
+      while (mcublks-- > 0) {
+	cinfo->MCU_membership[cinfo->blocks_in_MCU++] = ci;
+      }
+    }
+    
+  }
+}
+
+
+/*
+ * Save away a copy of the Q-table referenced by each component present
+ * in the current scan, unless already saved during a prior scan.
+ *
+ * In a multiple-scan JPEG file, the encoder could assign different components
+ * the same Q-table slot number, but change table definitions between scans
+ * so that each component uses a different Q-table.  (The IJG encoder is not
+ * currently capable of doing this, but other encoders might.)  Since we want
+ * to be able to dequantize all the components at the end of the file, this
+ * means that we have to save away the table actually used for each component.
+ * We do this by copying the table at the start of the first scan containing
+ * the component.
+ * The JPEG spec prohibits the encoder from changing the contents of a Q-table
+ * slot between scans of a component using that slot.  If the encoder does so
+ * anyway, this decoder will simply use the Q-table values that were current
+ * at the start of the first scan for the component.
+ *
+ * The decompressor output side looks only at the saved quant tables,
+ * not at the current Q-table slots.
+ */
+
+LOCAL(void)
+latch_quant_tables (j_decompress_ptr cinfo)
+{
+  int ci, qtblno;
+  jpeg_component_info *compptr;
+  JQUANT_TBL * qtbl;
+
+  for (ci = 0; ci < cinfo->comps_in_scan; ci++) {
+    compptr = cinfo->cur_comp_info[ci];
+    /* No work if we already saved Q-table for this component */
+    if (compptr->quant_table != NULL)
+      continue;
+    /* Make sure specified quantization table is present */
+    qtblno = compptr->quant_tbl_no;
+    if (qtblno < 0 || qtblno >= NUM_QUANT_TBLS ||
+	cinfo->quant_tbl_ptrs[qtblno] == NULL)
+      ERREXIT1(cinfo, JERR_NO_QUANT_TABLE, qtblno);
+    /* OK, save away the quantization table */
+    qtbl = (JQUANT_TBL *)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(JQUANT_TBL));
+    MEMCOPY(qtbl, cinfo->quant_tbl_ptrs[qtblno], SIZEOF(JQUANT_TBL));
+    compptr->quant_table = qtbl;
+  }
+}
+
+
+/*
+ * Initialize the input modules to read a scan of compressed data.
+ * The first call to this is done by jdmaster.c after initializing
+ * the entire decompressor (during jpeg_start_decompress).
+ * Subsequent calls come from consume_markers, below.
+ */
+
+METHODDEF(void)
+start_input_pass (j_decompress_ptr cinfo)
+{
+  per_scan_setup(cinfo);
+  latch_quant_tables(cinfo);
+  (*cinfo->entropy->start_pass) (cinfo);
+  (*cinfo->coef->start_input_pass) (cinfo);
+  cinfo->inputctl->consume_input = cinfo->coef->consume_data;
+}
+
+
+/*
+ * Finish up after inputting a compressed-data scan.
+ * This is called by the coefficient controller after it's read all
+ * the expected data of the scan.
+ */
+
+METHODDEF(void)
+finish_input_pass (j_decompress_ptr cinfo)
+{
+  cinfo->inputctl->consume_input = consume_markers;
+}
+
+
+/*
+ * Read JPEG markers before, between, or after compressed-data scans.
+ * Change state as necessary when a new scan is reached.
+ * Return value is JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+ *
+ * The consume_input method pointer points either here or to the
+ * coefficient controller's consume_data routine, depending on whether
+ * we are reading a compressed data segment or inter-segment markers.
+ *
+ * Note: This function should NOT return a pseudo SOS marker (with zero
+ * component number) to the caller.  A pseudo marker received by
+ * read_markers is processed and then skipped for other markers.
+ */
+
+METHODDEF(int)
+consume_markers (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl = (my_inputctl_ptr) cinfo->inputctl;
+  int val;
+
+  if (inputctl->pub.eoi_reached) /* After hitting EOI, read no further */
+    return JPEG_REACHED_EOI;
+
+  for (;;) {			/* Loop to pass pseudo SOS marker */
+    val = (*cinfo->marker->read_markers) (cinfo);
+
+    switch (val) {
+    case JPEG_REACHED_SOS:	/* Found SOS */
+      if (inputctl->inheaders) { /* 1st SOS */
+	if (inputctl->inheaders == 1)
+	  initial_setup(cinfo);
+	if (cinfo->comps_in_scan == 0) { /* pseudo SOS marker */
+	  inputctl->inheaders = 2;
+	  break;
+	}
+	inputctl->inheaders = 0;
+	/* Note: start_input_pass must be called by jdmaster.c
+	 * before any more input can be consumed.  jdapimin.c is
+	 * responsible for enforcing this sequencing.
+	 */
+      } else {			/* 2nd or later SOS marker */
+	if (! inputctl->pub.has_multiple_scans)
+	  ERREXIT(cinfo, JERR_EOI_EXPECTED); /* Oops, I wasn't expecting this! */
+	if (cinfo->comps_in_scan == 0) /* unexpected pseudo SOS marker */
+	  break;
+	start_input_pass(cinfo);
+      }
+      return val;
+    case JPEG_REACHED_EOI:	/* Found EOI */
+      inputctl->pub.eoi_reached = TRUE;
+      if (inputctl->inheaders) { /* Tables-only datastream, apparently */
+	if (cinfo->marker->saw_SOF)
+	  ERREXIT(cinfo, JERR_SOF_NO_SOS);
+      } else {
+	/* Prevent infinite loop in coef ctlr's decompress_data routine
+	 * if user set output_scan_number larger than number of scans.
+	 */
+	if (cinfo->output_scan_number > cinfo->input_scan_number)
+	  cinfo->output_scan_number = cinfo->input_scan_number;
+      }
+      return val;
+    case JPEG_SUSPENDED:
+      return val;
+    default:
+      return val;
+    }
+  }
+}
+
+
+/*
+ * Reset state to begin a fresh datastream.
+ */
+
+METHODDEF(void)
+reset_input_controller (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl = (my_inputctl_ptr) cinfo->inputctl;
+
+  inputctl->pub.consume_input = consume_markers;
+  inputctl->pub.has_multiple_scans = FALSE; /* "unknown" would be better */
+  inputctl->pub.eoi_reached = FALSE;
+  inputctl->inheaders = 1;
+  /* Reset other modules */
+  (*cinfo->err->reset_error_mgr) ((j_common_ptr) cinfo);
+  (*cinfo->marker->reset_marker_reader) (cinfo);
+  /* Reset progression state -- would be cleaner if entropy decoder did this */
+  cinfo->coef_bits = NULL;
+}
+
+
+/*
+ * Initialize the input controller module.
+ * This is called only once, when the decompression object is created.
+ */
+
+GLOBAL(void)
+jinit_input_controller (j_decompress_ptr cinfo)
+{
+  my_inputctl_ptr inputctl;
+
+  /* Create subobject in permanent pool */
+  inputctl = (my_inputctl_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				SIZEOF(my_input_controller));
+  cinfo->inputctl = (struct jpeg_input_controller *) inputctl;
+  /* Initialize method pointers */
+  inputctl->pub.consume_input = consume_markers;
+  inputctl->pub.reset_input_controller = reset_input_controller;
+  inputctl->pub.start_input_pass = start_input_pass;
+  inputctl->pub.finish_input_pass = finish_input_pass;
+  /* Initialize state: can't use reset_input_controller since we don't
+   * want to try to reset other modules yet.
+   */
+  inputctl->pub.has_multiple_scans = FALSE; /* "unknown" would be better */
+  inputctl->pub.eoi_reached = FALSE;
+  inputctl->inheaders = 1;
+}

Added: trunk/code/jpeg-8c/jdmainct.c
===================================================================
--- trunk/code/jpeg-8c/jdmainct.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdmainct.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,512 @@
+/*
+ * jdmainct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the main buffer controller for decompression.
+ * The main buffer lies between the JPEG decompressor proper and the
+ * post-processor; it holds downsampled data in the JPEG colorspace.
+ *
+ * Note that this code is bypassed in raw-data mode, since the application
+ * supplies the equivalent of the main buffer in that case.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/*
+ * In the current system design, the main buffer need never be a full-image
+ * buffer; any full-height buffers will be found inside the coefficient or
+ * postprocessing controllers.  Nonetheless, the main controller is not
+ * trivial.  Its responsibility is to provide context rows for upsampling/
+ * rescaling, and doing this in an efficient fashion is a bit tricky.
+ *
+ * Postprocessor input data is counted in "row groups".  A row group
+ * is defined to be (v_samp_factor * DCT_scaled_size / min_DCT_scaled_size)
+ * sample rows of each component.  (We require DCT_scaled_size values to be
+ * chosen such that these numbers are integers.  In practice DCT_scaled_size
+ * values will likely be powers of two, so we actually have the stronger
+ * condition that DCT_scaled_size / min_DCT_scaled_size is an integer.)
+ * Upsampling will typically produce max_v_samp_factor pixel rows from each
+ * row group (times any additional scale factor that the upsampler is
+ * applying).
+ *
+ * The coefficient controller will deliver data to us one iMCU row at a time;
+ * each iMCU row contains v_samp_factor * DCT_scaled_size sample rows, or
+ * exactly min_DCT_scaled_size row groups.  (This amount of data corresponds
+ * to one row of MCUs when the image is fully interleaved.)  Note that the
+ * number of sample rows varies across components, but the number of row
+ * groups does not.  Some garbage sample rows may be included in the last iMCU
+ * row at the bottom of the image.
+ *
+ * Depending on the vertical scaling algorithm used, the upsampler may need
+ * access to the sample row(s) above and below its current input row group.
+ * The upsampler is required to set need_context_rows TRUE at global selection
+ * time if so.  When need_context_rows is FALSE, this controller can simply
+ * obtain one iMCU row at a time from the coefficient controller and dole it
+ * out as row groups to the postprocessor.
+ *
+ * When need_context_rows is TRUE, this controller guarantees that the buffer
+ * passed to postprocessing contains at least one row group's worth of samples
+ * above and below the row group(s) being processed.  Note that the context
+ * rows "above" the first passed row group appear at negative row offsets in
+ * the passed buffer.  At the top and bottom of the image, the required
+ * context rows are manufactured by duplicating the first or last real sample
+ * row; this avoids having special cases in the upsampling inner loops.
+ *
+ * The amount of context is fixed at one row group just because that's a
+ * convenient number for this controller to work with.  The existing
+ * upsamplers really only need one sample row of context.  An upsampler
+ * supporting arbitrary output rescaling might wish for more than one row
+ * group of context when shrinking the image; tough, we don't handle that.
+ * (This is justified by the assumption that downsizing will be handled mostly
+ * by adjusting the DCT_scaled_size values, so that the actual scale factor at
+ * the upsample step needn't be much less than one.)
+ *
+ * To provide the desired context, we have to retain the last two row groups
+ * of one iMCU row while reading in the next iMCU row.  (The last row group
+ * can't be processed until we have another row group for its below-context,
+ * and so we have to save the next-to-last group too for its above-context.)
+ * We could do this most simply by copying data around in our buffer, but
+ * that'd be very slow.  We can avoid copying any data by creating a rather
+ * strange pointer structure.  Here's how it works.  We allocate a workspace
+ * consisting of M+2 row groups (where M = min_DCT_scaled_size is the number
+ * of row groups per iMCU row).  We create two sets of redundant pointers to
+ * the workspace.  Labeling the physical row groups 0 to M+1, the synthesized
+ * pointer lists look like this:
+ *                   M+1                          M-1
+ * master pointer --> 0         master pointer --> 0
+ *                    1                            1
+ *                   ...                          ...
+ *                   M-3                          M-3
+ *                   M-2                           M
+ *                   M-1                          M+1
+ *                    M                           M-2
+ *                   M+1                          M-1
+ *                    0                            0
+ * We read alternate iMCU rows using each master pointer; thus the last two
+ * row groups of the previous iMCU row remain un-overwritten in the workspace.
+ * The pointer lists are set up so that the required context rows appear to
+ * be adjacent to the proper places when we pass the pointer lists to the
+ * upsampler.
+ *
+ * The above pictures describe the normal state of the pointer lists.
+ * At top and bottom of the image, we diddle the pointer lists to duplicate
+ * the first or last sample row as necessary (this is cheaper than copying
+ * sample rows around).
+ *
+ * This scheme breaks down if M < 2, ie, min_DCT_scaled_size is 1.  In that
+ * situation each iMCU row provides only one row group so the buffering logic
+ * must be different (eg, we must read two iMCU rows before we can emit the
+ * first row group).  For now, we simply do not support providing context
+ * rows when min_DCT_scaled_size is 1.  That combination seems unlikely to
+ * be worth providing --- if someone wants a 1/8th-size preview, they probably
+ * want it quick and dirty, so a context-free upsampler is sufficient.
+ */
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_main_controller pub; /* public fields */
+
+  /* Pointer to allocated workspace (M or M+2 row groups). */
+  JSAMPARRAY buffer[MAX_COMPONENTS];
+
+  boolean buffer_full;		/* Have we gotten an iMCU row from decoder? */
+  JDIMENSION rowgroup_ctr;	/* counts row groups output to postprocessor */
+
+  /* Remaining fields are only used in the context case. */
+
+  /* These are the master pointers to the funny-order pointer lists. */
+  JSAMPIMAGE xbuffer[2];	/* pointers to weird pointer lists */
+
+  int whichptr;			/* indicates which pointer set is now in use */
+  int context_state;		/* process_data state machine status */
+  JDIMENSION rowgroups_avail;	/* row groups available to postprocessor */
+  JDIMENSION iMCU_row_ctr;	/* counts iMCU rows to detect image top/bot */
+} my_main_controller;
+
+typedef my_main_controller * my_main_ptr;
+
+/* context_state values: */
+#define CTX_PREPARE_FOR_IMCU	0	/* need to prepare for MCU row */
+#define CTX_PROCESS_IMCU	1	/* feeding iMCU to postprocessor */
+#define CTX_POSTPONED_ROW	2	/* feeding postponed row group */
+
+
+/* Forward declarations */
+METHODDEF(void) process_data_simple_main
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+METHODDEF(void) process_data_context_main
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+#ifdef QUANT_2PASS_SUPPORTED
+METHODDEF(void) process_data_crank_post
+	JPP((j_decompress_ptr cinfo, JSAMPARRAY output_buf,
+	     JDIMENSION *out_row_ctr, JDIMENSION out_rows_avail));
+#endif
+
+
+LOCAL(void)
+alloc_funny_pointers (j_decompress_ptr cinfo)
+/* Allocate space for the funny pointer lists.
+ * This is done only once, not once per pass.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, rgroup;
+  int M = cinfo->min_DCT_v_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf;
+
+  /* Get top-level space for component array pointers.
+   * We alloc both arrays with one call to save a few cycles.
+   */
+  main->xbuffer[0] = (JSAMPIMAGE)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				cinfo->num_components * 2 * SIZEOF(JSAMPARRAY));
+  main->xbuffer[1] = main->xbuffer[0] + cinfo->num_components;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->DCT_v_scaled_size) /
+      cinfo->min_DCT_v_scaled_size; /* height of a row group of component */
+    /* Get space for pointer lists --- M+4 row groups in each list.
+     * We alloc both pointer lists with one call to save a few cycles.
+     */
+    xbuf = (JSAMPARRAY)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  2 * (rgroup * (M + 4)) * SIZEOF(JSAMPROW));
+    xbuf += rgroup;		/* want one row group at negative offsets */
+    main->xbuffer[0][ci] = xbuf;
+    xbuf += rgroup * (M + 4);
+    main->xbuffer[1][ci] = xbuf;
+  }
+}
+
+
+LOCAL(void)
+make_funny_pointers (j_decompress_ptr cinfo)
+/* Create the funny pointer lists discussed in the comments above.
+ * The actual workspace is already allocated (in main->buffer),
+ * and the space for the pointer lists is allocated too.
+ * This routine just fills in the curiously ordered lists.
+ * This will be repeated at the beginning of each pass.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup;
+  int M = cinfo->min_DCT_v_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY buf, xbuf0, xbuf1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->DCT_v_scaled_size) /
+      cinfo->min_DCT_v_scaled_size; /* height of a row group of component */
+    xbuf0 = main->xbuffer[0][ci];
+    xbuf1 = main->xbuffer[1][ci];
+    /* First copy the workspace pointers as-is */
+    buf = main->buffer[ci];
+    for (i = 0; i < rgroup * (M + 2); i++) {
+      xbuf0[i] = xbuf1[i] = buf[i];
+    }
+    /* In the second list, put the last four row groups in swapped order */
+    for (i = 0; i < rgroup * 2; i++) {
+      xbuf1[rgroup*(M-2) + i] = buf[rgroup*M + i];
+      xbuf1[rgroup*M + i] = buf[rgroup*(M-2) + i];
+    }
+    /* The wraparound pointers at top and bottom will be filled later
+     * (see set_wraparound_pointers, below).  Initially we want the "above"
+     * pointers to duplicate the first actual data line.  This only needs
+     * to happen in xbuffer[0].
+     */
+    for (i = 0; i < rgroup; i++) {
+      xbuf0[i - rgroup] = xbuf0[0];
+    }
+  }
+}
+
+
+LOCAL(void)
+set_wraparound_pointers (j_decompress_ptr cinfo)
+/* Set up the "wraparound" pointers at top and bottom of the pointer lists.
+ * This changes the pointer list state from top-of-image to the normal state.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup;
+  int M = cinfo->min_DCT_v_scaled_size;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf0, xbuf1;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->DCT_v_scaled_size) /
+      cinfo->min_DCT_v_scaled_size; /* height of a row group of component */
+    xbuf0 = main->xbuffer[0][ci];
+    xbuf1 = main->xbuffer[1][ci];
+    for (i = 0; i < rgroup; i++) {
+      xbuf0[i - rgroup] = xbuf0[rgroup*(M+1) + i];
+      xbuf1[i - rgroup] = xbuf1[rgroup*(M+1) + i];
+      xbuf0[rgroup*(M+2) + i] = xbuf0[i];
+      xbuf1[rgroup*(M+2) + i] = xbuf1[i];
+    }
+  }
+}
+
+
+LOCAL(void)
+set_bottom_pointers (j_decompress_ptr cinfo)
+/* Change the pointer lists to duplicate the last sample row at the bottom
+ * of the image.  whichptr indicates which xbuffer holds the final iMCU row.
+ * Also sets rowgroups_avail to indicate number of nondummy row groups in row.
+ */
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  int ci, i, rgroup, iMCUheight, rows_left;
+  jpeg_component_info *compptr;
+  JSAMPARRAY xbuf;
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Count sample rows in one iMCU row and in one row group */
+    iMCUheight = compptr->v_samp_factor * compptr->DCT_v_scaled_size;
+    rgroup = iMCUheight / cinfo->min_DCT_v_scaled_size;
+    /* Count nondummy sample rows remaining for this component */
+    rows_left = (int) (compptr->downsampled_height % (JDIMENSION) iMCUheight);
+    if (rows_left == 0) rows_left = iMCUheight;
+    /* Count nondummy row groups.  Should get same answer for each component,
+     * so we need only do it once.
+     */
+    if (ci == 0) {
+      main->rowgroups_avail = (JDIMENSION) ((rows_left-1) / rgroup + 1);
+    }
+    /* Duplicate the last real sample row rgroup*2 times; this pads out the
+     * last partial rowgroup and ensures at least one full rowgroup of context.
+     */
+    xbuf = main->xbuffer[main->whichptr][ci];
+    for (i = 0; i < rgroup * 2; i++) {
+      xbuf[rows_left + i] = xbuf[rows_left-1];
+    }
+  }
+}
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_main (j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (cinfo->upsample->need_context_rows) {
+      main->pub.process_data = process_data_context_main;
+      make_funny_pointers(cinfo); /* Create the xbuffer[] lists */
+      main->whichptr = 0;	/* Read first iMCU row into xbuffer[0] */
+      main->context_state = CTX_PREPARE_FOR_IMCU;
+      main->iMCU_row_ctr = 0;
+    } else {
+      /* Simple case with no context needed */
+      main->pub.process_data = process_data_simple_main;
+    }
+    main->buffer_full = FALSE;	/* Mark buffer empty */
+    main->rowgroup_ctr = 0;
+    break;
+#ifdef QUANT_2PASS_SUPPORTED
+  case JBUF_CRANK_DEST:
+    /* For last pass of 2-pass quantization, just crank the postprocessor */
+    main->pub.process_data = process_data_crank_post;
+    break;
+#endif
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This handles the simple case where no context is required.
+ */
+
+METHODDEF(void)
+process_data_simple_main (j_decompress_ptr cinfo,
+			  JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			  JDIMENSION out_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+  JDIMENSION rowgroups_avail;
+
+  /* Read input data if we haven't filled the main buffer yet */
+  if (! main->buffer_full) {
+    if (! (*cinfo->coef->decompress_data) (cinfo, main->buffer))
+      return;			/* suspension forced, can do nothing more */
+    main->buffer_full = TRUE;	/* OK, we have an iMCU row to work with */
+  }
+
+  /* There are always min_DCT_scaled_size row groups in an iMCU row. */
+  rowgroups_avail = (JDIMENSION) cinfo->min_DCT_v_scaled_size;
+  /* Note: at the bottom of the image, we may pass extra garbage row groups
+   * to the postprocessor.  The postprocessor has to check for bottom
+   * of image anyway (at row resolution), so no point in us doing it too.
+   */
+
+  /* Feed the postprocessor */
+  (*cinfo->post->post_process_data) (cinfo, main->buffer,
+				     &main->rowgroup_ctr, rowgroups_avail,
+				     output_buf, out_row_ctr, out_rows_avail);
+
+  /* Has postprocessor consumed all the data yet? If so, mark buffer empty */
+  if (main->rowgroup_ctr >= rowgroups_avail) {
+    main->buffer_full = FALSE;
+    main->rowgroup_ctr = 0;
+  }
+}
+
+
+/*
+ * Process some data.
+ * This handles the case where context rows must be provided.
+ */
+
+METHODDEF(void)
+process_data_context_main (j_decompress_ptr cinfo,
+			   JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			   JDIMENSION out_rows_avail)
+{
+  my_main_ptr main = (my_main_ptr) cinfo->main;
+
+  /* Read input data if we haven't filled the main buffer yet */
+  if (! main->buffer_full) {
+    if (! (*cinfo->coef->decompress_data) (cinfo,
+					   main->xbuffer[main->whichptr]))
+      return;			/* suspension forced, can do nothing more */
+    main->buffer_full = TRUE;	/* OK, we have an iMCU row to work with */
+    main->iMCU_row_ctr++;	/* count rows received */
+  }
+
+  /* Postprocessor typically will not swallow all the input data it is handed
+   * in one call (due to filling the output buffer first).  Must be prepared
+   * to exit and restart.  This switch lets us keep track of how far we got.
+   * Note that each case falls through to the next on successful completion.
+   */
+  switch (main->context_state) {
+  case CTX_POSTPONED_ROW:
+    /* Call postprocessor using previously set pointers for postponed row */
+    (*cinfo->post->post_process_data) (cinfo, main->xbuffer[main->whichptr],
+			&main->rowgroup_ctr, main->rowgroups_avail,
+			output_buf, out_row_ctr, out_rows_avail);
+    if (main->rowgroup_ctr < main->rowgroups_avail)
+      return;			/* Need to suspend */
+    main->context_state = CTX_PREPARE_FOR_IMCU;
+    if (*out_row_ctr >= out_rows_avail)
+      return;			/* Postprocessor exactly filled output buf */
+    /*FALLTHROUGH*/
+  case CTX_PREPARE_FOR_IMCU:
+    /* Prepare to process first M-1 row groups of this iMCU row */
+    main->rowgroup_ctr = 0;
+    main->rowgroups_avail = (JDIMENSION) (cinfo->min_DCT_v_scaled_size - 1);
+    /* Check for bottom of image: if so, tweak pointers to "duplicate"
+     * the last sample row, and adjust rowgroups_avail to ignore padding rows.
+     */
+    if (main->iMCU_row_ctr == cinfo->total_iMCU_rows)
+      set_bottom_pointers(cinfo);
+    main->context_state = CTX_PROCESS_IMCU;
+    /*FALLTHROUGH*/
+  case CTX_PROCESS_IMCU:
+    /* Call postprocessor using previously set pointers */
+    (*cinfo->post->post_process_data) (cinfo, main->xbuffer[main->whichptr],
+			&main->rowgroup_ctr, main->rowgroups_avail,
+			output_buf, out_row_ctr, out_rows_avail);
+    if (main->rowgroup_ctr < main->rowgroups_avail)
+      return;			/* Need to suspend */
+    /* After the first iMCU, change wraparound pointers to normal state */
+    if (main->iMCU_row_ctr == 1)
+      set_wraparound_pointers(cinfo);
+    /* Prepare to load new iMCU row using other xbuffer list */
+    main->whichptr ^= 1;	/* 0=>1 or 1=>0 */
+    main->buffer_full = FALSE;
+    /* Still need to process last row group of this iMCU row, */
+    /* which is saved at index M+1 of the other xbuffer */
+    main->rowgroup_ctr = (JDIMENSION) (cinfo->min_DCT_v_scaled_size + 1);
+    main->rowgroups_avail = (JDIMENSION) (cinfo->min_DCT_v_scaled_size + 2);
+    main->context_state = CTX_POSTPONED_ROW;
+  }
+}
+
+
+/*
+ * Process some data.
+ * Final pass of two-pass quantization: just call the postprocessor.
+ * Source data will be the postprocessor controller's internal buffer.
+ */
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+METHODDEF(void)
+process_data_crank_post (j_decompress_ptr cinfo,
+			 JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+			 JDIMENSION out_rows_avail)
+{
+  (*cinfo->post->post_process_data) (cinfo, (JSAMPIMAGE) NULL,
+				     (JDIMENSION *) NULL, (JDIMENSION) 0,
+				     output_buf, out_row_ctr, out_rows_avail);
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
+
+
+/*
+ * Initialize main buffer controller.
+ */
+
+GLOBAL(void)
+jinit_d_main_controller (j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_main_ptr main;
+  int ci, rgroup, ngroups;
+  jpeg_component_info *compptr;
+
+  main = (my_main_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_main_controller));
+  cinfo->main = (struct jpeg_d_main_controller *) main;
+  main->pub.start_pass = start_pass_main;
+
+  if (need_full_buffer)		/* shouldn't happen */
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+
+  /* Allocate the workspace.
+   * ngroups is the number of row groups we need.
+   */
+  if (cinfo->upsample->need_context_rows) {
+    if (cinfo->min_DCT_v_scaled_size < 2) /* unsupported, see comments above */
+      ERREXIT(cinfo, JERR_NOTIMPL);
+    alloc_funny_pointers(cinfo); /* Alloc space for xbuffer[] lists */
+    ngroups = cinfo->min_DCT_v_scaled_size + 2;
+  } else {
+    ngroups = cinfo->min_DCT_v_scaled_size;
+  }
+
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    rgroup = (compptr->v_samp_factor * compptr->DCT_v_scaled_size) /
+      cinfo->min_DCT_v_scaled_size; /* height of a row group of component */
+    main->buffer[ci] = (*cinfo->mem->alloc_sarray)
+			((j_common_ptr) cinfo, JPOOL_IMAGE,
+			 compptr->width_in_blocks * compptr->DCT_h_scaled_size,
+			 (JDIMENSION) (rgroup * ngroups));
+  }
+}

Added: trunk/code/jpeg-8c/jdmarker.c
===================================================================
--- trunk/code/jpeg-8c/jdmarker.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdmarker.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,1406 @@
+/*
+ * jdmarker.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modified 2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains routines to decode JPEG datastream markers.
+ * Most of the complexity arises from our desire to support input
+ * suspension: if not all of the data for a marker is available,
+ * we must exit back to the application.  On resumption, we reprocess
+ * the marker.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+typedef enum {			/* JPEG marker codes */
+  M_SOF0  = 0xc0,
+  M_SOF1  = 0xc1,
+  M_SOF2  = 0xc2,
+  M_SOF3  = 0xc3,
+  
+  M_SOF5  = 0xc5,
+  M_SOF6  = 0xc6,
+  M_SOF7  = 0xc7,
+  
+  M_JPG   = 0xc8,
+  M_SOF9  = 0xc9,
+  M_SOF10 = 0xca,
+  M_SOF11 = 0xcb,
+  
+  M_SOF13 = 0xcd,
+  M_SOF14 = 0xce,
+  M_SOF15 = 0xcf,
+  
+  M_DHT   = 0xc4,
+  
+  M_DAC   = 0xcc,
+  
+  M_RST0  = 0xd0,
+  M_RST1  = 0xd1,
+  M_RST2  = 0xd2,
+  M_RST3  = 0xd3,
+  M_RST4  = 0xd4,
+  M_RST5  = 0xd5,
+  M_RST6  = 0xd6,
+  M_RST7  = 0xd7,
+  
+  M_SOI   = 0xd8,
+  M_EOI   = 0xd9,
+  M_SOS   = 0xda,
+  M_DQT   = 0xdb,
+  M_DNL   = 0xdc,
+  M_DRI   = 0xdd,
+  M_DHP   = 0xde,
+  M_EXP   = 0xdf,
+  
+  M_APP0  = 0xe0,
+  M_APP1  = 0xe1,
+  M_APP2  = 0xe2,
+  M_APP3  = 0xe3,
+  M_APP4  = 0xe4,
+  M_APP5  = 0xe5,
+  M_APP6  = 0xe6,
+  M_APP7  = 0xe7,
+  M_APP8  = 0xe8,
+  M_APP9  = 0xe9,
+  M_APP10 = 0xea,
+  M_APP11 = 0xeb,
+  M_APP12 = 0xec,
+  M_APP13 = 0xed,
+  M_APP14 = 0xee,
+  M_APP15 = 0xef,
+  
+  M_JPG0  = 0xf0,
+  M_JPG13 = 0xfd,
+  M_COM   = 0xfe,
+  
+  M_TEM   = 0x01,
+  
+  M_ERROR = 0x100
+} JPEG_MARKER;
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_marker_reader pub; /* public fields */
+
+  /* Application-overridable marker processing methods */
+  jpeg_marker_parser_method process_COM;
+  jpeg_marker_parser_method process_APPn[16];
+
+  /* Limit on marker data length to save for each marker type */
+  unsigned int length_limit_COM;
+  unsigned int length_limit_APPn[16];
+
+  /* Status of COM/APPn marker saving */
+  jpeg_saved_marker_ptr cur_marker;	/* NULL if not processing a marker */
+  unsigned int bytes_read;		/* data bytes read so far in marker */
+  /* Note: cur_marker is not linked into marker_list until it's all read. */
+} my_marker_reader;
+
+typedef my_marker_reader * my_marker_ptr;
+
+
+/*
+ * Macros for fetching data from the data source module.
+ *
+ * At all times, cinfo->src->next_input_byte and ->bytes_in_buffer reflect
+ * the current restart point; we update them only when we have reached a
+ * suitable place to restart if a suspension occurs.
+ */
+
+/* Declare and initialize local copies of input pointer/count */
+#define INPUT_VARS(cinfo)  \
+	struct jpeg_source_mgr * datasrc = (cinfo)->src;  \
+	const JOCTET * next_input_byte = datasrc->next_input_byte;  \
+	size_t bytes_in_buffer = datasrc->bytes_in_buffer
+
+/* Unload the local copies --- do this only at a restart boundary */
+#define INPUT_SYNC(cinfo)  \
+	( datasrc->next_input_byte = next_input_byte,  \
+	  datasrc->bytes_in_buffer = bytes_in_buffer )
+
+/* Reload the local copies --- used only in MAKE_BYTE_AVAIL */
+#define INPUT_RELOAD(cinfo)  \
+	( next_input_byte = datasrc->next_input_byte,  \
+	  bytes_in_buffer = datasrc->bytes_in_buffer )
+
+/* Internal macro for INPUT_BYTE and INPUT_2BYTES: make a byte available.
+ * Note we do *not* do INPUT_SYNC before calling fill_input_buffer,
+ * but we must reload the local copies after a successful fill.
+ */
+#define MAKE_BYTE_AVAIL(cinfo,action)  \
+	if (bytes_in_buffer == 0) {  \
+	  if (! (*datasrc->fill_input_buffer) (cinfo))  \
+	    { action; }  \
+	  INPUT_RELOAD(cinfo);  \
+	}
+
+/* Read a byte into variable V.
+ * If must suspend, take the specified action (typically "return FALSE").
+ */
+#define INPUT_BYTE(cinfo,V,action)  \
+	MAKESTMT( MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V = GETJOCTET(*next_input_byte++); )
+
+/* As above, but read two bytes interpreted as an unsigned 16-bit integer.
+ * V should be declared unsigned int or perhaps INT32.
+ */
+#define INPUT_2BYTES(cinfo,V,action)  \
+	MAKESTMT( MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V = ((unsigned int) GETJOCTET(*next_input_byte++)) << 8; \
+		  MAKE_BYTE_AVAIL(cinfo,action); \
+		  bytes_in_buffer--; \
+		  V += GETJOCTET(*next_input_byte++); )
+
+
+/*
+ * Routines to process JPEG markers.
+ *
+ * Entry condition: JPEG marker itself has been read and its code saved
+ *   in cinfo->unread_marker; input restart point is just after the marker.
+ *
+ * Exit: if return TRUE, have read and processed any parameters, and have
+ *   updated the restart point to point after the parameters.
+ *   If return FALSE, was forced to suspend before reaching end of
+ *   marker parameters; restart point has not been moved.  Same routine
+ *   will be called again after application supplies more input data.
+ *
+ * This approach to suspension assumes that all of a marker's parameters
+ * can fit into a single input bufferload.  This should hold for "normal"
+ * markers.  Some COM/APPn markers might have large parameter segments
+ * that might not fit.  If we are simply dropping such a marker, we use
+ * skip_input_data to get past it, and thereby put the problem on the
+ * source manager's shoulders.  If we are saving the marker's contents
+ * into memory, we use a slightly different convention: when forced to
+ * suspend, the marker processor updates the restart point to the end of
+ * what it's consumed (ie, the end of the buffer) before returning FALSE.
+ * On resumption, cinfo->unread_marker still contains the marker code,
+ * but the data source will point to the next chunk of marker data.
+ * The marker processor must retain internal state to deal with this.
+ *
+ * Note that we don't bother to avoid duplicate trace messages if a
+ * suspension occurs within marker parameters.  Other side effects
+ * require more care.
+ */
+
+
+LOCAL(boolean)
+get_soi (j_decompress_ptr cinfo)
+/* Process an SOI marker */
+{
+  int i;
+  
+  TRACEMS(cinfo, 1, JTRC_SOI);
+
+  if (cinfo->marker->saw_SOI)
+    ERREXIT(cinfo, JERR_SOI_DUPLICATE);
+
+  /* Reset all parameters that are defined to be reset by SOI */
+
+  for (i = 0; i < NUM_ARITH_TBLS; i++) {
+    cinfo->arith_dc_L[i] = 0;
+    cinfo->arith_dc_U[i] = 1;
+    cinfo->arith_ac_K[i] = 5;
+  }
+  cinfo->restart_interval = 0;
+
+  /* Set initial assumptions for colorspace etc */
+
+  cinfo->jpeg_color_space = JCS_UNKNOWN;
+  cinfo->CCIR601_sampling = FALSE; /* Assume non-CCIR sampling??? */
+
+  cinfo->saw_JFIF_marker = FALSE;
+  cinfo->JFIF_major_version = 1; /* set default JFIF APP0 values */
+  cinfo->JFIF_minor_version = 1;
+  cinfo->density_unit = 0;
+  cinfo->X_density = 1;
+  cinfo->Y_density = 1;
+  cinfo->saw_Adobe_marker = FALSE;
+  cinfo->Adobe_transform = 0;
+
+  cinfo->marker->saw_SOI = TRUE;
+
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_sof (j_decompress_ptr cinfo, boolean is_baseline, boolean is_prog,
+	 boolean is_arith)
+/* Process a SOFn marker */
+{
+  INT32 length;
+  int c, ci;
+  jpeg_component_info * compptr;
+  INPUT_VARS(cinfo);
+
+  cinfo->is_baseline = is_baseline;
+  cinfo->progressive_mode = is_prog;
+  cinfo->arith_code = is_arith;
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  INPUT_BYTE(cinfo, cinfo->data_precision, return FALSE);
+  INPUT_2BYTES(cinfo, cinfo->image_height, return FALSE);
+  INPUT_2BYTES(cinfo, cinfo->image_width, return FALSE);
+  INPUT_BYTE(cinfo, cinfo->num_components, return FALSE);
+
+  length -= 8;
+
+  TRACEMS4(cinfo, 1, JTRC_SOF, cinfo->unread_marker,
+	   (int) cinfo->image_width, (int) cinfo->image_height,
+	   cinfo->num_components);
+
+  if (cinfo->marker->saw_SOF)
+    ERREXIT(cinfo, JERR_SOF_DUPLICATE);
+
+  /* We don't support files in which the image height is initially specified */
+  /* as 0 and is later redefined by DNL.  As long as we have to check that,  */
+  /* might as well have a general sanity check. */
+  if (cinfo->image_height <= 0 || cinfo->image_width <= 0
+      || cinfo->num_components <= 0)
+    ERREXIT(cinfo, JERR_EMPTY_IMAGE);
+
+  if (length != (cinfo->num_components * 3))
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  if (cinfo->comp_info == NULL)	/* do only once, even if suspend */
+    cinfo->comp_info = (jpeg_component_info *) (*cinfo->mem->alloc_small)
+			((j_common_ptr) cinfo, JPOOL_IMAGE,
+			 cinfo->num_components * SIZEOF(jpeg_component_info));
+  
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    compptr->component_index = ci;
+    INPUT_BYTE(cinfo, compptr->component_id, return FALSE);
+    INPUT_BYTE(cinfo, c, return FALSE);
+    compptr->h_samp_factor = (c >> 4) & 15;
+    compptr->v_samp_factor = (c     ) & 15;
+    INPUT_BYTE(cinfo, compptr->quant_tbl_no, return FALSE);
+
+    TRACEMS4(cinfo, 1, JTRC_SOF_COMPONENT,
+	     compptr->component_id, compptr->h_samp_factor,
+	     compptr->v_samp_factor, compptr->quant_tbl_no);
+  }
+
+  cinfo->marker->saw_SOF = TRUE;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_sos (j_decompress_ptr cinfo)
+/* Process a SOS marker */
+{
+  INT32 length;
+  int i, ci, n, c, cc;
+  jpeg_component_info * compptr;
+  INPUT_VARS(cinfo);
+
+  if (! cinfo->marker->saw_SOF)
+    ERREXIT(cinfo, JERR_SOS_NO_SOF);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+
+  INPUT_BYTE(cinfo, n, return FALSE); /* Number of components */
+
+  TRACEMS1(cinfo, 1, JTRC_SOS, n);
+
+  if (length != (n * 2 + 6) || n > MAX_COMPS_IN_SCAN ||
+      (n == 0 && !cinfo->progressive_mode))
+      /* pseudo SOS marker only allowed in progressive mode */
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  cinfo->comps_in_scan = n;
+
+  /* Collect the component-spec parameters */
+
+  for (i = 0; i < n; i++) {
+    INPUT_BYTE(cinfo, cc, return FALSE);
+    INPUT_BYTE(cinfo, c, return FALSE);
+    
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      if (cc == compptr->component_id)
+	goto id_found;
+    }
+
+    ERREXIT1(cinfo, JERR_BAD_COMPONENT_ID, cc);
+
+  id_found:
+
+    cinfo->cur_comp_info[i] = compptr;
+    compptr->dc_tbl_no = (c >> 4) & 15;
+    compptr->ac_tbl_no = (c     ) & 15;
+    
+    TRACEMS3(cinfo, 1, JTRC_SOS_COMPONENT, cc,
+	     compptr->dc_tbl_no, compptr->ac_tbl_no);
+  }
+
+  /* Collect the additional scan parameters Ss, Se, Ah/Al. */
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Ss = c;
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Se = c;
+  INPUT_BYTE(cinfo, c, return FALSE);
+  cinfo->Ah = (c >> 4) & 15;
+  cinfo->Al = (c     ) & 15;
+
+  TRACEMS4(cinfo, 1, JTRC_SOS_PARAMS, cinfo->Ss, cinfo->Se,
+	   cinfo->Ah, cinfo->Al);
+
+  /* Prepare to scan data & restart markers */
+  cinfo->marker->next_restart_num = 0;
+
+  /* Count another (non-pseudo) SOS marker */
+  if (n) cinfo->input_scan_number++;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+#ifdef D_ARITH_CODING_SUPPORTED
+
+LOCAL(boolean)
+get_dac (j_decompress_ptr cinfo)
+/* Process a DAC marker */
+{
+  INT32 length;
+  int index, val;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  while (length > 0) {
+    INPUT_BYTE(cinfo, index, return FALSE);
+    INPUT_BYTE(cinfo, val, return FALSE);
+
+    length -= 2;
+
+    TRACEMS2(cinfo, 1, JTRC_DAC, index, val);
+
+    if (index < 0 || index >= (2*NUM_ARITH_TBLS))
+      ERREXIT1(cinfo, JERR_DAC_INDEX, index);
+
+    if (index >= NUM_ARITH_TBLS) { /* define AC table */
+      cinfo->arith_ac_K[index-NUM_ARITH_TBLS] = (UINT8) val;
+    } else {			/* define DC table */
+      cinfo->arith_dc_L[index] = (UINT8) (val & 0x0F);
+      cinfo->arith_dc_U[index] = (UINT8) (val >> 4);
+      if (cinfo->arith_dc_L[index] > cinfo->arith_dc_U[index])
+	ERREXIT1(cinfo, JERR_DAC_VALUE, val);
+    }
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+#else /* ! D_ARITH_CODING_SUPPORTED */
+
+#define get_dac(cinfo)  skip_variable(cinfo)
+
+#endif /* D_ARITH_CODING_SUPPORTED */
+
+
+LOCAL(boolean)
+get_dht (j_decompress_ptr cinfo)
+/* Process a DHT marker */
+{
+  INT32 length;
+  UINT8 bits[17];
+  UINT8 huffval[256];
+  int i, index, count;
+  JHUFF_TBL **htblptr;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  while (length > 16) {
+    INPUT_BYTE(cinfo, index, return FALSE);
+
+    TRACEMS1(cinfo, 1, JTRC_DHT, index);
+      
+    bits[0] = 0;
+    count = 0;
+    for (i = 1; i <= 16; i++) {
+      INPUT_BYTE(cinfo, bits[i], return FALSE);
+      count += bits[i];
+    }
+
+    length -= 1 + 16;
+
+    TRACEMS8(cinfo, 2, JTRC_HUFFBITS,
+	     bits[1], bits[2], bits[3], bits[4],
+	     bits[5], bits[6], bits[7], bits[8]);
+    TRACEMS8(cinfo, 2, JTRC_HUFFBITS,
+	     bits[9], bits[10], bits[11], bits[12],
+	     bits[13], bits[14], bits[15], bits[16]);
+
+    /* Here we just do minimal validation of the counts to avoid walking
+     * off the end of our table space.  jdhuff.c will check more carefully.
+     */
+    if (count > 256 || ((INT32) count) > length)
+      ERREXIT(cinfo, JERR_BAD_HUFF_TABLE);
+
+    for (i = 0; i < count; i++)
+      INPUT_BYTE(cinfo, huffval[i], return FALSE);
+
+    length -= count;
+
+    if (index & 0x10) {		/* AC table definition */
+      index -= 0x10;
+      htblptr = &cinfo->ac_huff_tbl_ptrs[index];
+    } else {			/* DC table definition */
+      htblptr = &cinfo->dc_huff_tbl_ptrs[index];
+    }
+
+    if (index < 0 || index >= NUM_HUFF_TBLS)
+      ERREXIT1(cinfo, JERR_DHT_INDEX, index);
+
+    if (*htblptr == NULL)
+      *htblptr = jpeg_alloc_huff_table((j_common_ptr) cinfo);
+  
+    MEMCOPY((*htblptr)->bits, bits, SIZEOF((*htblptr)->bits));
+    MEMCOPY((*htblptr)->huffval, huffval, SIZEOF((*htblptr)->huffval));
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_dqt (j_decompress_ptr cinfo)
+/* Process a DQT marker */
+{
+  INT32 length, count, i;
+  int n, prec;
+  unsigned int tmp;
+  JQUANT_TBL *quant_ptr;
+  const int *natural_order;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  while (length > 0) {
+    length--;
+    INPUT_BYTE(cinfo, n, return FALSE);
+    prec = n >> 4;
+    n &= 0x0F;
+
+    TRACEMS2(cinfo, 1, JTRC_DQT, n, prec);
+
+    if (n >= NUM_QUANT_TBLS)
+      ERREXIT1(cinfo, JERR_DQT_INDEX, n);
+      
+    if (cinfo->quant_tbl_ptrs[n] == NULL)
+      cinfo->quant_tbl_ptrs[n] = jpeg_alloc_quant_table((j_common_ptr) cinfo);
+    quant_ptr = cinfo->quant_tbl_ptrs[n];
+
+    if (prec) {
+      if (length < DCTSIZE2 * 2) {
+	/* Initialize full table for safety. */
+	for (i = 0; i < DCTSIZE2; i++) {
+	  quant_ptr->quantval[i] = 1;
+	}
+	count = length >> 1;
+      } else
+	count = DCTSIZE2;
+    } else {
+      if (length < DCTSIZE2) {
+	/* Initialize full table for safety. */
+	for (i = 0; i < DCTSIZE2; i++) {
+	  quant_ptr->quantval[i] = 1;
+	}
+	count = length;
+      } else
+	count = DCTSIZE2;
+    }
+
+    switch (count) {
+    case (2*2): natural_order = jpeg_natural_order2; break;
+    case (3*3): natural_order = jpeg_natural_order3; break;
+    case (4*4): natural_order = jpeg_natural_order4; break;
+    case (5*5): natural_order = jpeg_natural_order5; break;
+    case (6*6): natural_order = jpeg_natural_order6; break;
+    case (7*7): natural_order = jpeg_natural_order7; break;
+    default:    natural_order = jpeg_natural_order;  break;
+    }
+
+    for (i = 0; i < count; i++) {
+      if (prec)
+	INPUT_2BYTES(cinfo, tmp, return FALSE);
+      else
+	INPUT_BYTE(cinfo, tmp, return FALSE);
+      /* We convert the zigzag-order table to natural array order. */
+      quant_ptr->quantval[natural_order[i]] = (UINT16) tmp;
+    }
+
+    if (cinfo->err->trace_level >= 2) {
+      for (i = 0; i < DCTSIZE2; i += 8) {
+	TRACEMS8(cinfo, 2, JTRC_QUANTVALS,
+		 quant_ptr->quantval[i],   quant_ptr->quantval[i+1],
+		 quant_ptr->quantval[i+2], quant_ptr->quantval[i+3],
+		 quant_ptr->quantval[i+4], quant_ptr->quantval[i+5],
+		 quant_ptr->quantval[i+6], quant_ptr->quantval[i+7]);
+      }
+    }
+
+    length -= count;
+    if (prec) length -= count;
+  }
+
+  if (length != 0)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+get_dri (j_decompress_ptr cinfo)
+/* Process a DRI marker */
+{
+  INT32 length;
+  unsigned int tmp;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  
+  if (length != 4)
+    ERREXIT(cinfo, JERR_BAD_LENGTH);
+
+  INPUT_2BYTES(cinfo, tmp, return FALSE);
+
+  TRACEMS1(cinfo, 1, JTRC_DRI, tmp);
+
+  cinfo->restart_interval = tmp;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Routines for processing APPn and COM markers.
+ * These are either saved in memory or discarded, per application request.
+ * APP0 and APP14 are specially checked to see if they are
+ * JFIF and Adobe markers, respectively.
+ */
+
+#define APP0_DATA_LEN	14	/* Length of interesting data in APP0 */
+#define APP14_DATA_LEN	12	/* Length of interesting data in APP14 */
+#define APPN_DATA_LEN	14	/* Must be the largest of the above!! */
+
+
+LOCAL(void)
+examine_app0 (j_decompress_ptr cinfo, JOCTET FAR * data,
+	      unsigned int datalen, INT32 remaining)
+/* Examine first few bytes from an APP0.
+ * Take appropriate action if it is a JFIF marker.
+ * datalen is # of bytes at data[], remaining is length of rest of marker data.
+ */
+{
+  INT32 totallen = (INT32) datalen + remaining;
+
+  if (datalen >= APP0_DATA_LEN &&
+      GETJOCTET(data[0]) == 0x4A &&
+      GETJOCTET(data[1]) == 0x46 &&
+      GETJOCTET(data[2]) == 0x49 &&
+      GETJOCTET(data[3]) == 0x46 &&
+      GETJOCTET(data[4]) == 0) {
+    /* Found JFIF APP0 marker: save info */
+    cinfo->saw_JFIF_marker = TRUE;
+    cinfo->JFIF_major_version = GETJOCTET(data[5]);
+    cinfo->JFIF_minor_version = GETJOCTET(data[6]);
+    cinfo->density_unit = GETJOCTET(data[7]);
+    cinfo->X_density = (GETJOCTET(data[8]) << 8) + GETJOCTET(data[9]);
+    cinfo->Y_density = (GETJOCTET(data[10]) << 8) + GETJOCTET(data[11]);
+    /* Check version.
+     * Major version must be 1, anything else signals an incompatible change.
+     * (We used to treat this as an error, but now it's a nonfatal warning,
+     * because some bozo at Hijaak couldn't read the spec.)
+     * Minor version should be 0..2, but process anyway if newer.
+     */
+    if (cinfo->JFIF_major_version != 1)
+      WARNMS2(cinfo, JWRN_JFIF_MAJOR,
+	      cinfo->JFIF_major_version, cinfo->JFIF_minor_version);
+    /* Generate trace messages */
+    TRACEMS5(cinfo, 1, JTRC_JFIF,
+	     cinfo->JFIF_major_version, cinfo->JFIF_minor_version,
+	     cinfo->X_density, cinfo->Y_density, cinfo->density_unit);
+    /* Validate thumbnail dimensions and issue appropriate messages */
+    if (GETJOCTET(data[12]) | GETJOCTET(data[13]))
+      TRACEMS2(cinfo, 1, JTRC_JFIF_THUMBNAIL,
+	       GETJOCTET(data[12]), GETJOCTET(data[13]));
+    totallen -= APP0_DATA_LEN;
+    if (totallen !=
+	((INT32)GETJOCTET(data[12]) * (INT32)GETJOCTET(data[13]) * (INT32) 3))
+      TRACEMS1(cinfo, 1, JTRC_JFIF_BADTHUMBNAILSIZE, (int) totallen);
+  } else if (datalen >= 6 &&
+      GETJOCTET(data[0]) == 0x4A &&
+      GETJOCTET(data[1]) == 0x46 &&
+      GETJOCTET(data[2]) == 0x58 &&
+      GETJOCTET(data[3]) == 0x58 &&
+      GETJOCTET(data[4]) == 0) {
+    /* Found JFIF "JFXX" extension APP0 marker */
+    /* The library doesn't actually do anything with these,
+     * but we try to produce a helpful trace message.
+     */
+    switch (GETJOCTET(data[5])) {
+    case 0x10:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_JPEG, (int) totallen);
+      break;
+    case 0x11:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_PALETTE, (int) totallen);
+      break;
+    case 0x13:
+      TRACEMS1(cinfo, 1, JTRC_THUMB_RGB, (int) totallen);
+      break;
+    default:
+      TRACEMS2(cinfo, 1, JTRC_JFIF_EXTENSION,
+	       GETJOCTET(data[5]), (int) totallen);
+      break;
+    }
+  } else {
+    /* Start of APP0 does not match "JFIF" or "JFXX", or too short */
+    TRACEMS1(cinfo, 1, JTRC_APP0, (int) totallen);
+  }
+}
+
+
+LOCAL(void)
+examine_app14 (j_decompress_ptr cinfo, JOCTET FAR * data,
+	       unsigned int datalen, INT32 remaining)
+/* Examine first few bytes from an APP14.
+ * Take appropriate action if it is an Adobe marker.
+ * datalen is # of bytes at data[], remaining is length of rest of marker data.
+ */
+{
+  unsigned int version, flags0, flags1, transform;
+
+  if (datalen >= APP14_DATA_LEN &&
+      GETJOCTET(data[0]) == 0x41 &&
+      GETJOCTET(data[1]) == 0x64 &&
+      GETJOCTET(data[2]) == 0x6F &&
+      GETJOCTET(data[3]) == 0x62 &&
+      GETJOCTET(data[4]) == 0x65) {
+    /* Found Adobe APP14 marker */
+    version = (GETJOCTET(data[5]) << 8) + GETJOCTET(data[6]);
+    flags0 = (GETJOCTET(data[7]) << 8) + GETJOCTET(data[8]);
+    flags1 = (GETJOCTET(data[9]) << 8) + GETJOCTET(data[10]);
+    transform = GETJOCTET(data[11]);
+    TRACEMS4(cinfo, 1, JTRC_ADOBE, version, flags0, flags1, transform);
+    cinfo->saw_Adobe_marker = TRUE;
+    cinfo->Adobe_transform = (UINT8) transform;
+  } else {
+    /* Start of APP14 does not match "Adobe", or too short */
+    TRACEMS1(cinfo, 1, JTRC_APP14, (int) (datalen + remaining));
+  }
+}
+
+
+METHODDEF(boolean)
+get_interesting_appn (j_decompress_ptr cinfo)
+/* Process an APP0 or APP14 marker without saving it */
+{
+  INT32 length;
+  JOCTET b[APPN_DATA_LEN];
+  unsigned int i, numtoread;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+
+  /* get the interesting part of the marker data */
+  if (length >= APPN_DATA_LEN)
+    numtoread = APPN_DATA_LEN;
+  else if (length > 0)
+    numtoread = (unsigned int) length;
+  else
+    numtoread = 0;
+  for (i = 0; i < numtoread; i++)
+    INPUT_BYTE(cinfo, b[i], return FALSE);
+  length -= numtoread;
+
+  /* process it */
+  switch (cinfo->unread_marker) {
+  case M_APP0:
+    examine_app0(cinfo, (JOCTET FAR *) b, numtoread, length);
+    break;
+  case M_APP14:
+    examine_app14(cinfo, (JOCTET FAR *) b, numtoread, length);
+    break;
+  default:
+    /* can't get here unless jpeg_save_markers chooses wrong processor */
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, cinfo->unread_marker);
+    break;
+  }
+
+  /* skip any remaining data -- could be lots */
+  INPUT_SYNC(cinfo);
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+
+#ifdef SAVE_MARKERS_SUPPORTED
+
+METHODDEF(boolean)
+save_marker (j_decompress_ptr cinfo)
+/* Save an APPn or COM marker into the marker list */
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  jpeg_saved_marker_ptr cur_marker = marker->cur_marker;
+  unsigned int bytes_read, data_length;
+  JOCTET FAR * data;
+  INT32 length = 0;
+  INPUT_VARS(cinfo);
+
+  if (cur_marker == NULL) {
+    /* begin reading a marker */
+    INPUT_2BYTES(cinfo, length, return FALSE);
+    length -= 2;
+    if (length >= 0) {		/* watch out for bogus length word */
+      /* figure out how much we want to save */
+      unsigned int limit;
+      if (cinfo->unread_marker == (int) M_COM)
+	limit = marker->length_limit_COM;
+      else
+	limit = marker->length_limit_APPn[cinfo->unread_marker - (int) M_APP0];
+      if ((unsigned int) length < limit)
+	limit = (unsigned int) length;
+      /* allocate and initialize the marker item */
+      cur_marker = (jpeg_saved_marker_ptr)
+	(*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				    SIZEOF(struct jpeg_marker_struct) + limit);
+      cur_marker->next = NULL;
+      cur_marker->marker = (UINT8) cinfo->unread_marker;
+      cur_marker->original_length = (unsigned int) length;
+      cur_marker->data_length = limit;
+      /* data area is just beyond the jpeg_marker_struct */
+      data = cur_marker->data = (JOCTET FAR *) (cur_marker + 1);
+      marker->cur_marker = cur_marker;
+      marker->bytes_read = 0;
+      bytes_read = 0;
+      data_length = limit;
+    } else {
+      /* deal with bogus length word */
+      bytes_read = data_length = 0;
+      data = NULL;
+    }
+  } else {
+    /* resume reading a marker */
+    bytes_read = marker->bytes_read;
+    data_length = cur_marker->data_length;
+    data = cur_marker->data + bytes_read;
+  }
+
+  while (bytes_read < data_length) {
+    INPUT_SYNC(cinfo);		/* move the restart point to here */
+    marker->bytes_read = bytes_read;
+    /* If there's not at least one byte in buffer, suspend */
+    MAKE_BYTE_AVAIL(cinfo, return FALSE);
+    /* Copy bytes with reasonable rapidity */
+    while (bytes_read < data_length && bytes_in_buffer > 0) {
+      *data++ = *next_input_byte++;
+      bytes_in_buffer--;
+      bytes_read++;
+    }
+  }
+
+  /* Done reading what we want to read */
+  if (cur_marker != NULL) {	/* will be NULL if bogus length word */
+    /* Add new marker to end of list */
+    if (cinfo->marker_list == NULL) {
+      cinfo->marker_list = cur_marker;
+    } else {
+      jpeg_saved_marker_ptr prev = cinfo->marker_list;
+      while (prev->next != NULL)
+	prev = prev->next;
+      prev->next = cur_marker;
+    }
+    /* Reset pointer & calc remaining data length */
+    data = cur_marker->data;
+    length = cur_marker->original_length - data_length;
+  }
+  /* Reset to initial state for next marker */
+  marker->cur_marker = NULL;
+
+  /* Process the marker if interesting; else just make a generic trace msg */
+  switch (cinfo->unread_marker) {
+  case M_APP0:
+    examine_app0(cinfo, data, data_length, length);
+    break;
+  case M_APP14:
+    examine_app14(cinfo, data, data_length, length);
+    break;
+  default:
+    TRACEMS2(cinfo, 1, JTRC_MISC_MARKER, cinfo->unread_marker,
+	     (int) (data_length + length));
+    break;
+  }
+
+  /* skip any remaining data -- could be lots */
+  INPUT_SYNC(cinfo);		/* do before skip_input_data */
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+#endif /* SAVE_MARKERS_SUPPORTED */
+
+
+METHODDEF(boolean)
+skip_variable (j_decompress_ptr cinfo)
+/* Skip over an unknown or uninteresting variable-length marker */
+{
+  INT32 length;
+  INPUT_VARS(cinfo);
+
+  INPUT_2BYTES(cinfo, length, return FALSE);
+  length -= 2;
+  
+  TRACEMS2(cinfo, 1, JTRC_MISC_MARKER, cinfo->unread_marker, (int) length);
+
+  INPUT_SYNC(cinfo);		/* do before skip_input_data */
+  if (length > 0)
+    (*cinfo->src->skip_input_data) (cinfo, (long) length);
+
+  return TRUE;
+}
+
+
+/*
+ * Find the next JPEG marker, save it in cinfo->unread_marker.
+ * Returns FALSE if had to suspend before reaching a marker;
+ * in that case cinfo->unread_marker is unchanged.
+ *
+ * Note that the result might not be a valid marker code,
+ * but it will never be 0 or FF.
+ */
+
+LOCAL(boolean)
+next_marker (j_decompress_ptr cinfo)
+{
+  int c;
+  INPUT_VARS(cinfo);
+
+  for (;;) {
+    INPUT_BYTE(cinfo, c, return FALSE);
+    /* Skip any non-FF bytes.
+     * This may look a bit inefficient, but it will not occur in a valid file.
+     * We sync after each discarded byte so that a suspending data source
+     * can discard the byte from its buffer.
+     */
+    while (c != 0xFF) {
+      cinfo->marker->discarded_bytes++;
+      INPUT_SYNC(cinfo);
+      INPUT_BYTE(cinfo, c, return FALSE);
+    }
+    /* This loop swallows any duplicate FF bytes.  Extra FFs are legal as
+     * pad bytes, so don't count them in discarded_bytes.  We assume there
+     * will not be so many consecutive FF bytes as to overflow a suspending
+     * data source's input buffer.
+     */
+    do {
+      INPUT_BYTE(cinfo, c, return FALSE);
+    } while (c == 0xFF);
+    if (c != 0)
+      break;			/* found a valid marker, exit loop */
+    /* Reach here if we found a stuffed-zero data sequence (FF/00).
+     * Discard it and loop back to try again.
+     */
+    cinfo->marker->discarded_bytes += 2;
+    INPUT_SYNC(cinfo);
+  }
+
+  if (cinfo->marker->discarded_bytes != 0) {
+    WARNMS2(cinfo, JWRN_EXTRANEOUS_DATA, cinfo->marker->discarded_bytes, c);
+    cinfo->marker->discarded_bytes = 0;
+  }
+
+  cinfo->unread_marker = c;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+LOCAL(boolean)
+first_marker (j_decompress_ptr cinfo)
+/* Like next_marker, but used to obtain the initial SOI marker. */
+/* For this marker, we do not allow preceding garbage or fill; otherwise,
+ * we might well scan an entire input file before realizing it ain't JPEG.
+ * If an application wants to process non-JFIF files, it must seek to the
+ * SOI before calling the JPEG library.
+ */
+{
+  int c, c2;
+  INPUT_VARS(cinfo);
+
+  INPUT_BYTE(cinfo, c, return FALSE);
+  INPUT_BYTE(cinfo, c2, return FALSE);
+  if (c != 0xFF || c2 != (int) M_SOI)
+    ERREXIT2(cinfo, JERR_NO_SOI, c, c2);
+
+  cinfo->unread_marker = c2;
+
+  INPUT_SYNC(cinfo);
+  return TRUE;
+}
+
+
+/*
+ * Read markers until SOS or EOI.
+ *
+ * Returns same codes as are defined for jpeg_consume_input:
+ * JPEG_SUSPENDED, JPEG_REACHED_SOS, or JPEG_REACHED_EOI.
+ *
+ * Note: This function may return a pseudo SOS marker (with zero
+ * component number) for treat by input controller's consume_input.
+ * consume_input itself should filter out (skip) the pseudo marker
+ * after processing for the caller.
+ */
+
+METHODDEF(int)
+read_markers (j_decompress_ptr cinfo)
+{
+  /* Outer loop repeats once for each marker. */
+  for (;;) {
+    /* Collect the marker proper, unless we already did. */
+    /* NB: first_marker() enforces the requirement that SOI appear first. */
+    if (cinfo->unread_marker == 0) {
+      if (! cinfo->marker->saw_SOI) {
+	if (! first_marker(cinfo))
+	  return JPEG_SUSPENDED;
+      } else {
+	if (! next_marker(cinfo))
+	  return JPEG_SUSPENDED;
+      }
+    }
+    /* At this point cinfo->unread_marker contains the marker code and the
+     * input point is just past the marker proper, but before any parameters.
+     * A suspension will cause us to return with this state still true.
+     */
+    switch (cinfo->unread_marker) {
+    case M_SOI:
+      if (! get_soi(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF0:		/* Baseline */
+      if (! get_sof(cinfo, TRUE, FALSE, FALSE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF1:		/* Extended sequential, Huffman */
+      if (! get_sof(cinfo, FALSE, FALSE, FALSE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF2:		/* Progressive, Huffman */
+      if (! get_sof(cinfo, FALSE, TRUE, FALSE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF9:		/* Extended sequential, arithmetic */
+      if (! get_sof(cinfo, FALSE, FALSE, TRUE))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_SOF10:		/* Progressive, arithmetic */
+      if (! get_sof(cinfo, FALSE, TRUE, TRUE))
+	return JPEG_SUSPENDED;
+      break;
+
+    /* Currently unsupported SOFn types */
+    case M_SOF3:		/* Lossless, Huffman */
+    case M_SOF5:		/* Differential sequential, Huffman */
+    case M_SOF6:		/* Differential progressive, Huffman */
+    case M_SOF7:		/* Differential lossless, Huffman */
+    case M_JPG:			/* Reserved for JPEG extensions */
+    case M_SOF11:		/* Lossless, arithmetic */
+    case M_SOF13:		/* Differential sequential, arithmetic */
+    case M_SOF14:		/* Differential progressive, arithmetic */
+    case M_SOF15:		/* Differential lossless, arithmetic */
+      ERREXIT1(cinfo, JERR_SOF_UNSUPPORTED, cinfo->unread_marker);
+      break;
+
+    case M_SOS:
+      if (! get_sos(cinfo))
+	return JPEG_SUSPENDED;
+      cinfo->unread_marker = 0;	/* processed the marker */
+      return JPEG_REACHED_SOS;
+    
+    case M_EOI:
+      TRACEMS(cinfo, 1, JTRC_EOI);
+      cinfo->unread_marker = 0;	/* processed the marker */
+      return JPEG_REACHED_EOI;
+      
+    case M_DAC:
+      if (! get_dac(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DHT:
+      if (! get_dht(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DQT:
+      if (! get_dqt(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_DRI:
+      if (! get_dri(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_APP0:
+    case M_APP1:
+    case M_APP2:
+    case M_APP3:
+    case M_APP4:
+    case M_APP5:
+    case M_APP6:
+    case M_APP7:
+    case M_APP8:
+    case M_APP9:
+    case M_APP10:
+    case M_APP11:
+    case M_APP12:
+    case M_APP13:
+    case M_APP14:
+    case M_APP15:
+      if (! (*((my_marker_ptr) cinfo->marker)->process_APPn[
+		cinfo->unread_marker - (int) M_APP0]) (cinfo))
+	return JPEG_SUSPENDED;
+      break;
+      
+    case M_COM:
+      if (! (*((my_marker_ptr) cinfo->marker)->process_COM) (cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    case M_RST0:		/* these are all parameterless */
+    case M_RST1:
+    case M_RST2:
+    case M_RST3:
+    case M_RST4:
+    case M_RST5:
+    case M_RST6:
+    case M_RST7:
+    case M_TEM:
+      TRACEMS1(cinfo, 1, JTRC_PARMLESS_MARKER, cinfo->unread_marker);
+      break;
+
+    case M_DNL:			/* Ignore DNL ... perhaps the wrong thing */
+      if (! skip_variable(cinfo))
+	return JPEG_SUSPENDED;
+      break;
+
+    default:			/* must be DHP, EXP, JPGn, or RESn */
+      /* For now, we treat the reserved markers as fatal errors since they are
+       * likely to be used to signal incompatible JPEG Part 3 extensions.
+       * Once the JPEG 3 version-number marker is well defined, this code
+       * ought to change!
+       */
+      ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, cinfo->unread_marker);
+      break;
+    }
+    /* Successfully processed marker, so reset state variable */
+    cinfo->unread_marker = 0;
+  } /* end loop */
+}
+
+
+/*
+ * Read a restart marker, which is expected to appear next in the datastream;
+ * if the marker is not there, take appropriate recovery action.
+ * Returns FALSE if suspension is required.
+ *
+ * This is called by the entropy decoder after it has read an appropriate
+ * number of MCUs.  cinfo->unread_marker may be nonzero if the entropy decoder
+ * has already read a marker from the data source.  Under normal conditions
+ * cinfo->unread_marker will be reset to 0 before returning; if not reset,
+ * it holds a marker which the decoder will be unable to read past.
+ */
+
+METHODDEF(boolean)
+read_restart_marker (j_decompress_ptr cinfo)
+{
+  /* Obtain a marker unless we already did. */
+  /* Note that next_marker will complain if it skips any data. */
+  if (cinfo->unread_marker == 0) {
+    if (! next_marker(cinfo))
+      return FALSE;
+  }
+
+  if (cinfo->unread_marker ==
+      ((int) M_RST0 + cinfo->marker->next_restart_num)) {
+    /* Normal case --- swallow the marker and let entropy decoder continue */
+    TRACEMS1(cinfo, 3, JTRC_RST, cinfo->marker->next_restart_num);
+    cinfo->unread_marker = 0;
+  } else {
+    /* Uh-oh, the restart markers have been messed up. */
+    /* Let the data source manager determine how to resync. */
+    if (! (*cinfo->src->resync_to_restart) (cinfo,
+					    cinfo->marker->next_restart_num))
+      return FALSE;
+  }
+
+  /* Update next-restart state */
+  cinfo->marker->next_restart_num = (cinfo->marker->next_restart_num + 1) & 7;
+
+  return TRUE;
+}
+
+
+/*
+ * This is the default resync_to_restart method for data source managers
+ * to use if they don't have any better approach.  Some data source managers
+ * may be able to back up, or may have additional knowledge about the data
+ * which permits a more intelligent recovery strategy; such managers would
+ * presumably supply their own resync method.
+ *
+ * read_restart_marker calls resync_to_restart if it finds a marker other than
+ * the restart marker it was expecting.  (This code is *not* used unless
+ * a nonzero restart interval has been declared.)  cinfo->unread_marker is
+ * the marker code actually found (might be anything, except 0 or FF).
+ * The desired restart marker number (0..7) is passed as a parameter.
+ * This routine is supposed to apply whatever error recovery strategy seems
+ * appropriate in order to position the input stream to the next data segment.
+ * Note that cinfo->unread_marker is treated as a marker appearing before
+ * the current data-source input point; usually it should be reset to zero
+ * before returning.
+ * Returns FALSE if suspension is required.
+ *
+ * This implementation is substantially constrained by wanting to treat the
+ * input as a data stream; this means we can't back up.  Therefore, we have
+ * only the following actions to work with:
+ *   1. Simply discard the marker and let the entropy decoder resume at next
+ *      byte of file.
+ *   2. Read forward until we find another marker, discarding intervening
+ *      data.  (In theory we could look ahead within the current bufferload,
+ *      without having to discard data if we don't find the desired marker.
+ *      This idea is not implemented here, in part because it makes behavior
+ *      dependent on buffer size and chance buffer-boundary positions.)
+ *   3. Leave the marker unread (by failing to zero cinfo->unread_marker).
+ *      This will cause the entropy decoder to process an empty data segment,
+ *      inserting dummy zeroes, and then we will reprocess the marker.
+ *
+ * #2 is appropriate if we think the desired marker lies ahead, while #3 is
+ * appropriate if the found marker is a future restart marker (indicating
+ * that we have missed the desired restart marker, probably because it got
+ * corrupted).
+ * We apply #2 or #3 if the found marker is a restart marker no more than
+ * two counts behind or ahead of the expected one.  We also apply #2 if the
+ * found marker is not a legal JPEG marker code (it's certainly bogus data).
+ * If the found marker is a restart marker more than 2 counts away, we do #1
+ * (too much risk that the marker is erroneous; with luck we will be able to
+ * resync at some future point).
+ * For any valid non-restart JPEG marker, we apply #3.  This keeps us from
+ * overrunning the end of a scan.  An implementation limited to single-scan
+ * files might find it better to apply #2 for markers other than EOI, since
+ * any other marker would have to be bogus data in that case.
+ */
+
+GLOBAL(boolean)
+jpeg_resync_to_restart (j_decompress_ptr cinfo, int desired)
+{
+  int marker = cinfo->unread_marker;
+  int action = 1;
+  
+  /* Always put up a warning. */
+  WARNMS2(cinfo, JWRN_MUST_RESYNC, marker, desired);
+  
+  /* Outer loop handles repeated decision after scanning forward. */
+  for (;;) {
+    if (marker < (int) M_SOF0)
+      action = 2;		/* invalid marker */
+    else if (marker < (int) M_RST0 || marker > (int) M_RST7)
+      action = 3;		/* valid non-restart marker */
+    else {
+      if (marker == ((int) M_RST0 + ((desired+1) & 7)) ||
+	  marker == ((int) M_RST0 + ((desired+2) & 7)))
+	action = 3;		/* one of the next two expected restarts */
+      else if (marker == ((int) M_RST0 + ((desired-1) & 7)) ||
+	       marker == ((int) M_RST0 + ((desired-2) & 7)))
+	action = 2;		/* a prior restart, so advance */
+      else
+	action = 1;		/* desired restart or too far away */
+    }
+    TRACEMS2(cinfo, 4, JTRC_RECOVERY_ACTION, marker, action);
+    switch (action) {
+    case 1:
+      /* Discard marker and let entropy decoder resume processing. */
+      cinfo->unread_marker = 0;
+      return TRUE;
+    case 2:
+      /* Scan to the next marker, and repeat the decision loop. */
+      if (! next_marker(cinfo))
+	return FALSE;
+      marker = cinfo->unread_marker;
+      break;
+    case 3:
+      /* Return without advancing past this marker. */
+      /* Entropy decoder will be forced to process an empty segment. */
+      return TRUE;
+    }
+  } /* end loop */
+}
+
+
+/*
+ * Reset marker processing state to begin a fresh datastream.
+ */
+
+METHODDEF(void)
+reset_marker_reader (j_decompress_ptr cinfo)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  cinfo->comp_info = NULL;		/* until allocated by get_sof */
+  cinfo->input_scan_number = 0;		/* no SOS seen yet */
+  cinfo->unread_marker = 0;		/* no pending marker */
+  marker->pub.saw_SOI = FALSE;		/* set internal state too */
+  marker->pub.saw_SOF = FALSE;
+  marker->pub.discarded_bytes = 0;
+  marker->cur_marker = NULL;
+}
+
+
+/*
+ * Initialize the marker reader module.
+ * This is called only once, when the decompression object is created.
+ */
+
+GLOBAL(void)
+jinit_marker_reader (j_decompress_ptr cinfo)
+{
+  my_marker_ptr marker;
+  int i;
+
+  /* Create subobject in permanent pool */
+  marker = (my_marker_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_PERMANENT,
+				SIZEOF(my_marker_reader));
+  cinfo->marker = (struct jpeg_marker_reader *) marker;
+  /* Initialize public method pointers */
+  marker->pub.reset_marker_reader = reset_marker_reader;
+  marker->pub.read_markers = read_markers;
+  marker->pub.read_restart_marker = read_restart_marker;
+  /* Initialize COM/APPn processing.
+   * By default, we examine and then discard APP0 and APP14,
+   * but simply discard COM and all other APPn.
+   */
+  marker->process_COM = skip_variable;
+  marker->length_limit_COM = 0;
+  for (i = 0; i < 16; i++) {
+    marker->process_APPn[i] = skip_variable;
+    marker->length_limit_APPn[i] = 0;
+  }
+  marker->process_APPn[0] = get_interesting_appn;
+  marker->process_APPn[14] = get_interesting_appn;
+  /* Reset marker processing state */
+  reset_marker_reader(cinfo);
+}
+
+
+/*
+ * Control saving of COM and APPn markers into marker_list.
+ */
+
+#ifdef SAVE_MARKERS_SUPPORTED
+
+GLOBAL(void)
+jpeg_save_markers (j_decompress_ptr cinfo, int marker_code,
+		   unsigned int length_limit)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+  long maxlength;
+  jpeg_marker_parser_method processor;
+
+  /* Length limit mustn't be larger than what we can allocate
+   * (should only be a concern in a 16-bit environment).
+   */
+  maxlength = cinfo->mem->max_alloc_chunk - SIZEOF(struct jpeg_marker_struct);
+  if (((long) length_limit) > maxlength)
+    length_limit = (unsigned int) maxlength;
+
+  /* Choose processor routine to use.
+   * APP0/APP14 have special requirements.
+   */
+  if (length_limit) {
+    processor = save_marker;
+    /* If saving APP0/APP14, save at least enough for our internal use. */
+    if (marker_code == (int) M_APP0 && length_limit < APP0_DATA_LEN)
+      length_limit = APP0_DATA_LEN;
+    else if (marker_code == (int) M_APP14 && length_limit < APP14_DATA_LEN)
+      length_limit = APP14_DATA_LEN;
+  } else {
+    processor = skip_variable;
+    /* If discarding APP0/APP14, use our regular on-the-fly processor. */
+    if (marker_code == (int) M_APP0 || marker_code == (int) M_APP14)
+      processor = get_interesting_appn;
+  }
+
+  if (marker_code == (int) M_COM) {
+    marker->process_COM = processor;
+    marker->length_limit_COM = length_limit;
+  } else if (marker_code >= (int) M_APP0 && marker_code <= (int) M_APP15) {
+    marker->process_APPn[marker_code - (int) M_APP0] = processor;
+    marker->length_limit_APPn[marker_code - (int) M_APP0] = length_limit;
+  } else
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, marker_code);
+}
+
+#endif /* SAVE_MARKERS_SUPPORTED */
+
+
+/*
+ * Install a special processing method for COM or APPn markers.
+ */
+
+GLOBAL(void)
+jpeg_set_marker_processor (j_decompress_ptr cinfo, int marker_code,
+			   jpeg_marker_parser_method routine)
+{
+  my_marker_ptr marker = (my_marker_ptr) cinfo->marker;
+
+  if (marker_code == (int) M_COM)
+    marker->process_COM = routine;
+  else if (marker_code >= (int) M_APP0 && marker_code <= (int) M_APP15)
+    marker->process_APPn[marker_code - (int) M_APP0] = routine;
+  else
+    ERREXIT1(cinfo, JERR_UNKNOWN_MARKER, marker_code);
+}

