[quake3-bugzilla] [Bug 4988] New: [PATCH] micro-optimize VectorNormalize

[bugzilla-daemon at icculus.org]

https://bugzilla.icculus.org/show_bug.cgi?id=4988

           Summary: [PATCH] micro-optimize VectorNormalize
           Product: ioquake3
           Version: SVN HEAD
          Platform: PC
        OS/Version: Linux
            Status: NEW
          Severity: normal
          Priority: P3
         Component: Platform
        AssignedTo: zakk at icculus.org
        ReportedBy: mattst88 at gmail.com
         QAContact: quake3-bugzilla at icculus.org


Created attachment 2713
  --> https://bugzilla.icculus.org/attachment.cgi?id=2713
Micro-optimize VectorNormalize{,2}

The current VectorNormalize{,2} functions call sqrt unnecessarily.

vec_t VectorNormalize( vec3_t v ) {
    // NOTE: TTimo - Apple G4 altivec source uses double?
    float   length, ilength;

    length = v[0]*v[0] + v[1]*v[1] + v[2]*v[2];
    length = sqrt (length);

    if ( length ) {
        ilength = 1/length;
        v[0] *= ilength;
        v[1] *= ilength;
        v[2] *= ilength;
    }

    return length;
}

sqrt(length) == 0.0f if and only if length = 0.0f. Thus, the sqrt should be
moved inside the if statement.

Additionally, gcc is unable to recognize that it may use reciprocal-sqrt when
available to calculate ilength. As such, it's more efficient to do this

ilength = 1/sqrtf(length);
length *= ilength;

rather than

length = sqrt(length);
ilength = 1/length;

since the first consists of a reciprocal-sqrt and multiply and the second is a
sqrt and a reciprocal or divide.

Also note, sqrt_f_ should be used instead of sqrt. Otherwise gcc wants to
convert to and from double-precision which is silly. sqrtf isn't available to
QVMs though, so casting (float)sqrt(length) is used since it generates
identical code to sqrtf.

Attached is the final patch.

-- 
Configure bugmail: https://bugzilla.icculus.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the QA contact for the bug.