Added: trunk/code/jpeg-8c/jdmaster.c
===================================================================
--- trunk/code/jpeg-8c/jdmaster.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdmaster.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,533 @@
+/*
+ * jdmaster.c
+ *
+ * Copyright (C) 1991-1997, Thomas G. Lane.
+ * Modified 2002-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains master control logic for the JPEG decompressor.
+ * These routines are concerned with selecting the modules to be executed
+ * and with determining the number of passes and the work to be done in each
+ * pass.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private state */
+
+typedef struct {
+  struct jpeg_decomp_master pub; /* public fields */
+
+  int pass_number;		/* # of passes completed */
+
+  boolean using_merged_upsample; /* TRUE if using merged upsample/cconvert */
+
+  /* Saved references to initialized quantizer modules,
+   * in case we need to switch modes.
+   */
+  struct jpeg_color_quantizer * quantizer_1pass;
+  struct jpeg_color_quantizer * quantizer_2pass;
+} my_decomp_master;
+
+typedef my_decomp_master * my_master_ptr;
+
+
+/*
+ * Determine whether merged upsample/color conversion should be used.
+ * CRUCIAL: this must match the actual capabilities of jdmerge.c!
+ */
+
+LOCAL(boolean)
+use_merged_upsample (j_decompress_ptr cinfo)
+{
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+  /* Merging is the equivalent of plain box-filter upsampling */
+  if (cinfo->do_fancy_upsampling || cinfo->CCIR601_sampling)
+    return FALSE;
+  /* jdmerge.c only supports YCC=>RGB color conversion */
+  if (cinfo->jpeg_color_space != JCS_YCbCr || cinfo->num_components != 3 ||
+      cinfo->out_color_space != JCS_RGB ||
+      cinfo->out_color_components != RGB_PIXELSIZE)
+    return FALSE;
+  /* and it only handles 2h1v or 2h2v sampling ratios */
+  if (cinfo->comp_info[0].h_samp_factor != 2 ||
+      cinfo->comp_info[1].h_samp_factor != 1 ||
+      cinfo->comp_info[2].h_samp_factor != 1 ||
+      cinfo->comp_info[0].v_samp_factor >  2 ||
+      cinfo->comp_info[1].v_samp_factor != 1 ||
+      cinfo->comp_info[2].v_samp_factor != 1)
+    return FALSE;
+  /* furthermore, it doesn't work if we've scaled the IDCTs differently */
+  if (cinfo->comp_info[0].DCT_h_scaled_size != cinfo->min_DCT_h_scaled_size ||
+      cinfo->comp_info[1].DCT_h_scaled_size != cinfo->min_DCT_h_scaled_size ||
+      cinfo->comp_info[2].DCT_h_scaled_size != cinfo->min_DCT_h_scaled_size ||
+      cinfo->comp_info[0].DCT_v_scaled_size != cinfo->min_DCT_v_scaled_size ||
+      cinfo->comp_info[1].DCT_v_scaled_size != cinfo->min_DCT_v_scaled_size ||
+      cinfo->comp_info[2].DCT_v_scaled_size != cinfo->min_DCT_v_scaled_size)
+    return FALSE;
+  /* ??? also need to test for upsample-time rescaling, when & if supported */
+  return TRUE;			/* by golly, it'll work... */
+#else
+  return FALSE;
+#endif
+}
+
+
+/*
+ * Compute output image dimensions and related values.
+ * NOTE: this is exported for possible use by application.
+ * Hence it mustn't do anything that can't be done twice.
+ * Also note that it may be called before the master module is initialized!
+ */
+
+GLOBAL(void)
+jpeg_calc_output_dimensions (j_decompress_ptr cinfo)
+/* Do computations that are needed before master selection phase.
+ * This function is used for full decompression.
+ */
+{
+#ifdef IDCT_SCALING_SUPPORTED
+  int ci;
+  jpeg_component_info *compptr;
+#endif
+
+  /* Prevent application from calling me at wrong times */
+  if (cinfo->global_state != DSTATE_READY)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  /* Compute core output image dimensions and DCT scaling choices. */
+  jpeg_core_output_dimensions(cinfo);
+
+#ifdef IDCT_SCALING_SUPPORTED
+
+  /* In selecting the actual DCT scaling for each component, we try to
+   * scale up the chroma components via IDCT scaling rather than upsampling.
+   * This saves time if the upsampler gets to use 1:1 scaling.
+   * Note this code adapts subsampling ratios which are powers of 2.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    int ssize = 1;
+    while (cinfo->min_DCT_h_scaled_size * ssize <=
+	   (cinfo->do_fancy_upsampling ? DCTSIZE : DCTSIZE / 2) &&
+	   (cinfo->max_h_samp_factor % (compptr->h_samp_factor * ssize * 2)) == 0) {
+      ssize = ssize * 2;
+    }
+    compptr->DCT_h_scaled_size = cinfo->min_DCT_h_scaled_size * ssize;
+    ssize = 1;
+    while (cinfo->min_DCT_v_scaled_size * ssize <=
+	   (cinfo->do_fancy_upsampling ? DCTSIZE : DCTSIZE / 2) &&
+	   (cinfo->max_v_samp_factor % (compptr->v_samp_factor * ssize * 2)) == 0) {
+      ssize = ssize * 2;
+    }
+    compptr->DCT_v_scaled_size = cinfo->min_DCT_v_scaled_size * ssize;
+
+    /* We don't support IDCT ratios larger than 2. */
+    if (compptr->DCT_h_scaled_size > compptr->DCT_v_scaled_size * 2)
+	compptr->DCT_h_scaled_size = compptr->DCT_v_scaled_size * 2;
+    else if (compptr->DCT_v_scaled_size > compptr->DCT_h_scaled_size * 2)
+	compptr->DCT_v_scaled_size = compptr->DCT_h_scaled_size * 2;
+  }
+
+  /* Recompute downsampled dimensions of components;
+   * application needs to know these if using raw downsampled data.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Size in samples, after IDCT scaling */
+    compptr->downsampled_width = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_width *
+		    (long) (compptr->h_samp_factor * compptr->DCT_h_scaled_size),
+		    (long) (cinfo->max_h_samp_factor * cinfo->block_size));
+    compptr->downsampled_height = (JDIMENSION)
+      jdiv_round_up((long) cinfo->image_height *
+		    (long) (compptr->v_samp_factor * compptr->DCT_v_scaled_size),
+		    (long) (cinfo->max_v_samp_factor * cinfo->block_size));
+  }
+
+#endif /* IDCT_SCALING_SUPPORTED */
+
+  /* Report number of components in selected colorspace. */
+  /* Probably this should be in the color conversion module... */
+  switch (cinfo->out_color_space) {
+  case JCS_GRAYSCALE:
+    cinfo->out_color_components = 1;
+    break;
+  case JCS_RGB:
+#if RGB_PIXELSIZE != 3
+    cinfo->out_color_components = RGB_PIXELSIZE;
+    break;
+#endif /* else share code with YCbCr */
+  case JCS_YCbCr:
+    cinfo->out_color_components = 3;
+    break;
+  case JCS_CMYK:
+  case JCS_YCCK:
+    cinfo->out_color_components = 4;
+    break;
+  default:			/* else must be same colorspace as in file */
+    cinfo->out_color_components = cinfo->num_components;
+    break;
+  }
+  cinfo->output_components = (cinfo->quantize_colors ? 1 :
+			      cinfo->out_color_components);
+
+  /* See if upsampler will want to emit more than one row at a time */
+  if (use_merged_upsample(cinfo))
+    cinfo->rec_outbuf_height = cinfo->max_v_samp_factor;
+  else
+    cinfo->rec_outbuf_height = 1;
+}
+
+
+/*
+ * Several decompression processes need to range-limit values to the range
+ * 0..MAXJSAMPLE; the input value may fall somewhat outside this range
+ * due to noise introduced by quantization, roundoff error, etc.  These
+ * processes are inner loops and need to be as fast as possible.  On most
+ * machines, particularly CPUs with pipelines or instruction prefetch,
+ * a (subscript-check-less) C table lookup
+ *		x = sample_range_limit[x];
+ * is faster than explicit tests
+ *		if (x < 0)  x = 0;
+ *		else if (x > MAXJSAMPLE)  x = MAXJSAMPLE;
+ * These processes all use a common table prepared by the routine below.
+ *
+ * For most steps we can mathematically guarantee that the initial value
+ * of x is within MAXJSAMPLE+1 of the legal range, so a table running from
+ * -(MAXJSAMPLE+1) to 2*MAXJSAMPLE+1 is sufficient.  But for the initial
+ * limiting step (just after the IDCT), a wildly out-of-range value is 
+ * possible if the input data is corrupt.  To avoid any chance of indexing
+ * off the end of memory and getting a bad-pointer trap, we perform the
+ * post-IDCT limiting thus:
+ *		x = range_limit[x & MASK];
+ * where MASK is 2 bits wider than legal sample data, ie 10 bits for 8-bit
+ * samples.  Under normal circumstances this is more than enough range and
+ * a correct output will be generated; with bogus input data the mask will
+ * cause wraparound, and we will safely generate a bogus-but-in-range output.
+ * For the post-IDCT step, we want to convert the data from signed to unsigned
+ * representation by adding CENTERJSAMPLE at the same time that we limit it.
+ * So the post-IDCT limiting table ends up looking like this:
+ *   CENTERJSAMPLE,CENTERJSAMPLE+1,...,MAXJSAMPLE,
+ *   MAXJSAMPLE (repeat 2*(MAXJSAMPLE+1)-CENTERJSAMPLE times),
+ *   0          (repeat 2*(MAXJSAMPLE+1)-CENTERJSAMPLE times),
+ *   0,1,...,CENTERJSAMPLE-1
+ * Negative inputs select values from the upper half of the table after
+ * masking.
+ *
+ * We can save some space by overlapping the start of the post-IDCT table
+ * with the simpler range limiting table.  The post-IDCT table begins at
+ * sample_range_limit + CENTERJSAMPLE.
+ *
+ * Note that the table is allocated in near data space on PCs; it's small
+ * enough and used often enough to justify this.
+ */
+
+LOCAL(void)
+prepare_range_limit_table (j_decompress_ptr cinfo)
+/* Allocate and fill in the sample_range_limit table */
+{
+  JSAMPLE * table;
+  int i;
+
+  table = (JSAMPLE *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+		(5 * (MAXJSAMPLE+1) + CENTERJSAMPLE) * SIZEOF(JSAMPLE));
+  table += (MAXJSAMPLE+1);	/* allow negative subscripts of simple table */
+  cinfo->sample_range_limit = table;
+  /* First segment of "simple" table: limit[x] = 0 for x < 0 */
+  MEMZERO(table - (MAXJSAMPLE+1), (MAXJSAMPLE+1) * SIZEOF(JSAMPLE));
+  /* Main part of "simple" table: limit[x] = x */
+  for (i = 0; i <= MAXJSAMPLE; i++)
+    table[i] = (JSAMPLE) i;
+  table += CENTERJSAMPLE;	/* Point to where post-IDCT table starts */
+  /* End of simple table, rest of first half of post-IDCT table */
+  for (i = CENTERJSAMPLE; i < 2*(MAXJSAMPLE+1); i++)
+    table[i] = MAXJSAMPLE;
+  /* Second half of post-IDCT table */
+  MEMZERO(table + (2 * (MAXJSAMPLE+1)),
+	  (2 * (MAXJSAMPLE+1) - CENTERJSAMPLE) * SIZEOF(JSAMPLE));
+  MEMCOPY(table + (4 * (MAXJSAMPLE+1) - CENTERJSAMPLE),
+	  cinfo->sample_range_limit, CENTERJSAMPLE * SIZEOF(JSAMPLE));
+}
+
+
+/*
+ * Master selection of decompression modules.
+ * This is done once at jpeg_start_decompress time.  We determine
+ * which modules will be used and give them appropriate initialization calls.
+ * We also initialize the decompressor input side to begin consuming data.
+ *
+ * Since jpeg_read_header has finished, we know what is in the SOF
+ * and (first) SOS markers.  We also have all the application parameter
+ * settings.
+ */
+
+LOCAL(void)
+master_selection (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+  boolean use_c_buffer;
+  long samplesperrow;
+  JDIMENSION jd_samplesperrow;
+
+  /* Initialize dimensions and other stuff */
+  jpeg_calc_output_dimensions(cinfo);
+  prepare_range_limit_table(cinfo);
+
+  /* Width of an output scanline must be representable as JDIMENSION. */
+  samplesperrow = (long) cinfo->output_width * (long) cinfo->out_color_components;
+  jd_samplesperrow = (JDIMENSION) samplesperrow;
+  if ((long) jd_samplesperrow != samplesperrow)
+    ERREXIT(cinfo, JERR_WIDTH_OVERFLOW);
+
+  /* Initialize my private state */
+  master->pass_number = 0;
+  master->using_merged_upsample = use_merged_upsample(cinfo);
+
+  /* Color quantizer selection */
+  master->quantizer_1pass = NULL;
+  master->quantizer_2pass = NULL;
+  /* No mode changes if not using buffered-image mode. */
+  if (! cinfo->quantize_colors || ! cinfo->buffered_image) {
+    cinfo->enable_1pass_quant = FALSE;
+    cinfo->enable_external_quant = FALSE;
+    cinfo->enable_2pass_quant = FALSE;
+  }
+  if (cinfo->quantize_colors) {
+    if (cinfo->raw_data_out)
+      ERREXIT(cinfo, JERR_NOTIMPL);
+    /* 2-pass quantizer only works in 3-component color space. */
+    if (cinfo->out_color_components != 3) {
+      cinfo->enable_1pass_quant = TRUE;
+      cinfo->enable_external_quant = FALSE;
+      cinfo->enable_2pass_quant = FALSE;
+      cinfo->colormap = NULL;
+    } else if (cinfo->colormap != NULL) {
+      cinfo->enable_external_quant = TRUE;
+    } else if (cinfo->two_pass_quantize) {
+      cinfo->enable_2pass_quant = TRUE;
+    } else {
+      cinfo->enable_1pass_quant = TRUE;
+    }
+
+    if (cinfo->enable_1pass_quant) {
+#ifdef QUANT_1PASS_SUPPORTED
+      jinit_1pass_quantizer(cinfo);
+      master->quantizer_1pass = cinfo->cquantize;
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    }
+
+    /* We use the 2-pass code to map to external colormaps. */
+    if (cinfo->enable_2pass_quant || cinfo->enable_external_quant) {
+#ifdef QUANT_2PASS_SUPPORTED
+      jinit_2pass_quantizer(cinfo);
+      master->quantizer_2pass = cinfo->cquantize;
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    }
+    /* If both quantizers are initialized, the 2-pass one is left active;
+     * this is necessary for starting with quantization to an external map.
+     */
+  }
+
+  /* Post-processing: in particular, color conversion first */
+  if (! cinfo->raw_data_out) {
+    if (master->using_merged_upsample) {
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+      jinit_merged_upsampler(cinfo); /* does color conversion too */
+#else
+      ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif
+    } else {
+      jinit_color_deconverter(cinfo);
+      jinit_upsampler(cinfo);
+    }
+    jinit_d_post_controller(cinfo, cinfo->enable_2pass_quant);
+  }
+  /* Inverse DCT */
+  jinit_inverse_dct(cinfo);
+  /* Entropy decoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code)
+    jinit_arith_decoder(cinfo);
+  else {
+    jinit_huff_decoder(cinfo);
+  }
+
+  /* Initialize principal buffer controllers. */
+  use_c_buffer = cinfo->inputctl->has_multiple_scans || cinfo->buffered_image;
+  jinit_d_coef_controller(cinfo, use_c_buffer);
+
+  if (! cinfo->raw_data_out)
+    jinit_d_main_controller(cinfo, FALSE /* never need full buffer here */);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Initialize input side of decompressor to consume first scan. */
+  (*cinfo->inputctl->start_input_pass) (cinfo);
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+  /* If jpeg_start_decompress will read the whole file, initialize
+   * progress monitoring appropriately.  The input step is counted
+   * as one pass.
+   */
+  if (cinfo->progress != NULL && ! cinfo->buffered_image &&
+      cinfo->inputctl->has_multiple_scans) {
+    int nscans;
+    /* Estimate number of scans to set pass_limit. */
+    if (cinfo->progressive_mode) {
+      /* Arbitrarily estimate 2 interleaved DC scans + 3 AC scans/component. */
+      nscans = 2 + 3 * cinfo->num_components;
+    } else {
+      /* For a nonprogressive multiscan file, estimate 1 scan per component. */
+      nscans = cinfo->num_components;
+    }
+    cinfo->progress->pass_counter = 0L;
+    cinfo->progress->pass_limit = (long) cinfo->total_iMCU_rows * nscans;
+    cinfo->progress->completed_passes = 0;
+    cinfo->progress->total_passes = (cinfo->enable_2pass_quant ? 3 : 2);
+    /* Count the input pass as done */
+    master->pass_number++;
+  }
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+}
+
+
+/*
+ * Per-pass setup.
+ * This is called at the beginning of each output pass.  We determine which
+ * modules will be active during this pass and give them appropriate
+ * start_pass calls.  We also set is_dummy_pass to indicate whether this
+ * is a "real" output pass or a dummy pass for color quantization.
+ * (In the latter case, jdapistd.c will crank the pass to completion.)
+ */
+
+METHODDEF(void)
+prepare_for_output_pass (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  if (master->pub.is_dummy_pass) {
+#ifdef QUANT_2PASS_SUPPORTED
+    /* Final pass of 2-pass quantization */
+    master->pub.is_dummy_pass = FALSE;
+    (*cinfo->cquantize->start_pass) (cinfo, FALSE);
+    (*cinfo->post->start_pass) (cinfo, JBUF_CRANK_DEST);
+    (*cinfo->main->start_pass) (cinfo, JBUF_CRANK_DEST);
+#else
+    ERREXIT(cinfo, JERR_NOT_COMPILED);
+#endif /* QUANT_2PASS_SUPPORTED */
+  } else {
+    if (cinfo->quantize_colors && cinfo->colormap == NULL) {
+      /* Select new quantization method */
+      if (cinfo->two_pass_quantize && cinfo->enable_2pass_quant) {
+	cinfo->cquantize = master->quantizer_2pass;
+	master->pub.is_dummy_pass = TRUE;
+      } else if (cinfo->enable_1pass_quant) {
+	cinfo->cquantize = master->quantizer_1pass;
+      } else {
+	ERREXIT(cinfo, JERR_MODE_CHANGE);
+      }
+    }
+    (*cinfo->idct->start_pass) (cinfo);
+    (*cinfo->coef->start_output_pass) (cinfo);
+    if (! cinfo->raw_data_out) {
+      if (! master->using_merged_upsample)
+	(*cinfo->cconvert->start_pass) (cinfo);
+      (*cinfo->upsample->start_pass) (cinfo);
+      if (cinfo->quantize_colors)
+	(*cinfo->cquantize->start_pass) (cinfo, master->pub.is_dummy_pass);
+      (*cinfo->post->start_pass) (cinfo,
+	    (master->pub.is_dummy_pass ? JBUF_SAVE_AND_PASS : JBUF_PASS_THRU));
+      (*cinfo->main->start_pass) (cinfo, JBUF_PASS_THRU);
+    }
+  }
+
+  /* Set up progress monitor's pass info if present */
+  if (cinfo->progress != NULL) {
+    cinfo->progress->completed_passes = master->pass_number;
+    cinfo->progress->total_passes = master->pass_number +
+				    (master->pub.is_dummy_pass ? 2 : 1);
+    /* In buffered-image mode, we assume one more output pass if EOI not
+     * yet reached, but no more passes if EOI has been reached.
+     */
+    if (cinfo->buffered_image && ! cinfo->inputctl->eoi_reached) {
+      cinfo->progress->total_passes += (cinfo->enable_2pass_quant ? 2 : 1);
+    }
+  }
+}
+
+
+/*
+ * Finish up at end of an output pass.
+ */
+
+METHODDEF(void)
+finish_output_pass (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  if (cinfo->quantize_colors)
+    (*cinfo->cquantize->finish_pass) (cinfo);
+  master->pass_number++;
+}
+
+
+#ifdef D_MULTISCAN_FILES_SUPPORTED
+
+/*
+ * Switch to a new external colormap between output passes.
+ */
+
+GLOBAL(void)
+jpeg_new_colormap (j_decompress_ptr cinfo)
+{
+  my_master_ptr master = (my_master_ptr) cinfo->master;
+
+  /* Prevent application from calling me at wrong times */
+  if (cinfo->global_state != DSTATE_BUFIMAGE)
+    ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+
+  if (cinfo->quantize_colors && cinfo->enable_external_quant &&
+      cinfo->colormap != NULL) {
+    /* Select 2-pass quantizer for external colormap use */
+    cinfo->cquantize = master->quantizer_2pass;
+    /* Notify quantizer of colormap change */
+    (*cinfo->cquantize->new_color_map) (cinfo);
+    master->pub.is_dummy_pass = FALSE; /* just in case */
+  } else
+    ERREXIT(cinfo, JERR_MODE_CHANGE);
+}
+
+#endif /* D_MULTISCAN_FILES_SUPPORTED */
+
+
+/*
+ * Initialize master decompression control and select active modules.
+ * This is performed at the start of jpeg_start_decompress.
+ */
+
+GLOBAL(void)
+jinit_master_decompress (j_decompress_ptr cinfo)
+{
+  my_master_ptr master;
+
+  master = (my_master_ptr)
+      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(my_decomp_master));
+  cinfo->master = (struct jpeg_decomp_master *) master;
+  master->pub.prepare_for_output_pass = prepare_for_output_pass;
+  master->pub.finish_output_pass = finish_output_pass;
+
+  master->pub.is_dummy_pass = FALSE;
+
+  master_selection(cinfo);
+}

Added: trunk/code/jpeg-8c/jdmerge.c
===================================================================
--- trunk/code/jpeg-8c/jdmerge.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdmerge.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,400 @@
+/*
+ * jdmerge.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains code for merged upsampling/color conversion.
+ *
+ * This file combines functions from jdsample.c and jdcolor.c;
+ * read those files first to understand what's going on.
+ *
+ * When the chroma components are to be upsampled by simple replication
+ * (ie, box filtering), we can save some work in color conversion by
+ * calculating all the output pixels corresponding to a pair of chroma
+ * samples at one time.  In the conversion equations
+ *	R = Y           + K1 * Cr
+ *	G = Y + K2 * Cb + K3 * Cr
+ *	B = Y + K4 * Cb
+ * only the Y term varies among the group of pixels corresponding to a pair
+ * of chroma samples, so the rest of the terms can be calculated just once.
+ * At typical sampling ratios, this eliminates half or three-quarters of the
+ * multiplications needed for color conversion.
+ *
+ * This file currently provides implementations for the following cases:
+ *	YCbCr => RGB color conversion only.
+ *	Sampling ratios of 2h1v or 2h2v.
+ *	No scaling needed at upsample time.
+ *	Corner-aligned (non-CCIR601) sampling alignment.
+ * Other special cases could be added, but in most applications these are
+ * the only common cases.  (For uncommon cases we fall back on the more
+ * general code in jdsample.c and jdcolor.c.)
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+#ifdef UPSAMPLE_MERGING_SUPPORTED
+
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_upsampler pub;	/* public fields */
+
+  /* Pointer to routine to do actual upsampling/conversion of one row group */
+  JMETHOD(void, upmethod, (j_decompress_ptr cinfo,
+			   JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+			   JSAMPARRAY output_buf));
+
+  /* Private state for YCC->RGB conversion */
+  int * Cr_r_tab;		/* => table for Cr to R conversion */
+  int * Cb_b_tab;		/* => table for Cb to B conversion */
+  INT32 * Cr_g_tab;		/* => table for Cr to G conversion */
+  INT32 * Cb_g_tab;		/* => table for Cb to G conversion */
+
+  /* For 2:1 vertical sampling, we produce two output rows at a time.
+   * We need a "spare" row buffer to hold the second output row if the
+   * application provides just a one-row buffer; we also use the spare
+   * to discard the dummy last row if the image height is odd.
+   */
+  JSAMPROW spare_row;
+  boolean spare_full;		/* T if spare buffer is occupied */
+
+  JDIMENSION out_row_width;	/* samples per output row */
+  JDIMENSION rows_to_go;	/* counts rows remaining in image */
+} my_upsampler;
+
+typedef my_upsampler * my_upsample_ptr;
+
+#define SCALEBITS	16	/* speediest right-shift on some machines */
+#define ONE_HALF	((INT32) 1 << (SCALEBITS-1))
+#define FIX(x)		((INT32) ((x) * (1L<<SCALEBITS) + 0.5))
+
+
+/*
+ * Initialize tables for YCC->RGB colorspace conversion.
+ * This is taken directly from jdcolor.c; see that file for more info.
+ */
+
+LOCAL(void)
+build_ycc_rgb_table (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  int i;
+  INT32 x;
+  SHIFT_TEMPS
+
+  upsample->Cr_r_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  upsample->Cb_b_tab = (int *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(int));
+  upsample->Cr_g_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(INT32));
+  upsample->Cb_g_tab = (INT32 *)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				(MAXJSAMPLE+1) * SIZEOF(INT32));
+
+  for (i = 0, x = -CENTERJSAMPLE; i <= MAXJSAMPLE; i++, x++) {
+    /* i is the actual input pixel value, in the range 0..MAXJSAMPLE */
+    /* The Cb or Cr value we are thinking of is x = i - CENTERJSAMPLE */
+    /* Cr=>R value is nearest int to 1.40200 * x */
+    upsample->Cr_r_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.40200) * x + ONE_HALF, SCALEBITS);
+    /* Cb=>B value is nearest int to 1.77200 * x */
+    upsample->Cb_b_tab[i] = (int)
+		    RIGHT_SHIFT(FIX(1.77200) * x + ONE_HALF, SCALEBITS);
+    /* Cr=>G value is scaled-up -0.71414 * x */
+    upsample->Cr_g_tab[i] = (- FIX(0.71414)) * x;
+    /* Cb=>G value is scaled-up -0.34414 * x */
+    /* We also add in ONE_HALF so that need not do it in inner loop */
+    upsample->Cb_g_tab[i] = (- FIX(0.34414)) * x + ONE_HALF;
+  }
+}
+
+
+/*
+ * Initialize for an upsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_merged_upsample (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Mark the spare buffer empty */
+  upsample->spare_full = FALSE;
+  /* Initialize total-height counter for detecting bottom of image */
+  upsample->rows_to_go = cinfo->output_height;
+}
+
+
+/*
+ * Control routine to do upsampling (and color conversion).
+ *
+ * The control routine just handles the row buffering considerations.
+ */
+
+METHODDEF(void)
+merged_2v_upsample (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+/* 2:1 vertical sampling case: may need a spare row. */
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  JSAMPROW work_ptrs[2];
+  JDIMENSION num_rows;		/* number of rows returned to caller */
+
+  if (upsample->spare_full) {
+    /* If we have a spare row saved from a previous cycle, just return it. */
+    jcopy_sample_rows(& upsample->spare_row, 0, output_buf + *out_row_ctr, 0,
+		      1, upsample->out_row_width);
+    num_rows = 1;
+    upsample->spare_full = FALSE;
+  } else {
+    /* Figure number of rows to return to caller. */
+    num_rows = 2;
+    /* Not more than the distance to the end of the image. */
+    if (num_rows > upsample->rows_to_go)
+      num_rows = upsample->rows_to_go;
+    /* And not more than what the client can accept: */
+    out_rows_avail -= *out_row_ctr;
+    if (num_rows > out_rows_avail)
+      num_rows = out_rows_avail;
+    /* Create output pointer array for upsampler. */
+    work_ptrs[0] = output_buf[*out_row_ctr];
+    if (num_rows > 1) {
+      work_ptrs[1] = output_buf[*out_row_ctr + 1];
+    } else {
+      work_ptrs[1] = upsample->spare_row;
+      upsample->spare_full = TRUE;
+    }
+    /* Now do the upsampling. */
+    (*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr, work_ptrs);
+  }
+
+  /* Adjust counts */
+  *out_row_ctr += num_rows;
+  upsample->rows_to_go -= num_rows;
+  /* When the buffer is emptied, declare this input row group consumed */
+  if (! upsample->spare_full)
+    (*in_row_group_ctr)++;
+}
+
+
+METHODDEF(void)
+merged_1v_upsample (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+/* 1:1 vertical sampling case: much easier, never need a spare row. */
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Just do the upsampling. */
+  (*upsample->upmethod) (cinfo, input_buf, *in_row_group_ctr,
+			 output_buf + *out_row_ctr);
+  /* Adjust counts */
+  (*out_row_ctr)++;
+  (*in_row_group_ctr)++;
+}
+
+
+/*
+ * These are the routines invoked by the control routines to do
+ * the actual upsampling/conversion.  One row group is processed per call.
+ *
+ * Note: since we may be writing directly into application-supplied buffers,
+ * we have to be honest about the output width; we can't assume the buffer
+ * has been rounded up to an even width.
+ */
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
+ */
+
+METHODDEF(void)
+h2v1_merged_upsample (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+		      JSAMPARRAY output_buf)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register JSAMPROW outptr;
+  JSAMPROW inptr0, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register JSAMPLE * range_limit = cinfo->sample_range_limit;
+  int * Crrtab = upsample->Cr_r_tab;
+  int * Cbbtab = upsample->Cb_b_tab;
+  INT32 * Crgtab = upsample->Cr_g_tab;
+  INT32 * Cbgtab = upsample->Cb_g_tab;
+  SHIFT_TEMPS
+
+  inptr0 = input_buf[0][in_row_group_ctr];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr = output_buf[0];
+  /* Loop for each pair of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = GETJSAMPLE(*inptr1++);
+    cr = GETJSAMPLE(*inptr2++);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    /* Fetch 2 Y values and emit 2 pixels */
+    y  = GETJSAMPLE(*inptr0++);
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+    outptr += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr0++);
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+    outptr += RGB_PIXELSIZE;
+  }
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = GETJSAMPLE(*inptr1);
+    cr = GETJSAMPLE(*inptr2);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = GETJSAMPLE(*inptr0);
+    outptr[RGB_RED] =   range_limit[y + cred];
+    outptr[RGB_GREEN] = range_limit[y + cgreen];
+    outptr[RGB_BLUE] =  range_limit[y + cblue];
+  }
+}
+
+
+/*
+ * Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
+ */
+
+METHODDEF(void)
+h2v2_merged_upsample (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION in_row_group_ctr,
+		      JSAMPARRAY output_buf)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  register int y, cred, cgreen, cblue;
+  int cb, cr;
+  register JSAMPROW outptr0, outptr1;
+  JSAMPROW inptr00, inptr01, inptr1, inptr2;
+  JDIMENSION col;
+  /* copy these pointers into registers if possible */
+  register JSAMPLE * range_limit = cinfo->sample_range_limit;
+  int * Crrtab = upsample->Cr_r_tab;
+  int * Cbbtab = upsample->Cb_b_tab;
+  INT32 * Crgtab = upsample->Cr_g_tab;
+  INT32 * Cbgtab = upsample->Cb_g_tab;
+  SHIFT_TEMPS
+
+  inptr00 = input_buf[0][in_row_group_ctr*2];
+  inptr01 = input_buf[0][in_row_group_ctr*2 + 1];
+  inptr1 = input_buf[1][in_row_group_ctr];
+  inptr2 = input_buf[2][in_row_group_ctr];
+  outptr0 = output_buf[0];
+  outptr1 = output_buf[1];
+  /* Loop for each group of output pixels */
+  for (col = cinfo->output_width >> 1; col > 0; col--) {
+    /* Do the chroma part of the calculation */
+    cb = GETJSAMPLE(*inptr1++);
+    cr = GETJSAMPLE(*inptr2++);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    /* Fetch 4 Y values and emit 4 pixels */
+    y  = GETJSAMPLE(*inptr00++);
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+    outptr0 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr00++);
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+    outptr0 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr01++);
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+    outptr1 += RGB_PIXELSIZE;
+    y  = GETJSAMPLE(*inptr01++);
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+    outptr1 += RGB_PIXELSIZE;
+  }
+  /* If image width is odd, do the last output column separately */
+  if (cinfo->output_width & 1) {
+    cb = GETJSAMPLE(*inptr1);
+    cr = GETJSAMPLE(*inptr2);
+    cred = Crrtab[cr];
+    cgreen = (int) RIGHT_SHIFT(Cbgtab[cb] + Crgtab[cr], SCALEBITS);
+    cblue = Cbbtab[cb];
+    y  = GETJSAMPLE(*inptr00);
+    outptr0[RGB_RED] =   range_limit[y + cred];
+    outptr0[RGB_GREEN] = range_limit[y + cgreen];
+    outptr0[RGB_BLUE] =  range_limit[y + cblue];
+    y  = GETJSAMPLE(*inptr01);
+    outptr1[RGB_RED] =   range_limit[y + cred];
+    outptr1[RGB_GREEN] = range_limit[y + cgreen];
+    outptr1[RGB_BLUE] =  range_limit[y + cblue];
+  }
+}
+
+
+/*
+ * Module initialization routine for merged upsampling/color conversion.
+ *
+ * NB: this is called under the conditions determined by use_merged_upsample()
+ * in jdmaster.c.  That routine MUST correspond to the actual capabilities
+ * of this module; no safety checks are made here.
+ */
+
+GLOBAL(void)
+jinit_merged_upsampler (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample;
+
+  upsample = (my_upsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_upsampler));
+  cinfo->upsample = (struct jpeg_upsampler *) upsample;
+  upsample->pub.start_pass = start_pass_merged_upsample;
+  upsample->pub.need_context_rows = FALSE;
+
+  upsample->out_row_width = cinfo->output_width * cinfo->out_color_components;
+
+  if (cinfo->max_v_samp_factor == 2) {
+    upsample->pub.upsample = merged_2v_upsample;
+    upsample->upmethod = h2v2_merged_upsample;
+    /* Allocate a spare row buffer */
+    upsample->spare_row = (JSAMPROW)
+      (*cinfo->mem->alloc_large) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+		(size_t) (upsample->out_row_width * SIZEOF(JSAMPLE)));
+  } else {
+    upsample->pub.upsample = merged_1v_upsample;
+    upsample->upmethod = h2v1_merged_upsample;
+    /* No spare row needed */
+    upsample->spare_row = NULL;
+  }
+
+  build_ycc_rgb_table(cinfo);
+}
+
+#endif /* UPSAMPLE_MERGING_SUPPORTED */

Added: trunk/code/jpeg-8c/jdpostct.c
===================================================================
--- trunk/code/jpeg-8c/jdpostct.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdpostct.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,290 @@
+/*
+ * jdpostct.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains the decompression postprocessing controller.
+ * This controller manages the upsampling, color conversion, and color
+ * quantization/reduction steps; specifically, it controls the buffering
+ * between upsample/color conversion and color quantization/reduction.
+ *
+ * If no color quantization/reduction is required, then this module has no
+ * work to do, and it just hands off to the upsample/color conversion code.
+ * An integrated upsample/convert/quantize process would replace this module
+ * entirely.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Private buffer controller object */
+
+typedef struct {
+  struct jpeg_d_post_controller pub; /* public fields */
+
+  /* Color quantization source buffer: this holds output data from
+   * the upsample/color conversion step to be passed to the quantizer.
+   * For two-pass color quantization, we need a full-image buffer;
+   * for one-pass operation, a strip buffer is sufficient.
+   */
+  jvirt_sarray_ptr whole_image;	/* virtual array, or NULL if one-pass */
+  JSAMPARRAY buffer;		/* strip buffer, or current strip of virtual */
+  JDIMENSION strip_height;	/* buffer size in rows */
+  /* for two-pass mode only: */
+  JDIMENSION starting_row;	/* row # of first row in current strip */
+  JDIMENSION next_row;		/* index of next row to fill/empty in strip */
+} my_post_controller;
+
+typedef my_post_controller * my_post_ptr;
+
+
+/* Forward declarations */
+METHODDEF(void) post_process_1pass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+#ifdef QUANT_2PASS_SUPPORTED
+METHODDEF(void) post_process_prepass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+METHODDEF(void) post_process_2pass
+	JPP((j_decompress_ptr cinfo,
+	     JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	     JDIMENSION in_row_groups_avail,
+	     JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	     JDIMENSION out_rows_avail));
+#endif
+
+
+/*
+ * Initialize for a processing pass.
+ */
+
+METHODDEF(void)
+start_pass_dpost (j_decompress_ptr cinfo, J_BUF_MODE pass_mode)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+
+  switch (pass_mode) {
+  case JBUF_PASS_THRU:
+    if (cinfo->quantize_colors) {
+      /* Single-pass processing with color quantization. */
+      post->pub.post_process_data = post_process_1pass;
+      /* We could be doing buffered-image output before starting a 2-pass
+       * color quantization; in that case, jinit_d_post_controller did not
+       * allocate a strip buffer.  Use the virtual-array buffer as workspace.
+       */
+      if (post->buffer == NULL) {
+	post->buffer = (*cinfo->mem->access_virt_sarray)
+	  ((j_common_ptr) cinfo, post->whole_image,
+	   (JDIMENSION) 0, post->strip_height, TRUE);
+      }
+    } else {
+      /* For single-pass processing without color quantization,
+       * I have no work to do; just call the upsampler directly.
+       */
+      post->pub.post_process_data = cinfo->upsample->upsample;
+    }
+    break;
+#ifdef QUANT_2PASS_SUPPORTED
+  case JBUF_SAVE_AND_PASS:
+    /* First pass of 2-pass quantization */
+    if (post->whole_image == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    post->pub.post_process_data = post_process_prepass;
+    break;
+  case JBUF_CRANK_DEST:
+    /* Second pass of 2-pass quantization */
+    if (post->whole_image == NULL)
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    post->pub.post_process_data = post_process_2pass;
+    break;
+#endif /* QUANT_2PASS_SUPPORTED */
+  default:
+    ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+    break;
+  }
+  post->starting_row = post->next_row = 0;
+}
+
+
+/*
+ * Process some data in the one-pass (strip buffer) case.
+ * This is used for color precision reduction as well as one-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_1pass (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION num_rows, max_rows;
+
+  /* Fill the buffer, but not more than what we can dump out in one go. */
+  /* Note we rely on the upsampler to detect bottom of image. */
+  max_rows = out_rows_avail - *out_row_ctr;
+  if (max_rows > post->strip_height)
+    max_rows = post->strip_height;
+  num_rows = 0;
+  (*cinfo->upsample->upsample) (cinfo,
+		input_buf, in_row_group_ctr, in_row_groups_avail,
+		post->buffer, &num_rows, max_rows);
+  /* Quantize and emit data. */
+  (*cinfo->cquantize->color_quantize) (cinfo,
+		post->buffer, output_buf + *out_row_ctr, (int) num_rows);
+  *out_row_ctr += num_rows;
+}
+
+
+#ifdef QUANT_2PASS_SUPPORTED
+
+/*
+ * Process some data in the first pass of 2-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_prepass (j_decompress_ptr cinfo,
+		      JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		      JDIMENSION in_row_groups_avail,
+		      JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		      JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION old_next_row, num_rows;
+
+  /* Reposition virtual buffer if at start of strip. */
+  if (post->next_row == 0) {
+    post->buffer = (*cinfo->mem->access_virt_sarray)
+	((j_common_ptr) cinfo, post->whole_image,
+	 post->starting_row, post->strip_height, TRUE);
+  }
+
+  /* Upsample some data (up to a strip height's worth). */
+  old_next_row = post->next_row;
+  (*cinfo->upsample->upsample) (cinfo,
+		input_buf, in_row_group_ctr, in_row_groups_avail,
+		post->buffer, &post->next_row, post->strip_height);
+
+  /* Allow quantizer to scan new data.  No data is emitted, */
+  /* but we advance out_row_ctr so outer loop can tell when we're done. */
+  if (post->next_row > old_next_row) {
+    num_rows = post->next_row - old_next_row;
+    (*cinfo->cquantize->color_quantize) (cinfo, post->buffer + old_next_row,
+					 (JSAMPARRAY) NULL, (int) num_rows);
+    *out_row_ctr += num_rows;
+  }
+
+  /* Advance if we filled the strip. */
+  if (post->next_row >= post->strip_height) {
+    post->starting_row += post->strip_height;
+    post->next_row = 0;
+  }
+}
+
+
+/*
+ * Process some data in the second pass of 2-pass quantization.
+ */
+
+METHODDEF(void)
+post_process_2pass (j_decompress_ptr cinfo,
+		    JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+		    JDIMENSION in_row_groups_avail,
+		    JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+		    JDIMENSION out_rows_avail)
+{
+  my_post_ptr post = (my_post_ptr) cinfo->post;
+  JDIMENSION num_rows, max_rows;
+
+  /* Reposition virtual buffer if at start of strip. */
+  if (post->next_row == 0) {
+    post->buffer = (*cinfo->mem->access_virt_sarray)
+	((j_common_ptr) cinfo, post->whole_image,
+	 post->starting_row, post->strip_height, FALSE);
+  }
+
+  /* Determine number of rows to emit. */
+  num_rows = post->strip_height - post->next_row; /* available in strip */
+  max_rows = out_rows_avail - *out_row_ctr; /* available in output area */
+  if (num_rows > max_rows)
+    num_rows = max_rows;
+  /* We have to check bottom of image here, can't depend on upsampler. */
+  max_rows = cinfo->output_height - post->starting_row;
+  if (num_rows > max_rows)
+    num_rows = max_rows;
+
+  /* Quantize and emit data. */
+  (*cinfo->cquantize->color_quantize) (cinfo,
+		post->buffer + post->next_row, output_buf + *out_row_ctr,
+		(int) num_rows);
+  *out_row_ctr += num_rows;
+
+  /* Advance if we filled the strip. */
+  post->next_row += num_rows;
+  if (post->next_row >= post->strip_height) {
+    post->starting_row += post->strip_height;
+    post->next_row = 0;
+  }
+}
+
+#endif /* QUANT_2PASS_SUPPORTED */
+
+
+/*
+ * Initialize postprocessing controller.
+ */
+
+GLOBAL(void)
+jinit_d_post_controller (j_decompress_ptr cinfo, boolean need_full_buffer)
+{
+  my_post_ptr post;
+
+  post = (my_post_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_post_controller));
+  cinfo->post = (struct jpeg_d_post_controller *) post;
+  post->pub.start_pass = start_pass_dpost;
+  post->whole_image = NULL;	/* flag for no virtual arrays */
+  post->buffer = NULL;		/* flag for no strip buffer */
+
+  /* Create the quantization buffer, if needed */
+  if (cinfo->quantize_colors) {
+    /* The buffer strip height is max_v_samp_factor, which is typically
+     * an efficient number of rows for upsampling to return.
+     * (In the presence of output rescaling, we might want to be smarter?)
+     */
+    post->strip_height = (JDIMENSION) cinfo->max_v_samp_factor;
+    if (need_full_buffer) {
+      /* Two-pass color quantization: need full-image storage. */
+      /* We round up the number of rows to a multiple of the strip height. */
+#ifdef QUANT_2PASS_SUPPORTED
+      post->whole_image = (*cinfo->mem->request_virt_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE, FALSE,
+	 cinfo->output_width * cinfo->out_color_components,
+	 (JDIMENSION) jround_up((long) cinfo->output_height,
+				(long) post->strip_height),
+	 post->strip_height);
+#else
+      ERREXIT(cinfo, JERR_BAD_BUFFER_MODE);
+#endif /* QUANT_2PASS_SUPPORTED */
+    } else {
+      /* One-pass color quantization: just make a strip buffer. */
+      post->buffer = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 cinfo->output_width * cinfo->out_color_components,
+	 post->strip_height);
+    }
+  }
+}

Added: trunk/code/jpeg-8c/jdsample.c
===================================================================
--- trunk/code/jpeg-8c/jdsample.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdsample.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,361 @@
+/*
+ * jdsample.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Modified 2002-2008 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains upsampling routines.
+ *
+ * Upsampling input data is counted in "row groups".  A row group
+ * is defined to be (v_samp_factor * DCT_v_scaled_size / min_DCT_v_scaled_size)
+ * sample rows of each component.  Upsampling will normally produce
+ * max_v_samp_factor pixel rows from each row group (but this could vary
+ * if the upsampler is applying a scale factor of its own).
+ *
+ * An excellent reference for image resampling is
+ *   Digital Image Warping, George Wolberg, 1990.
+ *   Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Pointer to routine to upsample a single component */
+typedef JMETHOD(void, upsample1_ptr,
+		(j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr));
+
+/* Private subobject */
+
+typedef struct {
+  struct jpeg_upsampler pub;	/* public fields */
+
+  /* Color conversion buffer.  When using separate upsampling and color
+   * conversion steps, this buffer holds one upsampled row group until it
+   * has been color converted and output.
+   * Note: we do not allocate any storage for component(s) which are full-size,
+   * ie do not need rescaling.  The corresponding entry of color_buf[] is
+   * simply set to point to the input data array, thereby avoiding copying.
+   */
+  JSAMPARRAY color_buf[MAX_COMPONENTS];
+
+  /* Per-component upsampling method pointers */
+  upsample1_ptr methods[MAX_COMPONENTS];
+
+  int next_row_out;		/* counts rows emitted from color_buf */
+  JDIMENSION rows_to_go;	/* counts rows remaining in image */
+
+  /* Height of an input row group for each component. */
+  int rowgroup_height[MAX_COMPONENTS];
+
+  /* These arrays save pixel expansion factors so that int_expand need not
+   * recompute them each time.  They are unused for other upsampling methods.
+   */
+  UINT8 h_expand[MAX_COMPONENTS];
+  UINT8 v_expand[MAX_COMPONENTS];
+} my_upsampler;
+
+typedef my_upsampler * my_upsample_ptr;
+
+
+/*
+ * Initialize for an upsampling pass.
+ */
+
+METHODDEF(void)
+start_pass_upsample (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+
+  /* Mark the conversion buffer empty */
+  upsample->next_row_out = cinfo->max_v_samp_factor;
+  /* Initialize total-height counter for detecting bottom of image */
+  upsample->rows_to_go = cinfo->output_height;
+}
+
+
+/*
+ * Control routine to do upsampling (and color conversion).
+ *
+ * In this version we upsample each component independently.
+ * We upsample one row group into the conversion buffer, then apply
+ * color conversion a row at a time.
+ */
+
+METHODDEF(void)
+sep_upsample (j_decompress_ptr cinfo,
+	      JSAMPIMAGE input_buf, JDIMENSION *in_row_group_ctr,
+	      JDIMENSION in_row_groups_avail,
+	      JSAMPARRAY output_buf, JDIMENSION *out_row_ctr,
+	      JDIMENSION out_rows_avail)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  int ci;
+  jpeg_component_info * compptr;
+  JDIMENSION num_rows;
+
+  /* Fill the conversion buffer, if it's empty */
+  if (upsample->next_row_out >= cinfo->max_v_samp_factor) {
+    for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+	 ci++, compptr++) {
+      /* Invoke per-component upsample method.  Notice we pass a POINTER
+       * to color_buf[ci], so that fullsize_upsample can change it.
+       */
+      (*upsample->methods[ci]) (cinfo, compptr,
+	input_buf[ci] + (*in_row_group_ctr * upsample->rowgroup_height[ci]),
+	upsample->color_buf + ci);
+    }
+    upsample->next_row_out = 0;
+  }
+
+  /* Color-convert and emit rows */
+
+  /* How many we have in the buffer: */
+  num_rows = (JDIMENSION) (cinfo->max_v_samp_factor - upsample->next_row_out);
+  /* Not more than the distance to the end of the image.  Need this test
+   * in case the image height is not a multiple of max_v_samp_factor:
+   */
+  if (num_rows > upsample->rows_to_go) 
+    num_rows = upsample->rows_to_go;
+  /* And not more than what the client can accept: */
+  out_rows_avail -= *out_row_ctr;
+  if (num_rows > out_rows_avail)
+    num_rows = out_rows_avail;
+
+  (*cinfo->cconvert->color_convert) (cinfo, upsample->color_buf,
+				     (JDIMENSION) upsample->next_row_out,
+				     output_buf + *out_row_ctr,
+				     (int) num_rows);
+
+  /* Adjust counts */
+  *out_row_ctr += num_rows;
+  upsample->rows_to_go -= num_rows;
+  upsample->next_row_out += num_rows;
+  /* When the buffer is emptied, declare this input row group consumed */
+  if (upsample->next_row_out >= cinfo->max_v_samp_factor)
+    (*in_row_group_ctr)++;
+}
+
+
+/*
+ * These are the routines invoked by sep_upsample to upsample pixel values
+ * of a single component.  One row group is processed per call.
+ */
+
+
+/*
+ * For full-size components, we just make color_buf[ci] point at the
+ * input buffer, and thus avoid copying any data.  Note that this is
+ * safe only because sep_upsample doesn't declare the input row group
+ * "consumed" until we are done color converting and emitting it.
+ */
+
+METHODDEF(void)
+fullsize_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		   JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  *output_data_ptr = input_data;
+}
+
+
+/*
+ * This is a no-op version used for "uninteresting" components.
+ * These components will not be referenced by color conversion.
+ */
+
+METHODDEF(void)
+noop_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  *output_data_ptr = NULL;	/* safety check */
+}
+
+
+/*
+ * This version handles any integral sampling ratios.
+ * This is not used for typical JPEG files, so it need not be fast.
+ * Nor, for that matter, is it particularly accurate: the algorithm is
+ * simple replication of the input pixel onto the corresponding output
+ * pixels.  The hi-falutin sampling literature refers to this as a
+ * "box filter".  A box filter tends to introduce visible artifacts,
+ * so if you are actually going to use 3:1 or 4:1 sampling ratios
+ * you would be well advised to improve this code.
+ */
+
+METHODDEF(void)
+int_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	      JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample;
+  JSAMPARRAY output_data = *output_data_ptr;
+  register JSAMPROW inptr, outptr;
+  register JSAMPLE invalue;
+  register int h;
+  JSAMPROW outend;
+  int h_expand, v_expand;
+  int inrow, outrow;
+
+  h_expand = upsample->h_expand[compptr->component_index];
+  v_expand = upsample->v_expand[compptr->component_index];
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    /* Generate one output row with proper horizontal expansion */
+    inptr = input_data[inrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      for (h = h_expand; h > 0; h--) {
+	*outptr++ = invalue;
+      }
+    }
+    /* Generate any additional output rows by duplicating the first one */
+    if (v_expand > 1) {
+      jcopy_sample_rows(output_data, outrow, output_data, outrow+1,
+			v_expand-1, cinfo->output_width);
+    }
+    inrow++;
+    outrow += v_expand;
+  }
+}
+
+
+/*
+ * Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
+ * It's still a box filter.
+ */
+
+METHODDEF(void)
+h2v1_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  register JSAMPROW inptr, outptr;
+  register JSAMPLE invalue;
+  JSAMPROW outend;
+  int outrow;
+
+  for (outrow = 0; outrow < cinfo->max_v_samp_factor; outrow++) {
+    inptr = input_data[outrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      *outptr++ = invalue;
+      *outptr++ = invalue;
+    }
+  }
+}
+
+
+/*
+ * Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
+ * It's still a box filter.
+ */
+
+METHODDEF(void)
+h2v2_upsample (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JSAMPARRAY input_data, JSAMPARRAY * output_data_ptr)
+{
+  JSAMPARRAY output_data = *output_data_ptr;
+  register JSAMPROW inptr, outptr;
+  register JSAMPLE invalue;
+  JSAMPROW outend;
+  int inrow, outrow;
+
+  inrow = outrow = 0;
+  while (outrow < cinfo->max_v_samp_factor) {
+    inptr = input_data[inrow];
+    outptr = output_data[outrow];
+    outend = outptr + cinfo->output_width;
+    while (outptr < outend) {
+      invalue = *inptr++;	/* don't need GETJSAMPLE() here */
+      *outptr++ = invalue;
+      *outptr++ = invalue;
+    }
+    jcopy_sample_rows(output_data, outrow, output_data, outrow+1,
+		      1, cinfo->output_width);
+    inrow++;
+    outrow += 2;
+  }
+}
+
+
+/*
+ * Module initialization routine for upsampling.
+ */
+
+GLOBAL(void)
+jinit_upsampler (j_decompress_ptr cinfo)
+{
+  my_upsample_ptr upsample;
+  int ci;
+  jpeg_component_info * compptr;
+  boolean need_buffer;
+  int h_in_group, v_in_group, h_out_group, v_out_group;
+
+  upsample = (my_upsample_ptr)
+    (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				SIZEOF(my_upsampler));
+  cinfo->upsample = (struct jpeg_upsampler *) upsample;
+  upsample->pub.start_pass = start_pass_upsample;
+  upsample->pub.upsample = sep_upsample;
+  upsample->pub.need_context_rows = FALSE; /* until we find out differently */
+
+  if (cinfo->CCIR601_sampling)	/* this isn't supported */
+    ERREXIT(cinfo, JERR_CCIR601_NOTIMPL);
+
+  /* Verify we can handle the sampling factors, select per-component methods,
+   * and create storage as needed.
+   */
+  for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
+       ci++, compptr++) {
+    /* Compute size of an "input group" after IDCT scaling.  This many samples
+     * are to be converted to max_h_samp_factor * max_v_samp_factor pixels.
+     */
+    h_in_group = (compptr->h_samp_factor * compptr->DCT_h_scaled_size) /
+		 cinfo->min_DCT_h_scaled_size;
+    v_in_group = (compptr->v_samp_factor * compptr->DCT_v_scaled_size) /
+		 cinfo->min_DCT_v_scaled_size;
+    h_out_group = cinfo->max_h_samp_factor;
+    v_out_group = cinfo->max_v_samp_factor;
+    upsample->rowgroup_height[ci] = v_in_group; /* save for use later */
+    need_buffer = TRUE;
+    if (! compptr->component_needed) {
+      /* Don't bother to upsample an uninteresting component. */
+      upsample->methods[ci] = noop_upsample;
+      need_buffer = FALSE;
+    } else if (h_in_group == h_out_group && v_in_group == v_out_group) {
+      /* Fullsize components can be processed without any work. */
+      upsample->methods[ci] = fullsize_upsample;
+      need_buffer = FALSE;
+    } else if (h_in_group * 2 == h_out_group &&
+	       v_in_group == v_out_group) {
+      /* Special case for 2h1v upsampling */
+      upsample->methods[ci] = h2v1_upsample;
+    } else if (h_in_group * 2 == h_out_group &&
+	       v_in_group * 2 == v_out_group) {
+      /* Special case for 2h2v upsampling */
+      upsample->methods[ci] = h2v2_upsample;
+    } else if ((h_out_group % h_in_group) == 0 &&
+	       (v_out_group % v_in_group) == 0) {
+      /* Generic integral-factors upsampling method */
+      upsample->methods[ci] = int_upsample;
+      upsample->h_expand[ci] = (UINT8) (h_out_group / h_in_group);
+      upsample->v_expand[ci] = (UINT8) (v_out_group / v_in_group);
+    } else
+      ERREXIT(cinfo, JERR_FRACT_SAMPLE_NOTIMPL);
+    if (need_buffer) {
+      upsample->color_buf[ci] = (*cinfo->mem->alloc_sarray)
+	((j_common_ptr) cinfo, JPOOL_IMAGE,
+	 (JDIMENSION) jround_up((long) cinfo->output_width,
+				(long) cinfo->max_h_samp_factor),
+	 (JDIMENSION) cinfo->max_v_samp_factor);
+    }
+  }
+}

Added: trunk/code/jpeg-8c/jdtrans.c
===================================================================
--- trunk/code/jpeg-8c/jdtrans.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jdtrans.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,140 @@
+/*
+ * jdtrans.c
+ *
+ * Copyright (C) 1995-1997, Thomas G. Lane.
+ * Modified 2000-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains library routines for transcoding decompression,
+ * that is, reading raw DCT coefficient arrays from an input JPEG file.
+ * The routines in jdapimin.c will also be needed by a transcoder.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+
+
+/* Forward declarations */
+LOCAL(void) transdecode_master_selection JPP((j_decompress_ptr cinfo));
+
+
+/*
+ * Read the coefficient arrays from a JPEG file.
+ * jpeg_read_header must be completed before calling this.
+ *
+ * The entire image is read into a set of virtual coefficient-block arrays,
+ * one per component.  The return value is a pointer to the array of
+ * virtual-array descriptors.  These can be manipulated directly via the
+ * JPEG memory manager, or handed off to jpeg_write_coefficients().
+ * To release the memory occupied by the virtual arrays, call
+ * jpeg_finish_decompress() when done with the data.
+ *
+ * An alternative usage is to simply obtain access to the coefficient arrays
+ * during a buffered-image-mode decompression operation.  This is allowed
+ * after any jpeg_finish_output() call.  The arrays can be accessed until
+ * jpeg_finish_decompress() is called.  (Note that any call to the library
+ * may reposition the arrays, so don't rely on access_virt_barray() results
+ * to stay valid across library calls.)
+ *
+ * Returns NULL if suspended.  This case need be checked only if
+ * a suspending data source is used.
+ */
+
+GLOBAL(jvirt_barray_ptr *)
+jpeg_read_coefficients (j_decompress_ptr cinfo)
+{
+  if (cinfo->global_state == DSTATE_READY) {
+    /* First call: initialize active modules */
+    transdecode_master_selection(cinfo);
+    cinfo->global_state = DSTATE_RDCOEFS;
+  }
+  if (cinfo->global_state == DSTATE_RDCOEFS) {
+    /* Absorb whole file into the coef buffer */
+    for (;;) {
+      int retcode;
+      /* Call progress monitor hook if present */
+      if (cinfo->progress != NULL)
+	(*cinfo->progress->progress_monitor) ((j_common_ptr) cinfo);
+      /* Absorb some more input */
+      retcode = (*cinfo->inputctl->consume_input) (cinfo);
+      if (retcode == JPEG_SUSPENDED)
+	return NULL;
+      if (retcode == JPEG_REACHED_EOI)
+	break;
+      /* Advance progress counter if appropriate */
+      if (cinfo->progress != NULL &&
+	  (retcode == JPEG_ROW_COMPLETED || retcode == JPEG_REACHED_SOS)) {
+	if (++cinfo->progress->pass_counter >= cinfo->progress->pass_limit) {
+	  /* startup underestimated number of scans; ratchet up one scan */
+	  cinfo->progress->pass_limit += (long) cinfo->total_iMCU_rows;
+	}
+      }
+    }
+    /* Set state so that jpeg_finish_decompress does the right thing */
+    cinfo->global_state = DSTATE_STOPPING;
+  }
+  /* At this point we should be in state DSTATE_STOPPING if being used
+   * standalone, or in state DSTATE_BUFIMAGE if being invoked to get access
+   * to the coefficients during a full buffered-image-mode decompression.
+   */
+  if ((cinfo->global_state == DSTATE_STOPPING ||
+       cinfo->global_state == DSTATE_BUFIMAGE) && cinfo->buffered_image) {
+    return cinfo->coef->coef_arrays;
+  }
+  /* Oops, improper usage */
+  ERREXIT1(cinfo, JERR_BAD_STATE, cinfo->global_state);
+  return NULL;			/* keep compiler happy */
+}
+
+
+/*
+ * Master selection of decompression modules for transcoding.
+ * This substitutes for jdmaster.c's initialization of the full decompressor.
+ */
+
+LOCAL(void)
+transdecode_master_selection (j_decompress_ptr cinfo)
+{
+  /* This is effectively a buffered-image operation. */
+  cinfo->buffered_image = TRUE;
+
+  /* Compute output image dimensions and related values. */
+  jpeg_core_output_dimensions(cinfo);
+
+  /* Entropy decoding: either Huffman or arithmetic coding. */
+  if (cinfo->arith_code)
+    jinit_arith_decoder(cinfo);
+  else {
+    jinit_huff_decoder(cinfo);
+  }
+
+  /* Always get a full-image coefficient buffer. */
+  jinit_d_coef_controller(cinfo, TRUE);
+
+  /* We can now tell the memory manager to allocate virtual arrays. */
+  (*cinfo->mem->realize_virt_arrays) ((j_common_ptr) cinfo);
+
+  /* Initialize input side of decompressor to consume first scan. */
+  (*cinfo->inputctl->start_input_pass) (cinfo);
+
+  /* Initialize progress monitoring. */
+  if (cinfo->progress != NULL) {
+    int nscans;
+    /* Estimate number of scans to set pass_limit. */
+    if (cinfo->progressive_mode) {
+      /* Arbitrarily estimate 2 interleaved DC scans + 3 AC scans/component. */
+      nscans = 2 + 3 * cinfo->num_components;
+    } else if (cinfo->inputctl->has_multiple_scans) {
+      /* For a nonprogressive multiscan file, estimate 1 scan per component. */
+      nscans = cinfo->num_components;
+    } else {
+      nscans = 1;
+    }
+    cinfo->progress->pass_counter = 0L;
+    cinfo->progress->pass_limit = (long) cinfo->total_iMCU_rows * nscans;
+    cinfo->progress->completed_passes = 0;
+    cinfo->progress->total_passes = 1;
+  }
+}

Added: trunk/code/jpeg-8c/jerror.c
===================================================================
--- trunk/code/jpeg-8c/jerror.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jerror.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,252 @@
+/*
+ * jerror.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains simple error-reporting and trace-message routines.
+ * These are suitable for Unix-like systems and others where writing to
+ * stderr is the right thing to do.  Many applications will want to replace
+ * some or all of these routines.
+ *
+ * If you define USE_WINDOWS_MESSAGEBOX in jconfig.h or in the makefile,
+ * you get a Windows-specific hack to display error messages in a dialog box.
+ * It ain't much, but it beats dropping error messages into the bit bucket,
+ * which is what happens to output to stderr under most Windows C compilers.
+ *
+ * These routines are used by both the compression and decompression code.
+ */
+
+/* this is not a core library module, so it doesn't define JPEG_INTERNALS */
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jversion.h"
+#include "jerror.h"
+
+#ifdef USE_WINDOWS_MESSAGEBOX
+#include <windows.h>
+#endif
+
+#ifndef EXIT_FAILURE		/* define exit() codes if not provided */
+#define EXIT_FAILURE  1
+#endif
+
+
+/*
+ * Create the message string table.
+ * We do this from the master message list in jerror.h by re-reading
+ * jerror.h with a suitable definition for macro JMESSAGE.
+ * The message table is made an external symbol just in case any applications
+ * want to refer to it directly.
+ */
+
+#ifdef NEED_SHORT_EXTERNAL_NAMES
+#define jpeg_std_message_table	jMsgTable
+#endif
+
+#define JMESSAGE(code,string)	string ,
+
+const char * const jpeg_std_message_table[] = {
+#include "jerror.h"
+  NULL
+};
+
+
+/*
+ * Error exit handler: must not return to caller.
+ *
+ * Applications may override this if they want to get control back after
+ * an error.  Typically one would longjmp somewhere instead of exiting.
+ * The setjmp buffer can be made a private field within an expanded error
+ * handler object.  Note that the info needed to generate an error message
+ * is stored in the error object, so you can generate the message now or
+ * later, at your convenience.
+ * You should make sure that the JPEG object is cleaned up (with jpeg_abort
+ * or jpeg_destroy) at some point.
+ */
+
+METHODDEF(void)
+error_exit (j_common_ptr cinfo)
+{
+  /* Always display the message */
+  (*cinfo->err->output_message) (cinfo);
+
+  /* Let the memory manager delete any temp files before we die */
+  jpeg_destroy(cinfo);
+
+  exit(EXIT_FAILURE);
+}
+
+
+/*
+ * Actual output of an error or trace message.
+ * Applications may override this method to send JPEG messages somewhere
+ * other than stderr.
+ *
+ * On Windows, printing to stderr is generally completely useless,
+ * so we provide optional code to produce an error-dialog popup.
+ * Most Windows applications will still prefer to override this routine,
+ * but if they don't, it'll do something at least marginally useful.
+ *
+ * NOTE: to use the library in an environment that doesn't support the
+ * C stdio library, you may have to delete the call to fprintf() entirely,
+ * not just not use this routine.
+ */
+
+METHODDEF(void)
+output_message (j_common_ptr cinfo)
+{
+  char buffer[JMSG_LENGTH_MAX];
+
+  /* Create the message */
+  (*cinfo->err->format_message) (cinfo, buffer);
+
+#ifdef USE_WINDOWS_MESSAGEBOX
+  /* Display it in a message dialog box */
+  MessageBox(GetActiveWindow(), buffer, "JPEG Library Error",
+	     MB_OK | MB_ICONERROR);
+#else
+  /* Send it to stderr, adding a newline */
+  fprintf(stderr, "%s\n", buffer);
+#endif
+}
+
+
+/*
+ * Decide whether to emit a trace or warning message.
+ * msg_level is one of:
+ *   -1: recoverable corrupt-data warning, may want to abort.
+ *    0: important advisory messages (always display to user).
+ *    1: first level of tracing detail.
+ *    2,3,...: successively more detailed tracing messages.
+ * An application might override this method if it wanted to abort on warnings
+ * or change the policy about which messages to display.
+ */
+
+METHODDEF(void)
+emit_message (j_common_ptr cinfo, int msg_level)
+{
+  struct jpeg_error_mgr * err = cinfo->err;
+
+  if (msg_level < 0) {
+    /* It's a warning message.  Since corrupt files may generate many warnings,
+     * the policy implemented here is to show only the first warning,
+     * unless trace_level >= 3.
+     */
+    if (err->num_warnings == 0 || err->trace_level >= 3)
+      (*err->output_message) (cinfo);
+    /* Always count warnings in num_warnings. */
+    err->num_warnings++;
+  } else {
+    /* It's a trace message.  Show it if trace_level >= msg_level. */
+    if (err->trace_level >= msg_level)
+      (*err->output_message) (cinfo);
+  }
+}
+
+
+/*
+ * Format a message string for the most recent JPEG error or message.
+ * The message is stored into buffer, which should be at least JMSG_LENGTH_MAX
+ * characters.  Note that no '\n' character is added to the string.
+ * Few applications should need to override this method.
+ */
+
+METHODDEF(void)
+format_message (j_common_ptr cinfo, char * buffer)
+{
+  struct jpeg_error_mgr * err = cinfo->err;
+  int msg_code = err->msg_code;
+  const char * msgtext = NULL;
+  const char * msgptr;
+  char ch;
+  boolean isstring;
+
+  /* Look up message string in proper table */
+  if (msg_code > 0 && msg_code <= err->last_jpeg_message) {
+    msgtext = err->jpeg_message_table[msg_code];
+  } else if (err->addon_message_table != NULL &&
+	     msg_code >= err->first_addon_message &&
+	     msg_code <= err->last_addon_message) {
+    msgtext = err->addon_message_table[msg_code - err->first_addon_message];
+  }
+
+  /* Defend against bogus message number */
+  if (msgtext == NULL) {
+    err->msg_parm.i[0] = msg_code;
+    msgtext = err->jpeg_message_table[0];
+  }
+
+  /* Check for string parameter, as indicated by %s in the message text */
+  isstring = FALSE;
+  msgptr = msgtext;
+  while ((ch = *msgptr++) != '\0') {
+    if (ch == '%') {
+      if (*msgptr == 's') isstring = TRUE;
+      break;
+    }
+  }
+
+  /* Format the message into the passed buffer */
+  if (isstring)
+    sprintf(buffer, msgtext, err->msg_parm.s);
+  else
+    sprintf(buffer, msgtext,
+	    err->msg_parm.i[0], err->msg_parm.i[1],
+	    err->msg_parm.i[2], err->msg_parm.i[3],
+	    err->msg_parm.i[4], err->msg_parm.i[5],
+	    err->msg_parm.i[6], err->msg_parm.i[7]);
+}
+
+
+/*
+ * Reset error state variables at start of a new image.
+ * This is called during compression startup to reset trace/error
+ * processing to default state, without losing any application-specific
+ * method pointers.  An application might possibly want to override
+ * this method if it has additional error processing state.
+ */
+
+METHODDEF(void)
+reset_error_mgr (j_common_ptr cinfo)
+{
+  cinfo->err->num_warnings = 0;
+  /* trace_level is not reset since it is an application-supplied parameter */
+  cinfo->err->msg_code = 0;	/* may be useful as a flag for "no error" */
+}
+
+
+/*
+ * Fill in the standard error-handling methods in a jpeg_error_mgr object.
+ * Typical call is:
+ *	struct jpeg_compress_struct cinfo;
+ *	struct jpeg_error_mgr err;
+ *
+ *	cinfo.err = jpeg_std_error(&err);
+ * after which the application may override some of the methods.
+ */
+
+GLOBAL(struct jpeg_error_mgr *)
+jpeg_std_error (struct jpeg_error_mgr * err)
+{
+  err->error_exit = error_exit;
+  err->emit_message = emit_message;
+  err->output_message = output_message;
+  err->format_message = format_message;
+  err->reset_error_mgr = reset_error_mgr;
+
+  err->trace_level = 0;		/* default = no tracing */
+  err->num_warnings = 0;	/* no warnings emitted yet */
+  err->msg_code = 0;		/* may be useful as a flag for "no error" */
+
+  /* Initialize message table pointers */
+  err->jpeg_message_table = jpeg_std_message_table;
+  err->last_jpeg_message = (int) JMSG_LASTMSGCODE - 1;
+
+  err->addon_message_table = NULL;
+  err->first_addon_message = 0;	/* for safety */
+  err->last_addon_message = 0;
+
+  return err;
+}

Added: trunk/code/jpeg-8c/jerror.h
===================================================================
--- trunk/code/jpeg-8c/jerror.h	                        (rev 0)
+++ trunk/code/jpeg-8c/jerror.h	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,304 @@
+/*
+ * jerror.h
+ *
+ * Copyright (C) 1994-1997, Thomas G. Lane.
+ * Modified 1997-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file defines the error and message codes for the JPEG library.
+ * Edit this file to add new codes, or to translate the message strings to
+ * some other language.
+ * A set of error-reporting macros are defined too.  Some applications using
+ * the JPEG library may wish to include this file to get the error codes
+ * and/or the macros.
+ */
+
+/*
+ * To define the enum list of message codes, include this file without
+ * defining macro JMESSAGE.  To create a message string table, include it
+ * again with a suitable JMESSAGE definition (see jerror.c for an example).
+ */
+#ifndef JMESSAGE
+#ifndef JERROR_H
+/* First time through, define the enum list */
+#define JMAKE_ENUM_LIST
+#else
+/* Repeated inclusions of this file are no-ops unless JMESSAGE is defined */
+#define JMESSAGE(code,string)
+#endif /* JERROR_H */
+#endif /* JMESSAGE */
+
+#ifdef JMAKE_ENUM_LIST
+
+typedef enum {
+
+#define JMESSAGE(code,string)	code ,
+
+#endif /* JMAKE_ENUM_LIST */
+
+JMESSAGE(JMSG_NOMESSAGE, "Bogus message code %d") /* Must be first entry! */
+
+/* For maintenance convenience, list is alphabetical by message code name */
+JMESSAGE(JERR_BAD_ALIGN_TYPE, "ALIGN_TYPE is wrong, please fix")
+JMESSAGE(JERR_BAD_ALLOC_CHUNK, "MAX_ALLOC_CHUNK is wrong, please fix")
+JMESSAGE(JERR_BAD_BUFFER_MODE, "Bogus buffer control mode")
+JMESSAGE(JERR_BAD_COMPONENT_ID, "Invalid component ID %d in SOS")
+JMESSAGE(JERR_BAD_CROP_SPEC, "Invalid crop request")
+JMESSAGE(JERR_BAD_DCT_COEF, "DCT coefficient out of range")
+JMESSAGE(JERR_BAD_DCTSIZE, "DCT scaled block size %dx%d not supported")
+JMESSAGE(JERR_BAD_DROP_SAMPLING,
+	 "Component index %d: mismatching sampling ratio %d:%d, %d:%d, %c")
+JMESSAGE(JERR_BAD_HUFF_TABLE, "Bogus Huffman table definition")
+JMESSAGE(JERR_BAD_IN_COLORSPACE, "Bogus input colorspace")
+JMESSAGE(JERR_BAD_J_COLORSPACE, "Bogus JPEG colorspace")
+JMESSAGE(JERR_BAD_LENGTH, "Bogus marker length")
+JMESSAGE(JERR_BAD_LIB_VERSION,
+	 "Wrong JPEG library version: library is %d, caller expects %d")
+JMESSAGE(JERR_BAD_MCU_SIZE, "Sampling factors too large for interleaved scan")
+JMESSAGE(JERR_BAD_POOL_ID, "Invalid memory pool code %d")
+JMESSAGE(JERR_BAD_PRECISION, "Unsupported JPEG data precision %d")
+JMESSAGE(JERR_BAD_PROGRESSION,
+	 "Invalid progressive parameters Ss=%d Se=%d Ah=%d Al=%d")
+JMESSAGE(JERR_BAD_PROG_SCRIPT,
+	 "Invalid progressive parameters at scan script entry %d")
+JMESSAGE(JERR_BAD_SAMPLING, "Bogus sampling factors")
+JMESSAGE(JERR_BAD_SCAN_SCRIPT, "Invalid scan script at entry %d")
+JMESSAGE(JERR_BAD_STATE, "Improper call to JPEG library in state %d")
+JMESSAGE(JERR_BAD_STRUCT_SIZE,
+	 "JPEG parameter struct mismatch: library thinks size is %u, caller expects %u")
+JMESSAGE(JERR_BAD_VIRTUAL_ACCESS, "Bogus virtual array access")
+JMESSAGE(JERR_BUFFER_SIZE, "Buffer passed to JPEG library is too small")
+JMESSAGE(JERR_CANT_SUSPEND, "Suspension not allowed here")
+JMESSAGE(JERR_CCIR601_NOTIMPL, "CCIR601 sampling not implemented yet")
+JMESSAGE(JERR_COMPONENT_COUNT, "Too many color components: %d, max %d")
+JMESSAGE(JERR_CONVERSION_NOTIMPL, "Unsupported color conversion request")
+JMESSAGE(JERR_DAC_INDEX, "Bogus DAC index %d")
+JMESSAGE(JERR_DAC_VALUE, "Bogus DAC value 0x%x")
+JMESSAGE(JERR_DHT_INDEX, "Bogus DHT index %d")
+JMESSAGE(JERR_DQT_INDEX, "Bogus DQT index %d")
+JMESSAGE(JERR_EMPTY_IMAGE, "Empty JPEG image (DNL not supported)")
+JMESSAGE(JERR_EMS_READ, "Read from EMS failed")
+JMESSAGE(JERR_EMS_WRITE, "Write to EMS failed")
+JMESSAGE(JERR_EOI_EXPECTED, "Didn't expect more than one scan")
+JMESSAGE(JERR_FILE_READ, "Input file read error")
+JMESSAGE(JERR_FILE_WRITE, "Output file write error --- out of disk space?")
+JMESSAGE(JERR_FRACT_SAMPLE_NOTIMPL, "Fractional sampling not implemented yet")
+JMESSAGE(JERR_HUFF_CLEN_OVERFLOW, "Huffman code size table overflow")
+JMESSAGE(JERR_HUFF_MISSING_CODE, "Missing Huffman code table entry")
+JMESSAGE(JERR_IMAGE_TOO_BIG, "Maximum supported image dimension is %u pixels")
+JMESSAGE(JERR_INPUT_EMPTY, "Empty input file")
+JMESSAGE(JERR_INPUT_EOF, "Premature end of input file")
+JMESSAGE(JERR_MISMATCHED_QUANT_TABLE,
+	 "Cannot transcode due to multiple use of quantization table %d")
+JMESSAGE(JERR_MISSING_DATA, "Scan script does not transmit all data")
+JMESSAGE(JERR_MODE_CHANGE, "Invalid color quantization mode change")
+JMESSAGE(JERR_NOTIMPL, "Not implemented yet")
+JMESSAGE(JERR_NOT_COMPILED, "Requested feature was omitted at compile time")
+JMESSAGE(JERR_NO_ARITH_TABLE, "Arithmetic table 0x%02x was not defined")
+JMESSAGE(JERR_NO_BACKING_STORE, "Backing store not supported")
+JMESSAGE(JERR_NO_HUFF_TABLE, "Huffman table 0x%02x was not defined")
+JMESSAGE(JERR_NO_IMAGE, "JPEG datastream contains no image")
+JMESSAGE(JERR_NO_QUANT_TABLE, "Quantization table 0x%02x was not defined")
+JMESSAGE(JERR_NO_SOI, "Not a JPEG file: starts with 0x%02x 0x%02x")
+JMESSAGE(JERR_OUT_OF_MEMORY, "Insufficient memory (case %d)")
+JMESSAGE(JERR_QUANT_COMPONENTS,
+	 "Cannot quantize more than %d color components")
+JMESSAGE(JERR_QUANT_FEW_COLORS, "Cannot quantize to fewer than %d colors")
+JMESSAGE(JERR_QUANT_MANY_COLORS, "Cannot quantize to more than %d colors")
+JMESSAGE(JERR_SOF_DUPLICATE, "Invalid JPEG file structure: two SOF markers")
+JMESSAGE(JERR_SOF_NO_SOS, "Invalid JPEG file structure: missing SOS marker")
+JMESSAGE(JERR_SOF_UNSUPPORTED, "Unsupported JPEG process: SOF type 0x%02x")
+JMESSAGE(JERR_SOI_DUPLICATE, "Invalid JPEG file structure: two SOI markers")
+JMESSAGE(JERR_SOS_NO_SOF, "Invalid JPEG file structure: SOS before SOF")
+JMESSAGE(JERR_TFILE_CREATE, "Failed to create temporary file %s")
+JMESSAGE(JERR_TFILE_READ, "Read failed on temporary file")
+JMESSAGE(JERR_TFILE_SEEK, "Seek failed on temporary file")
+JMESSAGE(JERR_TFILE_WRITE,
+	 "Write failed on temporary file --- out of disk space?")
+JMESSAGE(JERR_TOO_LITTLE_DATA, "Application transferred too few scanlines")
+JMESSAGE(JERR_UNKNOWN_MARKER, "Unsupported marker type 0x%02x")
+JMESSAGE(JERR_VIRTUAL_BUG, "Virtual array controller messed up")
+JMESSAGE(JERR_WIDTH_OVERFLOW, "Image too wide for this implementation")
+JMESSAGE(JERR_XMS_READ, "Read from XMS failed")
+JMESSAGE(JERR_XMS_WRITE, "Write to XMS failed")
+JMESSAGE(JMSG_COPYRIGHT, JCOPYRIGHT)
+JMESSAGE(JMSG_VERSION, JVERSION)
+JMESSAGE(JTRC_16BIT_TABLES,
+	 "Caution: quantization tables are too coarse for baseline JPEG")
+JMESSAGE(JTRC_ADOBE,
+	 "Adobe APP14 marker: version %d, flags 0x%04x 0x%04x, transform %d")
+JMESSAGE(JTRC_APP0, "Unknown APP0 marker (not JFIF), length %u")
+JMESSAGE(JTRC_APP14, "Unknown APP14 marker (not Adobe), length %u")
+JMESSAGE(JTRC_DAC, "Define Arithmetic Table 0x%02x: 0x%02x")
+JMESSAGE(JTRC_DHT, "Define Huffman Table 0x%02x")
+JMESSAGE(JTRC_DQT, "Define Quantization Table %d  precision %d")
+JMESSAGE(JTRC_DRI, "Define Restart Interval %u")
+JMESSAGE(JTRC_EMS_CLOSE, "Freed EMS handle %u")
+JMESSAGE(JTRC_EMS_OPEN, "Obtained EMS handle %u")
+JMESSAGE(JTRC_EOI, "End Of Image")
+JMESSAGE(JTRC_HUFFBITS, "        %3d %3d %3d %3d %3d %3d %3d %3d")
+JMESSAGE(JTRC_JFIF, "JFIF APP0 marker: version %d.%02d, density %dx%d  %d")
+JMESSAGE(JTRC_JFIF_BADTHUMBNAILSIZE,
+	 "Warning: thumbnail image size does not match data length %u")
+JMESSAGE(JTRC_JFIF_EXTENSION,
+	 "JFIF extension marker: type 0x%02x, length %u")
+JMESSAGE(JTRC_JFIF_THUMBNAIL, "    with %d x %d thumbnail image")
+JMESSAGE(JTRC_MISC_MARKER, "Miscellaneous marker 0x%02x, length %u")
+JMESSAGE(JTRC_PARMLESS_MARKER, "Unexpected marker 0x%02x")
+JMESSAGE(JTRC_QUANTVALS, "        %4u %4u %4u %4u %4u %4u %4u %4u")
+JMESSAGE(JTRC_QUANT_3_NCOLORS, "Quantizing to %d = %d*%d*%d colors")
+JMESSAGE(JTRC_QUANT_NCOLORS, "Quantizing to %d colors")
+JMESSAGE(JTRC_QUANT_SELECTED, "Selected %d colors for quantization")
+JMESSAGE(JTRC_RECOVERY_ACTION, "At marker 0x%02x, recovery action %d")
+JMESSAGE(JTRC_RST, "RST%d")
+JMESSAGE(JTRC_SMOOTH_NOTIMPL,
+	 "Smoothing not supported with nonstandard sampling ratios")
+JMESSAGE(JTRC_SOF, "Start Of Frame 0x%02x: width=%u, height=%u, components=%d")
+JMESSAGE(JTRC_SOF_COMPONENT, "    Component %d: %dhx%dv q=%d")
+JMESSAGE(JTRC_SOI, "Start of Image")
+JMESSAGE(JTRC_SOS, "Start Of Scan: %d components")
+JMESSAGE(JTRC_SOS_COMPONENT, "    Component %d: dc=%d ac=%d")
+JMESSAGE(JTRC_SOS_PARAMS, "  Ss=%d, Se=%d, Ah=%d, Al=%d")
+JMESSAGE(JTRC_TFILE_CLOSE, "Closed temporary file %s")
+JMESSAGE(JTRC_TFILE_OPEN, "Opened temporary file %s")
+JMESSAGE(JTRC_THUMB_JPEG,
+	 "JFIF extension marker: JPEG-compressed thumbnail image, length %u")
+JMESSAGE(JTRC_THUMB_PALETTE,
+	 "JFIF extension marker: palette thumbnail image, length %u")
+JMESSAGE(JTRC_THUMB_RGB,
+	 "JFIF extension marker: RGB thumbnail image, length %u")
+JMESSAGE(JTRC_UNKNOWN_IDS,
+	 "Unrecognized component IDs %d %d %d, assuming YCbCr")
+JMESSAGE(JTRC_XMS_CLOSE, "Freed XMS handle %u")
+JMESSAGE(JTRC_XMS_OPEN, "Obtained XMS handle %u")
+JMESSAGE(JWRN_ADOBE_XFORM, "Unknown Adobe color transform code %d")
+JMESSAGE(JWRN_ARITH_BAD_CODE, "Corrupt JPEG data: bad arithmetic code")
+JMESSAGE(JWRN_BOGUS_PROGRESSION,
+	 "Inconsistent progression sequence for component %d coefficient %d")
+JMESSAGE(JWRN_EXTRANEOUS_DATA,
+	 "Corrupt JPEG data: %u extraneous bytes before marker 0x%02x")
+JMESSAGE(JWRN_HIT_MARKER, "Corrupt JPEG data: premature end of data segment")
+JMESSAGE(JWRN_HUFF_BAD_CODE, "Corrupt JPEG data: bad Huffman code")
+JMESSAGE(JWRN_JFIF_MAJOR, "Warning: unknown JFIF revision number %d.%02d")
+JMESSAGE(JWRN_JPEG_EOF, "Premature end of JPEG file")
+JMESSAGE(JWRN_MUST_RESYNC,
+	 "Corrupt JPEG data: found marker 0x%02x instead of RST%d")
+JMESSAGE(JWRN_NOT_SEQUENTIAL, "Invalid SOS parameters for sequential JPEG")
+JMESSAGE(JWRN_TOO_MUCH_DATA, "Application transferred too many scanlines")
+
+#ifdef JMAKE_ENUM_LIST
+
+  JMSG_LASTMSGCODE
+} J_MESSAGE_CODE;
+
+#undef JMAKE_ENUM_LIST
+#endif /* JMAKE_ENUM_LIST */
+
+/* Zap JMESSAGE macro so that future re-inclusions do nothing by default */
+#undef JMESSAGE
+
+
+#ifndef JERROR_H
+#define JERROR_H
+
+/* Macros to simplify using the error and trace message stuff */
+/* The first parameter is either type of cinfo pointer */
+
+/* Fatal errors (print message and exit) */
+#define ERREXIT(cinfo,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT1(cinfo,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT2(cinfo,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT3(cinfo,code,p1,p2,p3)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT4(cinfo,code,p1,p2,p3,p4)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (cinfo)->err->msg_parm.i[3] = (p4), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXIT6(cinfo,code,p1,p2,p3,p4,p5,p6)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (cinfo)->err->msg_parm.i[2] = (p3), \
+   (cinfo)->err->msg_parm.i[3] = (p4), \
+   (cinfo)->err->msg_parm.i[4] = (p5), \
+   (cinfo)->err->msg_parm.i[5] = (p6), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+#define ERREXITS(cinfo,code,str)  \
+  ((cinfo)->err->msg_code = (code), \
+   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (*(cinfo)->err->error_exit) ((j_common_ptr) (cinfo)))
+
+#define MAKESTMT(stuff)		do { stuff } while (0)
+
+/* Nonfatal errors (we can keep going, but the data is probably corrupt) */
+#define WARNMS(cinfo,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+#define WARNMS1(cinfo,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+#define WARNMS2(cinfo,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), -1))
+
+/* Informational/debugging messages */
+#define TRACEMS(cinfo,lvl,code)  \
+  ((cinfo)->err->msg_code = (code), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS1(cinfo,lvl,code,p1)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS2(cinfo,lvl,code,p1,p2)  \
+  ((cinfo)->err->msg_code = (code), \
+   (cinfo)->err->msg_parm.i[0] = (p1), \
+   (cinfo)->err->msg_parm.i[1] = (p2), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+#define TRACEMS3(cinfo,lvl,code,p1,p2,p3)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS4(cinfo,lvl,code,p1,p2,p3,p4)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS5(cinfo,lvl,code,p1,p2,p3,p4,p5)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   _mp[4] = (p5); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMS8(cinfo,lvl,code,p1,p2,p3,p4,p5,p6,p7,p8)  \
+  MAKESTMT(int * _mp = (cinfo)->err->msg_parm.i; \
+	   _mp[0] = (p1); _mp[1] = (p2); _mp[2] = (p3); _mp[3] = (p4); \
+	   _mp[4] = (p5); _mp[5] = (p6); _mp[6] = (p7); _mp[7] = (p8); \
+	   (cinfo)->err->msg_code = (code); \
+	   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)); )
+#define TRACEMSS(cinfo,lvl,code,str)  \
+  ((cinfo)->err->msg_code = (code), \
+   strncpy((cinfo)->err->msg_parm.s, (str), JMSG_STR_PARM_MAX), \
+   (*(cinfo)->err->emit_message) ((j_common_ptr) (cinfo), (lvl)))
+
+#endif /* JERROR_H */

Added: trunk/code/jpeg-8c/jfdctflt.c
===================================================================
--- trunk/code/jpeg-8c/jfdctflt.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jfdctflt.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,174 @@
+/*
+ * jfdctflt.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2003-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a floating-point implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * This implementation should be more accurate than either of the integer
+ * DCT implementations.  However, it may not give the same results on all
+ * machines because of differences in roundoff behavior.  Speed will depend
+ * on the hardware's floating point capacity.
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with a fixed-point
+ * implementation, accuracy is lost due to imprecise representation of the
+ * scaled quantization values.  However, that problem does not arise if
+ * we use floating point arithmetic.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_float (FAST_FLOAT * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
+  FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
+  FAST_FLOAT *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+
+  /* Pass 1: process rows. */
+
+  dataptr = data;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Load data into workspace */
+    tmp0 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]));
+    tmp7 = (FAST_FLOAT) (GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]));
+    tmp1 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]));
+    tmp6 = (FAST_FLOAT) (GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]));
+    tmp2 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]));
+    tmp5 = (FAST_FLOAT) (GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]));
+    tmp3 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]));
+    tmp4 = (FAST_FLOAT) (GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]));
+
+    /* Even part */
+
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */
+    dataptr[4] = tmp10 - tmp11;
+
+    z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
+    dataptr[2] = tmp13 + z1;	/* phase 5 */
+    dataptr[6] = tmp13 - z1;
+
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
+    z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
+    z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
+    z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[5] = z13 + z2;	/* phase 6 */
+    dataptr[3] = z13 - z2;
+    dataptr[1] = z11 + z4;
+    dataptr[7] = z11 - z4;
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+
+    /* Even part */
+
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[DCTSIZE*4] = tmp10 - tmp11;
+
+    z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
+    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
+    dataptr[DCTSIZE*6] = tmp13 - z1;
+
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
+    z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
+    z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
+    z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
+    dataptr[DCTSIZE*3] = z13 - z2;
+    dataptr[DCTSIZE*1] = z11 + z4;
+    dataptr[DCTSIZE*7] = z11 - z4;
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */

Added: trunk/code/jpeg-8c/jfdctfst.c
===================================================================
--- trunk/code/jpeg-8c/jfdctfst.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jfdctfst.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,230 @@
+/*
+ * jfdctfst.c
+ *
+ * Copyright (C) 1994-1996, Thomas G. Lane.
+ * Modified 2003-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a fast, not so accurate integer implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with fixed-point math,
+ * accuracy is lost due to imprecise representation of the scaled
+ * quantization values.  The smaller the quantization table entry, the less
+ * precise the scaled value, so this implementation does worse with high-
+ * quality-setting files than with low-quality ones.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_IFAST_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling decisions are generally the same as in the LL&M algorithm;
+ * see jfdctint.c for more details.  However, we choose to descale
+ * (right shift) multiplication products as soon as they are formed,
+ * rather than carrying additional fractional bits into subsequent additions.
+ * This compromises accuracy slightly, but it lets us save a few shifts.
+ * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
+ * everywhere except in the multiplications proper; this saves a good deal
+ * of work on 16-bit-int machines.
+ *
+ * Again to save a few shifts, the intermediate results between pass 1 and
+ * pass 2 are not upscaled, but are represented only to integral precision.
+ *
+ * A final compromise is to represent the multiplicative constants to only
+ * 8 fractional bits, rather than 13.  This saves some shifting work on some
+ * machines, and may also reduce the cost of multiplication (since there
+ * are fewer one-bits in the constants).
+ */
+
+#define CONST_BITS  8
+
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 8
+#define FIX_0_382683433  ((INT32)   98)		/* FIX(0.382683433) */
+#define FIX_0_541196100  ((INT32)  139)		/* FIX(0.541196100) */
+#define FIX_0_707106781  ((INT32)  181)		/* FIX(0.707106781) */
+#define FIX_1_306562965  ((INT32)  334)		/* FIX(1.306562965) */
+#else
+#define FIX_0_382683433  FIX(0.382683433)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_707106781  FIX(0.707106781)
+#define FIX_1_306562965  FIX(1.306562965)
+#endif
+
+
+/* We can gain a little more speed, with a further compromise in accuracy,
+ * by omitting the addition in a descaling shift.  This yields an incorrectly
+ * rounded result half the time...
+ */
+
+#ifndef USE_ACCURATE_ROUNDING
+#undef DESCALE
+#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
+#endif
+
+
+/* Multiply a DCTELEM variable by an INT32 constant, and immediately
+ * descale to yield a DCTELEM result.
+ */
+
+#define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_ifast (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  DCTELEM tmp10, tmp11, tmp12, tmp13;
+  DCTELEM z1, z2, z3, z4, z5, z11, z13;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+
+  dataptr = data;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Load data into workspace */
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
+    tmp7 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
+    tmp6 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
+    tmp5 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
+    tmp4 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
+
+    /* Even part */
+
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = tmp10 + tmp11 - 8 * CENTERJSAMPLE; /* phase 3 */
+    dataptr[4] = tmp10 - tmp11;
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
+    dataptr[2] = tmp13 + z1;	/* phase 5 */
+    dataptr[6] = tmp13 - z1;
+
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
+    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
+    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
+    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[5] = z13 + z2;	/* phase 6 */
+    dataptr[3] = z13 - z2;
+    dataptr[1] = z11 + z4;
+    dataptr[7] = z11 - z4;
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns. */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+
+    /* Even part */
+
+    tmp10 = tmp0 + tmp3;	/* phase 2 */
+    tmp13 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp1 - tmp2;
+
+    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
+    dataptr[DCTSIZE*4] = tmp10 - tmp11;
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
+    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
+    dataptr[DCTSIZE*6] = tmp13 - z1;
+
+    /* Odd part */
+
+    tmp10 = tmp4 + tmp5;	/* phase 2 */
+    tmp11 = tmp5 + tmp6;
+    tmp12 = tmp6 + tmp7;
+
+    /* The rotator is modified from fig 4-8 to avoid extra negations. */
+    z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
+    z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
+    z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
+    z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
+
+    z11 = tmp7 + z3;		/* phase 5 */
+    z13 = tmp7 - z3;
+
+    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
+    dataptr[DCTSIZE*3] = z13 - z2;
+    dataptr[DCTSIZE*1] = z11 + z4;
+    dataptr[DCTSIZE*7] = z11 - z4;
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#endif /* DCT_IFAST_SUPPORTED */

Added: trunk/code/jpeg-8c/jfdctint.c
===================================================================
--- trunk/code/jpeg-8c/jfdctint.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jfdctint.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,4348 @@
+/*
+ * jfdctint.c
+ *
+ * Copyright (C) 1991-1996, Thomas G. Lane.
+ * Modification developed 2003-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a slow-but-accurate integer implementation of the
+ * forward DCT (Discrete Cosine Transform).
+ *
+ * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
+ * on each column.  Direct algorithms are also available, but they are
+ * much more complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on an algorithm described in
+ *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
+ *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
+ *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
+ * The primary algorithm described there uses 11 multiplies and 29 adds.
+ * We use their alternate method with 12 multiplies and 32 adds.
+ * The advantage of this method is that no data path contains more than one
+ * multiplication; this allows a very simple and accurate implementation in
+ * scaled fixed-point arithmetic, with a minimal number of shifts.
+ *
+ * We also provide FDCT routines with various input sample block sizes for
+ * direct resolution reduction or enlargement and for direct resolving the
+ * common 2x1 and 1x2 subsampling cases without additional resampling: NxN
+ * (N=1...16), 2NxN, and Nx2N (N=1...8) pixels for one 8x8 output DCT block.
+ *
+ * For N<8 we fill the remaining block coefficients with zero.
+ * For N>8 we apply a partial N-point FDCT on the input samples, computing
+ * just the lower 8 frequency coefficients and discarding the rest.
+ *
+ * We must scale the output coefficients of the N-point FDCT appropriately
+ * to the standard 8-point FDCT level by 8/N per 1-D pass.  This scaling
+ * is folded into the constant multipliers (pass 2) and/or final/initial
+ * shifting.
+ *
+ * CAUTION: We rely on the FIX() macro except for the N=1,2,4,8 cases
+ * since there would be too many additional constants to pre-calculate.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_ISLOW_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCT blocks. /* deliberate syntax err */
+#endif
+
+
+/*
+ * The poop on this scaling stuff is as follows:
+ *
+ * Each 1-D DCT step produces outputs which are a factor of sqrt(N)
+ * larger than the true DCT outputs.  The final outputs are therefore
+ * a factor of N larger than desired; since N=8 this can be cured by
+ * a simple right shift at the end of the algorithm.  The advantage of
+ * this arrangement is that we save two multiplications per 1-D DCT,
+ * because the y0 and y4 outputs need not be divided by sqrt(N).
+ * In the IJG code, this factor of 8 is removed by the quantization step
+ * (in jcdctmgr.c), NOT in this module.
+ *
+ * We have to do addition and subtraction of the integer inputs, which
+ * is no problem, and multiplication by fractional constants, which is
+ * a problem to do in integer arithmetic.  We multiply all the constants
+ * by CONST_SCALE and convert them to integer constants (thus retaining
+ * CONST_BITS bits of precision in the constants).  After doing a
+ * multiplication we have to divide the product by CONST_SCALE, with proper
+ * rounding, to produce the correct output.  This division can be done
+ * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
+ * as long as possible so that partial sums can be added together with
+ * full fractional precision.
+ *
+ * The outputs of the first pass are scaled up by PASS1_BITS bits so that
+ * they are represented to better-than-integral precision.  These outputs
+ * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
+ * with the recommended scaling.  (For 12-bit sample data, the intermediate
+ * array is INT32 anyway.)
+ *
+ * To avoid overflow of the 32-bit intermediate results in pass 2, we must
+ * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
+ * shows that the values given below are the most effective.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_298631336  ((INT32)  2446)	/* FIX(0.298631336) */
+#define FIX_0_390180644  ((INT32)  3196)	/* FIX(0.390180644) */
+#define FIX_0_541196100  ((INT32)  4433)	/* FIX(0.541196100) */
+#define FIX_0_765366865  ((INT32)  6270)	/* FIX(0.765366865) */
+#define FIX_0_899976223  ((INT32)  7373)	/* FIX(0.899976223) */
+#define FIX_1_175875602  ((INT32)  9633)	/* FIX(1.175875602) */
+#define FIX_1_501321110  ((INT32)  12299)	/* FIX(1.501321110) */
+#define FIX_1_847759065  ((INT32)  15137)	/* FIX(1.847759065) */
+#define FIX_1_961570560  ((INT32)  16069)	/* FIX(1.961570560) */
+#define FIX_2_053119869  ((INT32)  16819)	/* FIX(2.053119869) */
+#define FIX_2_562915447  ((INT32)  20995)	/* FIX(2.562915447) */
+#define FIX_3_072711026  ((INT32)  25172)	/* FIX(3.072711026) */
+#else
+#define FIX_0_298631336  FIX(0.298631336)
+#define FIX_0_390180644  FIX(0.390180644)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_175875602  FIX(1.175875602)
+#define FIX_1_501321110  FIX(1.501321110)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_1_961570560  FIX(1.961570560)
+#define FIX_2_053119869  FIX(2.053119869)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_072711026  FIX(3.072711026)
+#endif
+
+
+/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
+#else
+#define MULTIPLY(var,const)  ((var) * (const))
+#endif
+
+
+/*
+ * Perform the forward DCT on one block of samples.
+ */
+
+GLOBAL(void)
+jpeg_fdct_islow (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3;
+  INT32 tmp10, tmp11, tmp12, tmp13;
+  INT32 z1;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  dataptr = data;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
+
+    tmp10 = tmp0 + tmp3;
+    tmp12 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp13 = tmp1 - tmp2;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM) ((tmp10 + tmp11 - 8 * CENTERJSAMPLE) << PASS1_BITS);
+    dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-1);
+    dataptr[2] = (DCTELEM) RIGHT_SHIFT(z1 + MULTIPLY(tmp12, FIX_0_765366865),
+				       CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM) RIGHT_SHIFT(z1 - MULTIPLY(tmp13, FIX_1_847759065),
+				       CONST_BITS-PASS1_BITS);
+
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * cK represents sqrt(2) * cos(K*pi/16).
+     * i0..i3 in the paper are tmp0..tmp3 here.
+     */
+
+    tmp10 = tmp0 + tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp1 + tmp3;
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
+    tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
+    tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
+    tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
+    tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
+    tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
+    tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
+    tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
+
+    tmp12 += z1;
+    tmp13 += z1;
+
+    dataptr[1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + tmp10 + tmp12, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM)
+      RIGHT_SHIFT(tmp1 + tmp11 + tmp13, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM)
+      RIGHT_SHIFT(tmp2 + tmp11 + tmp12, CONST_BITS-PASS1_BITS);
+    dataptr[7] = (DCTELEM)
+      RIGHT_SHIFT(tmp3 + tmp10 + tmp13, CONST_BITS-PASS1_BITS);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+
+    /* Add fudge factor here for final descale. */
+    tmp10 = tmp0 + tmp3 + (ONE << (PASS1_BITS-1));
+    tmp12 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp13 = tmp1 - tmp2;
+
+    tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+
+    dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp10 + tmp11, PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM) RIGHT_SHIFT(tmp10 - tmp11, PASS1_BITS);
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS+PASS1_BITS-1);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      RIGHT_SHIFT(z1 + MULTIPLY(tmp12, FIX_0_765366865), CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      RIGHT_SHIFT(z1 - MULTIPLY(tmp13, FIX_1_847759065), CONST_BITS+PASS1_BITS);
+
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * cK represents sqrt(2) * cos(K*pi/16).
+     * i0..i3 in the paper are tmp0..tmp3 here.
+     */
+
+    tmp10 = tmp0 + tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp1 + tmp3;
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS+PASS1_BITS-1);
+
+    tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
+    tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
+    tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
+    tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
+    tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
+    tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
+    tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
+    tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
+
+    tmp12 += z1;
+    tmp13 += z1;
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + tmp10 + tmp12, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      RIGHT_SHIFT(tmp1 + tmp11 + tmp13, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      RIGHT_SHIFT(tmp2 + tmp11 + tmp12, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*7] = (DCTELEM)
+      RIGHT_SHIFT(tmp3 + tmp10 + tmp13, CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+#ifdef DCT_SCALING_SUPPORTED
+
+
+/*
+ * Perform the forward DCT on a 7x7 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_7x7 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3;
+  INT32 tmp10, tmp11, tmp12;
+  INT32 z1, z2, z3;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* cK represents sqrt(2) * cos(K*pi/14). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 7; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[6]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[5]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[4]);
+    tmp3 = GETJSAMPLE(elemptr[3]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[6]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[5]);
+    tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[4]);
+
+    z1 = tmp0 + tmp2;
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((z1 + tmp1 + tmp3 - 7 * CENTERJSAMPLE) << PASS1_BITS);
+    tmp3 += tmp3;
+    z1 -= tmp3;
+    z1 -= tmp3;
+    z1 = MULTIPLY(z1, FIX(0.353553391));                /* (c2+c6-c4)/2 */
+    z2 = MULTIPLY(tmp0 - tmp2, FIX(0.920609002));       /* (c2+c4-c6)/2 */
+    z3 = MULTIPLY(tmp1 - tmp2, FIX(0.314692123));       /* c6 */
+    dataptr[2] = (DCTELEM) DESCALE(z1 + z2 + z3, CONST_BITS-PASS1_BITS);
+    z1 -= z2;
+    z2 = MULTIPLY(tmp0 - tmp1, FIX(0.881747734));       /* c4 */
+    dataptr[4] = (DCTELEM)
+      DESCALE(z2 + z3 - MULTIPLY(tmp1 - tmp3, FIX(0.707106781)), /* c2+c6-c4 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp1 = MULTIPLY(tmp10 + tmp11, FIX(0.935414347));   /* (c3+c1-c5)/2 */
+    tmp2 = MULTIPLY(tmp10 - tmp11, FIX(0.170262339));   /* (c3+c5-c1)/2 */
+    tmp0 = tmp1 - tmp2;
+    tmp1 += tmp2;
+    tmp2 = MULTIPLY(tmp11 + tmp12, - FIX(1.378756276)); /* -c1 */
+    tmp1 += tmp2;
+    tmp3 = MULTIPLY(tmp10 + tmp12, FIX(0.613604268));   /* c5 */
+    tmp0 += tmp3;
+    tmp2 += tmp3 + MULTIPLY(tmp12, FIX(1.870828693));   /* c3+c1-c5 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS-PASS1_BITS);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/7)**2 = 64/49, which we fold
+   * into the constant multipliers:
+   * cK now represents sqrt(2) * cos(K*pi/14) * 64/49.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 7; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*6];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*5];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*4];
+    tmp3 = dataptr[DCTSIZE*3];
+
+    tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*6];
+    tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*5];
+    tmp12 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*4];
+
+    z1 = tmp0 + tmp2;
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(z1 + tmp1 + tmp3, FIX(1.306122449)), /* 64/49 */
+	      CONST_BITS+PASS1_BITS);
+    tmp3 += tmp3;
+    z1 -= tmp3;
+    z1 -= tmp3;
+    z1 = MULTIPLY(z1, FIX(0.461784020));                /* (c2+c6-c4)/2 */
+    z2 = MULTIPLY(tmp0 - tmp2, FIX(1.202428084));       /* (c2+c4-c6)/2 */
+    z3 = MULTIPLY(tmp1 - tmp2, FIX(0.411026446));       /* c6 */
+    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + z2 + z3, CONST_BITS+PASS1_BITS);
+    z1 -= z2;
+    z2 = MULTIPLY(tmp0 - tmp1, FIX(1.151670509));       /* c4 */
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(z2 + z3 - MULTIPLY(tmp1 - tmp3, FIX(0.923568041)), /* c2+c6-c4 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.221765677));   /* (c3+c1-c5)/2 */
+    tmp2 = MULTIPLY(tmp10 - tmp11, FIX(0.222383464));   /* (c3+c5-c1)/2 */
+    tmp0 = tmp1 - tmp2;
+    tmp1 += tmp2;
+    tmp2 = MULTIPLY(tmp11 + tmp12, - FIX(1.800824523)); /* -c1 */
+    tmp1 += tmp2;
+    tmp3 = MULTIPLY(tmp10 + tmp12, FIX(0.801442310));   /* c5 */
+    tmp0 += tmp3;
+    tmp2 += tmp3 + MULTIPLY(tmp12, FIX(2.443531355));   /* c3+c1-c5 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 6x6 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_6x6 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2;
+  INT32 tmp10, tmp11, tmp12;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* cK represents sqrt(2) * cos(K*pi/12). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 6; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[5]);
+    tmp11 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[4]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[3]);
+
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[5]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[4]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[3]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 - 6 * CENTERJSAMPLE) << PASS1_BITS);
+    dataptr[2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp12, FIX(1.224744871)),                 /* c2 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(0.707106781)), /* c4 */
+	      CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = DESCALE(MULTIPLY(tmp0 + tmp2, FIX(0.366025404)),     /* c5 */
+		    CONST_BITS-PASS1_BITS);
+
+    dataptr[1] = (DCTELEM) (tmp10 + ((tmp0 + tmp1) << PASS1_BITS));
+    dataptr[3] = (DCTELEM) ((tmp0 - tmp1 - tmp2) << PASS1_BITS);
+    dataptr[5] = (DCTELEM) (tmp10 + ((tmp2 - tmp1) << PASS1_BITS));
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/6)**2 = 16/9, which we fold
+   * into the constant multipliers:
+   * cK now represents sqrt(2) * cos(K*pi/12) * 16/9.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 6; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*5];
+    tmp11 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
+
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+
+    tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*5];
+    tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11, FIX(1.777777778)),         /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp12, FIX(2.177324216)),                 /* c2 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(1.257078722)), /* c4 */
+	      CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp0 + tmp2, FIX(0.650711829));             /* c5 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),   /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp1 - tmp2, FIX(1.777777778)),    /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp2 - tmp1, FIX(1.777777778)),   /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 5x5 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_5x5 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2;
+  INT32 tmp10, tmp11;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* We scale the results further by 2 as part of output adaption */
+  /* scaling for different DCT size. */
+  /* cK represents sqrt(2) * cos(K*pi/10). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 5; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[4]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[3]);
+    tmp2 = GETJSAMPLE(elemptr[2]);
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[4]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[3]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp2 - 5 * CENTERJSAMPLE) << (PASS1_BITS+1));
+    tmp11 = MULTIPLY(tmp11, FIX(0.790569415));          /* (c2+c4)/2 */
+    tmp10 -= tmp2 << 2;
+    tmp10 = MULTIPLY(tmp10, FIX(0.353553391));          /* (c2-c4)/2 */
+    dataptr[2] = (DCTELEM) DESCALE(tmp11 + tmp10, CONST_BITS-PASS1_BITS-1);
+    dataptr[4] = (DCTELEM) DESCALE(tmp11 - tmp10, CONST_BITS-PASS1_BITS-1);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp0 + tmp1, FIX(0.831253876));    /* c3 */
+
+    dataptr[1] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp0, FIX(0.513743148)), /* c1-c3 */
+	      CONST_BITS-PASS1_BITS-1);
+    dataptr[3] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp1, FIX(2.176250899)), /* c1+c3 */
+	      CONST_BITS-PASS1_BITS-1);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/5)**2 = 64/25, which we partially
+   * fold into the constant multipliers (other part was done in pass 1):
+   * cK now represents sqrt(2) * cos(K*pi/10) * 32/25.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 5; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*4];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*3];
+    tmp2 = dataptr[DCTSIZE*2];
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+
+    tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*4];
+    tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*3];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp2, FIX(1.28)),        /* 32/25 */
+	      CONST_BITS+PASS1_BITS);
+    tmp11 = MULTIPLY(tmp11, FIX(1.011928851));          /* (c2+c4)/2 */
+    tmp10 -= tmp2 << 2;
+    tmp10 = MULTIPLY(tmp10, FIX(0.452548340));          /* (c2-c4)/2 */
+    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(tmp11 + tmp10, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp11 - tmp10, CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp0 + tmp1, FIX(1.064004961));    /* c3 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp0, FIX(0.657591230)), /* c1-c3 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp1, FIX(2.785601151)), /* c1+c3 */
+	      CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 4x4 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_4x4 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1;
+  INT32 tmp10, tmp11;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* We must also scale the output by (8/4)**2 = 2**2, which we add here. */
+  /* cK represents sqrt(2) * cos(K*pi/16) [refers to 8-point FDCT]. */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 4; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[3]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[2]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[3]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[2]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp0 + tmp1 - 4 * CENTERJSAMPLE) << (PASS1_BITS+2));
+    dataptr[2] = (DCTELEM) ((tmp0 - tmp1) << (PASS1_BITS+2));
+
+    /* Odd part */
+
+    tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS-PASS1_BITS-3);
+
+    dataptr[1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
+		  CONST_BITS-PASS1_BITS-2);
+    dataptr[3] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
+		  CONST_BITS-PASS1_BITS-2);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 4; ctr++) {
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*3] + (ONE << (PASS1_BITS-1));
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*2];
+
+    tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*3];
+    tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*2];
+
+    dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp0 + tmp1, PASS1_BITS);
+    dataptr[DCTSIZE*2] = (DCTELEM) RIGHT_SHIFT(tmp0 - tmp1, PASS1_BITS);
+
+    /* Odd part */
+
+    tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS+PASS1_BITS-1);
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
+		  CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
+		  CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 3x3 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_3x3 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* We scale the results further by 2**2 as part of output adaption */
+  /* scaling for different DCT size. */
+  /* cK represents sqrt(2) * cos(K*pi/6). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 3; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[2]);
+    tmp1 = GETJSAMPLE(elemptr[1]);
+
+    tmp2 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[2]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp0 + tmp1 - 3 * CENTERJSAMPLE) << (PASS1_BITS+2));
+    dataptr[2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp1 - tmp1, FIX(0.707106781)), /* c2 */
+	      CONST_BITS-PASS1_BITS-2);
+
+    /* Odd part */
+
+    dataptr[1] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp2, FIX(1.224744871)),               /* c1 */
+	      CONST_BITS-PASS1_BITS-2);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/3)**2 = 64/9, which we partially
+   * fold into the constant multipliers (other part was done in pass 1):
+   * cK now represents sqrt(2) * cos(K*pi/6) * 16/9.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 3; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*2];
+    tmp1 = dataptr[DCTSIZE*1];
+
+    tmp2 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*2];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),        /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp1 - tmp1, FIX(1.257078722)), /* c2 */
+	      CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp2, FIX(2.177324216)),               /* c1 */
+	      CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 2x2 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_2x2 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3;
+  JSAMPROW elemptr;
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT. */
+
+  /* Row 0 */
+  elemptr = sample_data[0] + start_col;
+
+  tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[1]);
+  tmp1 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[1]);
+
+  /* Row 1 */
+  elemptr = sample_data[1] + start_col;
+
+  tmp2 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[1]);
+  tmp3 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[1]);
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/2)**2 = 2**4.
+   */
+
+  /* Column 0 */
+  /* Apply unsigned->signed conversion */
+  data[DCTSIZE*0] = (DCTELEM) ((tmp0 + tmp2 - 4 * CENTERJSAMPLE) << 4);
+  data[DCTSIZE*1] = (DCTELEM) ((tmp0 - tmp2) << 4);
+
+  /* Column 1 */
+  data[DCTSIZE*0+1] = (DCTELEM) ((tmp1 + tmp3) << 4);
+  data[DCTSIZE*1+1] = (DCTELEM) ((tmp1 - tmp3) << 4);
+}
+
+
+/*
+ * Perform the forward DCT on a 1x1 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_1x1 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* We leave the result scaled up by an overall factor of 8. */
+  /* We must also scale the output by (8/1)**2 = 2**6. */
+  /* Apply unsigned->signed conversion */
+  data[0] = (DCTELEM)
+    ((GETJSAMPLE(sample_data[0][start_col]) - CENTERJSAMPLE) << 6);
+}
+
+
+/*
+ * Perform the forward DCT on a 9x9 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_9x9 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  INT32 tmp10, tmp11, tmp12, tmp13;
+  INT32 z1, z2;
+  DCTELEM workspace[8];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* we scale the results further by 2 as part of output adaption */
+  /* scaling for different DCT size. */
+  /* cK represents sqrt(2) * cos(K*pi/18). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[8]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[7]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[6]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[5]);
+    tmp4 = GETJSAMPLE(elemptr[4]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[8]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[7]);
+    tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[6]);
+    tmp13 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[5]);
+
+    z1 = tmp0 + tmp2 + tmp3;
+    z2 = tmp1 + tmp4;
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM) ((z1 + z2 - 9 * CENTERJSAMPLE) << 1);
+    dataptr[6] = (DCTELEM)
+      DESCALE(MULTIPLY(z1 - z2 - z2, FIX(0.707106781)),  /* c6 */
+	      CONST_BITS-1);
+    z1 = MULTIPLY(tmp0 - tmp2, FIX(1.328926049));        /* c2 */
+    z2 = MULTIPLY(tmp1 - tmp4 - tmp4, FIX(0.707106781)); /* c6 */
+    dataptr[2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp2 - tmp3, FIX(1.083350441))    /* c4 */
+	      + z1 + z2, CONST_BITS-1);
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp3 - tmp0, FIX(0.245575608))    /* c8 */
+	      + z1 - z2, CONST_BITS-1);
+
+    /* Odd part */
+
+    dataptr[3] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12 - tmp13, FIX(1.224744871)), /* c3 */
+	      CONST_BITS-1);
+
+    tmp11 = MULTIPLY(tmp11, FIX(1.224744871));        /* c3 */
+    tmp0 = MULTIPLY(tmp10 + tmp12, FIX(0.909038955)); /* c5 */
+    tmp1 = MULTIPLY(tmp10 + tmp13, FIX(0.483689525)); /* c7 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp11 + tmp0 + tmp1, CONST_BITS-1);
+
+    tmp2 = MULTIPLY(tmp12 - tmp13, FIX(1.392728481)); /* c1 */
+
+    dataptr[5] = (DCTELEM) DESCALE(tmp0 - tmp11 - tmp2, CONST_BITS-1);
+    dataptr[7] = (DCTELEM) DESCALE(tmp1 - tmp11 + tmp2, CONST_BITS-1);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 9)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/9)**2 = 64/81, which we partially
+   * fold into the constant multipliers and final/initial shifting:
+   * cK now represents sqrt(2) * cos(K*pi/18) * 128/81.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*0];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*7];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*6];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*5];
+    tmp4 = dataptr[DCTSIZE*4];
+
+    tmp10 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*0];
+    tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*7];
+    tmp12 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*6];
+    tmp13 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*5];
+
+    z1 = tmp0 + tmp2 + tmp3;
+    z2 = tmp1 + tmp4;
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(z1 + z2, FIX(1.580246914)),       /* 128/81 */
+	      CONST_BITS+2);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(MULTIPLY(z1 - z2 - z2, FIX(1.117403309)),  /* c6 */
+	      CONST_BITS+2);
+    z1 = MULTIPLY(tmp0 - tmp2, FIX(2.100031287));        /* c2 */
+    z2 = MULTIPLY(tmp1 - tmp4 - tmp4, FIX(1.117403309)); /* c6 */
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp2 - tmp3, FIX(1.711961190))    /* c4 */
+	      + z1 + z2, CONST_BITS+2);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp3 - tmp0, FIX(0.388070096))    /* c8 */
+	      + z1 - z2, CONST_BITS+2);
+
+    /* Odd part */
+
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12 - tmp13, FIX(1.935399303)), /* c3 */
+	      CONST_BITS+2);
+
+    tmp11 = MULTIPLY(tmp11, FIX(1.935399303));        /* c3 */
+    tmp0 = MULTIPLY(tmp10 + tmp12, FIX(1.436506004)); /* c5 */
+    tmp1 = MULTIPLY(tmp10 + tmp13, FIX(0.764348879)); /* c7 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(tmp11 + tmp0 + tmp1, CONST_BITS+2);
+
+    tmp2 = MULTIPLY(tmp12 - tmp13, FIX(2.200854883)); /* c1 */
+
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      DESCALE(tmp0 - tmp11 - tmp2, CONST_BITS+2);
+    dataptr[DCTSIZE*7] = (DCTELEM)
+      DESCALE(tmp1 - tmp11 + tmp2, CONST_BITS+2);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 10x10 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_10x10 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
+  DCTELEM workspace[8*2];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* we scale the results further by 2 as part of output adaption */
+  /* scaling for different DCT size. */
+  /* cK represents sqrt(2) * cos(K*pi/20). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[9]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[8]);
+    tmp12 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[7]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[6]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[5]);
+
+    tmp10 = tmp0 + tmp4;
+    tmp13 = tmp0 - tmp4;
+    tmp11 = tmp1 + tmp3;
+    tmp14 = tmp1 - tmp3;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[9]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[8]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[7]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[6]);
+    tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[5]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 + tmp12 - 10 * CENTERJSAMPLE) << 1);
+    tmp12 += tmp12;
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.144122806)) - /* c4 */
+	      MULTIPLY(tmp11 - tmp12, FIX(0.437016024)),  /* c8 */
+	      CONST_BITS-1);
+    tmp10 = MULTIPLY(tmp13 + tmp14, FIX(0.831253876));    /* c6 */
+    dataptr[2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp13, FIX(0.513743148)),  /* c2-c6 */
+	      CONST_BITS-1);
+    dataptr[6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp14, FIX(2.176250899)),  /* c2+c6 */
+	      CONST_BITS-1);
+
+    /* Odd part */
+
+    tmp10 = tmp0 + tmp4;
+    tmp11 = tmp1 - tmp3;
+    dataptr[5] = (DCTELEM) ((tmp10 - tmp11 - tmp2) << 1);
+    tmp2 <<= CONST_BITS;
+    dataptr[1] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0, FIX(1.396802247)) +          /* c1 */
+	      MULTIPLY(tmp1, FIX(1.260073511)) + tmp2 +   /* c3 */
+	      MULTIPLY(tmp3, FIX(0.642039522)) +          /* c7 */
+	      MULTIPLY(tmp4, FIX(0.221231742)),           /* c9 */
+	      CONST_BITS-1);
+    tmp12 = MULTIPLY(tmp0 - tmp4, FIX(0.951056516)) -     /* (c3+c7)/2 */
+	    MULTIPLY(tmp1 + tmp3, FIX(0.587785252));      /* (c1-c9)/2 */
+    tmp13 = MULTIPLY(tmp10 + tmp11, FIX(0.309016994)) +   /* (c3-c7)/2 */
+	    (tmp11 << (CONST_BITS - 1)) - tmp2;
+    dataptr[3] = (DCTELEM) DESCALE(tmp12 + tmp13, CONST_BITS-1);
+    dataptr[7] = (DCTELEM) DESCALE(tmp12 - tmp13, CONST_BITS-1);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 10)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/10)**2 = 16/25, which we partially
+   * fold into the constant multipliers and final/initial shifting:
+   * cK now represents sqrt(2) * cos(K*pi/20) * 32/25.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*1];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*0];
+    tmp12 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*7];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*6];
+    tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
+
+    tmp10 = tmp0 + tmp4;
+    tmp13 = tmp0 - tmp4;
+    tmp11 = tmp1 + tmp3;
+    tmp14 = tmp1 - tmp3;
+
+    tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*1];
+    tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*0];
+    tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*7];
+    tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*6];
+    tmp4 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12, FIX(1.28)), /* 32/25 */
+	      CONST_BITS+2);
+    tmp12 += tmp12;
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.464477191)) - /* c4 */
+	      MULTIPLY(tmp11 - tmp12, FIX(0.559380511)),  /* c8 */
+	      CONST_BITS+2);
+    tmp10 = MULTIPLY(tmp13 + tmp14, FIX(1.064004961));    /* c6 */
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp13, FIX(0.657591230)),  /* c2-c6 */
+	      CONST_BITS+2);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp14, FIX(2.785601151)),  /* c2+c6 */
+	      CONST_BITS+2);
+
+    /* Odd part */
+
+    tmp10 = tmp0 + tmp4;
+    tmp11 = tmp1 - tmp3;
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp11 - tmp2, FIX(1.28)),  /* 32/25 */
+	      CONST_BITS+2);
+    tmp2 = MULTIPLY(tmp2, FIX(1.28));                     /* 32/25 */
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0, FIX(1.787906876)) +          /* c1 */
+	      MULTIPLY(tmp1, FIX(1.612894094)) + tmp2 +   /* c3 */
+	      MULTIPLY(tmp3, FIX(0.821810588)) +          /* c7 */
+	      MULTIPLY(tmp4, FIX(0.283176630)),           /* c9 */
+	      CONST_BITS+2);
+    tmp12 = MULTIPLY(tmp0 - tmp4, FIX(1.217352341)) -     /* (c3+c7)/2 */
+	    MULTIPLY(tmp1 + tmp3, FIX(0.752365123));      /* (c1-c9)/2 */
+    tmp13 = MULTIPLY(tmp10 + tmp11, FIX(0.395541753)) +   /* (c3-c7)/2 */
+	    MULTIPLY(tmp11, FIX(0.64)) - tmp2;            /* 16/25 */
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp12 + tmp13, CONST_BITS+2);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp12 - tmp13, CONST_BITS+2);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on an 11x11 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_11x11 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
+  INT32 z1, z2, z3;
+  DCTELEM workspace[8*3];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* we scale the results further by 2 as part of output adaption */
+  /* scaling for different DCT size. */
+  /* cK represents sqrt(2) * cos(K*pi/22). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[10]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[9]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[8]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[7]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[6]);
+    tmp5 = GETJSAMPLE(elemptr[5]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[10]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[9]);
+    tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[8]);
+    tmp13 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[7]);
+    tmp14 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[6]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp0 + tmp1 + tmp2 + tmp3 + tmp4 + tmp5 - 11 * CENTERJSAMPLE) << 1);
+    tmp5 += tmp5;
+    tmp0 -= tmp5;
+    tmp1 -= tmp5;
+    tmp2 -= tmp5;
+    tmp3 -= tmp5;
+    tmp4 -= tmp5;
+    z1 = MULTIPLY(tmp0 + tmp3, FIX(1.356927976)) +       /* c2 */
+	 MULTIPLY(tmp2 + tmp4, FIX(0.201263574));        /* c10 */
+    z2 = MULTIPLY(tmp1 - tmp3, FIX(0.926112931));        /* c6 */
+    z3 = MULTIPLY(tmp0 - tmp1, FIX(1.189712156));        /* c4 */
+    dataptr[2] = (DCTELEM)
+      DESCALE(z1 + z2 - MULTIPLY(tmp3, FIX(1.018300590)) /* c2+c8-c6 */
+	      - MULTIPLY(tmp4, FIX(1.390975730)),        /* c4+c10 */
+	      CONST_BITS-1);
+    dataptr[4] = (DCTELEM)
+      DESCALE(z2 + z3 + MULTIPLY(tmp1, FIX(0.062335650)) /* c4-c6-c10 */
+	      - MULTIPLY(tmp2, FIX(1.356927976))         /* c2 */
+	      + MULTIPLY(tmp4, FIX(0.587485545)),        /* c8 */
+	      CONST_BITS-1);
+    dataptr[6] = (DCTELEM)
+      DESCALE(z1 + z3 - MULTIPLY(tmp0, FIX(1.620527200)) /* c2+c4-c6 */
+	      - MULTIPLY(tmp2, FIX(0.788749120)),        /* c8+c10 */
+	      CONST_BITS-1);
+
+    /* Odd part */
+
+    tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.286413905));    /* c3 */
+    tmp2 = MULTIPLY(tmp10 + tmp12, FIX(1.068791298));    /* c5 */
+    tmp3 = MULTIPLY(tmp10 + tmp13, FIX(0.764581576));    /* c7 */
+    tmp0 = tmp1 + tmp2 + tmp3 - MULTIPLY(tmp10, FIX(1.719967871)) /* c7+c5+c3-c1 */
+	   + MULTIPLY(tmp14, FIX(0.398430003));          /* c9 */
+    tmp4 = MULTIPLY(tmp11 + tmp12, - FIX(0.764581576));  /* -c7 */
+    tmp5 = MULTIPLY(tmp11 + tmp13, - FIX(1.399818907));  /* -c1 */
+    tmp1 += tmp4 + tmp5 + MULTIPLY(tmp11, FIX(1.276416582)) /* c9+c7+c1-c3 */
+	    - MULTIPLY(tmp14, FIX(1.068791298));         /* c5 */
+    tmp10 = MULTIPLY(tmp12 + tmp13, FIX(0.398430003));   /* c9 */
+    tmp2 += tmp4 + tmp10 - MULTIPLY(tmp12, FIX(1.989053629)) /* c9+c5+c3-c7 */
+	    + MULTIPLY(tmp14, FIX(1.399818907));         /* c1 */
+    tmp3 += tmp5 + tmp10 + MULTIPLY(tmp13, FIX(1.305598626)) /* c1+c5-c9-c7 */
+	    - MULTIPLY(tmp14, FIX(1.286413905));         /* c3 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS-1);
+    dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS-1);
+    dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS-1);
+    dataptr[7] = (DCTELEM) DESCALE(tmp3, CONST_BITS-1);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 11)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/11)**2 = 64/121, which we partially
+   * fold into the constant multipliers and final/initial shifting:
+   * cK now represents sqrt(2) * cos(K*pi/22) * 128/121.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*2];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*1];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*0];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*7];
+    tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*6];
+    tmp5 = dataptr[DCTSIZE*5];
+
+    tmp10 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*2];
+    tmp11 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*1];
+    tmp12 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*0];
+    tmp13 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*7];
+    tmp14 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*6];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 + tmp1 + tmp2 + tmp3 + tmp4 + tmp5,
+		       FIX(1.057851240)),                /* 128/121 */
+	      CONST_BITS+2);
+    tmp5 += tmp5;
+    tmp0 -= tmp5;
+    tmp1 -= tmp5;
+    tmp2 -= tmp5;
+    tmp3 -= tmp5;
+    tmp4 -= tmp5;
+    z1 = MULTIPLY(tmp0 + tmp3, FIX(1.435427942)) +       /* c2 */
+	 MULTIPLY(tmp2 + tmp4, FIX(0.212906922));        /* c10 */
+    z2 = MULTIPLY(tmp1 - tmp3, FIX(0.979689713));        /* c6 */
+    z3 = MULTIPLY(tmp0 - tmp1, FIX(1.258538479));        /* c4 */
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(z1 + z2 - MULTIPLY(tmp3, FIX(1.077210542)) /* c2+c8-c6 */
+	      - MULTIPLY(tmp4, FIX(1.471445400)),        /* c4+c10 */
+	      CONST_BITS+2);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(z2 + z3 + MULTIPLY(tmp1, FIX(0.065941844)) /* c4-c6-c10 */
+	      - MULTIPLY(tmp2, FIX(1.435427942))         /* c2 */
+	      + MULTIPLY(tmp4, FIX(0.621472312)),        /* c8 */
+	      CONST_BITS+2);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(z1 + z3 - MULTIPLY(tmp0, FIX(1.714276708)) /* c2+c4-c6 */
+	      - MULTIPLY(tmp2, FIX(0.834379234)),        /* c8+c10 */
+	      CONST_BITS+2);
+
+    /* Odd part */
+
+    tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.360834544));    /* c3 */
+    tmp2 = MULTIPLY(tmp10 + tmp12, FIX(1.130622199));    /* c5 */
+    tmp3 = MULTIPLY(tmp10 + tmp13, FIX(0.808813568));    /* c7 */
+    tmp0 = tmp1 + tmp2 + tmp3 - MULTIPLY(tmp10, FIX(1.819470145)) /* c7+c5+c3-c1 */
+	   + MULTIPLY(tmp14, FIX(0.421479672));          /* c9 */
+    tmp4 = MULTIPLY(tmp11 + tmp12, - FIX(0.808813568));  /* -c7 */
+    tmp5 = MULTIPLY(tmp11 + tmp13, - FIX(1.480800167));  /* -c1 */
+    tmp1 += tmp4 + tmp5 + MULTIPLY(tmp11, FIX(1.350258864)) /* c9+c7+c1-c3 */
+	    - MULTIPLY(tmp14, FIX(1.130622199));         /* c5 */
+    tmp10 = MULTIPLY(tmp12 + tmp13, FIX(0.421479672));   /* c9 */
+    tmp2 += tmp4 + tmp10 - MULTIPLY(tmp12, FIX(2.104122847)) /* c9+c5+c3-c7 */
+	    + MULTIPLY(tmp14, FIX(1.480800167));         /* c1 */
+    tmp3 += tmp5 + tmp10 + MULTIPLY(tmp13, FIX(1.381129125)) /* c1+c5-c9-c7 */
+	    - MULTIPLY(tmp14, FIX(1.360834544));         /* c3 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+2);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+2);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+2);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp3, CONST_BITS+2);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 12x12 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_12x12 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  DCTELEM workspace[8*4];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT. */
+  /* cK represents sqrt(2) * cos(K*pi/24). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[11]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[10]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[9]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[8]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[7]);
+    tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[6]);
+
+    tmp10 = tmp0 + tmp5;
+    tmp13 = tmp0 - tmp5;
+    tmp11 = tmp1 + tmp4;
+    tmp14 = tmp1 - tmp4;
+    tmp12 = tmp2 + tmp3;
+    tmp15 = tmp2 - tmp3;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[11]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[10]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[9]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[8]);
+    tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[7]);
+    tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[6]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM) (tmp10 + tmp11 + tmp12 - 12 * CENTERJSAMPLE);
+    dataptr[6] = (DCTELEM) (tmp13 - tmp14 - tmp15);
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.224744871)), /* c4 */
+	      CONST_BITS);
+    dataptr[2] = (DCTELEM)
+      DESCALE(tmp14 - tmp15 + MULTIPLY(tmp13 + tmp15, FIX(1.366025404)), /* c2 */
+	      CONST_BITS);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp1 + tmp4, FIX_0_541196100);    /* c9 */
+    tmp14 = tmp10 + MULTIPLY(tmp1, FIX_0_765366865);   /* c3-c9 */
+    tmp15 = tmp10 - MULTIPLY(tmp4, FIX_1_847759065);   /* c3+c9 */
+    tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.121971054));   /* c5 */
+    tmp13 = MULTIPLY(tmp0 + tmp3, FIX(0.860918669));   /* c7 */
+    tmp10 = tmp12 + tmp13 + tmp14 - MULTIPLY(tmp0, FIX(0.580774953)) /* c5+c7-c1 */
+	    + MULTIPLY(tmp5, FIX(0.184591911));        /* c11 */
+    tmp11 = MULTIPLY(tmp2 + tmp3, - FIX(0.184591911)); /* -c11 */
+    tmp12 += tmp11 - tmp15 - MULTIPLY(tmp2, FIX(2.339493912)) /* c1+c5-c11 */
+	    + MULTIPLY(tmp5, FIX(0.860918669));        /* c7 */
+    tmp13 += tmp11 - tmp14 + MULTIPLY(tmp3, FIX(0.725788011)) /* c1+c11-c7 */
+	    - MULTIPLY(tmp5, FIX(1.121971054));        /* c5 */
+    tmp11 = tmp15 + MULTIPLY(tmp0 - tmp3, FIX(1.306562965)) /* c3 */
+	    - MULTIPLY(tmp2 + tmp5, FIX_0_541196100);  /* c9 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp10, CONST_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp11, CONST_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp12, CONST_BITS);
+    dataptr[7] = (DCTELEM) DESCALE(tmp13, CONST_BITS);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 12)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/12)**2 = 4/9, which we partially
+   * fold into the constant multipliers and final shifting:
+   * cK now represents sqrt(2) * cos(K*pi/24) * 8/9.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*3];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*2];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*1];
+    tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*0];
+    tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*7];
+    tmp5 = dataptr[DCTSIZE*5] + dataptr[DCTSIZE*6];
+
+    tmp10 = tmp0 + tmp5;
+    tmp13 = tmp0 - tmp5;
+    tmp11 = tmp1 + tmp4;
+    tmp14 = tmp1 - tmp4;
+    tmp12 = tmp2 + tmp3;
+    tmp15 = tmp2 - tmp3;
+
+    tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*3];
+    tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*2];
+    tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*1];
+    tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*0];
+    tmp4 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*7];
+    tmp5 = dataptr[DCTSIZE*5] - dataptr[DCTSIZE*6];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12, FIX(0.888888889)), /* 8/9 */
+	      CONST_BITS+1);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp13 - tmp14 - tmp15, FIX(0.888888889)), /* 8/9 */
+	      CONST_BITS+1);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.088662108)),         /* c4 */
+	      CONST_BITS+1);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp14 - tmp15, FIX(0.888888889)) +        /* 8/9 */
+	      MULTIPLY(tmp13 + tmp15, FIX(1.214244803)),         /* c2 */
+	      CONST_BITS+1);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp1 + tmp4, FIX(0.481063200));   /* c9 */
+    tmp14 = tmp10 + MULTIPLY(tmp1, FIX(0.680326102));  /* c3-c9 */
+    tmp15 = tmp10 - MULTIPLY(tmp4, FIX(1.642452502));  /* c3+c9 */
+    tmp12 = MULTIPLY(tmp0 + tmp2, FIX(0.997307603));   /* c5 */
+    tmp13 = MULTIPLY(tmp0 + tmp3, FIX(0.765261039));   /* c7 */
+    tmp10 = tmp12 + tmp13 + tmp14 - MULTIPLY(tmp0, FIX(0.516244403)) /* c5+c7-c1 */
+	    + MULTIPLY(tmp5, FIX(0.164081699));        /* c11 */
+    tmp11 = MULTIPLY(tmp2 + tmp3, - FIX(0.164081699)); /* -c11 */
+    tmp12 += tmp11 - tmp15 - MULTIPLY(tmp2, FIX(2.079550144)) /* c1+c5-c11 */
+	    + MULTIPLY(tmp5, FIX(0.765261039));        /* c7 */
+    tmp13 += tmp11 - tmp14 + MULTIPLY(tmp3, FIX(0.645144899)) /* c1+c11-c7 */
+	    - MULTIPLY(tmp5, FIX(0.997307603));        /* c5 */
+    tmp11 = tmp15 + MULTIPLY(tmp0 - tmp3, FIX(1.161389302)) /* c3 */
+	    - MULTIPLY(tmp2 + tmp5, FIX(0.481063200)); /* c9 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10, CONST_BITS+1);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp11, CONST_BITS+1);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp12, CONST_BITS+1);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp13, CONST_BITS+1);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 13x13 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_13x13 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  INT32 z1, z2;
+  DCTELEM workspace[8*5];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT. */
+  /* cK represents sqrt(2) * cos(K*pi/26). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[12]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[11]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[10]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[9]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[8]);
+    tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[7]);
+    tmp6 = GETJSAMPLE(elemptr[6]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[12]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[11]);
+    tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[10]);
+    tmp13 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[9]);
+    tmp14 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[8]);
+    tmp15 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[7]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      (tmp0 + tmp1 + tmp2 + tmp3 + tmp4 + tmp5 + tmp6 - 13 * CENTERJSAMPLE);
+    tmp6 += tmp6;
+    tmp0 -= tmp6;
+    tmp1 -= tmp6;
+    tmp2 -= tmp6;
+    tmp3 -= tmp6;
+    tmp4 -= tmp6;
+    tmp5 -= tmp6;
+    dataptr[2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0, FIX(1.373119086)) +   /* c2 */
+	      MULTIPLY(tmp1, FIX(1.058554052)) +   /* c6 */
+	      MULTIPLY(tmp2, FIX(0.501487041)) -   /* c10 */
+	      MULTIPLY(tmp3, FIX(0.170464608)) -   /* c12 */
+	      MULTIPLY(tmp4, FIX(0.803364869)) -   /* c8 */
+	      MULTIPLY(tmp5, FIX(1.252223920)),    /* c4 */
+	      CONST_BITS);
+    z1 = MULTIPLY(tmp0 - tmp2, FIX(1.155388986)) - /* (c4+c6)/2 */
+	 MULTIPLY(tmp3 - tmp4, FIX(0.435816023)) - /* (c2-c10)/2 */
+	 MULTIPLY(tmp1 - tmp5, FIX(0.316450131));  /* (c8-c12)/2 */
+    z2 = MULTIPLY(tmp0 + tmp2, FIX(0.096834934)) - /* (c4-c6)/2 */
+	 MULTIPLY(tmp3 + tmp4, FIX(0.937303064)) + /* (c2+c10)/2 */
+	 MULTIPLY(tmp1 + tmp5, FIX(0.486914739));  /* (c8+c12)/2 */
+
+    dataptr[4] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS);
+    dataptr[6] = (DCTELEM) DESCALE(z1 - z2, CONST_BITS);
+
+    /* Odd part */
+
+    tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.322312651));   /* c3 */
+    tmp2 = MULTIPLY(tmp10 + tmp12, FIX(1.163874945));   /* c5 */
+    tmp3 = MULTIPLY(tmp10 + tmp13, FIX(0.937797057)) +  /* c7 */
+	   MULTIPLY(tmp14 + tmp15, FIX(0.338443458));   /* c11 */
+    tmp0 = tmp1 + tmp2 + tmp3 -
+	   MULTIPLY(tmp10, FIX(2.020082300)) +          /* c3+c5+c7-c1 */
+	   MULTIPLY(tmp14, FIX(0.318774355));           /* c9-c11 */
+    tmp4 = MULTIPLY(tmp14 - tmp15, FIX(0.937797057)) -  /* c7 */
+	   MULTIPLY(tmp11 + tmp12, FIX(0.338443458));   /* c11 */
+    tmp5 = MULTIPLY(tmp11 + tmp13, - FIX(1.163874945)); /* -c5 */
+    tmp1 += tmp4 + tmp5 +
+	    MULTIPLY(tmp11, FIX(0.837223564)) -         /* c5+c9+c11-c3 */
+	    MULTIPLY(tmp14, FIX(2.341699410));          /* c1+c7 */
+    tmp6 = MULTIPLY(tmp12 + tmp13, - FIX(0.657217813)); /* -c9 */
+    tmp2 += tmp4 + tmp6 -
+	    MULTIPLY(tmp12, FIX(1.572116027)) +         /* c1+c5-c9-c11 */
+	    MULTIPLY(tmp15, FIX(2.260109708));          /* c3+c7 */
+    tmp3 += tmp5 + tmp6 +
+	    MULTIPLY(tmp13, FIX(2.205608352)) -         /* c3+c5+c9-c7 */
+	    MULTIPLY(tmp15, FIX(1.742345811));          /* c1+c11 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS);
+    dataptr[7] = (DCTELEM) DESCALE(tmp3, CONST_BITS);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 13)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/13)**2 = 64/169, which we partially
+   * fold into the constant multipliers and final shifting:
+   * cK now represents sqrt(2) * cos(K*pi/26) * 128/169.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*4];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*3];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*2];
+    tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*1];
+    tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*0];
+    tmp5 = dataptr[DCTSIZE*5] + dataptr[DCTSIZE*7];
+    tmp6 = dataptr[DCTSIZE*6];
+
+    tmp10 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*4];
+    tmp11 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*3];
+    tmp12 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*2];
+    tmp13 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*1];
+    tmp14 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*0];
+    tmp15 = dataptr[DCTSIZE*5] - dataptr[DCTSIZE*7];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 + tmp1 + tmp2 + tmp3 + tmp4 + tmp5 + tmp6,
+		       FIX(0.757396450)),          /* 128/169 */
+	      CONST_BITS+1);
+    tmp6 += tmp6;
+    tmp0 -= tmp6;
+    tmp1 -= tmp6;
+    tmp2 -= tmp6;
+    tmp3 -= tmp6;
+    tmp4 -= tmp6;
+    tmp5 -= tmp6;
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0, FIX(1.039995521)) +   /* c2 */
+	      MULTIPLY(tmp1, FIX(0.801745081)) +   /* c6 */
+	      MULTIPLY(tmp2, FIX(0.379824504)) -   /* c10 */
+	      MULTIPLY(tmp3, FIX(0.129109289)) -   /* c12 */
+	      MULTIPLY(tmp4, FIX(0.608465700)) -   /* c8 */
+	      MULTIPLY(tmp5, FIX(0.948429952)),    /* c4 */
+	      CONST_BITS+1);
+    z1 = MULTIPLY(tmp0 - tmp2, FIX(0.875087516)) - /* (c4+c6)/2 */
+	 MULTIPLY(tmp3 - tmp4, FIX(0.330085509)) - /* (c2-c10)/2 */
+	 MULTIPLY(tmp1 - tmp5, FIX(0.239678205));  /* (c8-c12)/2 */
+    z2 = MULTIPLY(tmp0 + tmp2, FIX(0.073342435)) - /* (c4-c6)/2 */
+	 MULTIPLY(tmp3 + tmp4, FIX(0.709910013)) + /* (c2+c10)/2 */
+	 MULTIPLY(tmp1 + tmp5, FIX(0.368787494));  /* (c8+c12)/2 */
+
+    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS+1);
+    dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 - z2, CONST_BITS+1);
+
+    /* Odd part */
+
+    tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.001514908));   /* c3 */
+    tmp2 = MULTIPLY(tmp10 + tmp12, FIX(0.881514751));   /* c5 */
+    tmp3 = MULTIPLY(tmp10 + tmp13, FIX(0.710284161)) +  /* c7 */
+	   MULTIPLY(tmp14 + tmp15, FIX(0.256335874));   /* c11 */
+    tmp0 = tmp1 + tmp2 + tmp3 -
+	   MULTIPLY(tmp10, FIX(1.530003162)) +          /* c3+c5+c7-c1 */
+	   MULTIPLY(tmp14, FIX(0.241438564));           /* c9-c11 */
+    tmp4 = MULTIPLY(tmp14 - tmp15, FIX(0.710284161)) -  /* c7 */
+	   MULTIPLY(tmp11 + tmp12, FIX(0.256335874));   /* c11 */
+    tmp5 = MULTIPLY(tmp11 + tmp13, - FIX(0.881514751)); /* -c5 */
+    tmp1 += tmp4 + tmp5 +
+	    MULTIPLY(tmp11, FIX(0.634110155)) -         /* c5+c9+c11-c3 */
+	    MULTIPLY(tmp14, FIX(1.773594819));          /* c1+c7 */
+    tmp6 = MULTIPLY(tmp12 + tmp13, - FIX(0.497774438)); /* -c9 */
+    tmp2 += tmp4 + tmp6 -
+	    MULTIPLY(tmp12, FIX(1.190715098)) +         /* c1+c5-c9-c11 */
+	    MULTIPLY(tmp15, FIX(1.711799069));          /* c3+c7 */
+    tmp3 += tmp5 + tmp6 +
+	    MULTIPLY(tmp13, FIX(1.670519935)) -         /* c3+c5+c9-c7 */
+	    MULTIPLY(tmp15, FIX(1.319646532));          /* c1+c11 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+1);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+1);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+1);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp3, CONST_BITS+1);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 14x14 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_14x14 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  DCTELEM workspace[8*6];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT. */
+  /* cK represents sqrt(2) * cos(K*pi/28). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[13]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[12]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[11]);
+    tmp13 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[10]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[9]);
+    tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[8]);
+    tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[7]);
+
+    tmp10 = tmp0 + tmp6;
+    tmp14 = tmp0 - tmp6;
+    tmp11 = tmp1 + tmp5;
+    tmp15 = tmp1 - tmp5;
+    tmp12 = tmp2 + tmp4;
+    tmp16 = tmp2 - tmp4;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[13]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[12]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[11]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[10]);
+    tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[9]);
+    tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[8]);
+    tmp6 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[7]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      (tmp10 + tmp11 + tmp12 + tmp13 - 14 * CENTERJSAMPLE);
+    tmp13 += tmp13;
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.274162392)) + /* c4 */
+	      MULTIPLY(tmp11 - tmp13, FIX(0.314692123)) - /* c12 */
+	      MULTIPLY(tmp12 - tmp13, FIX(0.881747734)),  /* c8 */
+	      CONST_BITS);
+
+    tmp10 = MULTIPLY(tmp14 + tmp15, FIX(1.105676686));    /* c6 */
+
+    dataptr[2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp14, FIX(0.273079590))   /* c2-c6 */
+	      + MULTIPLY(tmp16, FIX(0.613604268)),        /* c10 */
+	      CONST_BITS);
+    dataptr[6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp15, FIX(1.719280954))   /* c6+c10 */
+	      - MULTIPLY(tmp16, FIX(1.378756276)),        /* c2 */
+	      CONST_BITS);
+
+    /* Odd part */
+
+    tmp10 = tmp1 + tmp2;
+    tmp11 = tmp5 - tmp4;
+    dataptr[7] = (DCTELEM) (tmp0 - tmp10 + tmp3 - tmp11 - tmp6);
+    tmp3 <<= CONST_BITS;
+    tmp10 = MULTIPLY(tmp10, - FIX(0.158341681));          /* -c13 */
+    tmp11 = MULTIPLY(tmp11, FIX(1.405321284));            /* c1 */
+    tmp10 += tmp11 - tmp3;
+    tmp11 = MULTIPLY(tmp0 + tmp2, FIX(1.197448846)) +     /* c5 */
+	    MULTIPLY(tmp4 + tmp6, FIX(0.752406978));      /* c9 */
+    dataptr[5] = (DCTELEM)
+      DESCALE(tmp10 + tmp11 - MULTIPLY(tmp2, FIX(2.373959773)) /* c3+c5-c13 */
+	      + MULTIPLY(tmp4, FIX(1.119999435)),         /* c1+c11-c9 */
+	      CONST_BITS);
+    tmp12 = MULTIPLY(tmp0 + tmp1, FIX(1.334852607)) +     /* c3 */
+	    MULTIPLY(tmp5 - tmp6, FIX(0.467085129));      /* c11 */
+    dataptr[3] = (DCTELEM)
+      DESCALE(tmp10 + tmp12 - MULTIPLY(tmp1, FIX(0.424103948)) /* c3-c9-c13 */
+	      - MULTIPLY(tmp5, FIX(3.069855259)),         /* c1+c5+c11 */
+	      CONST_BITS);
+    dataptr[1] = (DCTELEM)
+      DESCALE(tmp11 + tmp12 + tmp3 + tmp6 -
+	      MULTIPLY(tmp0 + tmp6, FIX(1.126980169)),    /* c3+c5-c1 */
+	      CONST_BITS);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 14)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/14)**2 = 16/49, which we partially
+   * fold into the constant multipliers and final shifting:
+   * cK now represents sqrt(2) * cos(K*pi/28) * 32/49.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*5];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*3];
+    tmp13 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*2];
+    tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*1];
+    tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*0];
+    tmp6 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
+
+    tmp10 = tmp0 + tmp6;
+    tmp14 = tmp0 - tmp6;
+    tmp11 = tmp1 + tmp5;
+    tmp15 = tmp1 - tmp5;
+    tmp12 = tmp2 + tmp4;
+    tmp16 = tmp2 - tmp4;
+
+    tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*5];
+    tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*3];
+    tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*2];
+    tmp4 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*1];
+    tmp5 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*0];
+    tmp6 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12 + tmp13,
+		       FIX(0.653061224)),                 /* 32/49 */
+	      CONST_BITS+1);
+    tmp13 += tmp13;
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp13, FIX(0.832106052)) + /* c4 */
+	      MULTIPLY(tmp11 - tmp13, FIX(0.205513223)) - /* c12 */
+	      MULTIPLY(tmp12 - tmp13, FIX(0.575835255)),  /* c8 */
+	      CONST_BITS+1);
+
+    tmp10 = MULTIPLY(tmp14 + tmp15, FIX(0.722074570));    /* c6 */
+
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp14, FIX(0.178337691))   /* c2-c6 */
+	      + MULTIPLY(tmp16, FIX(0.400721155)),        /* c10 */
+	      CONST_BITS+1);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp15, FIX(1.122795725))   /* c6+c10 */
+	      - MULTIPLY(tmp16, FIX(0.900412262)),        /* c2 */
+	      CONST_BITS+1);
+
+    /* Odd part */
+
+    tmp10 = tmp1 + tmp2;
+    tmp11 = tmp5 - tmp4;
+    dataptr[DCTSIZE*7] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp10 + tmp3 - tmp11 - tmp6,
+		       FIX(0.653061224)),                 /* 32/49 */
+	      CONST_BITS+1);
+    tmp3  = MULTIPLY(tmp3 , FIX(0.653061224));            /* 32/49 */
+    tmp10 = MULTIPLY(tmp10, - FIX(0.103406812));          /* -c13 */
+    tmp11 = MULTIPLY(tmp11, FIX(0.917760839));            /* c1 */
+    tmp10 += tmp11 - tmp3;
+    tmp11 = MULTIPLY(tmp0 + tmp2, FIX(0.782007410)) +     /* c5 */
+	    MULTIPLY(tmp4 + tmp6, FIX(0.491367823));      /* c9 */
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      DESCALE(tmp10 + tmp11 - MULTIPLY(tmp2, FIX(1.550341076)) /* c3+c5-c13 */
+	      + MULTIPLY(tmp4, FIX(0.731428202)),         /* c1+c11-c9 */
+	      CONST_BITS+1);
+    tmp12 = MULTIPLY(tmp0 + tmp1, FIX(0.871740478)) +     /* c3 */
+	    MULTIPLY(tmp5 - tmp6, FIX(0.305035186));      /* c11 */
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      DESCALE(tmp10 + tmp12 - MULTIPLY(tmp1, FIX(0.276965844)) /* c3-c9-c13 */
+	      - MULTIPLY(tmp5, FIX(2.004803435)),         /* c1+c5+c11 */
+	      CONST_BITS+1);
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(tmp11 + tmp12 + tmp3
+	      - MULTIPLY(tmp0, FIX(0.735987049))          /* c3+c5-c1 */
+	      - MULTIPLY(tmp6, FIX(0.082925825)),         /* c9-c11-c13 */
+	      CONST_BITS+1);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 15x15 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_15x15 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  INT32 z1, z2, z3;
+  DCTELEM workspace[8*7];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT. */
+  /* cK represents sqrt(2) * cos(K*pi/30). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[14]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[13]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[12]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[11]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[10]);
+    tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[9]);
+    tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[8]);
+    tmp7 = GETJSAMPLE(elemptr[7]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[14]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[13]);
+    tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[12]);
+    tmp13 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[11]);
+    tmp14 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[10]);
+    tmp15 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[9]);
+    tmp16 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[8]);
+
+    z1 = tmp0 + tmp4 + tmp5;
+    z2 = tmp1 + tmp3 + tmp6;
+    z3 = tmp2 + tmp7;
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM) (z1 + z2 + z3 - 15 * CENTERJSAMPLE);
+    z3 += z3;
+    dataptr[6] = (DCTELEM)
+      DESCALE(MULTIPLY(z1 - z3, FIX(1.144122806)) - /* c6 */
+	      MULTIPLY(z2 - z3, FIX(0.437016024)),  /* c12 */
+	      CONST_BITS);
+    tmp2 += ((tmp1 + tmp4) >> 1) - tmp7 - tmp7;
+    z1 = MULTIPLY(tmp3 - tmp2, FIX(1.531135173)) -  /* c2+c14 */
+         MULTIPLY(tmp6 - tmp2, FIX(2.238241955));   /* c4+c8 */
+    z2 = MULTIPLY(tmp5 - tmp2, FIX(0.798468008)) -  /* c8-c14 */
+	 MULTIPLY(tmp0 - tmp2, FIX(0.091361227));   /* c2-c4 */
+    z3 = MULTIPLY(tmp0 - tmp3, FIX(1.383309603)) +  /* c2 */
+	 MULTIPLY(tmp6 - tmp5, FIX(0.946293579)) +  /* c8 */
+	 MULTIPLY(tmp1 - tmp4, FIX(0.790569415));   /* (c6+c12)/2 */
+
+    dataptr[2] = (DCTELEM) DESCALE(z1 + z3, CONST_BITS);
+    dataptr[4] = (DCTELEM) DESCALE(z2 + z3, CONST_BITS);
+
+    /* Odd part */
+
+    tmp2 = MULTIPLY(tmp10 - tmp12 - tmp13 + tmp15 + tmp16,
+		    FIX(1.224744871));                         /* c5 */
+    tmp1 = MULTIPLY(tmp10 - tmp14 - tmp15, FIX(1.344997024)) + /* c3 */
+	   MULTIPLY(tmp11 - tmp13 - tmp16, FIX(0.831253876));  /* c9 */
+    tmp12 = MULTIPLY(tmp12, FIX(1.224744871));                 /* c5 */
+    tmp4 = MULTIPLY(tmp10 - tmp16, FIX(1.406466353)) +         /* c1 */
+	   MULTIPLY(tmp11 + tmp14, FIX(1.344997024)) +         /* c3 */
+	   MULTIPLY(tmp13 + tmp15, FIX(0.575212477));          /* c11 */
+    tmp0 = MULTIPLY(tmp13, FIX(0.475753014)) -                 /* c7-c11 */
+	   MULTIPLY(tmp14, FIX(0.513743148)) +                 /* c3-c9 */
+	   MULTIPLY(tmp16, FIX(1.700497885)) + tmp4 + tmp12;   /* c1+c13 */
+    tmp3 = MULTIPLY(tmp10, - FIX(0.355500862)) -               /* -(c1-c7) */
+	   MULTIPLY(tmp11, FIX(2.176250899)) -                 /* c3+c9 */
+	   MULTIPLY(tmp15, FIX(0.869244010)) + tmp4 - tmp12;   /* c11+c13 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS);
+    dataptr[7] = (DCTELEM) DESCALE(tmp3, CONST_BITS);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 15)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/15)**2 = 64/225, which we partially
+   * fold into the constant multipliers and final shifting:
+   * cK now represents sqrt(2) * cos(K*pi/30) * 256/225.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*6];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*5];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*4];
+    tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*3];
+    tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*2];
+    tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*1];
+    tmp6 = dataptr[DCTSIZE*6] + wsptr[DCTSIZE*0];
+    tmp7 = dataptr[DCTSIZE*7];
+
+    tmp10 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*6];
+    tmp11 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*5];
+    tmp12 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*4];
+    tmp13 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*3];
+    tmp14 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*2];
+    tmp15 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*1];
+    tmp16 = dataptr[DCTSIZE*6] - wsptr[DCTSIZE*0];
+
+    z1 = tmp0 + tmp4 + tmp5;
+    z2 = tmp1 + tmp3 + tmp6;
+    z3 = tmp2 + tmp7;
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(z1 + z2 + z3, FIX(1.137777778)), /* 256/225 */
+	      CONST_BITS+2);
+    z3 += z3;
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(MULTIPLY(z1 - z3, FIX(1.301757503)) - /* c6 */
+	      MULTIPLY(z2 - z3, FIX(0.497227121)),  /* c12 */
+	      CONST_BITS+2);
+    tmp2 += ((tmp1 + tmp4) >> 1) - tmp7 - tmp7;
+    z1 = MULTIPLY(tmp3 - tmp2, FIX(1.742091575)) -  /* c2+c14 */
+         MULTIPLY(tmp6 - tmp2, FIX(2.546621957));   /* c4+c8 */
+    z2 = MULTIPLY(tmp5 - tmp2, FIX(0.908479156)) -  /* c8-c14 */
+	 MULTIPLY(tmp0 - tmp2, FIX(0.103948774));   /* c2-c4 */
+    z3 = MULTIPLY(tmp0 - tmp3, FIX(1.573898926)) +  /* c2 */
+	 MULTIPLY(tmp6 - tmp5, FIX(1.076671805)) +  /* c8 */
+	 MULTIPLY(tmp1 - tmp4, FIX(0.899492312));   /* (c6+c12)/2 */
+
+    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + z3, CONST_BITS+2);
+    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(z2 + z3, CONST_BITS+2);
+
+    /* Odd part */
+
+    tmp2 = MULTIPLY(tmp10 - tmp12 - tmp13 + tmp15 + tmp16,
+		    FIX(1.393487498));                         /* c5 */
+    tmp1 = MULTIPLY(tmp10 - tmp14 - tmp15, FIX(1.530307725)) + /* c3 */
+	   MULTIPLY(tmp11 - tmp13 - tmp16, FIX(0.945782187));  /* c9 */
+    tmp12 = MULTIPLY(tmp12, FIX(1.393487498));                 /* c5 */
+    tmp4 = MULTIPLY(tmp10 - tmp16, FIX(1.600246161)) +         /* c1 */
+	   MULTIPLY(tmp11 + tmp14, FIX(1.530307725)) +         /* c3 */
+	   MULTIPLY(tmp13 + tmp15, FIX(0.654463974));          /* c11 */
+    tmp0 = MULTIPLY(tmp13, FIX(0.541301207)) -                 /* c7-c11 */
+	   MULTIPLY(tmp14, FIX(0.584525538)) +                 /* c3-c9 */
+	   MULTIPLY(tmp16, FIX(1.934788705)) + tmp4 + tmp12;   /* c1+c13 */
+    tmp3 = MULTIPLY(tmp10, - FIX(0.404480980)) -               /* -(c1-c7) */
+	   MULTIPLY(tmp11, FIX(2.476089912)) -                 /* c3+c9 */
+	   MULTIPLY(tmp15, FIX(0.989006518)) + tmp4 - tmp12;   /* c11+c13 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+2);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+2);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+2);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp3, CONST_BITS+2);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 16x16 sample block.
+ */
+
+GLOBAL(void)
+jpeg_fdct_16x16 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16, tmp17;
+  DCTELEM workspace[DCTSIZE2];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* cK represents sqrt(2) * cos(K*pi/32). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[15]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[14]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[13]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[12]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[11]);
+    tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[10]);
+    tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[9]);
+    tmp7 = GETJSAMPLE(elemptr[7]) + GETJSAMPLE(elemptr[8]);
+
+    tmp10 = tmp0 + tmp7;
+    tmp14 = tmp0 - tmp7;
+    tmp11 = tmp1 + tmp6;
+    tmp15 = tmp1 - tmp6;
+    tmp12 = tmp2 + tmp5;
+    tmp16 = tmp2 - tmp5;
+    tmp13 = tmp3 + tmp4;
+    tmp17 = tmp3 - tmp4;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[15]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[14]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[13]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[12]);
+    tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[11]);
+    tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[10]);
+    tmp6 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[9]);
+    tmp7 = GETJSAMPLE(elemptr[7]) - GETJSAMPLE(elemptr[8]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 + tmp12 + tmp13 - 16 * CENTERJSAMPLE) << PASS1_BITS);
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.306562965)) + /* c4[16] = c2[8] */
+	      MULTIPLY(tmp11 - tmp12, FIX_0_541196100),   /* c12[16] = c6[8] */
+	      CONST_BITS-PASS1_BITS);
+
+    tmp10 = MULTIPLY(tmp17 - tmp15, FIX(0.275899379)) +   /* c14[16] = c7[8] */
+	    MULTIPLY(tmp14 - tmp16, FIX(1.387039845));    /* c2[16] = c1[8] */
+
+    dataptr[2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp15, FIX(1.451774982))   /* c6+c14 */
+	      + MULTIPLY(tmp16, FIX(2.172734804)),        /* c2+c10 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp14, FIX(0.211164243))   /* c2-c6 */
+	      - MULTIPLY(tmp17, FIX(1.061594338)),        /* c10+c14 */
+	      CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp11 = MULTIPLY(tmp0 + tmp1, FIX(1.353318001)) +         /* c3 */
+	    MULTIPLY(tmp6 - tmp7, FIX(0.410524528));          /* c13 */
+    tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.247225013)) +         /* c5 */
+	    MULTIPLY(tmp5 + tmp7, FIX(0.666655658));          /* c11 */
+    tmp13 = MULTIPLY(tmp0 + tmp3, FIX(1.093201867)) +         /* c7 */
+	    MULTIPLY(tmp4 - tmp7, FIX(0.897167586));          /* c9 */
+    tmp14 = MULTIPLY(tmp1 + tmp2, FIX(0.138617169)) +         /* c15 */
+	    MULTIPLY(tmp6 - tmp5, FIX(1.407403738));          /* c1 */
+    tmp15 = MULTIPLY(tmp1 + tmp3, - FIX(0.666655658)) +       /* -c11 */
+	    MULTIPLY(tmp4 + tmp6, - FIX(1.247225013));        /* -c5 */
+    tmp16 = MULTIPLY(tmp2 + tmp3, - FIX(1.353318001)) +       /* -c3 */
+	    MULTIPLY(tmp5 - tmp4, FIX(0.410524528));          /* c13 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+	    MULTIPLY(tmp0, FIX(2.286341144)) +                /* c7+c5+c3-c1 */
+	    MULTIPLY(tmp7, FIX(0.779653625));                 /* c15+c13-c11+c9 */
+    tmp11 += tmp14 + tmp15 + MULTIPLY(tmp1, FIX(0.071888074)) /* c9-c3-c15+c11 */
+	     - MULTIPLY(tmp6, FIX(1.663905119));              /* c7+c13+c1-c5 */
+    tmp12 += tmp14 + tmp16 - MULTIPLY(tmp2, FIX(1.125726048)) /* c7+c5+c15-c3 */
+	     + MULTIPLY(tmp5, FIX(1.227391138));              /* c9-c11+c1-c13 */
+    tmp13 += tmp15 + tmp16 + MULTIPLY(tmp3, FIX(1.065388962)) /* c15+c3+c11-c7 */
+	     + MULTIPLY(tmp4, FIX(2.167985692));              /* c1+c13+c5-c9 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp10, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp11, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp12, CONST_BITS-PASS1_BITS);
+    dataptr[7] = (DCTELEM) DESCALE(tmp13, CONST_BITS-PASS1_BITS);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == DCTSIZE * 2)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/16)**2 = 1/2**2.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*3];
+    tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*2];
+    tmp6 = dataptr[DCTSIZE*6] + wsptr[DCTSIZE*1];
+    tmp7 = dataptr[DCTSIZE*7] + wsptr[DCTSIZE*0];
+
+    tmp10 = tmp0 + tmp7;
+    tmp14 = tmp0 - tmp7;
+    tmp11 = tmp1 + tmp6;
+    tmp15 = tmp1 - tmp6;
+    tmp12 = tmp2 + tmp5;
+    tmp16 = tmp2 - tmp5;
+    tmp13 = tmp3 + tmp4;
+    tmp17 = tmp3 - tmp4;
+
+    tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*3];
+    tmp5 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*2];
+    tmp6 = dataptr[DCTSIZE*6] - wsptr[DCTSIZE*1];
+    tmp7 = dataptr[DCTSIZE*7] - wsptr[DCTSIZE*0];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(tmp10 + tmp11 + tmp12 + tmp13, PASS1_BITS+2);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.306562965)) + /* c4[16] = c2[8] */
+	      MULTIPLY(tmp11 - tmp12, FIX_0_541196100),   /* c12[16] = c6[8] */
+	      CONST_BITS+PASS1_BITS+2);
+
+    tmp10 = MULTIPLY(tmp17 - tmp15, FIX(0.275899379)) +   /* c14[16] = c7[8] */
+	    MULTIPLY(tmp14 - tmp16, FIX(1.387039845));    /* c2[16] = c1[8] */
+
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp15, FIX(1.451774982))   /* c6+c14 */
+	      + MULTIPLY(tmp16, FIX(2.172734804)),        /* c2+10 */
+	      CONST_BITS+PASS1_BITS+2);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp14, FIX(0.211164243))   /* c2-c6 */
+	      - MULTIPLY(tmp17, FIX(1.061594338)),        /* c10+c14 */
+	      CONST_BITS+PASS1_BITS+2);
+
+    /* Odd part */
+
+    tmp11 = MULTIPLY(tmp0 + tmp1, FIX(1.353318001)) +         /* c3 */
+	    MULTIPLY(tmp6 - tmp7, FIX(0.410524528));          /* c13 */
+    tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.247225013)) +         /* c5 */
+	    MULTIPLY(tmp5 + tmp7, FIX(0.666655658));          /* c11 */
+    tmp13 = MULTIPLY(tmp0 + tmp3, FIX(1.093201867)) +         /* c7 */
+	    MULTIPLY(tmp4 - tmp7, FIX(0.897167586));          /* c9 */
+    tmp14 = MULTIPLY(tmp1 + tmp2, FIX(0.138617169)) +         /* c15 */
+	    MULTIPLY(tmp6 - tmp5, FIX(1.407403738));          /* c1 */
+    tmp15 = MULTIPLY(tmp1 + tmp3, - FIX(0.666655658)) +       /* -c11 */
+	    MULTIPLY(tmp4 + tmp6, - FIX(1.247225013));        /* -c5 */
+    tmp16 = MULTIPLY(tmp2 + tmp3, - FIX(1.353318001)) +       /* -c3 */
+	    MULTIPLY(tmp5 - tmp4, FIX(0.410524528));          /* c13 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+	    MULTIPLY(tmp0, FIX(2.286341144)) +                /* c7+c5+c3-c1 */
+	    MULTIPLY(tmp7, FIX(0.779653625));                 /* c15+c13-c11+c9 */
+    tmp11 += tmp14 + tmp15 + MULTIPLY(tmp1, FIX(0.071888074)) /* c9-c3-c15+c11 */
+	     - MULTIPLY(tmp6, FIX(1.663905119));              /* c7+c13+c1-c5 */
+    tmp12 += tmp14 + tmp16 - MULTIPLY(tmp2, FIX(1.125726048)) /* c7+c5+c15-c3 */
+	     + MULTIPLY(tmp5, FIX(1.227391138));              /* c9-c11+c1-c13 */
+    tmp13 += tmp15 + tmp16 + MULTIPLY(tmp3, FIX(1.065388962)) /* c15+c3+c11-c7 */
+	     + MULTIPLY(tmp4, FIX(2.167985692));              /* c1+c13+c5-c9 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10, CONST_BITS+PASS1_BITS+2);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp11, CONST_BITS+PASS1_BITS+2);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp12, CONST_BITS+PASS1_BITS+2);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp13, CONST_BITS+PASS1_BITS+2);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 16x8 sample block.
+ *
+ * 16-point FDCT in pass 1 (rows), 8-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_16x8 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16, tmp17;
+  INT32 z1;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* 16-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/32). */
+
+  dataptr = data;
+  ctr = 0;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[15]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[14]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[13]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[12]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[11]);
+    tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[10]);
+    tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[9]);
+    tmp7 = GETJSAMPLE(elemptr[7]) + GETJSAMPLE(elemptr[8]);
+
+    tmp10 = tmp0 + tmp7;
+    tmp14 = tmp0 - tmp7;
+    tmp11 = tmp1 + tmp6;
+    tmp15 = tmp1 - tmp6;
+    tmp12 = tmp2 + tmp5;
+    tmp16 = tmp2 - tmp5;
+    tmp13 = tmp3 + tmp4;
+    tmp17 = tmp3 - tmp4;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[15]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[14]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[13]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[12]);
+    tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[11]);
+    tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[10]);
+    tmp6 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[9]);
+    tmp7 = GETJSAMPLE(elemptr[7]) - GETJSAMPLE(elemptr[8]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 + tmp12 + tmp13 - 16 * CENTERJSAMPLE) << PASS1_BITS);
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.306562965)) + /* c4[16] = c2[8] */
+	      MULTIPLY(tmp11 - tmp12, FIX_0_541196100),   /* c12[16] = c6[8] */
+	      CONST_BITS-PASS1_BITS);
+
+    tmp10 = MULTIPLY(tmp17 - tmp15, FIX(0.275899379)) +   /* c14[16] = c7[8] */
+	    MULTIPLY(tmp14 - tmp16, FIX(1.387039845));    /* c2[16] = c1[8] */
+
+    dataptr[2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp15, FIX(1.451774982))   /* c6+c14 */
+	      + MULTIPLY(tmp16, FIX(2.172734804)),        /* c2+c10 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp14, FIX(0.211164243))   /* c2-c6 */
+	      - MULTIPLY(tmp17, FIX(1.061594338)),        /* c10+c14 */
+	      CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp11 = MULTIPLY(tmp0 + tmp1, FIX(1.353318001)) +         /* c3 */
+	    MULTIPLY(tmp6 - tmp7, FIX(0.410524528));          /* c13 */
+    tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.247225013)) +         /* c5 */
+	    MULTIPLY(tmp5 + tmp7, FIX(0.666655658));          /* c11 */
+    tmp13 = MULTIPLY(tmp0 + tmp3, FIX(1.093201867)) +         /* c7 */
+	    MULTIPLY(tmp4 - tmp7, FIX(0.897167586));          /* c9 */
+    tmp14 = MULTIPLY(tmp1 + tmp2, FIX(0.138617169)) +         /* c15 */
+	    MULTIPLY(tmp6 - tmp5, FIX(1.407403738));          /* c1 */
+    tmp15 = MULTIPLY(tmp1 + tmp3, - FIX(0.666655658)) +       /* -c11 */
+	    MULTIPLY(tmp4 + tmp6, - FIX(1.247225013));        /* -c5 */
+    tmp16 = MULTIPLY(tmp2 + tmp3, - FIX(1.353318001)) +       /* -c3 */
+	    MULTIPLY(tmp5 - tmp4, FIX(0.410524528));          /* c13 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+	    MULTIPLY(tmp0, FIX(2.286341144)) +                /* c7+c5+c3-c1 */
+	    MULTIPLY(tmp7, FIX(0.779653625));                 /* c15+c13-c11+c9 */
+    tmp11 += tmp14 + tmp15 + MULTIPLY(tmp1, FIX(0.071888074)) /* c9-c3-c15+c11 */
+	     - MULTIPLY(tmp6, FIX(1.663905119));              /* c7+c13+c1-c5 */
+    tmp12 += tmp14 + tmp16 - MULTIPLY(tmp2, FIX(1.125726048)) /* c7+c5+c15-c3 */
+	     + MULTIPLY(tmp5, FIX(1.227391138));              /* c9-c11+c1-c13 */
+    tmp13 += tmp15 + tmp16 + MULTIPLY(tmp3, FIX(1.065388962)) /* c15+c3+c11-c7 */
+	     + MULTIPLY(tmp4, FIX(2.167985692));              /* c1+c13+c5-c9 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp10, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp11, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp12, CONST_BITS-PASS1_BITS);
+    dataptr[7] = (DCTELEM) DESCALE(tmp13, CONST_BITS-PASS1_BITS);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by 8/16 = 1/2.
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+
+    tmp10 = tmp0 + tmp3;
+    tmp12 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp13 = tmp1 - tmp2;
+
+    tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+
+    dataptr[DCTSIZE*0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS+1);
+    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS+1);
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, FIX_0_765366865),
+					   CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 - MULTIPLY(tmp13, FIX_1_847759065),
+					   CONST_BITS+PASS1_BITS+1);
+
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * 8-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
+     * i0..i3 in the paper are tmp0..tmp3 here.
+     */
+
+    tmp10 = tmp0 + tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp1 + tmp3;
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
+
+    tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
+    tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
+    tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
+    tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
+    tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
+    tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
+    tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
+    tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
+
+    tmp12 += z1;
+    tmp13 += z1;
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0 + tmp10 + tmp12,
+					   CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1 + tmp11 + tmp13,
+					   CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2 + tmp11 + tmp12,
+					   CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp3 + tmp10 + tmp13,
+					   CONST_BITS+PASS1_BITS+1);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 14x7 sample block.
+ *
+ * 14-point FDCT in pass 1 (rows), 7-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_14x7 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  INT32 z1, z2, z3;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Zero bottom row of output coefficient block. */
+  MEMZERO(&data[DCTSIZE*7], SIZEOF(DCTELEM) * DCTSIZE);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* 14-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/28). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 7; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[13]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[12]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[11]);
+    tmp13 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[10]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[9]);
+    tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[8]);
+    tmp6 = GETJSAMPLE(elemptr[6]) + GETJSAMPLE(elemptr[7]);
+
+    tmp10 = tmp0 + tmp6;
+    tmp14 = tmp0 - tmp6;
+    tmp11 = tmp1 + tmp5;
+    tmp15 = tmp1 - tmp5;
+    tmp12 = tmp2 + tmp4;
+    tmp16 = tmp2 - tmp4;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[13]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[12]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[11]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[10]);
+    tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[9]);
+    tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[8]);
+    tmp6 = GETJSAMPLE(elemptr[6]) - GETJSAMPLE(elemptr[7]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 + tmp12 + tmp13 - 14 * CENTERJSAMPLE) << PASS1_BITS);
+    tmp13 += tmp13;
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.274162392)) + /* c4 */
+	      MULTIPLY(tmp11 - tmp13, FIX(0.314692123)) - /* c12 */
+	      MULTIPLY(tmp12 - tmp13, FIX(0.881747734)),  /* c8 */
+	      CONST_BITS-PASS1_BITS);
+
+    tmp10 = MULTIPLY(tmp14 + tmp15, FIX(1.105676686));    /* c6 */
+
+    dataptr[2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp14, FIX(0.273079590))   /* c2-c6 */
+	      + MULTIPLY(tmp16, FIX(0.613604268)),        /* c10 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp15, FIX(1.719280954))   /* c6+c10 */
+	      - MULTIPLY(tmp16, FIX(1.378756276)),        /* c2 */
+	      CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = tmp1 + tmp2;
+    tmp11 = tmp5 - tmp4;
+    dataptr[7] = (DCTELEM) ((tmp0 - tmp10 + tmp3 - tmp11 - tmp6) << PASS1_BITS);
+    tmp3 <<= CONST_BITS;
+    tmp10 = MULTIPLY(tmp10, - FIX(0.158341681));          /* -c13 */
+    tmp11 = MULTIPLY(tmp11, FIX(1.405321284));            /* c1 */
+    tmp10 += tmp11 - tmp3;
+    tmp11 = MULTIPLY(tmp0 + tmp2, FIX(1.197448846)) +     /* c5 */
+	    MULTIPLY(tmp4 + tmp6, FIX(0.752406978));      /* c9 */
+    dataptr[5] = (DCTELEM)
+      DESCALE(tmp10 + tmp11 - MULTIPLY(tmp2, FIX(2.373959773)) /* c3+c5-c13 */
+	      + MULTIPLY(tmp4, FIX(1.119999435)),         /* c1+c11-c9 */
+	      CONST_BITS-PASS1_BITS);
+    tmp12 = MULTIPLY(tmp0 + tmp1, FIX(1.334852607)) +     /* c3 */
+	    MULTIPLY(tmp5 - tmp6, FIX(0.467085129));      /* c11 */
+    dataptr[3] = (DCTELEM)
+      DESCALE(tmp10 + tmp12 - MULTIPLY(tmp1, FIX(0.424103948)) /* c3-c9-c13 */
+	      - MULTIPLY(tmp5, FIX(3.069855259)),         /* c1+c5+c11 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[1] = (DCTELEM)
+      DESCALE(tmp11 + tmp12 + tmp3 + tmp6 -
+	      MULTIPLY(tmp0 + tmp6, FIX(1.126980169)),    /* c3+c5-c1 */
+	      CONST_BITS-PASS1_BITS);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/14)*(8/7) = 32/49, which we
+   * partially fold into the constant multipliers and final shifting:
+   * 7-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/14) * 64/49.
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*6];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*5];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*4];
+    tmp3 = dataptr[DCTSIZE*3];
+
+    tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*6];
+    tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*5];
+    tmp12 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*4];
+
+    z1 = tmp0 + tmp2;
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(z1 + tmp1 + tmp3, FIX(1.306122449)), /* 64/49 */
+	      CONST_BITS+PASS1_BITS+1);
+    tmp3 += tmp3;
+    z1 -= tmp3;
+    z1 -= tmp3;
+    z1 = MULTIPLY(z1, FIX(0.461784020));                /* (c2+c6-c4)/2 */
+    z2 = MULTIPLY(tmp0 - tmp2, FIX(1.202428084));       /* (c2+c4-c6)/2 */
+    z3 = MULTIPLY(tmp1 - tmp2, FIX(0.411026446));       /* c6 */
+    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(z1 + z2 + z3, CONST_BITS+PASS1_BITS+1);
+    z1 -= z2;
+    z2 = MULTIPLY(tmp0 - tmp1, FIX(1.151670509));       /* c4 */
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(z2 + z3 - MULTIPLY(tmp1 - tmp3, FIX(0.923568041)), /* c2+c6-c4 */
+	      CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*6] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS+PASS1_BITS+1);
+
+    /* Odd part */
+
+    tmp1 = MULTIPLY(tmp10 + tmp11, FIX(1.221765677));   /* (c3+c1-c5)/2 */
+    tmp2 = MULTIPLY(tmp10 - tmp11, FIX(0.222383464));   /* (c3+c5-c1)/2 */
+    tmp0 = tmp1 - tmp2;
+    tmp1 += tmp2;
+    tmp2 = MULTIPLY(tmp11 + tmp12, - FIX(1.800824523)); /* -c1 */
+    tmp1 += tmp2;
+    tmp3 = MULTIPLY(tmp10 + tmp12, FIX(0.801442310));   /* c5 */
+    tmp0 += tmp3;
+    tmp2 += tmp3 + MULTIPLY(tmp12, FIX(2.443531355));   /* c3+c1-c5 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp0, CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp1, CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp2, CONST_BITS+PASS1_BITS+1);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 12x6 sample block.
+ *
+ * 12-point FDCT in pass 1 (rows), 6-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_12x6 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Zero 2 bottom rows of output coefficient block. */
+  MEMZERO(&data[DCTSIZE*6], SIZEOF(DCTELEM) * DCTSIZE * 2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* 12-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/24). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 6; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[11]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[10]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[9]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[8]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[7]);
+    tmp5 = GETJSAMPLE(elemptr[5]) + GETJSAMPLE(elemptr[6]);
+
+    tmp10 = tmp0 + tmp5;
+    tmp13 = tmp0 - tmp5;
+    tmp11 = tmp1 + tmp4;
+    tmp14 = tmp1 - tmp4;
+    tmp12 = tmp2 + tmp3;
+    tmp15 = tmp2 - tmp3;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[11]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[10]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[9]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[8]);
+    tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[7]);
+    tmp5 = GETJSAMPLE(elemptr[5]) - GETJSAMPLE(elemptr[6]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 + tmp12 - 12 * CENTERJSAMPLE) << PASS1_BITS);
+    dataptr[6] = (DCTELEM) ((tmp13 - tmp14 - tmp15) << PASS1_BITS);
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.224744871)), /* c4 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[2] = (DCTELEM)
+      DESCALE(tmp14 - tmp15 + MULTIPLY(tmp13 + tmp15, FIX(1.366025404)), /* c2 */
+	      CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp1 + tmp4, FIX_0_541196100);    /* c9 */
+    tmp14 = tmp10 + MULTIPLY(tmp1, FIX_0_765366865);   /* c3-c9 */
+    tmp15 = tmp10 - MULTIPLY(tmp4, FIX_1_847759065);   /* c3+c9 */
+    tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.121971054));   /* c5 */
+    tmp13 = MULTIPLY(tmp0 + tmp3, FIX(0.860918669));   /* c7 */
+    tmp10 = tmp12 + tmp13 + tmp14 - MULTIPLY(tmp0, FIX(0.580774953)) /* c5+c7-c1 */
+	    + MULTIPLY(tmp5, FIX(0.184591911));        /* c11 */
+    tmp11 = MULTIPLY(tmp2 + tmp3, - FIX(0.184591911)); /* -c11 */
+    tmp12 += tmp11 - tmp15 - MULTIPLY(tmp2, FIX(2.339493912)) /* c1+c5-c11 */
+	    + MULTIPLY(tmp5, FIX(0.860918669));        /* c7 */
+    tmp13 += tmp11 - tmp14 + MULTIPLY(tmp3, FIX(0.725788011)) /* c1+c11-c7 */
+	    - MULTIPLY(tmp5, FIX(1.121971054));        /* c5 */
+    tmp11 = tmp15 + MULTIPLY(tmp0 - tmp3, FIX(1.306562965)) /* c3 */
+	    - MULTIPLY(tmp2 + tmp5, FIX_0_541196100);  /* c9 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp10, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp11, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp12, CONST_BITS-PASS1_BITS);
+    dataptr[7] = (DCTELEM) DESCALE(tmp13, CONST_BITS-PASS1_BITS);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/12)*(8/6) = 8/9, which we
+   * partially fold into the constant multipliers and final shifting:
+   * 6-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/12) * 16/9.
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*5];
+    tmp11 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
+
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+
+    tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*5];
+    tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11, FIX(1.777777778)),         /* 16/9 */
+	      CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp12, FIX(2.177324216)),                 /* c2 */
+	      CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(1.257078722)), /* c4 */
+	      CONST_BITS+PASS1_BITS+1);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp0 + tmp2, FIX(0.650711829));             /* c5 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),   /* 16/9 */
+	      CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp1 - tmp2, FIX(1.777777778)),    /* 16/9 */
+	      CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp2 - tmp1, FIX(1.777777778)),   /* 16/9 */
+	      CONST_BITS+PASS1_BITS+1);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 10x5 sample block.
+ *
+ * 10-point FDCT in pass 1 (rows), 5-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_10x5 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Zero 3 bottom rows of output coefficient block. */
+  MEMZERO(&data[DCTSIZE*5], SIZEOF(DCTELEM) * DCTSIZE * 3);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* 10-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/20). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 5; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[9]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[8]);
+    tmp12 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[7]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[6]);
+    tmp4 = GETJSAMPLE(elemptr[4]) + GETJSAMPLE(elemptr[5]);
+
+    tmp10 = tmp0 + tmp4;
+    tmp13 = tmp0 - tmp4;
+    tmp11 = tmp1 + tmp3;
+    tmp14 = tmp1 - tmp3;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[9]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[8]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[7]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[6]);
+    tmp4 = GETJSAMPLE(elemptr[4]) - GETJSAMPLE(elemptr[5]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 + tmp12 - 10 * CENTERJSAMPLE) << PASS1_BITS);
+    tmp12 += tmp12;
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.144122806)) - /* c4 */
+	      MULTIPLY(tmp11 - tmp12, FIX(0.437016024)),  /* c8 */
+	      CONST_BITS-PASS1_BITS);
+    tmp10 = MULTIPLY(tmp13 + tmp14, FIX(0.831253876));    /* c6 */
+    dataptr[2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp13, FIX(0.513743148)),  /* c2-c6 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp14, FIX(2.176250899)),  /* c2+c6 */
+	      CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = tmp0 + tmp4;
+    tmp11 = tmp1 - tmp3;
+    dataptr[5] = (DCTELEM) ((tmp10 - tmp11 - tmp2) << PASS1_BITS);
+    tmp2 <<= CONST_BITS;
+    dataptr[1] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0, FIX(1.396802247)) +          /* c1 */
+	      MULTIPLY(tmp1, FIX(1.260073511)) + tmp2 +   /* c3 */
+	      MULTIPLY(tmp3, FIX(0.642039522)) +          /* c7 */
+	      MULTIPLY(tmp4, FIX(0.221231742)),           /* c9 */
+	      CONST_BITS-PASS1_BITS);
+    tmp12 = MULTIPLY(tmp0 - tmp4, FIX(0.951056516)) -     /* (c3+c7)/2 */
+	    MULTIPLY(tmp1 + tmp3, FIX(0.587785252));      /* (c1-c9)/2 */
+    tmp13 = MULTIPLY(tmp10 + tmp11, FIX(0.309016994)) +   /* (c3-c7)/2 */
+	    (tmp11 << (CONST_BITS - 1)) - tmp2;
+    dataptr[3] = (DCTELEM) DESCALE(tmp12 + tmp13, CONST_BITS-PASS1_BITS);
+    dataptr[7] = (DCTELEM) DESCALE(tmp12 - tmp13, CONST_BITS-PASS1_BITS);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/10)*(8/5) = 32/25, which we
+   * fold into the constant multipliers:
+   * 5-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/10) * 32/25.
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*4];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*3];
+    tmp2 = dataptr[DCTSIZE*2];
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+
+    tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*4];
+    tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*3];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp2, FIX(1.28)),        /* 32/25 */
+	      CONST_BITS+PASS1_BITS);
+    tmp11 = MULTIPLY(tmp11, FIX(1.011928851));          /* (c2+c4)/2 */
+    tmp10 -= tmp2 << 2;
+    tmp10 = MULTIPLY(tmp10, FIX(0.452548340));          /* (c2-c4)/2 */
+    dataptr[DCTSIZE*2] = (DCTELEM) DESCALE(tmp11 + tmp10, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM) DESCALE(tmp11 - tmp10, CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp0 + tmp1, FIX(1.064004961));    /* c3 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp0, FIX(0.657591230)), /* c1-c3 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp1, FIX(2.785601151)), /* c1+c3 */
+	      CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on an 8x4 sample block.
+ *
+ * 8-point FDCT in pass 1 (rows), 4-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_8x4 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3;
+  INT32 tmp10, tmp11, tmp12, tmp13;
+  INT32 z1;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Zero 4 bottom rows of output coefficient block. */
+  MEMZERO(&data[DCTSIZE*4], SIZEOF(DCTELEM) * DCTSIZE * 4);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* We must also scale the output by 8/4 = 2, which we add here. */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 4; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
+
+    tmp10 = tmp0 + tmp3;
+    tmp12 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp13 = tmp1 - tmp2;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 - 8 * CENTERJSAMPLE) << (PASS1_BITS+1));
+    dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << (PASS1_BITS+1));
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-2);
+    dataptr[2] = (DCTELEM) RIGHT_SHIFT(z1 + MULTIPLY(tmp12, FIX_0_765366865),
+				       CONST_BITS-PASS1_BITS-1);
+    dataptr[6] = (DCTELEM) RIGHT_SHIFT(z1 - MULTIPLY(tmp13, FIX_1_847759065),
+				       CONST_BITS-PASS1_BITS-1);
+
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * 8-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
+     * i0..i3 in the paper are tmp0..tmp3 here.
+     */
+
+    tmp10 = tmp0 + tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp1 + tmp3;
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-2);
+
+    tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
+    tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
+    tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
+    tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
+    tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
+    tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
+    tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
+    tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
+
+    tmp12 += z1;
+    tmp13 += z1;
+
+    dataptr[1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + tmp10 + tmp12, CONST_BITS-PASS1_BITS-1);
+    dataptr[3] = (DCTELEM)
+      RIGHT_SHIFT(tmp1 + tmp11 + tmp13, CONST_BITS-PASS1_BITS-1);
+    dataptr[5] = (DCTELEM)
+      RIGHT_SHIFT(tmp2 + tmp11 + tmp12, CONST_BITS-PASS1_BITS-1);
+    dataptr[7] = (DCTELEM)
+      RIGHT_SHIFT(tmp3 + tmp10 + tmp13, CONST_BITS-PASS1_BITS-1);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * 4-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
+   */
+
+  dataptr = data;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*3] + (ONE << (PASS1_BITS-1));
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*2];
+
+    tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*3];
+    tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*2];
+
+    dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp0 + tmp1, PASS1_BITS);
+    dataptr[DCTSIZE*2] = (DCTELEM) RIGHT_SHIFT(tmp0 - tmp1, PASS1_BITS);
+
+    /* Odd part */
+
+    tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);   /* c6 */
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS+PASS1_BITS-1);
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
+		  CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
+		  CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 6x3 sample block.
+ *
+ * 6-point FDCT in pass 1 (rows), 3-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_6x3 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2;
+  INT32 tmp10, tmp11, tmp12;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* We scale the results further by 2 as part of output adaption */
+  /* scaling for different DCT size. */
+  /* 6-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/12). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 3; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[5]);
+    tmp11 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[4]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[3]);
+
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[5]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[4]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[3]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 - 6 * CENTERJSAMPLE) << (PASS1_BITS+1));
+    dataptr[2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp12, FIX(1.224744871)),                 /* c2 */
+	      CONST_BITS-PASS1_BITS-1);
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(0.707106781)), /* c4 */
+	      CONST_BITS-PASS1_BITS-1);
+
+    /* Odd part */
+
+    tmp10 = DESCALE(MULTIPLY(tmp0 + tmp2, FIX(0.366025404)),     /* c5 */
+		    CONST_BITS-PASS1_BITS-1);
+
+    dataptr[1] = (DCTELEM) (tmp10 + ((tmp0 + tmp1) << (PASS1_BITS+1)));
+    dataptr[3] = (DCTELEM) ((tmp0 - tmp1 - tmp2) << (PASS1_BITS+1));
+    dataptr[5] = (DCTELEM) (tmp10 + ((tmp2 - tmp1) << (PASS1_BITS+1)));
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/6)*(8/3) = 32/9, which we partially
+   * fold into the constant multipliers (other part was done in pass 1):
+   * 3-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/6) * 16/9.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 6; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*2];
+    tmp1 = dataptr[DCTSIZE*1];
+
+    tmp2 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*2];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),        /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp1 - tmp1, FIX(1.257078722)), /* c2 */
+	      CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp2, FIX(2.177324216)),               /* c1 */
+	      CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 4x2 sample block.
+ *
+ * 4-point FDCT in pass 1 (rows), 2-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_4x2 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1;
+  INT32 tmp10, tmp11;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* We must also scale the output by (8/4)*(8/2) = 2**3, which we add here. */
+  /* 4-point FDCT kernel, */
+  /* cK represents sqrt(2) * cos(K*pi/16) [refers to 8-point FDCT]. */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 2; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[3]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[2]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[3]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[2]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp0 + tmp1 - 4 * CENTERJSAMPLE) << (PASS1_BITS+3));
+    dataptr[2] = (DCTELEM) ((tmp0 - tmp1) << (PASS1_BITS+3));
+
+    /* Odd part */
+
+    tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS-PASS1_BITS-4);
+
+    dataptr[1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
+		  CONST_BITS-PASS1_BITS-3);
+    dataptr[3] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
+		  CONST_BITS-PASS1_BITS-3);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 4; ctr++) {
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = dataptr[DCTSIZE*0] + (ONE << (PASS1_BITS-1));
+    tmp1 = dataptr[DCTSIZE*1];
+
+    dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp0 + tmp1, PASS1_BITS);
+
+    /* Odd part */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) RIGHT_SHIFT(tmp0 - tmp1, PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 2x1 sample block.
+ *
+ * 2-point FDCT in pass 1 (rows), 1-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_2x1 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1;
+  JSAMPROW elemptr;
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  elemptr = sample_data[0] + start_col;
+
+  tmp0 = GETJSAMPLE(elemptr[0]);
+  tmp1 = GETJSAMPLE(elemptr[1]);
+
+  /* We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/2)*(8/1) = 2**5.
+   */
+
+  /* Even part */
+  /* Apply unsigned->signed conversion */
+  data[0] = (DCTELEM) ((tmp0 + tmp1 - 2 * CENTERJSAMPLE) << 5);
+
+  /* Odd part */
+  data[1] = (DCTELEM) ((tmp0 - tmp1) << 5);
+}
+
+
+/*
+ * Perform the forward DCT on an 8x16 sample block.
+ *
+ * 8-point FDCT in pass 1 (rows), 16-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_8x16 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16, tmp17;
+  INT32 z1;
+  DCTELEM workspace[DCTSIZE2];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[7]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[6]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[5]);
+    tmp3 = GETJSAMPLE(elemptr[3]) + GETJSAMPLE(elemptr[4]);
+
+    tmp10 = tmp0 + tmp3;
+    tmp12 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp13 = tmp1 - tmp2;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[7]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[6]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[5]);
+    tmp3 = GETJSAMPLE(elemptr[3]) - GETJSAMPLE(elemptr[4]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM) ((tmp10 + tmp11 - 8 * CENTERJSAMPLE) << PASS1_BITS);
+    dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, FIX_0_765366865),
+				   CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM) DESCALE(z1 - MULTIPLY(tmp13, FIX_1_847759065),
+				   CONST_BITS-PASS1_BITS);
+
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * 8-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
+     * i0..i3 in the paper are tmp0..tmp3 here.
+     */
+
+    tmp10 = tmp0 + tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp1 + tmp3;
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
+
+    tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
+    tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
+    tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
+    tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
+    tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
+    tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
+    tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
+    tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
+
+    tmp12 += z1;
+    tmp13 += z1;
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp0 + tmp10 + tmp12, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp1 + tmp11 + tmp13, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp2 + tmp11 + tmp12, CONST_BITS-PASS1_BITS);
+    dataptr[7] = (DCTELEM) DESCALE(tmp3 + tmp10 + tmp13, CONST_BITS-PASS1_BITS);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == DCTSIZE * 2)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by 8/16 = 1/2.
+   * 16-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/32).
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*3];
+    tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*2];
+    tmp6 = dataptr[DCTSIZE*6] + wsptr[DCTSIZE*1];
+    tmp7 = dataptr[DCTSIZE*7] + wsptr[DCTSIZE*0];
+
+    tmp10 = tmp0 + tmp7;
+    tmp14 = tmp0 - tmp7;
+    tmp11 = tmp1 + tmp6;
+    tmp15 = tmp1 - tmp6;
+    tmp12 = tmp2 + tmp5;
+    tmp16 = tmp2 - tmp5;
+    tmp13 = tmp3 + tmp4;
+    tmp17 = tmp3 - tmp4;
+
+    tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*4];
+    tmp4 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*3];
+    tmp5 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*2];
+    tmp6 = dataptr[DCTSIZE*6] - wsptr[DCTSIZE*1];
+    tmp7 = dataptr[DCTSIZE*7] - wsptr[DCTSIZE*0];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(tmp10 + tmp11 + tmp12 + tmp13, PASS1_BITS+1);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp13, FIX(1.306562965)) + /* c4[16] = c2[8] */
+	      MULTIPLY(tmp11 - tmp12, FIX_0_541196100),   /* c12[16] = c6[8] */
+	      CONST_BITS+PASS1_BITS+1);
+
+    tmp10 = MULTIPLY(tmp17 - tmp15, FIX(0.275899379)) +   /* c14[16] = c7[8] */
+	    MULTIPLY(tmp14 - tmp16, FIX(1.387039845));    /* c2[16] = c1[8] */
+
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp15, FIX(1.451774982))   /* c6+c14 */
+	      + MULTIPLY(tmp16, FIX(2.172734804)),        /* c2+c10 */
+	      CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp14, FIX(0.211164243))   /* c2-c6 */
+	      - MULTIPLY(tmp17, FIX(1.061594338)),        /* c10+c14 */
+	      CONST_BITS+PASS1_BITS+1);
+
+    /* Odd part */
+
+    tmp11 = MULTIPLY(tmp0 + tmp1, FIX(1.353318001)) +         /* c3 */
+	    MULTIPLY(tmp6 - tmp7, FIX(0.410524528));          /* c13 */
+    tmp12 = MULTIPLY(tmp0 + tmp2, FIX(1.247225013)) +         /* c5 */
+	    MULTIPLY(tmp5 + tmp7, FIX(0.666655658));          /* c11 */
+    tmp13 = MULTIPLY(tmp0 + tmp3, FIX(1.093201867)) +         /* c7 */
+	    MULTIPLY(tmp4 - tmp7, FIX(0.897167586));          /* c9 */
+    tmp14 = MULTIPLY(tmp1 + tmp2, FIX(0.138617169)) +         /* c15 */
+	    MULTIPLY(tmp6 - tmp5, FIX(1.407403738));          /* c1 */
+    tmp15 = MULTIPLY(tmp1 + tmp3, - FIX(0.666655658)) +       /* -c11 */
+	    MULTIPLY(tmp4 + tmp6, - FIX(1.247225013));        /* -c5 */
+    tmp16 = MULTIPLY(tmp2 + tmp3, - FIX(1.353318001)) +       /* -c3 */
+	    MULTIPLY(tmp5 - tmp4, FIX(0.410524528));          /* c13 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+	    MULTIPLY(tmp0, FIX(2.286341144)) +                /* c7+c5+c3-c1 */
+	    MULTIPLY(tmp7, FIX(0.779653625));                 /* c15+c13-c11+c9 */
+    tmp11 += tmp14 + tmp15 + MULTIPLY(tmp1, FIX(0.071888074)) /* c9-c3-c15+c11 */
+	     - MULTIPLY(tmp6, FIX(1.663905119));              /* c7+c13+c1-c5 */
+    tmp12 += tmp14 + tmp16 - MULTIPLY(tmp2, FIX(1.125726048)) /* c7+c5+c15-c3 */
+	     + MULTIPLY(tmp5, FIX(1.227391138));              /* c9-c11+c1-c13 */
+    tmp13 += tmp15 + tmp16 + MULTIPLY(tmp3, FIX(1.065388962)) /* c15+c3+c11-c7 */
+	     + MULTIPLY(tmp4, FIX(2.167985692));              /* c1+c13+c5-c9 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10, CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp11, CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp12, CONST_BITS+PASS1_BITS+1);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp13, CONST_BITS+PASS1_BITS+1);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 7x14 sample block.
+ *
+ * 7-point FDCT in pass 1 (rows), 14-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_7x14 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  INT32 z1, z2, z3;
+  DCTELEM workspace[8*6];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* 7-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/14). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[6]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[5]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[4]);
+    tmp3 = GETJSAMPLE(elemptr[3]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[6]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[5]);
+    tmp12 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[4]);
+
+    z1 = tmp0 + tmp2;
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((z1 + tmp1 + tmp3 - 7 * CENTERJSAMPLE) << PASS1_BITS);
+    tmp3 += tmp3;
+    z1 -= tmp3;
+    z1 -= tmp3;
+    z1 = MULTIPLY(z1, FIX(0.353553391));                /* (c2+c6-c4)/2 */
+    z2 = MULTIPLY(tmp0 - tmp2, FIX(0.920609002));       /* (c2+c4-c6)/2 */
+    z3 = MULTIPLY(tmp1 - tmp2, FIX(0.314692123));       /* c6 */
+    dataptr[2] = (DCTELEM) DESCALE(z1 + z2 + z3, CONST_BITS-PASS1_BITS);
+    z1 -= z2;
+    z2 = MULTIPLY(tmp0 - tmp1, FIX(0.881747734));       /* c4 */
+    dataptr[4] = (DCTELEM)
+      DESCALE(z2 + z3 - MULTIPLY(tmp1 - tmp3, FIX(0.707106781)), /* c2+c6-c4 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[6] = (DCTELEM) DESCALE(z1 + z2, CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp1 = MULTIPLY(tmp10 + tmp11, FIX(0.935414347));   /* (c3+c1-c5)/2 */
+    tmp2 = MULTIPLY(tmp10 - tmp11, FIX(0.170262339));   /* (c3+c5-c1)/2 */
+    tmp0 = tmp1 - tmp2;
+    tmp1 += tmp2;
+    tmp2 = MULTIPLY(tmp11 + tmp12, - FIX(1.378756276)); /* -c1 */
+    tmp1 += tmp2;
+    tmp3 = MULTIPLY(tmp10 + tmp12, FIX(0.613604268));   /* c5 */
+    tmp0 += tmp3;
+    tmp2 += tmp3 + MULTIPLY(tmp12, FIX(1.870828693));   /* c3+c1-c5 */
+
+    dataptr[1] = (DCTELEM) DESCALE(tmp0, CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM) DESCALE(tmp1, CONST_BITS-PASS1_BITS);
+    dataptr[5] = (DCTELEM) DESCALE(tmp2, CONST_BITS-PASS1_BITS);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 14)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/7)*(8/14) = 32/49, which we
+   * fold into the constant multipliers:
+   * 14-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/28) * 32/49.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 7; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*5];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*3];
+    tmp13 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*2];
+    tmp4 = dataptr[DCTSIZE*4] + wsptr[DCTSIZE*1];
+    tmp5 = dataptr[DCTSIZE*5] + wsptr[DCTSIZE*0];
+    tmp6 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7];
+
+    tmp10 = tmp0 + tmp6;
+    tmp14 = tmp0 - tmp6;
+    tmp11 = tmp1 + tmp5;
+    tmp15 = tmp1 - tmp5;
+    tmp12 = tmp2 + tmp4;
+    tmp16 = tmp2 - tmp4;
+
+    tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*5];
+    tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*3];
+    tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*2];
+    tmp4 = dataptr[DCTSIZE*4] - wsptr[DCTSIZE*1];
+    tmp5 = dataptr[DCTSIZE*5] - wsptr[DCTSIZE*0];
+    tmp6 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12 + tmp13,
+		       FIX(0.653061224)),                 /* 32/49 */
+	      CONST_BITS+PASS1_BITS);
+    tmp13 += tmp13;
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp13, FIX(0.832106052)) + /* c4 */
+	      MULTIPLY(tmp11 - tmp13, FIX(0.205513223)) - /* c12 */
+	      MULTIPLY(tmp12 - tmp13, FIX(0.575835255)),  /* c8 */
+	      CONST_BITS+PASS1_BITS);
+
+    tmp10 = MULTIPLY(tmp14 + tmp15, FIX(0.722074570));    /* c6 */
+
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp14, FIX(0.178337691))   /* c2-c6 */
+	      + MULTIPLY(tmp16, FIX(0.400721155)),        /* c10 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp15, FIX(1.122795725))   /* c6+c10 */
+	      - MULTIPLY(tmp16, FIX(0.900412262)),        /* c2 */
+	      CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = tmp1 + tmp2;
+    tmp11 = tmp5 - tmp4;
+    dataptr[DCTSIZE*7] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp10 + tmp3 - tmp11 - tmp6,
+		       FIX(0.653061224)),                 /* 32/49 */
+	      CONST_BITS+PASS1_BITS);
+    tmp3  = MULTIPLY(tmp3 , FIX(0.653061224));            /* 32/49 */
+    tmp10 = MULTIPLY(tmp10, - FIX(0.103406812));          /* -c13 */
+    tmp11 = MULTIPLY(tmp11, FIX(0.917760839));            /* c1 */
+    tmp10 += tmp11 - tmp3;
+    tmp11 = MULTIPLY(tmp0 + tmp2, FIX(0.782007410)) +     /* c5 */
+	    MULTIPLY(tmp4 + tmp6, FIX(0.491367823));      /* c9 */
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      DESCALE(tmp10 + tmp11 - MULTIPLY(tmp2, FIX(1.550341076)) /* c3+c5-c13 */
+	      + MULTIPLY(tmp4, FIX(0.731428202)),         /* c1+c11-c9 */
+	      CONST_BITS+PASS1_BITS);
+    tmp12 = MULTIPLY(tmp0 + tmp1, FIX(0.871740478)) +     /* c3 */
+	    MULTIPLY(tmp5 - tmp6, FIX(0.305035186));      /* c11 */
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      DESCALE(tmp10 + tmp12 - MULTIPLY(tmp1, FIX(0.276965844)) /* c3-c9-c13 */
+	      - MULTIPLY(tmp5, FIX(2.004803435)),         /* c1+c5+c11 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(tmp11 + tmp12 + tmp3
+	      - MULTIPLY(tmp0, FIX(0.735987049))          /* c3+c5-c1 */
+	      - MULTIPLY(tmp6, FIX(0.082925825)),         /* c9-c11-c13 */
+	      CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 6x12 sample block.
+ *
+ * 6-point FDCT in pass 1 (rows), 12-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_6x12 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  DCTELEM workspace[8*4];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* 6-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/12). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[5]);
+    tmp11 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[4]);
+    tmp2 = GETJSAMPLE(elemptr[2]) + GETJSAMPLE(elemptr[3]);
+
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[5]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[4]);
+    tmp2 = GETJSAMPLE(elemptr[2]) - GETJSAMPLE(elemptr[3]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp11 - 6 * CENTERJSAMPLE) << PASS1_BITS);
+    dataptr[2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp12, FIX(1.224744871)),                 /* c2 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(0.707106781)), /* c4 */
+	      CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = DESCALE(MULTIPLY(tmp0 + tmp2, FIX(0.366025404)),     /* c5 */
+		    CONST_BITS-PASS1_BITS);
+
+    dataptr[1] = (DCTELEM) (tmp10 + ((tmp0 + tmp1) << PASS1_BITS));
+    dataptr[3] = (DCTELEM) ((tmp0 - tmp1 - tmp2) << PASS1_BITS);
+    dataptr[5] = (DCTELEM) (tmp10 + ((tmp2 - tmp1) << PASS1_BITS));
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 12)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/6)*(8/12) = 8/9, which we
+   * fold into the constant multipliers:
+   * 12-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/24) * 8/9.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 6; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*3];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*2];
+    tmp2 = dataptr[DCTSIZE*2] + wsptr[DCTSIZE*1];
+    tmp3 = dataptr[DCTSIZE*3] + wsptr[DCTSIZE*0];
+    tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*7];
+    tmp5 = dataptr[DCTSIZE*5] + dataptr[DCTSIZE*6];
+
+    tmp10 = tmp0 + tmp5;
+    tmp13 = tmp0 - tmp5;
+    tmp11 = tmp1 + tmp4;
+    tmp14 = tmp1 - tmp4;
+    tmp12 = tmp2 + tmp3;
+    tmp15 = tmp2 - tmp3;
+
+    tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*3];
+    tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*2];
+    tmp2 = dataptr[DCTSIZE*2] - wsptr[DCTSIZE*1];
+    tmp3 = dataptr[DCTSIZE*3] - wsptr[DCTSIZE*0];
+    tmp4 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*7];
+    tmp5 = dataptr[DCTSIZE*5] - dataptr[DCTSIZE*6];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12, FIX(0.888888889)), /* 8/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp13 - tmp14 - tmp15, FIX(0.888888889)), /* 8/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.088662108)),         /* c4 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp14 - tmp15, FIX(0.888888889)) +        /* 8/9 */
+	      MULTIPLY(tmp13 + tmp15, FIX(1.214244803)),         /* c2 */
+	      CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp1 + tmp4, FIX(0.481063200));   /* c9 */
+    tmp14 = tmp10 + MULTIPLY(tmp1, FIX(0.680326102));  /* c3-c9 */
+    tmp15 = tmp10 - MULTIPLY(tmp4, FIX(1.642452502));  /* c3+c9 */
+    tmp12 = MULTIPLY(tmp0 + tmp2, FIX(0.997307603));   /* c5 */
+    tmp13 = MULTIPLY(tmp0 + tmp3, FIX(0.765261039));   /* c7 */
+    tmp10 = tmp12 + tmp13 + tmp14 - MULTIPLY(tmp0, FIX(0.516244403)) /* c5+c7-c1 */
+	    + MULTIPLY(tmp5, FIX(0.164081699));        /* c11 */
+    tmp11 = MULTIPLY(tmp2 + tmp3, - FIX(0.164081699)); /* -c11 */
+    tmp12 += tmp11 - tmp15 - MULTIPLY(tmp2, FIX(2.079550144)) /* c1+c5-c11 */
+	    + MULTIPLY(tmp5, FIX(0.765261039));        /* c7 */
+    tmp13 += tmp11 - tmp14 + MULTIPLY(tmp3, FIX(0.645144899)) /* c1+c11-c7 */
+	    - MULTIPLY(tmp5, FIX(0.997307603));        /* c5 */
+    tmp11 = tmp15 + MULTIPLY(tmp0 - tmp3, FIX(1.161389302)) /* c3 */
+	    - MULTIPLY(tmp2 + tmp5, FIX(0.481063200)); /* c9 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM) DESCALE(tmp10, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp11, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*5] = (DCTELEM) DESCALE(tmp12, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp13, CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 5x10 sample block.
+ *
+ * 5-point FDCT in pass 1 (rows), 10-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_5x10 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4;
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
+  DCTELEM workspace[8*2];
+  DCTELEM *dataptr;
+  DCTELEM *wsptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* 5-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/10). */
+
+  dataptr = data;
+  ctr = 0;
+  for (;;) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[4]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[3]);
+    tmp2 = GETJSAMPLE(elemptr[2]);
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+
+    tmp0 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[4]);
+    tmp1 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[3]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp10 + tmp2 - 5 * CENTERJSAMPLE) << PASS1_BITS);
+    tmp11 = MULTIPLY(tmp11, FIX(0.790569415));          /* (c2+c4)/2 */
+    tmp10 -= tmp2 << 2;
+    tmp10 = MULTIPLY(tmp10, FIX(0.353553391));          /* (c2-c4)/2 */
+    dataptr[2] = (DCTELEM) DESCALE(tmp11 + tmp10, CONST_BITS-PASS1_BITS);
+    dataptr[4] = (DCTELEM) DESCALE(tmp11 - tmp10, CONST_BITS-PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp0 + tmp1, FIX(0.831253876));    /* c3 */
+
+    dataptr[1] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp0, FIX(0.513743148)), /* c1-c3 */
+	      CONST_BITS-PASS1_BITS);
+    dataptr[3] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp1, FIX(2.176250899)), /* c1+c3 */
+	      CONST_BITS-PASS1_BITS);
+
+    ctr++;
+
+    if (ctr != DCTSIZE) {
+      if (ctr == 10)
+	break;			/* Done. */
+      dataptr += DCTSIZE;	/* advance pointer to next row */
+    } else
+      dataptr = workspace;	/* switch pointer to extended workspace */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/5)*(8/10) = 32/25, which we
+   * fold into the constant multipliers:
+   * 10-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/20) * 32/25.
+   */
+
+  dataptr = data;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 5; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + wsptr[DCTSIZE*1];
+    tmp1 = dataptr[DCTSIZE*1] + wsptr[DCTSIZE*0];
+    tmp12 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*7];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*6];
+    tmp4 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5];
+
+    tmp10 = tmp0 + tmp4;
+    tmp13 = tmp0 - tmp4;
+    tmp11 = tmp1 + tmp3;
+    tmp14 = tmp1 - tmp3;
+
+    tmp0 = dataptr[DCTSIZE*0] - wsptr[DCTSIZE*1];
+    tmp1 = dataptr[DCTSIZE*1] - wsptr[DCTSIZE*0];
+    tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*7];
+    tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*6];
+    tmp4 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11 + tmp12, FIX(1.28)), /* 32/25 */
+	      CONST_BITS+PASS1_BITS);
+    tmp12 += tmp12;
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp12, FIX(1.464477191)) - /* c4 */
+	      MULTIPLY(tmp11 - tmp12, FIX(0.559380511)),  /* c8 */
+	      CONST_BITS+PASS1_BITS);
+    tmp10 = MULTIPLY(tmp13 + tmp14, FIX(1.064004961));    /* c6 */
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp13, FIX(0.657591230)),  /* c2-c6 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      DESCALE(tmp10 - MULTIPLY(tmp14, FIX(2.785601151)),  /* c2+c6 */
+	      CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = tmp0 + tmp4;
+    tmp11 = tmp1 - tmp3;
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp11 - tmp2, FIX(1.28)),  /* 32/25 */
+	      CONST_BITS+PASS1_BITS);
+    tmp2 = MULTIPLY(tmp2, FIX(1.28));                     /* 32/25 */
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0, FIX(1.787906876)) +          /* c1 */
+	      MULTIPLY(tmp1, FIX(1.612894094)) + tmp2 +   /* c3 */
+	      MULTIPLY(tmp3, FIX(0.821810588)) +          /* c7 */
+	      MULTIPLY(tmp4, FIX(0.283176630)),           /* c9 */
+	      CONST_BITS+PASS1_BITS);
+    tmp12 = MULTIPLY(tmp0 - tmp4, FIX(1.217352341)) -     /* (c3+c7)/2 */
+	    MULTIPLY(tmp1 + tmp3, FIX(0.752365123));      /* (c1-c9)/2 */
+    tmp13 = MULTIPLY(tmp10 + tmp11, FIX(0.395541753)) +   /* (c3-c7)/2 */
+	    MULTIPLY(tmp11, FIX(0.64)) - tmp2;            /* 16/25 */
+    dataptr[DCTSIZE*3] = (DCTELEM) DESCALE(tmp12 + tmp13, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*7] = (DCTELEM) DESCALE(tmp12 - tmp13, CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+    wsptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 4x8 sample block.
+ *
+ * 4-point FDCT in pass 1 (rows), 8-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_4x8 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3;
+  INT32 tmp10, tmp11, tmp12, tmp13;
+  INT32 z1;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* We must also scale the output by 8/4 = 2, which we add here. */
+  /* 4-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[3]);
+    tmp1 = GETJSAMPLE(elemptr[1]) + GETJSAMPLE(elemptr[2]);
+
+    tmp10 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[3]);
+    tmp11 = GETJSAMPLE(elemptr[1]) - GETJSAMPLE(elemptr[2]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp0 + tmp1 - 4 * CENTERJSAMPLE) << (PASS1_BITS+1));
+    dataptr[2] = (DCTELEM) ((tmp0 - tmp1) << (PASS1_BITS+1));
+
+    /* Odd part */
+
+    tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS-PASS1_BITS-2);
+
+    dataptr[1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
+		  CONST_BITS-PASS1_BITS-1);
+    dataptr[3] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
+		  CONST_BITS-PASS1_BITS-1);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 4; ctr++) {
+    /* Even part per LL&M figure 1 --- note that published figure is faulty;
+     * rotator "sqrt(2)*c1" should be "sqrt(2)*c6".
+     */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
+
+    /* Add fudge factor here for final descale. */
+    tmp10 = tmp0 + tmp3 + (ONE << (PASS1_BITS-1));
+    tmp12 = tmp0 - tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp13 = tmp1 - tmp2;
+
+    tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
+    tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
+    tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
+    tmp3 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
+
+    dataptr[DCTSIZE*0] = (DCTELEM) RIGHT_SHIFT(tmp10 + tmp11, PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM) RIGHT_SHIFT(tmp10 - tmp11, PASS1_BITS);
+
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS+PASS1_BITS-1);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      RIGHT_SHIFT(z1 + MULTIPLY(tmp12, FIX_0_765366865), CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*6] = (DCTELEM)
+      RIGHT_SHIFT(z1 - MULTIPLY(tmp13, FIX_1_847759065), CONST_BITS+PASS1_BITS);
+
+    /* Odd part per figure 8 --- note paper omits factor of sqrt(2).
+     * 8-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
+     * i0..i3 in the paper are tmp0..tmp3 here.
+     */
+
+    tmp10 = tmp0 + tmp3;
+    tmp11 = tmp1 + tmp2;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp1 + tmp3;
+    z1 = MULTIPLY(tmp12 + tmp13, FIX_1_175875602); /*  c3 */
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS+PASS1_BITS-1);
+
+    tmp0  = MULTIPLY(tmp0,    FIX_1_501321110);    /*  c1+c3-c5-c7 */
+    tmp1  = MULTIPLY(tmp1,    FIX_3_072711026);    /*  c1+c3+c5-c7 */
+    tmp2  = MULTIPLY(tmp2,    FIX_2_053119869);    /*  c1+c3-c5+c7 */
+    tmp3  = MULTIPLY(tmp3,    FIX_0_298631336);    /* -c1+c3+c5-c7 */
+    tmp10 = MULTIPLY(tmp10, - FIX_0_899976223);    /*  c7-c3 */
+    tmp11 = MULTIPLY(tmp11, - FIX_2_562915447);    /* -c1-c3 */
+    tmp12 = MULTIPLY(tmp12, - FIX_0_390180644);    /*  c5-c3 */
+    tmp13 = MULTIPLY(tmp13, - FIX_1_961570560);    /* -c3-c5 */
+
+    tmp12 += z1;
+    tmp13 += z1;
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + tmp10 + tmp12, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      RIGHT_SHIFT(tmp1 + tmp11 + tmp13, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      RIGHT_SHIFT(tmp2 + tmp11 + tmp12, CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*7] = (DCTELEM)
+      RIGHT_SHIFT(tmp3 + tmp10 + tmp13, CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 3x6 sample block.
+ *
+ * 3-point FDCT in pass 1 (rows), 6-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_3x6 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1, tmp2;
+  INT32 tmp10, tmp11, tmp12;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+  /* We scale the results further by 2 as part of output adaption */
+  /* scaling for different DCT size. */
+  /* 3-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/6). */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 6; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]) + GETJSAMPLE(elemptr[2]);
+    tmp1 = GETJSAMPLE(elemptr[1]);
+
+    tmp2 = GETJSAMPLE(elemptr[0]) - GETJSAMPLE(elemptr[2]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM)
+      ((tmp0 + tmp1 - 3 * CENTERJSAMPLE) << (PASS1_BITS+1));
+    dataptr[2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp1 - tmp1, FIX(0.707106781)), /* c2 */
+	      CONST_BITS-PASS1_BITS-1);
+
+    /* Odd part */
+
+    dataptr[1] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp2, FIX(1.224744871)),               /* c1 */
+	      CONST_BITS-PASS1_BITS-1);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We remove the PASS1_BITS scaling, but leave the results scaled up
+   * by an overall factor of 8.
+   * We must also scale the output by (8/6)*(8/3) = 32/9, which we partially
+   * fold into the constant multipliers (other part was done in pass 1):
+   * 6-point FDCT kernel, cK represents sqrt(2) * cos(K*pi/12) * 16/9.
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 3; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*5];
+    tmp11 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3];
+
+    tmp10 = tmp0 + tmp2;
+    tmp12 = tmp0 - tmp2;
+
+    tmp0 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*5];
+    tmp1 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*4];
+    tmp2 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3];
+
+    dataptr[DCTSIZE*0] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 + tmp11, FIX(1.777777778)),         /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*2] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp12, FIX(2.177324216)),                 /* c2 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*4] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp10 - tmp11 - tmp11, FIX(1.257078722)), /* c4 */
+	      CONST_BITS+PASS1_BITS);
+
+    /* Odd part */
+
+    tmp10 = MULTIPLY(tmp0 + tmp2, FIX(0.650711829));             /* c5 */
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp0 + tmp1, FIX(1.777777778)),   /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      DESCALE(MULTIPLY(tmp0 - tmp1 - tmp2, FIX(1.777777778)),    /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+    dataptr[DCTSIZE*5] = (DCTELEM)
+      DESCALE(tmp10 + MULTIPLY(tmp2 - tmp1, FIX(1.777777778)),   /* 16/9 */
+	      CONST_BITS+PASS1_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 2x4 sample block.
+ *
+ * 2-point FDCT in pass 1 (rows), 4-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_2x4 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1;
+  INT32 tmp10, tmp11;
+  DCTELEM *dataptr;
+  JSAMPROW elemptr;
+  int ctr;
+  SHIFT_TEMPS
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  /* Pass 1: process rows. */
+  /* Note results are scaled up by sqrt(8) compared to a true DCT. */
+  /* We must also scale the output by (8/2)*(8/4) = 2**3, which we add here. */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 4; ctr++) {
+    elemptr = sample_data[ctr] + start_col;
+
+    /* Even part */
+
+    tmp0 = GETJSAMPLE(elemptr[0]);
+    tmp1 = GETJSAMPLE(elemptr[1]);
+
+    /* Apply unsigned->signed conversion */
+    dataptr[0] = (DCTELEM) ((tmp0 + tmp1 - 2 * CENTERJSAMPLE) << 3);
+
+    /* Odd part */
+
+    dataptr[1] = (DCTELEM) ((tmp0 - tmp1) << 3);
+
+    dataptr += DCTSIZE;		/* advance pointer to next row */
+  }
+
+  /* Pass 2: process columns.
+   * We leave the results scaled up by an overall factor of 8.
+   * 4-point FDCT kernel,
+   * cK represents sqrt(2) * cos(K*pi/16) [refers to 8-point FDCT].
+   */
+
+  dataptr = data;
+  for (ctr = 0; ctr < 2; ctr++) {
+    /* Even part */
+
+    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*3];
+    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*2];
+
+    tmp10 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*3];
+    tmp11 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*2];
+
+    dataptr[DCTSIZE*0] = (DCTELEM) (tmp0 + tmp1);
+    dataptr[DCTSIZE*2] = (DCTELEM) (tmp0 - tmp1);
+
+    /* Odd part */
+
+    tmp0 = MULTIPLY(tmp10 + tmp11, FIX_0_541196100);       /* c6 */
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS-1);
+
+    dataptr[DCTSIZE*1] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 + MULTIPLY(tmp10, FIX_0_765366865), /* c2-c6 */
+		  CONST_BITS);
+    dataptr[DCTSIZE*3] = (DCTELEM)
+      RIGHT_SHIFT(tmp0 - MULTIPLY(tmp11, FIX_1_847759065), /* c2+c6 */
+		  CONST_BITS);
+
+    dataptr++;			/* advance pointer to next column */
+  }
+}
+
+
+/*
+ * Perform the forward DCT on a 1x2 sample block.
+ *
+ * 1-point FDCT in pass 1 (rows), 2-point in pass 2 (columns).
+ */
+
+GLOBAL(void)
+jpeg_fdct_1x2 (DCTELEM * data, JSAMPARRAY sample_data, JDIMENSION start_col)
+{
+  INT32 tmp0, tmp1;
+
+  /* Pre-zero output coefficient block. */
+  MEMZERO(data, SIZEOF(DCTELEM) * DCTSIZE2);
+
+  tmp0 = GETJSAMPLE(sample_data[0][start_col]);
+  tmp1 = GETJSAMPLE(sample_data[1][start_col]);
+
+  /* We leave the results scaled up by an overall factor of 8.
+   * We must also scale the output by (8/1)*(8/2) = 2**5.
+   */
+
+  /* Even part */
+  /* Apply unsigned->signed conversion */
+  data[DCTSIZE*0] = (DCTELEM) ((tmp0 + tmp1 - 2 * CENTERJSAMPLE) << 5);
+
+  /* Odd part */
+  data[DCTSIZE*1] = (DCTELEM) ((tmp0 - tmp1) << 5);
+}
+
+#endif /* DCT_SCALING_SUPPORTED */
+#endif /* DCT_ISLOW_SUPPORTED */

Added: trunk/code/jpeg-8c/jidctflt.c
===================================================================
--- trunk/code/jpeg-8c/jidctflt.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jidctflt.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,235 @@
+/*
+ * jidctflt.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * Modified 2010 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a floating-point implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * This implementation should be more accurate than either of the integer
+ * IDCT implementations.  However, it may not give the same results on all
+ * machines because of differences in roundoff behavior.  Speed will depend
+ * on the hardware's floating point capacity.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with a fixed-point
+ * implementation, accuracy is lost due to imprecise representation of the
+ * scaled quantization values.  However, that problem does not arise if
+ * we use floating point arithmetic.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_FLOAT_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce a float result.
+ */
+
+#define DEQUANTIZE(coef,quantval)  (((FAST_FLOAT) (coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
+  FAST_FLOAT z5, z10, z11, z12, z13;
+  JCOEFPTR inptr;
+  FLOAT_MULT_TYPE * quantptr;
+  FAST_FLOAT * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = cinfo->sample_range_limit;
+  int ctr;
+  FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (FLOAT_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = tmp0 + tmp2;	/* phase 3 */
+    tmp11 = tmp0 - tmp2;
+
+    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
+    tmp12 = (tmp1 - tmp3) * ((FAST_FLOAT) 1.414213562) - tmp13; /* 2*c4 */
+
+    tmp0 = tmp10 + tmp13;	/* phase 2 */
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+    
+    /* Odd part */
+
+    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    z13 = tmp6 + tmp5;		/* phase 6 */
+    z10 = tmp6 - tmp5;
+    z11 = tmp4 + tmp7;
+    z12 = tmp4 - tmp7;
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */
+
+    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
+    tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */
+    tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 - tmp5;
+
+    wsptr[DCTSIZE*0] = tmp0 + tmp7;
+    wsptr[DCTSIZE*7] = tmp0 - tmp7;
+    wsptr[DCTSIZE*1] = tmp1 + tmp6;
+    wsptr[DCTSIZE*6] = tmp1 - tmp6;
+    wsptr[DCTSIZE*2] = tmp2 + tmp5;
+    wsptr[DCTSIZE*5] = tmp2 - tmp5;
+    wsptr[DCTSIZE*3] = tmp3 + tmp4;
+    wsptr[DCTSIZE*4] = tmp3 - tmp4;
+
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+  
+  /* Pass 2: process rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * And testing floats for zero is relatively expensive, so we don't bother.
+     */
+    
+    /* Even part */
+
+    /* Apply signed->unsigned and prepare float->int conversion */
+    z5 = wsptr[0] + ((FAST_FLOAT) CENTERJSAMPLE + (FAST_FLOAT) 0.5);
+    tmp10 = z5 + wsptr[4];
+    tmp11 = z5 - wsptr[4];
+
+    tmp13 = wsptr[2] + wsptr[6];
+    tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13;
+
+    tmp0 = tmp10 + tmp13;
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z13 = wsptr[5] + wsptr[3];
+    z10 = wsptr[5] - wsptr[3];
+    z11 = wsptr[1] + wsptr[7];
+    z12 = wsptr[1] - wsptr[7];
+
+    tmp7 = z11 + z13;
+    tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562);
+
+    z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */
+    tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */
+    tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 - tmp5;
+
+    /* Final output stage: float->int conversion and range-limit */
+
+    outptr[0] = range_limit[((int) (tmp0 + tmp7)) & RANGE_MASK];
+    outptr[7] = range_limit[((int) (tmp0 - tmp7)) & RANGE_MASK];
+    outptr[1] = range_limit[((int) (tmp1 + tmp6)) & RANGE_MASK];
+    outptr[6] = range_limit[((int) (tmp1 - tmp6)) & RANGE_MASK];
+    outptr[2] = range_limit[((int) (tmp2 + tmp5)) & RANGE_MASK];
+    outptr[5] = range_limit[((int) (tmp2 - tmp5)) & RANGE_MASK];
+    outptr[3] = range_limit[((int) (tmp3 + tmp4)) & RANGE_MASK];
+    outptr[4] = range_limit[((int) (tmp3 - tmp4)) & RANGE_MASK];
+    
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_FLOAT_SUPPORTED */

Added: trunk/code/jpeg-8c/jidctfst.c
===================================================================
--- trunk/code/jpeg-8c/jidctfst.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jidctfst.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,368 @@
+/*
+ * jidctfst.c
+ *
+ * Copyright (C) 1994-1998, Thomas G. Lane.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a fast, not so accurate integer implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on Arai, Agui, and Nakajima's algorithm for
+ * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
+ * Japanese, but the algorithm is described in the Pennebaker & Mitchell
+ * JPEG textbook (see REFERENCES section in file README).  The following code
+ * is based directly on figure 4-8 in P&M.
+ * While an 8-point DCT cannot be done in less than 11 multiplies, it is
+ * possible to arrange the computation so that many of the multiplies are
+ * simple scalings of the final outputs.  These multiplies can then be
+ * folded into the multiplications or divisions by the JPEG quantization
+ * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
+ * to be done in the DCT itself.
+ * The primary disadvantage of this method is that with fixed-point math,
+ * accuracy is lost due to imprecise representation of the scaled
+ * quantization values.  The smaller the quantization table entry, the less
+ * precise the scaled value, so this implementation does worse with high-
+ * quality-setting files than with low-quality ones.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_IFAST_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
+#endif
+
+
+/* Scaling decisions are generally the same as in the LL&M algorithm;
+ * see jidctint.c for more details.  However, we choose to descale
+ * (right shift) multiplication products as soon as they are formed,
+ * rather than carrying additional fractional bits into subsequent additions.
+ * This compromises accuracy slightly, but it lets us save a few shifts.
+ * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
+ * everywhere except in the multiplications proper; this saves a good deal
+ * of work on 16-bit-int machines.
+ *
+ * The dequantized coefficients are not integers because the AA&N scaling
+ * factors have been incorporated.  We represent them scaled up by PASS1_BITS,
+ * so that the first and second IDCT rounds have the same input scaling.
+ * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to
+ * avoid a descaling shift; this compromises accuracy rather drastically
+ * for small quantization table entries, but it saves a lot of shifts.
+ * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway,
+ * so we use a much larger scaling factor to preserve accuracy.
+ *
+ * A final compromise is to represent the multiplicative constants to only
+ * 8 fractional bits, rather than 13.  This saves some shifting work on some
+ * machines, and may also reduce the cost of multiplication (since there
+ * are fewer one-bits in the constants).
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  8
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  8
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 8
+#define FIX_1_082392200  ((INT32)  277)		/* FIX(1.082392200) */
+#define FIX_1_414213562  ((INT32)  362)		/* FIX(1.414213562) */
+#define FIX_1_847759065  ((INT32)  473)		/* FIX(1.847759065) */
+#define FIX_2_613125930  ((INT32)  669)		/* FIX(2.613125930) */
+#else
+#define FIX_1_082392200  FIX(1.082392200)
+#define FIX_1_414213562  FIX(1.414213562)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_2_613125930  FIX(2.613125930)
+#endif
+
+
+/* We can gain a little more speed, with a further compromise in accuracy,
+ * by omitting the addition in a descaling shift.  This yields an incorrectly
+ * rounded result half the time...
+ */
+
+#ifndef USE_ACCURATE_ROUNDING
+#undef DESCALE
+#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
+#endif
+
+
+/* Multiply a DCTELEM variable by an INT32 constant, and immediately
+ * descale to yield a DCTELEM result.
+ */
+
+#define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce a DCTELEM result.  For 8-bit data a 16x16->16
+ * multiplication will do.  For 12-bit data, the multiplier table is
+ * declared INT32, so a 32-bit multiply will be used.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define DEQUANTIZE(coef,quantval)  (((IFAST_MULT_TYPE) (coef)) * (quantval))
+#else
+#define DEQUANTIZE(coef,quantval)  \
+	DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS)
+#endif
+
+
+/* Like DESCALE, but applies to a DCTELEM and produces an int.
+ * We assume that int right shift is unsigned if INT32 right shift is.
+ */
+
+#ifdef RIGHT_SHIFT_IS_UNSIGNED
+#define ISHIFT_TEMPS	DCTELEM ishift_temp;
+#if BITS_IN_JSAMPLE == 8
+#define DCTELEMBITS  16		/* DCTELEM may be 16 or 32 bits */
+#else
+#define DCTELEMBITS  32		/* DCTELEM must be 32 bits */
+#endif
+#define IRIGHT_SHIFT(x,shft)  \
+    ((ishift_temp = (x)) < 0 ? \
+     (ishift_temp >> (shft)) | ((~((DCTELEM) 0)) << (DCTELEMBITS-(shft))) : \
+     (ishift_temp >> (shft)))
+#else
+#define ISHIFT_TEMPS
+#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
+#endif
+
+#ifdef USE_ACCURATE_ROUNDING
+#define IDESCALE(x,n)  ((int) IRIGHT_SHIFT((x) + (1 << ((n)-1)), n))
+#else
+#define IDESCALE(x,n)  ((int) IRIGHT_SHIFT(x, n))
+#endif
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  DCTELEM tmp10, tmp11, tmp12, tmp13;
+  DCTELEM z5, z10, z11, z12, z13;
+  JCOEFPTR inptr;
+  IFAST_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE2];	/* buffers data between passes */
+  SHIFT_TEMPS			/* for DESCALE */
+  ISHIFT_TEMPS			/* for IDESCALE */
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (IFAST_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      int dcval = (int) DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = tmp0 + tmp2;	/* phase 3 */
+    tmp11 = tmp0 - tmp2;
+
+    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
+    tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
+
+    tmp0 = tmp10 + tmp13;	/* phase 2 */
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+    
+    /* Odd part */
+
+    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    z13 = tmp6 + tmp5;		/* phase 6 */
+    z10 = tmp6 - tmp5;
+    z11 = tmp4 + tmp7;
+    z12 = tmp4 - tmp7;
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
+
+    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
+    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
+    tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);
+    wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);
+    wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);
+    wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);
+    wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5);
+    wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);
+    wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4);
+    wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4);
+
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+  
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * On machines with very fast multiplication, it's possible that the
+     * test takes more time than it's worth.  In that case this section
+     * may be commented out.
+     */
+    
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
+	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[IDESCALE(wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+      
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      outptr[4] = dcval;
+      outptr[5] = dcval;
+      outptr[6] = dcval;
+      outptr[7] = dcval;
+
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+    
+    /* Even part */
+
+    tmp10 = ((DCTELEM) wsptr[0] + (DCTELEM) wsptr[4]);
+    tmp11 = ((DCTELEM) wsptr[0] - (DCTELEM) wsptr[4]);
+
+    tmp13 = ((DCTELEM) wsptr[2] + (DCTELEM) wsptr[6]);
+    tmp12 = MULTIPLY((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6], FIX_1_414213562)
+	    - tmp13;
+
+    tmp0 = tmp10 + tmp13;
+    tmp3 = tmp10 - tmp13;
+    tmp1 = tmp11 + tmp12;
+    tmp2 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3];
+    z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3];
+    z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7];
+    z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7];
+
+    tmp7 = z11 + z13;		/* phase 5 */
+    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
+
+    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
+    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
+    tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
+
+    tmp6 = tmp12 - tmp7;	/* phase 2 */
+    tmp5 = tmp11 - tmp6;
+    tmp4 = tmp10 + tmp5;
+
+    /* Final output stage: scale down by a factor of 8 and range-limit */
+
+    outptr[0] = range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#endif /* DCT_IFAST_SUPPORTED */

Added: trunk/code/jpeg-8c/jidctint.c
===================================================================
--- trunk/code/jpeg-8c/jidctint.c	                        (rev 0)
+++ trunk/code/jpeg-8c/jidctint.c	2011-03-12 16:45:15 UTC (rev 1926)
@@ -0,0 +1,5137 @@
+/*
+ * jidctint.c
+ *
+ * Copyright (C) 1991-1998, Thomas G. Lane.
+ * Modification developed 2002-2009 by Guido Vollbeding.
+ * This file is part of the Independent JPEG Group's software.
+ * For conditions of distribution and use, see the accompanying README file.
+ *
+ * This file contains a slow-but-accurate integer implementation of the
+ * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
+ * must also perform dequantization of the input coefficients.
+ *
+ * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
+ * on each row (or vice versa, but it's more convenient to emit a row at
+ * a time).  Direct algorithms are also available, but they are much more
+ * complex and seem not to be any faster when reduced to code.
+ *
+ * This implementation is based on an algorithm described in
+ *   C. Loeffler, A. Ligtenberg and G. Moschytz, "Practical Fast 1-D DCT
+ *   Algorithms with 11 Multiplications", Proc. Int'l. Conf. on Acoustics,
+ *   Speech, and Signal Processing 1989 (ICASSP '89), pp. 988-991.
+ * The primary algorithm described there uses 11 multiplies and 29 adds.
+ * We use their alternate method with 12 multiplies and 32 adds.
+ * The advantage of this method is that no data path contains more than one
+ * multiplication; this allows a very simple and accurate implementation in
+ * scaled fixed-point arithmetic, with a minimal number of shifts.
+ *
+ * We also provide IDCT routines with various output sample block sizes for
+ * direct resolution reduction or enlargement and for direct resolving the
+ * common 2x1 and 1x2 subsampling cases without additional resampling: NxN
+ * (N=1...16), 2NxN, and Nx2N (N=1...8) pixels for one 8x8 input DCT block.
+ *
+ * For N<8 we simply take the corresponding low-frequency coefficients of
+ * the 8x8 input DCT block and apply an NxN point IDCT on the sub-block
+ * to yield the downscaled outputs.
+ * This can be seen as direct low-pass downsampling from the DCT domain
+ * point of view rather than the usual spatial domain point of view,
+ * yielding significant computational savings and results at least
+ * as good as common bilinear (averaging) spatial downsampling.
+ *
+ * For N>8 we apply a partial NxN IDCT on the 8 input coefficients as
+ * lower frequencies and higher frequencies assumed to be zero.
+ * It turns out that the computational effort is similar to the 8x8 IDCT
+ * regarding the output size.
+ * Furthermore, the scaling and descaling is the same for all IDCT sizes.
+ *
+ * CAUTION: We rely on the FIX() macro except for the N=1,2,4,8 cases
+ * since there would be too many additional constants to pre-calculate.
+ */
+
+#define JPEG_INTERNALS
+#include "jinclude.h"
+#include "jpeglib.h"
+#include "jdct.h"		/* Private declarations for DCT subsystem */
+
+#ifdef DCT_ISLOW_SUPPORTED
+
+
+/*
+ * This module is specialized to the case DCTSIZE = 8.
+ */
+
+#if DCTSIZE != 8
+  Sorry, this code only copes with 8x8 DCT blocks. /* deliberate syntax err */
+#endif
+
+
+/*
+ * The poop on this scaling stuff is as follows:
+ *
+ * Each 1-D IDCT step produces outputs which are a factor of sqrt(N)
+ * larger than the true IDCT outputs.  The final outputs are therefore
+ * a factor of N larger than desired; since N=8 this can be cured by
+ * a simple right shift at the end of the algorithm.  The advantage of
+ * this arrangement is that we save two multiplications per 1-D IDCT,
+ * because the y0 and y4 inputs need not be divided by sqrt(N).
+ *
+ * We have to do addition and subtraction of the integer inputs, which
+ * is no problem, and multiplication by fractional constants, which is
+ * a problem to do in integer arithmetic.  We multiply all the constants
+ * by CONST_SCALE and convert them to integer constants (thus retaining
+ * CONST_BITS bits of precision in the constants).  After doing a
+ * multiplication we have to divide the product by CONST_SCALE, with proper
+ * rounding, to produce the correct output.  This division can be done
+ * cheaply as a right shift of CONST_BITS bits.  We postpone shifting
+ * as long as possible so that partial sums can be added together with
+ * full fractional precision.
+ *
+ * The outputs of the first pass are scaled up by PASS1_BITS bits so that
+ * they are represented to better-than-integral precision.  These outputs
+ * require BITS_IN_JSAMPLE + PASS1_BITS + 3 bits; this fits in a 16-bit word
+ * with the recommended scaling.  (To scale up 12-bit sample data further, an
+ * intermediate INT32 array would be needed.)
+ *
+ * To avoid overflow of the 32-bit intermediate results in pass 2, we must
+ * have BITS_IN_JSAMPLE + CONST_BITS + PASS1_BITS <= 26.  Error analysis
+ * shows that the values given below are the most effective.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define CONST_BITS  13
+#define PASS1_BITS  2
+#else
+#define CONST_BITS  13
+#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
+#endif
+
+/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
+ * causing a lot of useless floating-point operations at run time.
+ * To get around this we use the following pre-calculated constants.
+ * If you change CONST_BITS you may want to add appropriate values.
+ * (With a reasonable C compiler, you can just rely on the FIX() macro...)
+ */
+
+#if CONST_BITS == 13
+#define FIX_0_298631336  ((INT32)  2446)	/* FIX(0.298631336) */
+#define FIX_0_390180644  ((INT32)  3196)	/* FIX(0.390180644) */
+#define FIX_0_541196100  ((INT32)  4433)	/* FIX(0.541196100) */
+#define FIX_0_765366865  ((INT32)  6270)	/* FIX(0.765366865) */
+#define FIX_0_899976223  ((INT32)  7373)	/* FIX(0.899976223) */
+#define FIX_1_175875602  ((INT32)  9633)	/* FIX(1.175875602) */
+#define FIX_1_501321110  ((INT32)  12299)	/* FIX(1.501321110) */
+#define FIX_1_847759065  ((INT32)  15137)	/* FIX(1.847759065) */
+#define FIX_1_961570560  ((INT32)  16069)	/* FIX(1.961570560) */
+#define FIX_2_053119869  ((INT32)  16819)	/* FIX(2.053119869) */
+#define FIX_2_562915447  ((INT32)  20995)	/* FIX(2.562915447) */
+#define FIX_3_072711026  ((INT32)  25172)	/* FIX(3.072711026) */
+#else
+#define FIX_0_298631336  FIX(0.298631336)
+#define FIX_0_390180644  FIX(0.390180644)
+#define FIX_0_541196100  FIX(0.541196100)
+#define FIX_0_765366865  FIX(0.765366865)
+#define FIX_0_899976223  FIX(0.899976223)
+#define FIX_1_175875602  FIX(1.175875602)
+#define FIX_1_501321110  FIX(1.501321110)
+#define FIX_1_847759065  FIX(1.847759065)
+#define FIX_1_961570560  FIX(1.961570560)
+#define FIX_2_053119869  FIX(2.053119869)
+#define FIX_2_562915447  FIX(2.562915447)
+#define FIX_3_072711026  FIX(3.072711026)
+#endif
+
+
+/* Multiply an INT32 variable by an INT32 constant to yield an INT32 result.
+ * For 8-bit samples with the recommended scaling, all the variable
+ * and constant values involved are no more than 16 bits wide, so a
+ * 16x16->32 bit multiply can be used instead of a full 32x32 multiply.
+ * For 12-bit samples, a full 32-bit multiplication will be needed.
+ */
+
+#if BITS_IN_JSAMPLE == 8
+#define MULTIPLY(var,const)  MULTIPLY16C16(var,const)
+#else
+#define MULTIPLY(var,const)  ((var) * (const))
+#endif
+
+
+/* Dequantize a coefficient by multiplying it by the multiplier-table
+ * entry; produce an int result.  In this module, both inputs and result
+ * are 16 bits or less, so either int or short multiply will work.
+ */
+
+#define DEQUANTIZE(coef,quantval)  (((ISLOW_MULT_TYPE) (coef)) * (quantval))
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients.
+ */
+
+GLOBAL(void)
+jpeg_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3;
+  INT32 tmp10, tmp11, tmp12, tmp13;
+  INT32 z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[DCTSIZE2];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+  /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
+
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z2, FIX_0_765366865);
+    tmp3 = z1 - MULTIPLY(z3, FIX_1_847759065);
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z2 <<= CONST_BITS;
+    z3 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    z2 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    tmp0 = z2 + z3;
+    tmp1 = z2 - z3;
+
+    tmp10 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+    tmp11 = tmp1 + tmp3;
+    tmp12 = tmp1 - tmp3;
+
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    
+    z2 = tmp0 + tmp2;
+    z3 = tmp1 + tmp3;
+
+    z1 = MULTIPLY(z2 + z3, FIX_1_175875602); /* sqrt(2) * c3 */
+    z2 = MULTIPLY(z2, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z3 = MULTIPLY(z3, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    z2 += z1;
+    z3 += z1;
+
+    z1 = MULTIPLY(tmp0 + tmp3, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    tmp0 += z1 + z2;
+    tmp3 += z1 + z3;
+
+    z1 = MULTIPLY(tmp1 + tmp2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp1 += z1 + z3;
+    tmp2 += z1 + z2;
+
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+
+    wsptr[DCTSIZE*0] = (int) RIGHT_SHIFT(tmp10 + tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*7] = (int) RIGHT_SHIFT(tmp10 - tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*1] = (int) RIGHT_SHIFT(tmp11 + tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*6] = (int) RIGHT_SHIFT(tmp11 - tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*2] = (int) RIGHT_SHIFT(tmp12 + tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*5] = (int) RIGHT_SHIFT(tmp12 - tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*3] = (int) RIGHT_SHIFT(tmp13 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*4] = (int) RIGHT_SHIFT(tmp13 - tmp0, CONST_BITS-PASS1_BITS);
+    
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < DCTSIZE; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+    /* Rows of zeroes can be exploited in the same way as we did with columns.
+     * However, the column calculation has created many nonzero AC terms, so
+     * the simplification applies less often (typically 5% to 10% of the time).
+     * On machines with very fast multiplication, it's possible that the
+     * test takes more time than it's worth.  In that case this section
+     * may be commented out.
+     */
+
+#ifndef NO_ZERO_ROW_TEST
+    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
+	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
+      /* AC terms all zero */
+      JSAMPLE dcval = range_limit[(int) DESCALE((INT32) wsptr[0], PASS1_BITS+3)
+				  & RANGE_MASK];
+
+      outptr[0] = dcval;
+      outptr[1] = dcval;
+      outptr[2] = dcval;
+      outptr[3] = dcval;
+      outptr[4] = dcval;
+      outptr[5] = dcval;
+      outptr[6] = dcval;
+      outptr[7] = dcval;
+
+      wsptr += DCTSIZE;		/* advance pointer to next row */
+      continue;
+    }
+#endif
+
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+    
+    z2 = (INT32) wsptr[2];
+    z3 = (INT32) wsptr[6];
+
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z2, FIX_0_765366865);
+    tmp3 = z1 - MULTIPLY(z3, FIX_1_847759065);
+
+    /* Add fudge factor here for final descale. */
+    z2 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z3 = (INT32) wsptr[4];
+
+    tmp0 = (z2 + z3) << CONST_BITS;
+    tmp1 = (z2 - z3) << CONST_BITS;
+    
+    tmp10 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+    tmp11 = tmp1 + tmp3;
+    tmp12 = tmp1 - tmp3;
+
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+
+    tmp0 = (INT32) wsptr[7];
+    tmp1 = (INT32) wsptr[5];
+    tmp2 = (INT32) wsptr[3];
+    tmp3 = (INT32) wsptr[1];
+
+    z2 = tmp0 + tmp2;
+    z3 = tmp1 + tmp3;
+
+    z1 = MULTIPLY(z2 + z3, FIX_1_175875602); /* sqrt(2) * c3 */
+    z2 = MULTIPLY(z2, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z3 = MULTIPLY(z3, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    z2 += z1;
+    z3 += z1;
+
+    z1 = MULTIPLY(tmp0 + tmp3, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    tmp0 += z1 + z2;
+    tmp3 += z1 + z3;
+
+    z1 = MULTIPLY(tmp1 + tmp2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp1 += z1 + z3;
+    tmp2 += z1 + z2;
+
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp10 + tmp3,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) RIGHT_SHIFT(tmp10 - tmp3,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp11 + tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) RIGHT_SHIFT(tmp11 - tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp12 + tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) RIGHT_SHIFT(tmp12 - tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp13 + tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) RIGHT_SHIFT(tmp13 - tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += DCTSIZE;		/* advance pointer to next row */
+  }
+}
+
+#ifdef IDCT_SCALING_SUPPORTED
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 7x7 output block.
+ *
+ * Optimized algorithm with 12 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/14).
+ */
+
+GLOBAL(void)
+jpeg_idct_7x7 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp10, tmp11, tmp12, tmp13;
+  INT32 z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[7*7];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 7; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp13 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp13 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp13 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = MULTIPLY(z2 - z3, FIX(0.881747734));     /* c4 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.314692123));     /* c6 */
+    tmp11 = tmp10 + tmp12 + tmp13 - MULTIPLY(z2, FIX(1.841218003)); /* c2+c4-c6 */
+    tmp0 = z1 + z3;
+    z2 -= tmp0;
+    tmp0 = MULTIPLY(tmp0, FIX(1.274162392)) + tmp13; /* c2 */
+    tmp10 += tmp0 - MULTIPLY(z3, FIX(0.077722536));  /* c2-c4-c6 */
+    tmp12 += tmp0 - MULTIPLY(z1, FIX(2.470602249));  /* c2+c4+c6 */
+    tmp13 += MULTIPLY(z2, FIX(1.414213562));         /* c0 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+
+    tmp1 = MULTIPLY(z1 + z2, FIX(0.935414347));      /* (c3+c1-c5)/2 */
+    tmp2 = MULTIPLY(z1 - z2, FIX(0.170262339));      /* (c3+c5-c1)/2 */
+    tmp0 = tmp1 - tmp2;
+    tmp1 += tmp2;
+    tmp2 = MULTIPLY(z2 + z3, - FIX(1.378756276));    /* -c1 */
+    tmp1 += tmp2;
+    z2 = MULTIPLY(z1 + z3, FIX(0.613604268));        /* c5 */
+    tmp0 += z2;
+    tmp2 += z2 + MULTIPLY(z3, FIX(1.870828693));     /* c3+c1-c5 */
+
+    /* Final output stage */
+
+    wsptr[7*0] = (int) RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[7*6] = (int) RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[7*1] = (int) RIGHT_SHIFT(tmp11 + tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[7*5] = (int) RIGHT_SHIFT(tmp11 - tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[7*2] = (int) RIGHT_SHIFT(tmp12 + tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[7*4] = (int) RIGHT_SHIFT(tmp12 - tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[7*3] = (int) RIGHT_SHIFT(tmp13, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 7 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 7; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp13 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp13 <<= CONST_BITS;
+
+    z1 = (INT32) wsptr[2];
+    z2 = (INT32) wsptr[4];
+    z3 = (INT32) wsptr[6];
+
+    tmp10 = MULTIPLY(z2 - z3, FIX(0.881747734));     /* c4 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.314692123));     /* c6 */
+    tmp11 = tmp10 + tmp12 + tmp13 - MULTIPLY(z2, FIX(1.841218003)); /* c2+c4-c6 */
+    tmp0 = z1 + z3;
+    z2 -= tmp0;
+    tmp0 = MULTIPLY(tmp0, FIX(1.274162392)) + tmp13; /* c2 */
+    tmp10 += tmp0 - MULTIPLY(z3, FIX(0.077722536));  /* c2-c4-c6 */
+    tmp12 += tmp0 - MULTIPLY(z1, FIX(2.470602249));  /* c2+c4+c6 */
+    tmp13 += MULTIPLY(z2, FIX(1.414213562));         /* c0 */
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+
+    tmp1 = MULTIPLY(z1 + z2, FIX(0.935414347));      /* (c3+c1-c5)/2 */
+    tmp2 = MULTIPLY(z1 - z2, FIX(0.170262339));      /* (c3+c5-c1)/2 */
+    tmp0 = tmp1 - tmp2;
+    tmp1 += tmp2;
+    tmp2 = MULTIPLY(z2 + z3, - FIX(1.378756276));    /* -c1 */
+    tmp1 += tmp2;
+    z2 = MULTIPLY(z1 + z3, FIX(0.613604268));        /* c5 */
+    tmp0 += z2;
+    tmp2 += z2 + MULTIPLY(z3, FIX(1.870828693));     /* c3+c1-c5 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp10 + tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) RIGHT_SHIFT(tmp10 - tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp11 + tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) RIGHT_SHIFT(tmp11 - tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp12 + tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) RIGHT_SHIFT(tmp12 - tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp13,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += 7;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 6x6 output block.
+ *
+ * Optimized algorithm with 3 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/12).
+ */
+
+GLOBAL(void)
+jpeg_idct_6x6 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp10, tmp11, tmp12;
+  INT32 z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[6*6];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 6; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp0 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS-PASS1_BITS-1);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp10 = MULTIPLY(tmp2, FIX(0.707106781));   /* c4 */
+    tmp1 = tmp0 + tmp10;
+    tmp11 = RIGHT_SHIFT(tmp0 - tmp10 - tmp10, CONST_BITS-PASS1_BITS);
+    tmp10 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp0 = MULTIPLY(tmp10, FIX(1.224744871));   /* c2 */
+    tmp10 = tmp1 + tmp0;
+    tmp12 = tmp1 - tmp0;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp1 = MULTIPLY(z1 + z3, FIX(0.366025404)); /* c5 */
+    tmp0 = tmp1 + ((z1 + z2) << CONST_BITS);
+    tmp2 = tmp1 + ((z3 - z2) << CONST_BITS);
+    tmp1 = (z1 - z2 - z3) << PASS1_BITS;
+
+    /* Final output stage */
+
+    wsptr[6*0] = (int) RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[6*5] = (int) RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[6*1] = (int) (tmp11 + tmp1);
+    wsptr[6*4] = (int) (tmp11 - tmp1);
+    wsptr[6*2] = (int) RIGHT_SHIFT(tmp12 + tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[6*3] = (int) RIGHT_SHIFT(tmp12 - tmp2, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 6 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 6; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp0 <<= CONST_BITS;
+    tmp2 = (INT32) wsptr[4];
+    tmp10 = MULTIPLY(tmp2, FIX(0.707106781));   /* c4 */
+    tmp1 = tmp0 + tmp10;
+    tmp11 = tmp0 - tmp10 - tmp10;
+    tmp10 = (INT32) wsptr[2];
+    tmp0 = MULTIPLY(tmp10, FIX(1.224744871));   /* c2 */
+    tmp10 = tmp1 + tmp0;
+    tmp12 = tmp1 - tmp0;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    tmp1 = MULTIPLY(z1 + z3, FIX(0.366025404)); /* c5 */
+    tmp0 = tmp1 + ((z1 + z2) << CONST_BITS);
+    tmp2 = tmp1 + ((z3 - z2) << CONST_BITS);
+    tmp1 = (z1 - z2 - z3) << CONST_BITS;
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp10 + tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) RIGHT_SHIFT(tmp10 - tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp11 + tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) RIGHT_SHIFT(tmp11 - tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp12 + tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp12 - tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += 6;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 5x5 output block.
+ *
+ * Optimized algorithm with 5 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/10).
+ */
+
+GLOBAL(void)
+jpeg_idct_5x5 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp10, tmp11, tmp12;
+  INT32 z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[5*5];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 5; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp12 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp12 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp12 += ONE << (CONST_BITS-PASS1_BITS-1);
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z1 = MULTIPLY(tmp0 + tmp1, FIX(0.790569415)); /* (c2+c4)/2 */
+    z2 = MULTIPLY(tmp0 - tmp1, FIX(0.353553391)); /* (c2-c4)/2 */
+    z3 = tmp12 + z2;
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z1;
+    tmp12 -= z2 << 2;
+
+    /* Odd part */
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));     /* c3 */
+    tmp0 = z1 + MULTIPLY(z2, FIX(0.513743148));   /* c1-c3 */
+    tmp1 = z1 - MULTIPLY(z3, FIX(2.176250899));   /* c1+c3 */
+
+    /* Final output stage */
+
+    wsptr[5*0] = (int) RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[5*4] = (int) RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[5*1] = (int) RIGHT_SHIFT(tmp11 + tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[5*3] = (int) RIGHT_SHIFT(tmp11 - tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[5*2] = (int) RIGHT_SHIFT(tmp12, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 5 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 5; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp12 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp12 <<= CONST_BITS;
+    tmp0 = (INT32) wsptr[2];
+    tmp1 = (INT32) wsptr[4];
+    z1 = MULTIPLY(tmp0 + tmp1, FIX(0.790569415)); /* (c2+c4)/2 */
+    z2 = MULTIPLY(tmp0 - tmp1, FIX(0.353553391)); /* (c2-c4)/2 */
+    z3 = tmp12 + z2;
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z1;
+    tmp12 -= z2 << 2;
+
+    /* Odd part */
+
+    z2 = (INT32) wsptr[1];
+    z3 = (INT32) wsptr[3];
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));     /* c3 */
+    tmp0 = z1 + MULTIPLY(z2, FIX(0.513743148));   /* c1-c3 */
+    tmp1 = z1 - MULTIPLY(z3, FIX(2.176250899));   /* c1+c3 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp10 + tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) RIGHT_SHIFT(tmp10 - tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp11 + tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp11 - tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp12,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += 5;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 4x4 output block.
+ *
+ * Optimized algorithm with 3 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/16) [refers to 8-point IDCT].
+ */
+
+GLOBAL(void)
+jpeg_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp2, tmp10, tmp12;
+  INT32 z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[4*4];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 4; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    
+    tmp10 = (tmp0 + tmp2) << PASS1_BITS;
+    tmp12 = (tmp0 - tmp2) << PASS1_BITS;
+
+    /* Odd part */
+    /* Same rotation as in the even part of the 8x8 LL&M IDCT */
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);               /* c6 */
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-1);
+    tmp0 = RIGHT_SHIFT(z1 + MULTIPLY(z2, FIX_0_765366865), /* c2-c6 */
+		       CONST_BITS-PASS1_BITS);
+    tmp2 = RIGHT_SHIFT(z1 - MULTIPLY(z3, FIX_1_847759065), /* c2+c6 */
+		       CONST_BITS-PASS1_BITS);
+
+    /* Final output stage */
+
+    wsptr[4*0] = (int) (tmp10 + tmp0);
+    wsptr[4*3] = (int) (tmp10 - tmp0);
+    wsptr[4*1] = (int) (tmp12 + tmp2);
+    wsptr[4*2] = (int) (tmp12 - tmp2);
+  }
+
+  /* Pass 2: process 4 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 4; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp2 = (INT32) wsptr[2];
+
+    tmp10 = (tmp0 + tmp2) << CONST_BITS;
+    tmp12 = (tmp0 - tmp2) << CONST_BITS;
+
+    /* Odd part */
+    /* Same rotation as in the even part of the 8x8 LL&M IDCT */
+
+    z2 = (INT32) wsptr[1];
+    z3 = (INT32) wsptr[3];
+
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);   /* c6 */
+    tmp0 = z1 + MULTIPLY(z2, FIX_0_765366865); /* c2-c6 */
+    tmp2 = z1 - MULTIPLY(z3, FIX_1_847759065); /* c2+c6 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp10 + tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp10 - tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp12 + tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp12 - tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += 4;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 3x3 output block.
+ *
+ * Optimized algorithm with 2 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/6).
+ */
+
+GLOBAL(void)
+jpeg_idct_3x3 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp2, tmp10, tmp12;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[3*3];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 3; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp0 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS-PASS1_BITS-1);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp12 = MULTIPLY(tmp2, FIX(0.707106781)); /* c2 */
+    tmp10 = tmp0 + tmp12;
+    tmp2 = tmp0 - tmp12 - tmp12;
+
+    /* Odd part */
+
+    tmp12 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    tmp0 = MULTIPLY(tmp12, FIX(1.224744871)); /* c1 */
+
+    /* Final output stage */
+
+    wsptr[3*0] = (int) RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[3*2] = (int) RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[3*1] = (int) RIGHT_SHIFT(tmp2, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 3 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 3; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp0 <<= CONST_BITS;
+    tmp2 = (INT32) wsptr[2];
+    tmp12 = MULTIPLY(tmp2, FIX(0.707106781)); /* c2 */
+    tmp10 = tmp0 + tmp12;
+    tmp2 = tmp0 - tmp12 - tmp12;
+
+    /* Odd part */
+
+    tmp12 = (INT32) wsptr[1];
+    tmp0 = MULTIPLY(tmp12, FIX(1.224744871)); /* c1 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp10 + tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp10 - tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += 3;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 2x2 output block.
+ *
+ * Multiplication-less algorithm.
+ */
+
+GLOBAL(void)
+jpeg_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+  ISLOW_MULT_TYPE * quantptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input. */
+
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+
+  /* Column 0 */
+  tmp4 = DEQUANTIZE(coef_block[DCTSIZE*0], quantptr[DCTSIZE*0]);
+  tmp5 = DEQUANTIZE(coef_block[DCTSIZE*1], quantptr[DCTSIZE*1]);
+  /* Add fudge factor here for final descale. */
+  tmp4 += ONE << 2;
+
+  tmp0 = tmp4 + tmp5;
+  tmp2 = tmp4 - tmp5;
+
+  /* Column 1 */
+  tmp4 = DEQUANTIZE(coef_block[DCTSIZE*0+1], quantptr[DCTSIZE*0+1]);
+  tmp5 = DEQUANTIZE(coef_block[DCTSIZE*1+1], quantptr[DCTSIZE*1+1]);
+
+  tmp1 = tmp4 + tmp5;
+  tmp3 = tmp4 - tmp5;
+
+  /* Pass 2: process 2 rows, store into output array. */
+
+  /* Row 0 */
+  outptr = output_buf[0] + output_col;
+
+  outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp0 + tmp1, 3) & RANGE_MASK];
+  outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp0 - tmp1, 3) & RANGE_MASK];
+
+  /* Row 1 */
+  outptr = output_buf[1] + output_col;
+
+  outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp2 + tmp3, 3) & RANGE_MASK];
+  outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp2 - tmp3, 3) & RANGE_MASK];
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a reduced-size 1x1 output block.
+ *
+ * We hardly need an inverse DCT routine for this: just take the
+ * average pixel value, which is one-eighth of the DC coefficient.
+ */
+
+GLOBAL(void)
+jpeg_idct_1x1 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  int dcval;
+  ISLOW_MULT_TYPE * quantptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  SHIFT_TEMPS
+
+  /* 1x1 is trivial: just take the DC coefficient divided by 8. */
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  dcval = DEQUANTIZE(coef_block[0], quantptr[0]);
+  dcval = (int) DESCALE((INT32) dcval, 3);
+
+  output_buf[0][output_col] = range_limit[dcval & RANGE_MASK];
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 9x9 output block.
+ *
+ * Optimized algorithm with 10 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/18).
+ */
+
+GLOBAL(void)
+jpeg_idct_9x9 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp10, tmp11, tmp12, tmp13, tmp14;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*9];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp0 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp0 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp3 = MULTIPLY(z3, FIX(0.707106781));      /* c6 */
+    tmp1 = tmp0 + tmp3;
+    tmp2 = tmp0 - tmp3 - tmp3;
+
+    tmp0 = MULTIPLY(z1 - z2, FIX(0.707106781)); /* c6 */
+    tmp11 = tmp2 + tmp0;
+    tmp14 = tmp2 - tmp0 - tmp0;
+
+    tmp0 = MULTIPLY(z1 + z2, FIX(1.328926049)); /* c2 */
+    tmp2 = MULTIPLY(z1, FIX(1.083350441));      /* c4 */
+    tmp3 = MULTIPLY(z2, FIX(0.245575608));      /* c8 */
+
+    tmp10 = tmp1 + tmp0 - tmp3;
+    tmp12 = tmp1 - tmp0 + tmp2;
+    tmp13 = tmp1 - tmp2 + tmp3;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    z2 = MULTIPLY(z2, - FIX(1.224744871));           /* -c3 */
+
+    tmp2 = MULTIPLY(z1 + z3, FIX(0.909038955));      /* c5 */
+    tmp3 = MULTIPLY(z1 + z4, FIX(0.483689525));      /* c7 */
+    tmp0 = tmp2 + tmp3 - z2;
+    tmp1 = MULTIPLY(z3 - z4, FIX(1.392728481));      /* c1 */
+    tmp2 += z2 - tmp1;
+    tmp3 += z2 + tmp1;
+    tmp1 = MULTIPLY(z1 - z3 - z4, FIX(1.224744871)); /* c3 */
+
+    /* Final output stage */
+
+    wsptr[8*0] = (int) RIGHT_SHIFT(tmp10 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[8*8] = (int) RIGHT_SHIFT(tmp10 - tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[8*1] = (int) RIGHT_SHIFT(tmp11 + tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[8*7] = (int) RIGHT_SHIFT(tmp11 - tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[8*2] = (int) RIGHT_SHIFT(tmp12 + tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[8*6] = (int) RIGHT_SHIFT(tmp12 - tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[8*3] = (int) RIGHT_SHIFT(tmp13 + tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[8*5] = (int) RIGHT_SHIFT(tmp13 - tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[8*4] = (int) RIGHT_SHIFT(tmp14, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 9 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 9; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp0 <<= CONST_BITS;
+
+    z1 = (INT32) wsptr[2];
+    z2 = (INT32) wsptr[4];
+    z3 = (INT32) wsptr[6];
+
+    tmp3 = MULTIPLY(z3, FIX(0.707106781));      /* c6 */
+    tmp1 = tmp0 + tmp3;
+    tmp2 = tmp0 - tmp3 - tmp3;
+
+    tmp0 = MULTIPLY(z1 - z2, FIX(0.707106781)); /* c6 */
+    tmp11 = tmp2 + tmp0;
+    tmp14 = tmp2 - tmp0 - tmp0;
+
+    tmp0 = MULTIPLY(z1 + z2, FIX(1.328926049)); /* c2 */
+    tmp2 = MULTIPLY(z1, FIX(1.083350441));      /* c4 */
+    tmp3 = MULTIPLY(z2, FIX(0.245575608));      /* c8 */
+
+    tmp10 = tmp1 + tmp0 - tmp3;
+    tmp12 = tmp1 - tmp0 + tmp2;
+    tmp13 = tmp1 - tmp2 + tmp3;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+
+    z2 = MULTIPLY(z2, - FIX(1.224744871));           /* -c3 */
+
+    tmp2 = MULTIPLY(z1 + z3, FIX(0.909038955));      /* c5 */
+    tmp3 = MULTIPLY(z1 + z4, FIX(0.483689525));      /* c7 */
+    tmp0 = tmp2 + tmp3 - z2;
+    tmp1 = MULTIPLY(z3 - z4, FIX(1.392728481));      /* c1 */
+    tmp2 += z2 - tmp1;
+    tmp3 += z2 + tmp1;
+    tmp1 = MULTIPLY(z1 - z3 - z4, FIX(1.224744871)); /* c3 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp10 + tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[8] = range_limit[(int) RIGHT_SHIFT(tmp10 - tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp11 + tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) RIGHT_SHIFT(tmp11 - tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp12 + tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) RIGHT_SHIFT(tmp12 - tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp13 + tmp3,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) RIGHT_SHIFT(tmp13 - tmp3,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) RIGHT_SHIFT(tmp14,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 10x10 output block.
+ *
+ * Optimized algorithm with 12 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/20).
+ */
+
+GLOBAL(void)
+jpeg_idct_10x10 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24;
+  INT32 z1, z2, z3, z4, z5;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*10];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z3 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    z3 += ONE << (CONST_BITS-PASS1_BITS-1);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z1 = MULTIPLY(z4, FIX(1.144122806));         /* c4 */
+    z2 = MULTIPLY(z4, FIX(0.437016024));         /* c8 */
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z2;
+
+    tmp22 = RIGHT_SHIFT(z3 - ((z1 - z2) << 1),   /* c0 = (c4-c8)*2 */
+			CONST_BITS-PASS1_BITS);
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));    /* c6 */
+    tmp12 = z1 + MULTIPLY(z2, FIX(0.513743148)); /* c2-c6 */
+    tmp13 = z1 - MULTIPLY(z3, FIX(2.176250899)); /* c2+c6 */
+
+    tmp20 = tmp10 + tmp12;
+    tmp24 = tmp10 - tmp12;
+    tmp21 = tmp11 + tmp13;
+    tmp23 = tmp11 - tmp13;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    tmp11 = z2 + z4;
+    tmp13 = z2 - z4;
+
+    tmp12 = MULTIPLY(tmp13, FIX(0.309016994));        /* (c3-c7)/2 */
+    z5 = z3 << CONST_BITS;
+
+    z2 = MULTIPLY(tmp11, FIX(0.951056516));           /* (c3+c7)/2 */
+    z4 = z5 + tmp12;
+
+    tmp10 = MULTIPLY(z1, FIX(1.396802247)) + z2 + z4; /* c1 */
+    tmp14 = MULTIPLY(z1, FIX(0.221231742)) - z2 + z4; /* c9 */
+
+    z2 = MULTIPLY(tmp11, FIX(0.587785252));           /* (c1-c9)/2 */
+    z4 = z5 - tmp12 - (tmp13 << (CONST_BITS - 1));
+
+    tmp12 = (z1 - tmp13 - z3) << PASS1_BITS;
+
+    tmp11 = MULTIPLY(z1, FIX(1.260073511)) - z2 - z4; /* c3 */
+    tmp13 = MULTIPLY(z1, FIX(0.642039522)) - z2 + z4; /* c7 */
+
+    /* Final output stage */
+
+    wsptr[8*0] = (int) RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*9] = (int) RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*1] = (int) RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*8] = (int) RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*2] = (int) (tmp22 + tmp12);
+    wsptr[8*7] = (int) (tmp22 - tmp12);
+    wsptr[8*3] = (int) RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*6] = (int) RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*4] = (int) RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*5] = (int) RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 10 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 10; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z3 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z3 <<= CONST_BITS;
+    z4 = (INT32) wsptr[4];
+    z1 = MULTIPLY(z4, FIX(1.144122806));         /* c4 */
+    z2 = MULTIPLY(z4, FIX(0.437016024));         /* c8 */
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z2;
+
+    tmp22 = z3 - ((z1 - z2) << 1);               /* c0 = (c4-c8)*2 */
+
+    z2 = (INT32) wsptr[2];
+    z3 = (INT32) wsptr[6];
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));    /* c6 */
+    tmp12 = z1 + MULTIPLY(z2, FIX(0.513743148)); /* c2-c6 */
+    tmp13 = z1 - MULTIPLY(z3, FIX(2.176250899)); /* c2+c6 */
+
+    tmp20 = tmp10 + tmp12;
+    tmp24 = tmp10 - tmp12;
+    tmp21 = tmp11 + tmp13;
+    tmp23 = tmp11 - tmp13;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z3 <<= CONST_BITS;
+    z4 = (INT32) wsptr[7];
+
+    tmp11 = z2 + z4;
+    tmp13 = z2 - z4;
+
+    tmp12 = MULTIPLY(tmp13, FIX(0.309016994));        /* (c3-c7)/2 */
+
+    z2 = MULTIPLY(tmp11, FIX(0.951056516));           /* (c3+c7)/2 */
+    z4 = z3 + tmp12;
+
+    tmp10 = MULTIPLY(z1, FIX(1.396802247)) + z2 + z4; /* c1 */
+    tmp14 = MULTIPLY(z1, FIX(0.221231742)) - z2 + z4; /* c9 */
+
+    z2 = MULTIPLY(tmp11, FIX(0.587785252));           /* (c1-c9)/2 */
+    z4 = z3 - tmp12 - (tmp13 << (CONST_BITS - 1));
+
+    tmp12 = ((z1 - tmp13) << CONST_BITS) - z3;
+
+    tmp11 = MULTIPLY(z1, FIX(1.260073511)) - z2 - z4; /* c3 */
+    tmp13 = MULTIPLY(z1, FIX(0.642039522)) - z2 + z4; /* c7 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[9] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[8] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 11x11 output block.
+ *
+ * Optimized algorithm with 24 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/22).
+ */
+
+GLOBAL(void)
+jpeg_idct_11x11 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*11];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp10 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp10 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp10 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp20 = MULTIPLY(z2 - z3, FIX(2.546640132));     /* c2+c4 */
+    tmp23 = MULTIPLY(z2 - z1, FIX(0.430815045));     /* c2-c6 */
+    z4 = z1 + z3;
+    tmp24 = MULTIPLY(z4, - FIX(1.155664402));        /* -(c2-c10) */
+    z4 -= z2;
+    tmp25 = tmp10 + MULTIPLY(z4, FIX(1.356927976));  /* c2 */
+    tmp21 = tmp20 + tmp23 + tmp25 -
+	    MULTIPLY(z2, FIX(1.821790775));          /* c2+c4+c10-c6 */
+    tmp20 += tmp25 + MULTIPLY(z3, FIX(2.115825087)); /* c4+c6 */
+    tmp23 += tmp25 - MULTIPLY(z1, FIX(1.513598477)); /* c6+c8 */
+    tmp24 += tmp25;
+    tmp22 = tmp24 - MULTIPLY(z3, FIX(0.788749120));  /* c8+c10 */
+    tmp24 += MULTIPLY(z2, FIX(1.944413522)) -        /* c2+c8 */
+	     MULTIPLY(z1, FIX(1.390975730));         /* c4+c10 */
+    tmp25 = tmp10 - MULTIPLY(z4, FIX(1.414213562));  /* c0 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    tmp11 = z1 + z2;
+    tmp14 = MULTIPLY(tmp11 + z3 + z4, FIX(0.398430003)); /* c9 */
+    tmp11 = MULTIPLY(tmp11, FIX(0.887983902));           /* c3-c9 */
+    tmp12 = MULTIPLY(z1 + z3, FIX(0.670361295));         /* c5-c9 */
+    tmp13 = tmp14 + MULTIPLY(z1 + z4, FIX(0.366151574)); /* c7-c9 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+	    MULTIPLY(z1, FIX(0.923107866));              /* c7+c5+c3-c1-2*c9 */
+    z1    = tmp14 - MULTIPLY(z2 + z3, FIX(1.163011579)); /* c7+c9 */
+    tmp11 += z1 + MULTIPLY(z2, FIX(2.073276588));        /* c1+c7+3*c9-c3 */
+    tmp12 += z1 - MULTIPLY(z3, FIX(1.192193623));        /* c3+c5-c7-c9 */
+    z1    = MULTIPLY(z2 + z4, - FIX(1.798248910));       /* -(c1+c9) */
+    tmp11 += z1;
+    tmp13 += z1 + MULTIPLY(z4, FIX(2.102458632));        /* c1+c5+c9-c7 */
+    tmp14 += MULTIPLY(z2, - FIX(1.467221301)) +          /* -(c5+c9) */
+	     MULTIPLY(z3, FIX(1.001388905)) -            /* c1-c9 */
+	     MULTIPLY(z4, FIX(1.684843907));             /* c3+c9 */
+
+    /* Final output stage */
+
+    wsptr[8*0]  = (int) RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*10] = (int) RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*1]  = (int) RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*9]  = (int) RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*2]  = (int) RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*8]  = (int) RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*3]  = (int) RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*7]  = (int) RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*4]  = (int) RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*6]  = (int) RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*5]  = (int) RIGHT_SHIFT(tmp25, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 11 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 11; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp10 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp10 <<= CONST_BITS;
+
+    z1 = (INT32) wsptr[2];
+    z2 = (INT32) wsptr[4];
+    z3 = (INT32) wsptr[6];
+
+    tmp20 = MULTIPLY(z2 - z3, FIX(2.546640132));     /* c2+c4 */
+    tmp23 = MULTIPLY(z2 - z1, FIX(0.430815045));     /* c2-c6 */
+    z4 = z1 + z3;
+    tmp24 = MULTIPLY(z4, - FIX(1.155664402));        /* -(c2-c10) */
+    z4 -= z2;
+    tmp25 = tmp10 + MULTIPLY(z4, FIX(1.356927976));  /* c2 */
+    tmp21 = tmp20 + tmp23 + tmp25 -
+	    MULTIPLY(z2, FIX(1.821790775));          /* c2+c4+c10-c6 */
+    tmp20 += tmp25 + MULTIPLY(z3, FIX(2.115825087)); /* c4+c6 */
+    tmp23 += tmp25 - MULTIPLY(z1, FIX(1.513598477)); /* c6+c8 */
+    tmp24 += tmp25;
+    tmp22 = tmp24 - MULTIPLY(z3, FIX(0.788749120));  /* c8+c10 */
+    tmp24 += MULTIPLY(z2, FIX(1.944413522)) -        /* c2+c8 */
+	     MULTIPLY(z1, FIX(1.390975730));         /* c4+c10 */
+    tmp25 = tmp10 - MULTIPLY(z4, FIX(1.414213562));  /* c0 */
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+
+    tmp11 = z1 + z2;
+    tmp14 = MULTIPLY(tmp11 + z3 + z4, FIX(0.398430003)); /* c9 */
+    tmp11 = MULTIPLY(tmp11, FIX(0.887983902));           /* c3-c9 */
+    tmp12 = MULTIPLY(z1 + z3, FIX(0.670361295));         /* c5-c9 */
+    tmp13 = tmp14 + MULTIPLY(z1 + z4, FIX(0.366151574)); /* c7-c9 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+	    MULTIPLY(z1, FIX(0.923107866));              /* c7+c5+c3-c1-2*c9 */
+    z1    = tmp14 - MULTIPLY(z2 + z3, FIX(1.163011579)); /* c7+c9 */
+    tmp11 += z1 + MULTIPLY(z2, FIX(2.073276588));        /* c1+c7+3*c9-c3 */
+    tmp12 += z1 - MULTIPLY(z3, FIX(1.192193623));        /* c3+c5-c7-c9 */
+    z1    = MULTIPLY(z2 + z4, - FIX(1.798248910));       /* -(c1+c9) */
+    tmp11 += z1;
+    tmp13 += z1 + MULTIPLY(z4, FIX(2.102458632));        /* c1+c5+c9-c7 */
+    tmp14 += MULTIPLY(z2, - FIX(1.467221301)) +          /* -(c5+c9) */
+	     MULTIPLY(z3, FIX(1.001388905)) -            /* c1-c9 */
+	     MULTIPLY(z4, FIX(1.684843907));             /* c3+c9 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 12x12 output block.
+ *
+ * Optimized algorithm with 15 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/24).
+ */
+
+GLOBAL(void)
+jpeg_idct_12x12 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*12];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z3 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    z3 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    z4 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z4 = MULTIPLY(z4, FIX(1.224744871)); /* c4 */
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z4 = MULTIPLY(z1, FIX(1.366025404)); /* c2 */
+    z1 <<= CONST_BITS;
+    z2 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+    z2 <<= CONST_BITS;
+
+    tmp12 = z1 - z2;
+
+    tmp21 = z3 + tmp12;
+    tmp24 = z3 - tmp12;
+
+    tmp12 = z4 + z2;
+
+    tmp20 = tmp10 + tmp12;
+    tmp25 = tmp10 - tmp12;
+
+    tmp12 = z4 - z1 - z2;
+
+    tmp22 = tmp11 + tmp12;
+    tmp23 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    tmp11 = MULTIPLY(z2, FIX(1.306562965));                  /* c3 */
+    tmp14 = MULTIPLY(z2, - FIX_0_541196100);                 /* -c9 */
+
+    tmp10 = z1 + z3;
+    tmp15 = MULTIPLY(tmp10 + z4, FIX(0.860918669));          /* c7 */
+    tmp12 = tmp15 + MULTIPLY(tmp10, FIX(0.261052384));       /* c5-c7 */
+    tmp10 = tmp12 + tmp11 + MULTIPLY(z1, FIX(0.280143716));  /* c1-c5 */
+    tmp13 = MULTIPLY(z3 + z4, - FIX(1.045510580));           /* -(c7+c11) */
+    tmp12 += tmp13 + tmp14 - MULTIPLY(z3, FIX(1.478575242)); /* c1+c5-c7-c11 */
+    tmp13 += tmp15 - tmp11 + MULTIPLY(z4, FIX(1.586706681)); /* c1+c11 */
+    tmp15 += tmp14 - MULTIPLY(z1, FIX(0.676326758)) -        /* c7-c11 */
+	     MULTIPLY(z4, FIX(1.982889723));                 /* c5+c7 */
+
+    z1 -= z4;
+    z2 -= z3;
+    z3 = MULTIPLY(z1 + z2, FIX_0_541196100);                 /* c9 */
+    tmp11 = z3 + MULTIPLY(z1, FIX_0_765366865);              /* c3-c9 */
+    tmp14 = z3 - MULTIPLY(z2, FIX_1_847759065);              /* c3+c9 */
+
+    /* Final output stage */
+
+    wsptr[8*0]  = (int) RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*11] = (int) RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*1]  = (int) RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*10] = (int) RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*2]  = (int) RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*9]  = (int) RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*3]  = (int) RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*8]  = (int) RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*4]  = (int) RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*7]  = (int) RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*5]  = (int) RIGHT_SHIFT(tmp25 + tmp15, CONST_BITS-PASS1_BITS);
+    wsptr[8*6]  = (int) RIGHT_SHIFT(tmp25 - tmp15, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 12 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 12; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z3 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z3 <<= CONST_BITS;
+
+    z4 = (INT32) wsptr[4];
+    z4 = MULTIPLY(z4, FIX(1.224744871)); /* c4 */
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    z1 = (INT32) wsptr[2];
+    z4 = MULTIPLY(z1, FIX(1.366025404)); /* c2 */
+    z1 <<= CONST_BITS;
+    z2 = (INT32) wsptr[6];
+    z2 <<= CONST_BITS;
+
+    tmp12 = z1 - z2;
+
+    tmp21 = z3 + tmp12;
+    tmp24 = z3 - tmp12;
+
+    tmp12 = z4 + z2;
+
+    tmp20 = tmp10 + tmp12;
+    tmp25 = tmp10 - tmp12;
+
+    tmp12 = z4 - z1 - z2;
+
+    tmp22 = tmp11 + tmp12;
+    tmp23 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+
+    tmp11 = MULTIPLY(z2, FIX(1.306562965));                  /* c3 */
+    tmp14 = MULTIPLY(z2, - FIX_0_541196100);                 /* -c9 */
+
+    tmp10 = z1 + z3;
+    tmp15 = MULTIPLY(tmp10 + z4, FIX(0.860918669));          /* c7 */
+    tmp12 = tmp15 + MULTIPLY(tmp10, FIX(0.261052384));       /* c5-c7 */
+    tmp10 = tmp12 + tmp11 + MULTIPLY(z1, FIX(0.280143716));  /* c1-c5 */
+    tmp13 = MULTIPLY(z3 + z4, - FIX(1.045510580));           /* -(c7+c11) */
+    tmp12 += tmp13 + tmp14 - MULTIPLY(z3, FIX(1.478575242)); /* c1+c5-c7-c11 */
+    tmp13 += tmp15 - tmp11 + MULTIPLY(z4, FIX(1.586706681)); /* c1+c11 */
+    tmp15 += tmp14 - MULTIPLY(z1, FIX(0.676326758)) -        /* c7-c11 */
+	     MULTIPLY(z4, FIX(1.982889723));                 /* c5+c7 */
+
+    z1 -= z4;
+    z2 -= z3;
+    z3 = MULTIPLY(z1 + z2, FIX_0_541196100);                 /* c9 */
+    tmp11 = z3 + MULTIPLY(z1, FIX_0_765366865);              /* c3-c9 */
+    tmp14 = z3 - MULTIPLY(z2, FIX_1_847759065);              /* c3+c9 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[11] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25 + tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp25 - tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 13x13 output block.
+ *
+ * Optimized algorithm with 29 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/26).
+ */
+
+GLOBAL(void)
+jpeg_idct_13x13 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*13];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z1 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    tmp12 = MULTIPLY(tmp10, FIX(1.155388986));                /* (c4+c6)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.096834934)) + z1;           /* (c4-c6)/2 */
+
+    tmp20 = MULTIPLY(z2, FIX(1.373119086)) + tmp12 + tmp13;   /* c2 */
+    tmp22 = MULTIPLY(z2, FIX(0.501487041)) - tmp12 + tmp13;   /* c10 */
+
+    tmp12 = MULTIPLY(tmp10, FIX(0.316450131));                /* (c8-c12)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.486914739)) + z1;           /* (c8+c12)/2 */
+
+    tmp21 = MULTIPLY(z2, FIX(1.058554052)) - tmp12 + tmp13;   /* c6 */
+    tmp25 = MULTIPLY(z2, - FIX(1.252223920)) + tmp12 + tmp13; /* c4 */
+
+    tmp12 = MULTIPLY(tmp10, FIX(0.435816023));                /* (c2-c10)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.937303064)) - z1;           /* (c2+c10)/2 */
+
+    tmp23 = MULTIPLY(z2, - FIX(0.170464608)) - tmp12 - tmp13; /* c12 */
+    tmp24 = MULTIPLY(z2, - FIX(0.803364869)) + tmp12 - tmp13; /* c8 */
+
+    tmp26 = MULTIPLY(tmp11 - z2, FIX(1.414213562)) + z1;      /* c0 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.322312651));     /* c3 */
+    tmp12 = MULTIPLY(z1 + z3, FIX(1.163874945));     /* c5 */
+    tmp15 = z1 + z4;
+    tmp13 = MULTIPLY(tmp15, FIX(0.937797057));       /* c7 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+	    MULTIPLY(z1, FIX(2.020082300));          /* c7+c5+c3-c1 */
+    tmp14 = MULTIPLY(z2 + z3, - FIX(0.338443458));   /* -c11 */
+    tmp11 += tmp14 + MULTIPLY(z2, FIX(0.837223564)); /* c5+c9+c11-c3 */
+    tmp12 += tmp14 - MULTIPLY(z3, FIX(1.572116027)); /* c1+c5-c9-c11 */
+    tmp14 = MULTIPLY(z2 + z4, - FIX(1.163874945));   /* -c5 */
+    tmp11 += tmp14;
+    tmp13 += tmp14 + MULTIPLY(z4, FIX(2.205608352)); /* c3+c5+c9-c7 */
+    tmp14 = MULTIPLY(z3 + z4, - FIX(0.657217813));   /* -c9 */
+    tmp12 += tmp14;
+    tmp13 += tmp14;
+    tmp15 = MULTIPLY(tmp15, FIX(0.338443458));       /* c11 */
+    tmp14 = tmp15 + MULTIPLY(z1, FIX(0.318774355)) - /* c9-c11 */
+	    MULTIPLY(z2, FIX(0.466105296));          /* c1-c7 */
+    z1    = MULTIPLY(z3 - z2, FIX(0.937797057));     /* c7 */
+    tmp14 += z1;
+    tmp15 += z1 + MULTIPLY(z3, FIX(0.384515595)) -   /* c3-c7 */
+	     MULTIPLY(z4, FIX(1.742345811));         /* c1+c11 */
+
+    /* Final output stage */
+
+    wsptr[8*0]  = (int) RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*12] = (int) RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*1]  = (int) RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*11] = (int) RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*2]  = (int) RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*10] = (int) RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*3]  = (int) RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*9]  = (int) RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*4]  = (int) RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*8]  = (int) RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*5]  = (int) RIGHT_SHIFT(tmp25 + tmp15, CONST_BITS-PASS1_BITS);
+    wsptr[8*7]  = (int) RIGHT_SHIFT(tmp25 - tmp15, CONST_BITS-PASS1_BITS);
+    wsptr[8*6]  = (int) RIGHT_SHIFT(tmp26, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 13 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 13; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z1 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z1 <<= CONST_BITS;
+
+    z2 = (INT32) wsptr[2];
+    z3 = (INT32) wsptr[4];
+    z4 = (INT32) wsptr[6];
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    tmp12 = MULTIPLY(tmp10, FIX(1.155388986));                /* (c4+c6)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.096834934)) + z1;           /* (c4-c6)/2 */
+
+    tmp20 = MULTIPLY(z2, FIX(1.373119086)) + tmp12 + tmp13;   /* c2 */
+    tmp22 = MULTIPLY(z2, FIX(0.501487041)) - tmp12 + tmp13;   /* c10 */
+
+    tmp12 = MULTIPLY(tmp10, FIX(0.316450131));                /* (c8-c12)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.486914739)) + z1;           /* (c8+c12)/2 */
+
+    tmp21 = MULTIPLY(z2, FIX(1.058554052)) - tmp12 + tmp13;   /* c6 */
+    tmp25 = MULTIPLY(z2, - FIX(1.252223920)) + tmp12 + tmp13; /* c4 */
+
+    tmp12 = MULTIPLY(tmp10, FIX(0.435816023));                /* (c2-c10)/2 */
+    tmp13 = MULTIPLY(tmp11, FIX(0.937303064)) - z1;           /* (c2+c10)/2 */
+
+    tmp23 = MULTIPLY(z2, - FIX(0.170464608)) - tmp12 - tmp13; /* c12 */
+    tmp24 = MULTIPLY(z2, - FIX(0.803364869)) + tmp12 - tmp13; /* c8 */
+
+    tmp26 = MULTIPLY(tmp11 - z2, FIX(1.414213562)) + z1;      /* c0 */
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.322312651));     /* c3 */
+    tmp12 = MULTIPLY(z1 + z3, FIX(1.163874945));     /* c5 */
+    tmp15 = z1 + z4;
+    tmp13 = MULTIPLY(tmp15, FIX(0.937797057));       /* c7 */
+    tmp10 = tmp11 + tmp12 + tmp13 -
+	    MULTIPLY(z1, FIX(2.020082300));          /* c7+c5+c3-c1 */
+    tmp14 = MULTIPLY(z2 + z3, - FIX(0.338443458));   /* -c11 */
+    tmp11 += tmp14 + MULTIPLY(z2, FIX(0.837223564)); /* c5+c9+c11-c3 */
+    tmp12 += tmp14 - MULTIPLY(z3, FIX(1.572116027)); /* c1+c5-c9-c11 */
+    tmp14 = MULTIPLY(z2 + z4, - FIX(1.163874945));   /* -c5 */
+    tmp11 += tmp14;
+    tmp13 += tmp14 + MULTIPLY(z4, FIX(2.205608352)); /* c3+c5+c9-c7 */
+    tmp14 = MULTIPLY(z3 + z4, - FIX(0.657217813));   /* -c9 */
+    tmp12 += tmp14;
+    tmp13 += tmp14;
+    tmp15 = MULTIPLY(tmp15, FIX(0.338443458));       /* c11 */
+    tmp14 = tmp15 + MULTIPLY(z1, FIX(0.318774355)) - /* c9-c11 */
+	    MULTIPLY(z2, FIX(0.466105296));          /* c1-c7 */
+    z1    = MULTIPLY(z3 - z2, FIX(0.937797057));     /* c7 */
+    tmp14 += z1;
+    tmp15 += z1 + MULTIPLY(z3, FIX(0.384515595)) -   /* c3-c7 */
+	     MULTIPLY(z4, FIX(1.742345811));         /* c1+c11 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[12] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[11] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25 + tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp25 - tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp26,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 14x14 output block.
+ *
+ * Optimized algorithm with 20 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/28).
+ */
+
+GLOBAL(void)
+jpeg_idct_14x14 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*14];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z1 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-1);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z2 = MULTIPLY(z4, FIX(1.274162392));         /* c4 */
+    z3 = MULTIPLY(z4, FIX(0.314692123));         /* c12 */
+    z4 = MULTIPLY(z4, FIX(0.881747734));         /* c8 */
+
+    tmp10 = z1 + z2;
+    tmp11 = z1 + z3;
+    tmp12 = z1 - z4;
+
+    tmp23 = RIGHT_SHIFT(z1 - ((z2 + z3 - z4) << 1), /* c0 = (c4+c12-c8)*2 */
+			CONST_BITS-PASS1_BITS);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    z3 = MULTIPLY(z1 + z2, FIX(1.105676686));    /* c6 */
+
+    tmp13 = z3 + MULTIPLY(z1, FIX(0.273079590)); /* c2-c6 */
+    tmp14 = z3 - MULTIPLY(z2, FIX(1.719280954)); /* c6+c10 */
+    tmp15 = MULTIPLY(z1, FIX(0.613604268)) -     /* c10 */
+	    MULTIPLY(z2, FIX(1.378756276));      /* c2 */
+
+    tmp20 = tmp10 + tmp13;
+    tmp26 = tmp10 - tmp13;
+    tmp21 = tmp11 + tmp14;
+    tmp25 = tmp11 - tmp14;
+    tmp22 = tmp12 + tmp15;
+    tmp24 = tmp12 - tmp15;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    tmp13 = z4 << CONST_BITS;
+
+    tmp14 = z1 + z3;
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.334852607));           /* c3 */
+    tmp12 = MULTIPLY(tmp14, FIX(1.197448846));             /* c5 */
+    tmp10 = tmp11 + tmp12 + tmp13 - MULTIPLY(z1, FIX(1.126980169)); /* c3+c5-c1 */
+    tmp14 = MULTIPLY(tmp14, FIX(0.752406978));             /* c9 */
+    tmp16 = tmp14 - MULTIPLY(z1, FIX(1.061150426));        /* c9+c11-c13 */
+    z1    -= z2;
+    tmp15 = MULTIPLY(z1, FIX(0.467085129)) - tmp13;        /* c11 */
+    tmp16 += tmp15;
+    z1    += z4;
+    z4    = MULTIPLY(z2 + z3, - FIX(0.158341681)) - tmp13; /* -c13 */
+    tmp11 += z4 - MULTIPLY(z2, FIX(0.424103948));          /* c3-c9-c13 */
+    tmp12 += z4 - MULTIPLY(z3, FIX(2.373959773));          /* c3+c5-c13 */
+    z4    = MULTIPLY(z3 - z2, FIX(1.405321284));           /* c1 */
+    tmp14 += z4 + tmp13 - MULTIPLY(z3, FIX(1.6906431334)); /* c1+c9-c11 */
+    tmp15 += z4 + MULTIPLY(z2, FIX(0.674957567));          /* c1+c11-c5 */
+
+    tmp13 = (z1 - z3) << PASS1_BITS;
+
+    /* Final output stage */
+
+    wsptr[8*0]  = (int) RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*13] = (int) RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*1]  = (int) RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*12] = (int) RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*2]  = (int) RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*11] = (int) RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*3]  = (int) (tmp23 + tmp13);
+    wsptr[8*10] = (int) (tmp23 - tmp13);
+    wsptr[8*4]  = (int) RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*9]  = (int) RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*5]  = (int) RIGHT_SHIFT(tmp25 + tmp15, CONST_BITS-PASS1_BITS);
+    wsptr[8*8]  = (int) RIGHT_SHIFT(tmp25 - tmp15, CONST_BITS-PASS1_BITS);
+    wsptr[8*6]  = (int) RIGHT_SHIFT(tmp26 + tmp16, CONST_BITS-PASS1_BITS);
+    wsptr[8*7]  = (int) RIGHT_SHIFT(tmp26 - tmp16, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 14 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 14; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z1 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z1 <<= CONST_BITS;
+    z4 = (INT32) wsptr[4];
+    z2 = MULTIPLY(z4, FIX(1.274162392));         /* c4 */
+    z3 = MULTIPLY(z4, FIX(0.314692123));         /* c12 */
+    z4 = MULTIPLY(z4, FIX(0.881747734));         /* c8 */
+
+    tmp10 = z1 + z2;
+    tmp11 = z1 + z3;
+    tmp12 = z1 - z4;
+
+    tmp23 = z1 - ((z2 + z3 - z4) << 1);          /* c0 = (c4+c12-c8)*2 */
+
+    z1 = (INT32) wsptr[2];
+    z2 = (INT32) wsptr[6];
+
+    z3 = MULTIPLY(z1 + z2, FIX(1.105676686));    /* c6 */
+
+    tmp13 = z3 + MULTIPLY(z1, FIX(0.273079590)); /* c2-c6 */
+    tmp14 = z3 - MULTIPLY(z2, FIX(1.719280954)); /* c6+c10 */
+    tmp15 = MULTIPLY(z1, FIX(0.613604268)) -     /* c10 */
+	    MULTIPLY(z2, FIX(1.378756276));      /* c2 */
+
+    tmp20 = tmp10 + tmp13;
+    tmp26 = tmp10 - tmp13;
+    tmp21 = tmp11 + tmp14;
+    tmp25 = tmp11 - tmp14;
+    tmp22 = tmp12 + tmp15;
+    tmp24 = tmp12 - tmp15;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+    z4 <<= CONST_BITS;
+
+    tmp14 = z1 + z3;
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.334852607));           /* c3 */
+    tmp12 = MULTIPLY(tmp14, FIX(1.197448846));             /* c5 */
+    tmp10 = tmp11 + tmp12 + z4 - MULTIPLY(z1, FIX(1.126980169)); /* c3+c5-c1 */
+    tmp14 = MULTIPLY(tmp14, FIX(0.752406978));             /* c9 */
+    tmp16 = tmp14 - MULTIPLY(z1, FIX(1.061150426));        /* c9+c11-c13 */
+    z1    -= z2;
+    tmp15 = MULTIPLY(z1, FIX(0.467085129)) - z4;           /* c11 */
+    tmp16 += tmp15;
+    tmp13 = MULTIPLY(z2 + z3, - FIX(0.158341681)) - z4;    /* -c13 */
+    tmp11 += tmp13 - MULTIPLY(z2, FIX(0.424103948));       /* c3-c9-c13 */
+    tmp12 += tmp13 - MULTIPLY(z3, FIX(2.373959773));       /* c3+c5-c13 */
+    tmp13 = MULTIPLY(z3 - z2, FIX(1.405321284));           /* c1 */
+    tmp14 += tmp13 + z4 - MULTIPLY(z3, FIX(1.6906431334)); /* c1+c9-c11 */
+    tmp15 += tmp13 + MULTIPLY(z2, FIX(0.674957567));       /* c1+c11-c5 */
+
+    tmp13 = ((z1 - z3) << CONST_BITS) + z4;
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[13] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[12] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[11] = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25 + tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp25 - tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp26 + tmp16,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp26 - tmp16,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 15x15 output block.
+ *
+ * Optimized algorithm with 22 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/30).
+ */
+
+GLOBAL(void)
+jpeg_idct_15x15 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26, tmp27;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*15];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z1 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp10 = MULTIPLY(z4, FIX(0.437016024)); /* c12 */
+    tmp11 = MULTIPLY(z4, FIX(1.144122806)); /* c6 */
+
+    tmp12 = z1 - tmp10;
+    tmp13 = z1 + tmp11;
+    z1 -= (tmp11 - tmp10) << 1;             /* c0 = (c6-c12)*2 */
+
+    z4 = z2 - z3;
+    z3 += z2;
+    tmp10 = MULTIPLY(z3, FIX(1.337628990)); /* (c2+c4)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.045680613)); /* (c2-c4)/2 */
+    z2 = MULTIPLY(z2, FIX(1.439773946));    /* c4+c14 */
+
+    tmp20 = tmp13 + tmp10 + tmp11;
+    tmp23 = tmp12 - tmp10 + tmp11 + z2;
+
+    tmp10 = MULTIPLY(z3, FIX(0.547059574)); /* (c8+c14)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.399234004)); /* (c8-c14)/2 */
+
+    tmp25 = tmp13 - tmp10 - tmp11;
+    tmp26 = tmp12 + tmp10 - tmp11 - z2;
+
+    tmp10 = MULTIPLY(z3, FIX(0.790569415)); /* (c6+c12)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.353553391)); /* (c6-c12)/2 */
+
+    tmp21 = tmp12 + tmp10 + tmp11;
+    tmp24 = tmp13 - tmp10 + tmp11;
+    tmp11 += tmp11;
+    tmp22 = z1 + tmp11;                     /* c10 = c6-c12 */
+    tmp27 = z1 - tmp11 - tmp11;             /* c0 = (c6-c12)*2 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z3 = MULTIPLY(z4, FIX(1.224744871));                    /* c5 */
+    z4 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    tmp13 = z2 - z4;
+    tmp15 = MULTIPLY(z1 + tmp13, FIX(0.831253876));         /* c9 */
+    tmp11 = tmp15 + MULTIPLY(z1, FIX(0.513743148));         /* c3-c9 */
+    tmp14 = tmp15 - MULTIPLY(tmp13, FIX(2.176250899));      /* c3+c9 */
+
+    tmp13 = MULTIPLY(z2, - FIX(0.831253876));               /* -c9 */
+    tmp15 = MULTIPLY(z2, - FIX(1.344997024));               /* -c3 */
+    z2 = z1 - z4;
+    tmp12 = z3 + MULTIPLY(z2, FIX(1.406466353));            /* c1 */
+
+    tmp10 = tmp12 + MULTIPLY(z4, FIX(2.457431844)) - tmp15; /* c1+c7 */
+    tmp16 = tmp12 - MULTIPLY(z1, FIX(1.112434820)) + tmp13; /* c1-c13 */
+    tmp12 = MULTIPLY(z2, FIX(1.224744871)) - z3;            /* c5 */
+    z2 = MULTIPLY(z1 + z4, FIX(0.575212477));               /* c11 */
+    tmp13 += z2 + MULTIPLY(z1, FIX(0.475753014)) - z3;      /* c7-c11 */
+    tmp15 += z2 - MULTIPLY(z4, FIX(0.869244010)) + z3;      /* c11+c13 */
+
+    /* Final output stage */
+
+    wsptr[8*0]  = (int) RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*14] = (int) RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*1]  = (int) RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*13] = (int) RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*2]  = (int) RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*12] = (int) RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*3]  = (int) RIGHT_SHIFT(tmp23 + tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*11] = (int) RIGHT_SHIFT(tmp23 - tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*4]  = (int) RIGHT_SHIFT(tmp24 + tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*10] = (int) RIGHT_SHIFT(tmp24 - tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*5]  = (int) RIGHT_SHIFT(tmp25 + tmp15, CONST_BITS-PASS1_BITS);
+    wsptr[8*9]  = (int) RIGHT_SHIFT(tmp25 - tmp15, CONST_BITS-PASS1_BITS);
+    wsptr[8*6]  = (int) RIGHT_SHIFT(tmp26 + tmp16, CONST_BITS-PASS1_BITS);
+    wsptr[8*8]  = (int) RIGHT_SHIFT(tmp26 - tmp16, CONST_BITS-PASS1_BITS);
+    wsptr[8*7]  = (int) RIGHT_SHIFT(tmp27, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 15 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 15; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z1 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z1 <<= CONST_BITS;
+
+    z2 = (INT32) wsptr[2];
+    z3 = (INT32) wsptr[4];
+    z4 = (INT32) wsptr[6];
+
+    tmp10 = MULTIPLY(z4, FIX(0.437016024)); /* c12 */
+    tmp11 = MULTIPLY(z4, FIX(1.144122806)); /* c6 */
+
+    tmp12 = z1 - tmp10;
+    tmp13 = z1 + tmp11;
+    z1 -= (tmp11 - tmp10) << 1;             /* c0 = (c6-c12)*2 */
+
+    z4 = z2 - z3;
+    z3 += z2;
+    tmp10 = MULTIPLY(z3, FIX(1.337628990)); /* (c2+c4)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.045680613)); /* (c2-c4)/2 */
+    z2 = MULTIPLY(z2, FIX(1.439773946));    /* c4+c14 */
+
+    tmp20 = tmp13 + tmp10 + tmp11;
+    tmp23 = tmp12 - tmp10 + tmp11 + z2;
+
+    tmp10 = MULTIPLY(z3, FIX(0.547059574)); /* (c8+c14)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.399234004)); /* (c8-c14)/2 */
+
+    tmp25 = tmp13 - tmp10 - tmp11;
+    tmp26 = tmp12 + tmp10 - tmp11 - z2;
+
+    tmp10 = MULTIPLY(z3, FIX(0.790569415)); /* (c6+c12)/2 */
+    tmp11 = MULTIPLY(z4, FIX(0.353553391)); /* (c6-c12)/2 */
+
+    tmp21 = tmp12 + tmp10 + tmp11;
+    tmp24 = tmp13 - tmp10 + tmp11;
+    tmp11 += tmp11;
+    tmp22 = z1 + tmp11;                     /* c10 = c6-c12 */
+    tmp27 = z1 - tmp11 - tmp11;             /* c0 = (c6-c12)*2 */
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z4 = (INT32) wsptr[5];
+    z3 = MULTIPLY(z4, FIX(1.224744871));                    /* c5 */
+    z4 = (INT32) wsptr[7];
+
+    tmp13 = z2 - z4;
+    tmp15 = MULTIPLY(z1 + tmp13, FIX(0.831253876));         /* c9 */
+    tmp11 = tmp15 + MULTIPLY(z1, FIX(0.513743148));         /* c3-c9 */
+    tmp14 = tmp15 - MULTIPLY(tmp13, FIX(2.176250899));      /* c3+c9 */
+
+    tmp13 = MULTIPLY(z2, - FIX(0.831253876));               /* -c9 */
+    tmp15 = MULTIPLY(z2, - FIX(1.344997024));               /* -c3 */
+    z2 = z1 - z4;
+    tmp12 = z3 + MULTIPLY(z2, FIX(1.406466353));            /* c1 */
+
+    tmp10 = tmp12 + MULTIPLY(z4, FIX(2.457431844)) - tmp15; /* c1+c7 */
+    tmp16 = tmp12 - MULTIPLY(z1, FIX(1.112434820)) + tmp13; /* c1-c13 */
+    tmp12 = MULTIPLY(z2, FIX(1.224744871)) - z3;            /* c5 */
+    z2 = MULTIPLY(z1 + z4, FIX(0.575212477));               /* c11 */
+    tmp13 += z2 + MULTIPLY(z1, FIX(0.475753014)) - z3;      /* c7-c11 */
+    tmp15 += z2 - MULTIPLY(z4, FIX(0.869244010)) + z3;      /* c11+c13 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[14] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[13] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[12] = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[11] = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25 + tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp25 - tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp26 + tmp16,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp26 - tmp16,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp27,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 16x16 output block.
+ *
+ * Optimized algorithm with 28 multiplications in the 1-D kernel.
+ * cK represents sqrt(2) * cos(K*pi/32).
+ */
+
+GLOBAL(void)
+jpeg_idct_16x16 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		 JCOEFPTR coef_block,
+		 JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp10, tmp11, tmp12, tmp13;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26, tmp27;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*16];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp0 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp0 += 1 << (CONST_BITS-PASS1_BITS-1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp1 = MULTIPLY(z1, FIX(1.306562965));      /* c4[16] = c2[8] */
+    tmp2 = MULTIPLY(z1, FIX_0_541196100);       /* c12[16] = c6[8] */
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+    z3 = z1 - z2;
+    z4 = MULTIPLY(z3, FIX(0.275899379));        /* c14[16] = c7[8] */
+    z3 = MULTIPLY(z3, FIX(1.387039845));        /* c2[16] = c1[8] */
+
+    tmp0 = z3 + MULTIPLY(z2, FIX_2_562915447);  /* (c6+c2)[16] = (c3+c1)[8] */
+    tmp1 = z4 + MULTIPLY(z1, FIX_0_899976223);  /* (c6-c14)[16] = (c3-c7)[8] */
+    tmp2 = z3 - MULTIPLY(z1, FIX(0.601344887)); /* (c2-c10)[16] = (c1-c5)[8] */
+    tmp3 = z4 - MULTIPLY(z2, FIX(0.509795579)); /* (c10-c14)[16] = (c5-c7)[8] */
+
+    tmp20 = tmp10 + tmp0;
+    tmp27 = tmp10 - tmp0;
+    tmp21 = tmp12 + tmp1;
+    tmp26 = tmp12 - tmp1;
+    tmp22 = tmp13 + tmp2;
+    tmp25 = tmp13 - tmp2;
+    tmp23 = tmp11 + tmp3;
+    tmp24 = tmp11 - tmp3;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    z4 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+
+    tmp11 = z1 + z3;
+
+    tmp1  = MULTIPLY(z1 + z2, FIX(1.353318001));   /* c3 */
+    tmp2  = MULTIPLY(tmp11,   FIX(1.247225013));   /* c5 */
+    tmp3  = MULTIPLY(z1 + z4, FIX(1.093201867));   /* c7 */
+    tmp10 = MULTIPLY(z1 - z4, FIX(0.897167586));   /* c9 */
+    tmp11 = MULTIPLY(tmp11,   FIX(0.666655658));   /* c11 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.410524528));   /* c13 */
+    tmp0  = tmp1 + tmp2 + tmp3 -
+	    MULTIPLY(z1, FIX(2.286341144));        /* c7+c5+c3-c1 */
+    tmp13 = tmp10 + tmp11 + tmp12 -
+	    MULTIPLY(z1, FIX(1.835730603));        /* c9+c11+c13-c15 */
+    z1    = MULTIPLY(z2 + z3, FIX(0.138617169));   /* c15 */
+    tmp1  += z1 + MULTIPLY(z2, FIX(0.071888074));  /* c9+c11-c3-c15 */
+    tmp2  += z1 - MULTIPLY(z3, FIX(1.125726048));  /* c5+c7+c15-c3 */
+    z1    = MULTIPLY(z3 - z2, FIX(1.407403738));   /* c1 */
+    tmp11 += z1 - MULTIPLY(z3, FIX(0.766367282));  /* c1+c11-c9-c13 */
+    tmp12 += z1 + MULTIPLY(z2, FIX(1.971951411));  /* c1+c5+c13-c7 */
+    z2    += z4;
+    z1    = MULTIPLY(z2, - FIX(0.666655658));      /* -c11 */
+    tmp1  += z1;
+    tmp3  += z1 + MULTIPLY(z4, FIX(1.065388962));  /* c3+c11+c15-c7 */
+    z2    = MULTIPLY(z2, - FIX(1.247225013));      /* -c5 */
+    tmp10 += z2 + MULTIPLY(z4, FIX(3.141271809));  /* c1+c5+c9-c13 */
+    tmp12 += z2;
+    z2    = MULTIPLY(z3 + z4, - FIX(1.353318001)); /* -c3 */
+    tmp2  += z2;
+    tmp3  += z2;
+    z2    = MULTIPLY(z4 - z3, FIX(0.410524528));   /* c13 */
+    tmp10 += z2;
+    tmp11 += z2;
+
+    /* Final output stage */
+
+    wsptr[8*0]  = (int) RIGHT_SHIFT(tmp20 + tmp0,  CONST_BITS-PASS1_BITS);
+    wsptr[8*15] = (int) RIGHT_SHIFT(tmp20 - tmp0,  CONST_BITS-PASS1_BITS);
+    wsptr[8*1]  = (int) RIGHT_SHIFT(tmp21 + tmp1,  CONST_BITS-PASS1_BITS);
+    wsptr[8*14] = (int) RIGHT_SHIFT(tmp21 - tmp1,  CONST_BITS-PASS1_BITS);
+    wsptr[8*2]  = (int) RIGHT_SHIFT(tmp22 + tmp2,  CONST_BITS-PASS1_BITS);
+    wsptr[8*13] = (int) RIGHT_SHIFT(tmp22 - tmp2,  CONST_BITS-PASS1_BITS);
+    wsptr[8*3]  = (int) RIGHT_SHIFT(tmp23 + tmp3,  CONST_BITS-PASS1_BITS);
+    wsptr[8*12] = (int) RIGHT_SHIFT(tmp23 - tmp3,  CONST_BITS-PASS1_BITS);
+    wsptr[8*4]  = (int) RIGHT_SHIFT(tmp24 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*11] = (int) RIGHT_SHIFT(tmp24 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*5]  = (int) RIGHT_SHIFT(tmp25 + tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*10] = (int) RIGHT_SHIFT(tmp25 - tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*6]  = (int) RIGHT_SHIFT(tmp26 + tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*9]  = (int) RIGHT_SHIFT(tmp26 - tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*7]  = (int) RIGHT_SHIFT(tmp27 + tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*8]  = (int) RIGHT_SHIFT(tmp27 - tmp13, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 16 rows from work array, store into output array. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 16; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp0 <<= CONST_BITS;
+
+    z1 = (INT32) wsptr[4];
+    tmp1 = MULTIPLY(z1, FIX(1.306562965));      /* c4[16] = c2[8] */
+    tmp2 = MULTIPLY(z1, FIX_0_541196100);       /* c12[16] = c6[8] */
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+
+    z1 = (INT32) wsptr[2];
+    z2 = (INT32) wsptr[6];
+    z3 = z1 - z2;
+    z4 = MULTIPLY(z3, FIX(0.275899379));        /* c14[16] = c7[8] */
+    z3 = MULTIPLY(z3, FIX(1.387039845));        /* c2[16] = c1[8] */
+
+    tmp0 = z3 + MULTIPLY(z2, FIX_2_562915447);  /* (c6+c2)[16] = (c3+c1)[8] */
+    tmp1 = z4 + MULTIPLY(z1, FIX_0_899976223);  /* (c6-c14)[16] = (c3-c7)[8] */
+    tmp2 = z3 - MULTIPLY(z1, FIX(0.601344887)); /* (c2-c10)[16] = (c1-c5)[8] */
+    tmp3 = z4 - MULTIPLY(z2, FIX(0.509795579)); /* (c10-c14)[16] = (c5-c7)[8] */
+
+    tmp20 = tmp10 + tmp0;
+    tmp27 = tmp10 - tmp0;
+    tmp21 = tmp12 + tmp1;
+    tmp26 = tmp12 - tmp1;
+    tmp22 = tmp13 + tmp2;
+    tmp25 = tmp13 - tmp2;
+    tmp23 = tmp11 + tmp3;
+    tmp24 = tmp11 - tmp3;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+
+    tmp11 = z1 + z3;
+
+    tmp1  = MULTIPLY(z1 + z2, FIX(1.353318001));   /* c3 */
+    tmp2  = MULTIPLY(tmp11,   FIX(1.247225013));   /* c5 */
+    tmp3  = MULTIPLY(z1 + z4, FIX(1.093201867));   /* c7 */
+    tmp10 = MULTIPLY(z1 - z4, FIX(0.897167586));   /* c9 */
+    tmp11 = MULTIPLY(tmp11,   FIX(0.666655658));   /* c11 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.410524528));   /* c13 */
+    tmp0  = tmp1 + tmp2 + tmp3 -
+	    MULTIPLY(z1, FIX(2.286341144));        /* c7+c5+c3-c1 */
+    tmp13 = tmp10 + tmp11 + tmp12 -
+	    MULTIPLY(z1, FIX(1.835730603));        /* c9+c11+c13-c15 */
+    z1    = MULTIPLY(z2 + z3, FIX(0.138617169));   /* c15 */
+    tmp1  += z1 + MULTIPLY(z2, FIX(0.071888074));  /* c9+c11-c3-c15 */
+    tmp2  += z1 - MULTIPLY(z3, FIX(1.125726048));  /* c5+c7+c15-c3 */
+    z1    = MULTIPLY(z3 - z2, FIX(1.407403738));   /* c1 */
+    tmp11 += z1 - MULTIPLY(z3, FIX(0.766367282));  /* c1+c11-c9-c13 */
+    tmp12 += z1 + MULTIPLY(z2, FIX(1.971951411));  /* c1+c5+c13-c7 */
+    z2    += z4;
+    z1    = MULTIPLY(z2, - FIX(0.666655658));      /* -c11 */
+    tmp1  += z1;
+    tmp3  += z1 + MULTIPLY(z4, FIX(1.065388962));  /* c3+c11+c15-c7 */
+    z2    = MULTIPLY(z2, - FIX(1.247225013));      /* -c5 */
+    tmp10 += z2 + MULTIPLY(z4, FIX(3.141271809));  /* c1+c5+c9-c13 */
+    tmp12 += z2;
+    z2    = MULTIPLY(z3 + z4, - FIX(1.353318001)); /* -c3 */
+    tmp2  += z2;
+    tmp3  += z2;
+    z2    = MULTIPLY(z4 - z3, FIX(0.410524528));   /* c13 */
+    tmp10 += z2;
+    tmp11 += z2;
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp0,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[15] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp0,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp1,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[14] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp1,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp2,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[13] = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp2,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp3,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[12] = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp3,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[11] = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp25 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp26 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp26 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp27 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp27 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 16x8 output block.
+ *
+ * 8-point IDCT in pass 1 (columns), 16-point in pass 2 (rows).
+ */
+
+GLOBAL(void)
+jpeg_idct_16x8 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		JCOEFPTR coef_block,
+		JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3, tmp10, tmp11, tmp12, tmp13;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26, tmp27;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*8];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array. */
+  /* Note results are scaled up by sqrt(8) compared to a true IDCT; */
+  /* furthermore, we scale the results by 2**PASS1_BITS. */
+
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = DCTSIZE; ctr > 0; ctr--) {
+    /* Due to quantization, we will usually find that many of the input
+     * coefficients are zero, especially the AC terms.  We can exploit this
+     * by short-circuiting the IDCT calculation for any column in which all
+     * the AC terms are zero.  In that case each output is equal to the
+     * DC coefficient (with scale factor as needed).
+     * With typical images and quantization tables, half or more of the
+     * column DCT calculations can be simplified this way.
+     */
+    
+    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
+	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
+	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
+	inptr[DCTSIZE*7] == 0) {
+      /* AC terms all zero */
+      int dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) << PASS1_BITS;
+      
+      wsptr[DCTSIZE*0] = dcval;
+      wsptr[DCTSIZE*1] = dcval;
+      wsptr[DCTSIZE*2] = dcval;
+      wsptr[DCTSIZE*3] = dcval;
+      wsptr[DCTSIZE*4] = dcval;
+      wsptr[DCTSIZE*5] = dcval;
+      wsptr[DCTSIZE*6] = dcval;
+      wsptr[DCTSIZE*7] = dcval;
+      
+      inptr++;			/* advance pointers to next column */
+      quantptr++;
+      wsptr++;
+      continue;
+    }
+    
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+    
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z2, FIX_0_765366865);
+    tmp3 = z1 - MULTIPLY(z3, FIX_1_847759065);
+    
+    z2 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z2 <<= CONST_BITS;
+    z3 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    z2 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    tmp0 = z2 + z3;
+    tmp1 = z2 - z3;
+    
+    tmp10 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+    tmp11 = tmp1 + tmp3;
+    tmp12 = tmp1 - tmp3;
+    
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+    
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
+    tmp1 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    tmp3 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    
+    z2 = tmp0 + tmp2;
+    z3 = tmp1 + tmp3;
+
+    z1 = MULTIPLY(z2 + z3, FIX_1_175875602); /* sqrt(2) * c3 */
+    z2 = MULTIPLY(z2, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z3 = MULTIPLY(z3, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    z2 += z1;
+    z3 += z1;
+
+    z1 = MULTIPLY(tmp0 + tmp3, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    tmp0 += z1 + z2;
+    tmp3 += z1 + z3;
+
+    z1 = MULTIPLY(tmp1 + tmp2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp1 += z1 + z3;
+    tmp2 += z1 + z2;
+    
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+    
+    wsptr[DCTSIZE*0] = (int) RIGHT_SHIFT(tmp10 + tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*7] = (int) RIGHT_SHIFT(tmp10 - tmp3, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*1] = (int) RIGHT_SHIFT(tmp11 + tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*6] = (int) RIGHT_SHIFT(tmp11 - tmp2, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*2] = (int) RIGHT_SHIFT(tmp12 + tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*5] = (int) RIGHT_SHIFT(tmp12 - tmp1, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*3] = (int) RIGHT_SHIFT(tmp13 + tmp0, CONST_BITS-PASS1_BITS);
+    wsptr[DCTSIZE*4] = (int) RIGHT_SHIFT(tmp13 - tmp0, CONST_BITS-PASS1_BITS);
+    
+    inptr++;			/* advance pointers to next column */
+    quantptr++;
+    wsptr++;
+  }
+
+  /* Pass 2: process 8 rows from work array, store into output array.
+   * 16-point IDCT kernel, cK represents sqrt(2) * cos(K*pi/32).
+   */
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    tmp0 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    tmp0 <<= CONST_BITS;
+
+    z1 = (INT32) wsptr[4];
+    tmp1 = MULTIPLY(z1, FIX(1.306562965));      /* c4[16] = c2[8] */
+    tmp2 = MULTIPLY(z1, FIX_0_541196100);       /* c12[16] = c6[8] */
+
+    tmp10 = tmp0 + tmp1;
+    tmp11 = tmp0 - tmp1;
+    tmp12 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+
+    z1 = (INT32) wsptr[2];
+    z2 = (INT32) wsptr[6];
+    z3 = z1 - z2;
+    z4 = MULTIPLY(z3, FIX(0.275899379));        /* c14[16] = c7[8] */
+    z3 = MULTIPLY(z3, FIX(1.387039845));        /* c2[16] = c1[8] */
+
+    tmp0 = z3 + MULTIPLY(z2, FIX_2_562915447);  /* (c6+c2)[16] = (c3+c1)[8] */
+    tmp1 = z4 + MULTIPLY(z1, FIX_0_899976223);  /* (c6-c14)[16] = (c3-c7)[8] */
+    tmp2 = z3 - MULTIPLY(z1, FIX(0.601344887)); /* (c2-c10)[16] = (c1-c5)[8] */
+    tmp3 = z4 - MULTIPLY(z2, FIX(0.509795579)); /* (c10-c14)[16] = (c5-c7)[8] */
+
+    tmp20 = tmp10 + tmp0;
+    tmp27 = tmp10 - tmp0;
+    tmp21 = tmp12 + tmp1;
+    tmp26 = tmp12 - tmp1;
+    tmp22 = tmp13 + tmp2;
+    tmp25 = tmp13 - tmp2;
+    tmp23 = tmp11 + tmp3;
+    tmp24 = tmp11 - tmp3;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+
+    tmp11 = z1 + z3;
+
+    tmp1  = MULTIPLY(z1 + z2, FIX(1.353318001));   /* c3 */
+    tmp2  = MULTIPLY(tmp11,   FIX(1.247225013));   /* c5 */
+    tmp3  = MULTIPLY(z1 + z4, FIX(1.093201867));   /* c7 */
+    tmp10 = MULTIPLY(z1 - z4, FIX(0.897167586));   /* c9 */
+    tmp11 = MULTIPLY(tmp11,   FIX(0.666655658));   /* c11 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.410524528));   /* c13 */
+    tmp0  = tmp1 + tmp2 + tmp3 -
+	    MULTIPLY(z1, FIX(2.286341144));        /* c7+c5+c3-c1 */
+    tmp13 = tmp10 + tmp11 + tmp12 -
+	    MULTIPLY(z1, FIX(1.835730603));        /* c9+c11+c13-c15 */
+    z1    = MULTIPLY(z2 + z3, FIX(0.138617169));   /* c15 */
+    tmp1  += z1 + MULTIPLY(z2, FIX(0.071888074));  /* c9+c11-c3-c15 */
+    tmp2  += z1 - MULTIPLY(z3, FIX(1.125726048));  /* c5+c7+c15-c3 */
+    z1    = MULTIPLY(z3 - z2, FIX(1.407403738));   /* c1 */
+    tmp11 += z1 - MULTIPLY(z3, FIX(0.766367282));  /* c1+c11-c9-c13 */
+    tmp12 += z1 + MULTIPLY(z2, FIX(1.971951411));  /* c1+c5+c13-c7 */
+    z2    += z4;
+    z1    = MULTIPLY(z2, - FIX(0.666655658));      /* -c11 */
+    tmp1  += z1;
+    tmp3  += z1 + MULTIPLY(z4, FIX(1.065388962));  /* c3+c11+c15-c7 */
+    z2    = MULTIPLY(z2, - FIX(1.247225013));      /* -c5 */
+    tmp10 += z2 + MULTIPLY(z4, FIX(3.141271809));  /* c1+c5+c9-c13 */
+    tmp12 += z2;
+    z2    = MULTIPLY(z3 + z4, - FIX(1.353318001)); /* -c3 */
+    tmp2  += z2;
+    tmp3  += z2;
+    z2    = MULTIPLY(z4 - z3, FIX(0.410524528));   /* c13 */
+    tmp10 += z2;
+    tmp11 += z2;
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp0,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[15] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp0,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp1,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[14] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp1,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp2,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[13] = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp2,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp3,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[12] = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp3,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[11] = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp25 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp26 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp26 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp27 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp27 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 14x7 output block.
+ *
+ * 7-point IDCT in pass 1 (columns), 14-point in pass 2 (rows).
+ */
+
+GLOBAL(void)
+jpeg_idct_14x7 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		JCOEFPTR coef_block,
+		JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25, tmp26;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*7];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array.
+   * 7-point IDCT kernel, cK represents sqrt(2) * cos(K*pi/14).
+   */
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp23 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp23 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp23 += ONE << (CONST_BITS-PASS1_BITS-1);
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
+
+    tmp20 = MULTIPLY(z2 - z3, FIX(0.881747734));       /* c4 */
+    tmp22 = MULTIPLY(z1 - z2, FIX(0.314692123));       /* c6 */
+    tmp21 = tmp20 + tmp22 + tmp23 - MULTIPLY(z2, FIX(1.841218003)); /* c2+c4-c6 */
+    tmp10 = z1 + z3;
+    z2 -= tmp10;
+    tmp10 = MULTIPLY(tmp10, FIX(1.274162392)) + tmp23; /* c2 */
+    tmp20 += tmp10 - MULTIPLY(z3, FIX(0.077722536));   /* c2-c4-c6 */
+    tmp22 += tmp10 - MULTIPLY(z1, FIX(2.470602249));   /* c2+c4+c6 */
+    tmp23 += MULTIPLY(z2, FIX(1.414213562));           /* c0 */
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+
+    tmp11 = MULTIPLY(z1 + z2, FIX(0.935414347));       /* (c3+c1-c5)/2 */
+    tmp12 = MULTIPLY(z1 - z2, FIX(0.170262339));       /* (c3+c5-c1)/2 */
+    tmp10 = tmp11 - tmp12;
+    tmp11 += tmp12;
+    tmp12 = MULTIPLY(z2 + z3, - FIX(1.378756276));     /* -c1 */
+    tmp11 += tmp12;
+    z2 = MULTIPLY(z1 + z3, FIX(0.613604268));          /* c5 */
+    tmp10 += z2;
+    tmp12 += z2 + MULTIPLY(z3, FIX(1.870828693));      /* c3+c1-c5 */
+
+    /* Final output stage */
+
+    wsptr[8*0] = (int) RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*6] = (int) RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*1] = (int) RIGHT_SHIFT(tmp21 + tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*5] = (int) RIGHT_SHIFT(tmp21 - tmp11, CONST_BITS-PASS1_BITS);
+    wsptr[8*2] = (int) RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*4] = (int) RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*3] = (int) RIGHT_SHIFT(tmp23, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 7 rows from work array, store into output array.
+   * 14-point IDCT kernel, cK represents sqrt(2) * cos(K*pi/28).
+   */
+  wsptr = workspace;
+  for (ctr = 0; ctr < 7; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z1 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z1 <<= CONST_BITS;
+    z4 = (INT32) wsptr[4];
+    z2 = MULTIPLY(z4, FIX(1.274162392));         /* c4 */
+    z3 = MULTIPLY(z4, FIX(0.314692123));         /* c12 */
+    z4 = MULTIPLY(z4, FIX(0.881747734));         /* c8 */
+
+    tmp10 = z1 + z2;
+    tmp11 = z1 + z3;
+    tmp12 = z1 - z4;
+
+    tmp23 = z1 - ((z2 + z3 - z4) << 1);          /* c0 = (c4+c12-c8)*2 */
+
+    z1 = (INT32) wsptr[2];
+    z2 = (INT32) wsptr[6];
+
+    z3 = MULTIPLY(z1 + z2, FIX(1.105676686));    /* c6 */
+
+    tmp13 = z3 + MULTIPLY(z1, FIX(0.273079590)); /* c2-c6 */
+    tmp14 = z3 - MULTIPLY(z2, FIX(1.719280954)); /* c6+c10 */
+    tmp15 = MULTIPLY(z1, FIX(0.613604268)) -     /* c10 */
+	    MULTIPLY(z2, FIX(1.378756276));      /* c2 */
+
+    tmp20 = tmp10 + tmp13;
+    tmp26 = tmp10 - tmp13;
+    tmp21 = tmp11 + tmp14;
+    tmp25 = tmp11 - tmp14;
+    tmp22 = tmp12 + tmp15;
+    tmp24 = tmp12 - tmp15;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+    z4 <<= CONST_BITS;
+
+    tmp14 = z1 + z3;
+    tmp11 = MULTIPLY(z1 + z2, FIX(1.334852607));           /* c3 */
+    tmp12 = MULTIPLY(tmp14, FIX(1.197448846));             /* c5 */
+    tmp10 = tmp11 + tmp12 + z4 - MULTIPLY(z1, FIX(1.126980169)); /* c3+c5-c1 */
+    tmp14 = MULTIPLY(tmp14, FIX(0.752406978));             /* c9 */
+    tmp16 = tmp14 - MULTIPLY(z1, FIX(1.061150426));        /* c9+c11-c13 */
+    z1    -= z2;
+    tmp15 = MULTIPLY(z1, FIX(0.467085129)) - z4;           /* c11 */
+    tmp16 += tmp15;
+    tmp13 = MULTIPLY(z2 + z3, - FIX(0.158341681)) - z4;    /* -c13 */
+    tmp11 += tmp13 - MULTIPLY(z2, FIX(0.424103948));       /* c3-c9-c13 */
+    tmp12 += tmp13 - MULTIPLY(z3, FIX(2.373959773));       /* c3+c5-c13 */
+    tmp13 = MULTIPLY(z3 - z2, FIX(1.405321284));           /* c1 */
+    tmp14 += tmp13 + z4 - MULTIPLY(z3, FIX(1.6906431334)); /* c1+c9-c11 */
+    tmp15 += tmp13 + MULTIPLY(z2, FIX(0.674957567));       /* c1+c11-c5 */
+
+    tmp13 = ((z1 - z3) << CONST_BITS) + z4;
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[13] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[12] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[11] = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25 + tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp25 - tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp26 + tmp16,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp26 - tmp16,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 12x6 output block.
+ *
+ * 6-point IDCT in pass 1 (columns), 12-point in pass 2 (rows).
+ */
+
+GLOBAL(void)
+jpeg_idct_12x6 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		JCOEFPTR coef_block,
+		JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14, tmp15;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24, tmp25;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*6];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array.
+   * 6-point IDCT kernel, cK represents sqrt(2) * cos(K*pi/12).
+   */
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp10 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp10 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp10 += ONE << (CONST_BITS-PASS1_BITS-1);
+    tmp12 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    tmp20 = MULTIPLY(tmp12, FIX(0.707106781));   /* c4 */
+    tmp11 = tmp10 + tmp20;
+    tmp21 = RIGHT_SHIFT(tmp10 - tmp20 - tmp20, CONST_BITS-PASS1_BITS);
+    tmp20 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp10 = MULTIPLY(tmp20, FIX(1.224744871));   /* c2 */
+    tmp20 = tmp11 + tmp10;
+    tmp22 = tmp11 - tmp10;
+
+    /* Odd part */
+
+    z1 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z2 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
+    tmp11 = MULTIPLY(z1 + z3, FIX(0.366025404)); /* c5 */
+    tmp10 = tmp11 + ((z1 + z2) << CONST_BITS);
+    tmp12 = tmp11 + ((z3 - z2) << CONST_BITS);
+    tmp11 = (z1 - z2 - z3) << PASS1_BITS;
+
+    /* Final output stage */
+
+    wsptr[8*0] = (int) RIGHT_SHIFT(tmp20 + tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*5] = (int) RIGHT_SHIFT(tmp20 - tmp10, CONST_BITS-PASS1_BITS);
+    wsptr[8*1] = (int) (tmp21 + tmp11);
+    wsptr[8*4] = (int) (tmp21 - tmp11);
+    wsptr[8*2] = (int) RIGHT_SHIFT(tmp22 + tmp12, CONST_BITS-PASS1_BITS);
+    wsptr[8*3] = (int) RIGHT_SHIFT(tmp22 - tmp12, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 6 rows from work array, store into output array.
+   * 12-point IDCT kernel, cK represents sqrt(2) * cos(K*pi/24).
+   */
+  wsptr = workspace;
+  for (ctr = 0; ctr < 6; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z3 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z3 <<= CONST_BITS;
+
+    z4 = (INT32) wsptr[4];
+    z4 = MULTIPLY(z4, FIX(1.224744871)); /* c4 */
+
+    tmp10 = z3 + z4;
+    tmp11 = z3 - z4;
+
+    z1 = (INT32) wsptr[2];
+    z4 = MULTIPLY(z1, FIX(1.366025404)); /* c2 */
+    z1 <<= CONST_BITS;
+    z2 = (INT32) wsptr[6];
+    z2 <<= CONST_BITS;
+
+    tmp12 = z1 - z2;
+
+    tmp21 = z3 + tmp12;
+    tmp24 = z3 - tmp12;
+
+    tmp12 = z4 + z2;
+
+    tmp20 = tmp10 + tmp12;
+    tmp25 = tmp10 - tmp12;
+
+    tmp12 = z4 - z1 - z2;
+
+    tmp22 = tmp11 + tmp12;
+    tmp23 = tmp11 - tmp12;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z4 = (INT32) wsptr[7];
+
+    tmp11 = MULTIPLY(z2, FIX(1.306562965));                  /* c3 */
+    tmp14 = MULTIPLY(z2, - FIX_0_541196100);                 /* -c9 */
+
+    tmp10 = z1 + z3;
+    tmp15 = MULTIPLY(tmp10 + z4, FIX(0.860918669));          /* c7 */
+    tmp12 = tmp15 + MULTIPLY(tmp10, FIX(0.261052384));       /* c5-c7 */
+    tmp10 = tmp12 + tmp11 + MULTIPLY(z1, FIX(0.280143716));  /* c1-c5 */
+    tmp13 = MULTIPLY(z3 + z4, - FIX(1.045510580));           /* -(c7+c11) */
+    tmp12 += tmp13 + tmp14 - MULTIPLY(z3, FIX(1.478575242)); /* c1+c5-c7-c11 */
+    tmp13 += tmp15 - tmp11 + MULTIPLY(z4, FIX(1.586706681)); /* c1+c11 */
+    tmp15 += tmp14 - MULTIPLY(z1, FIX(0.676326758)) -        /* c7-c11 */
+	     MULTIPLY(z4, FIX(1.982889723));                 /* c5+c7 */
+
+    z1 -= z4;
+    z2 -= z3;
+    z3 = MULTIPLY(z1 + z2, FIX_0_541196100);                 /* c9 */
+    tmp11 = z3 + MULTIPLY(z1, FIX_0_765366865);              /* c3-c9 */
+    tmp14 = z3 - MULTIPLY(z2, FIX_1_847759065);              /* c3+c9 */
+
+    /* Final output stage */
+
+    outptr[0]  = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[11] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[1]  = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[10] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[2]  = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[9]  = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[3]  = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[8]  = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[4]  = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[7]  = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[5]  = range_limit[(int) RIGHT_SHIFT(tmp25 + tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+    outptr[6]  = range_limit[(int) RIGHT_SHIFT(tmp25 - tmp15,
+					       CONST_BITS+PASS1_BITS+3)
+			     & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 10x5 output block.
+ *
+ * 5-point IDCT in pass 1 (columns), 10-point in pass 2 (rows).
+ */
+
+GLOBAL(void)
+jpeg_idct_10x5 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+		JCOEFPTR coef_block,
+		JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp10, tmp11, tmp12, tmp13, tmp14;
+  INT32 tmp20, tmp21, tmp22, tmp23, tmp24;
+  INT32 z1, z2, z3, z4;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*5];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array.
+   * 5-point IDCT kernel, cK represents sqrt(2) * cos(K*pi/10).
+   */
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp12 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp12 <<= CONST_BITS;
+    /* Add fudge factor here for final descale. */
+    tmp12 += ONE << (CONST_BITS-PASS1_BITS-1);
+    tmp13 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+    tmp14 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
+    z1 = MULTIPLY(tmp13 + tmp14, FIX(0.790569415)); /* (c2+c4)/2 */
+    z2 = MULTIPLY(tmp13 - tmp14, FIX(0.353553391)); /* (c2-c4)/2 */
+    z3 = tmp12 + z2;
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z1;
+    tmp12 -= z2 << 2;
+
+    /* Odd part */
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));       /* c3 */
+    tmp13 = z1 + MULTIPLY(z2, FIX(0.513743148));    /* c1-c3 */
+    tmp14 = z1 - MULTIPLY(z3, FIX(2.176250899));    /* c1+c3 */
+
+    /* Final output stage */
+
+    wsptr[8*0] = (int) RIGHT_SHIFT(tmp10 + tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*4] = (int) RIGHT_SHIFT(tmp10 - tmp13, CONST_BITS-PASS1_BITS);
+    wsptr[8*1] = (int) RIGHT_SHIFT(tmp11 + tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*3] = (int) RIGHT_SHIFT(tmp11 - tmp14, CONST_BITS-PASS1_BITS);
+    wsptr[8*2] = (int) RIGHT_SHIFT(tmp12, CONST_BITS-PASS1_BITS);
+  }
+
+  /* Pass 2: process 5 rows from work array, store into output array.
+   * 10-point IDCT kernel, cK represents sqrt(2) * cos(K*pi/20).
+   */
+  wsptr = workspace;
+  for (ctr = 0; ctr < 5; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part */
+
+    /* Add fudge factor here for final descale. */
+    z3 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z3 <<= CONST_BITS;
+    z4 = (INT32) wsptr[4];
+    z1 = MULTIPLY(z4, FIX(1.144122806));         /* c4 */
+    z2 = MULTIPLY(z4, FIX(0.437016024));         /* c8 */
+    tmp10 = z3 + z1;
+    tmp11 = z3 - z2;
+
+    tmp22 = z3 - ((z1 - z2) << 1);               /* c0 = (c4-c8)*2 */
+
+    z2 = (INT32) wsptr[2];
+    z3 = (INT32) wsptr[6];
+
+    z1 = MULTIPLY(z2 + z3, FIX(0.831253876));    /* c6 */
+    tmp12 = z1 + MULTIPLY(z2, FIX(0.513743148)); /* c2-c6 */
+    tmp13 = z1 - MULTIPLY(z3, FIX(2.176250899)); /* c2+c6 */
+
+    tmp20 = tmp10 + tmp12;
+    tmp24 = tmp10 - tmp12;
+    tmp21 = tmp11 + tmp13;
+    tmp23 = tmp11 - tmp13;
+
+    /* Odd part */
+
+    z1 = (INT32) wsptr[1];
+    z2 = (INT32) wsptr[3];
+    z3 = (INT32) wsptr[5];
+    z3 <<= CONST_BITS;
+    z4 = (INT32) wsptr[7];
+
+    tmp11 = z2 + z4;
+    tmp13 = z2 - z4;
+
+    tmp12 = MULTIPLY(tmp13, FIX(0.309016994));        /* (c3-c7)/2 */
+
+    z2 = MULTIPLY(tmp11, FIX(0.951056516));           /* (c3+c7)/2 */
+    z4 = z3 + tmp12;
+
+    tmp10 = MULTIPLY(z1, FIX(1.396802247)) + z2 + z4; /* c1 */
+    tmp14 = MULTIPLY(z1, FIX(0.221231742)) - z2 + z4; /* c9 */
+
+    z2 = MULTIPLY(tmp11, FIX(0.587785252));           /* (c1-c9)/2 */
+    z4 = z3 - tmp12 - (tmp13 << (CONST_BITS - 1));
+
+    tmp12 = ((z1 - tmp13) << CONST_BITS) - z3;
+
+    tmp11 = MULTIPLY(z1, FIX(1.260073511)) - z2 - z4; /* c3 */
+    tmp13 = MULTIPLY(z1, FIX(0.642039522)) - z2 + z4; /* c7 */
+
+    /* Final output stage */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp20 + tmp10,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[9] = range_limit[(int) RIGHT_SHIFT(tmp20 - tmp10,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp21 + tmp11,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[8] = range_limit[(int) RIGHT_SHIFT(tmp21 - tmp11,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp22 + tmp12,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) RIGHT_SHIFT(tmp22 - tmp12,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp23 + tmp13,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) RIGHT_SHIFT(tmp23 - tmp13,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) RIGHT_SHIFT(tmp24 + tmp14,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) RIGHT_SHIFT(tmp24 - tmp14,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+
+    wsptr += 8;		/* advance pointer to next row */
+  }
+}
+
+
+/*
+ * Perform dequantization and inverse DCT on one block of coefficients,
+ * producing a 8x4 output block.
+ *
+ * 4-point IDCT in pass 1 (columns), 8-point in pass 2 (rows).
+ */
+
+GLOBAL(void)
+jpeg_idct_8x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
+	       JCOEFPTR coef_block,
+	       JSAMPARRAY output_buf, JDIMENSION output_col)
+{
+  INT32 tmp0, tmp1, tmp2, tmp3;
+  INT32 tmp10, tmp11, tmp12, tmp13;
+  INT32 z1, z2, z3;
+  JCOEFPTR inptr;
+  ISLOW_MULT_TYPE * quantptr;
+  int * wsptr;
+  JSAMPROW outptr;
+  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
+  int ctr;
+  int workspace[8*4];	/* buffers data between passes */
+  SHIFT_TEMPS
+
+  /* Pass 1: process columns from input, store into work array.
+   * 4-point IDCT kernel, cK represents sqrt(2) * cos(K*pi/16).
+   */
+  inptr = coef_block;
+  quantptr = (ISLOW_MULT_TYPE *) compptr->dct_table;
+  wsptr = workspace;
+  for (ctr = 0; ctr < 8; ctr++, inptr++, quantptr++, wsptr++) {
+    /* Even part */
+
+    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
+    tmp2 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
+
+    tmp10 = (tmp0 + tmp2) << PASS1_BITS;
+    tmp12 = (tmp0 - tmp2) << PASS1_BITS;
+
+    /* Odd part */
+    /* Same rotation as in the even part of the 8x8 LL&M IDCT */
+
+    z2 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
+    z3 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
+
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);               /* c6 */
+    /* Add fudge factor here for final descale. */
+    z1 += ONE << (CONST_BITS-PASS1_BITS-1);
+    tmp0 = RIGHT_SHIFT(z1 + MULTIPLY(z2, FIX_0_765366865), /* c2-c6 */
+		       CONST_BITS-PASS1_BITS);
+    tmp2 = RIGHT_SHIFT(z1 - MULTIPLY(z3, FIX_1_847759065), /* c2+c6 */
+		       CONST_BITS-PASS1_BITS);
+
+    /* Final output stage */
+
+    wsptr[8*0] = (int) (tmp10 + tmp0);
+    wsptr[8*3] = (int) (tmp10 - tmp0);
+    wsptr[8*1] = (int) (tmp12 + tmp2);
+    wsptr[8*2] = (int) (tmp12 - tmp2);
+  }
+
+  /* Pass 2: process rows from work array, store into output array. */
+  /* Note that we must descale the results by a factor of 8 == 2**3, */
+  /* and also undo the PASS1_BITS scaling. */
+
+  wsptr = workspace;
+  for (ctr = 0; ctr < 4; ctr++) {
+    outptr = output_buf[ctr] + output_col;
+
+    /* Even part: reverse the even part of the forward DCT. */
+    /* The rotator is sqrt(2)*c(-6). */
+
+    z2 = (INT32) wsptr[2];
+    z3 = (INT32) wsptr[6];
+    
+    z1 = MULTIPLY(z2 + z3, FIX_0_541196100);
+    tmp2 = z1 + MULTIPLY(z2, FIX_0_765366865);
+    tmp3 = z1 - MULTIPLY(z3, FIX_1_847759065);
+    
+    /* Add fudge factor here for final descale. */
+    z2 = (INT32) wsptr[0] + (ONE << (PASS1_BITS+2));
+    z3 = (INT32) wsptr[4];
+    
+    tmp0 = (z2 + z3) << CONST_BITS;
+    tmp1 = (z2 - z3) << CONST_BITS;
+    
+    tmp10 = tmp0 + tmp2;
+    tmp13 = tmp0 - tmp2;
+    tmp11 = tmp1 + tmp3;
+    tmp12 = tmp1 - tmp3;
+
+    /* Odd part per figure 8; the matrix is unitary and hence its
+     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
+     */
+
+    tmp0 = (INT32) wsptr[7];
+    tmp1 = (INT32) wsptr[5];
+    tmp2 = (INT32) wsptr[3];
+    tmp3 = (INT32) wsptr[1];
+
+    z2 = tmp0 + tmp2;
+    z3 = tmp1 + tmp3;
+
+    z1 = MULTIPLY(z2 + z3, FIX_1_175875602); /* sqrt(2) * c3 */
+    z2 = MULTIPLY(z2, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+    z3 = MULTIPLY(z3, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+    z2 += z1;
+    z3 += z1;
+
+    z1 = MULTIPLY(tmp0 + tmp3, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+    tmp0 = MULTIPLY(tmp0, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+    tmp3 = MULTIPLY(tmp3, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+    tmp0 += z1 + z2;
+    tmp3 += z1 + z3;
+
+    z1 = MULTIPLY(tmp1 + tmp2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+    tmp1 = MULTIPLY(tmp1, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+    tmp2 = MULTIPLY(tmp2, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+    tmp1 += z1 + z3;
+    tmp2 += z1 + z2;
+
+    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */
+
+    outptr[0] = range_limit[(int) RIGHT_SHIFT(tmp10 + tmp3,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[7] = range_limit[(int) RIGHT_SHIFT(tmp10 - tmp3,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[1] = range_limit[(int) RIGHT_SHIFT(tmp11 + tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[6] = range_limit[(int) RIGHT_SHIFT(tmp11 - tmp2,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[2] = range_limit[(int) RIGHT_SHIFT(tmp12 + tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[5] = range_limit[(int) RIGHT_SHIFT(tmp12 - tmp1,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[3] = range_limit[(int) RIGHT_SHIFT(tmp13 + tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    & RANGE_MASK];
+    outptr[4] = range_limit[(int) RIGHT_SHIFT(tmp13 - tmp0,
+					      CONST_BITS+PASS1_BITS+3)
+			    &