Bug 55153 - Mesa could optimize repeated multiplication of a varying by a constant factor in the fragment shader
Summary: Mesa could optimize repeated multiplication of a varying by a constant factor...
Status: RESOLVED MOVED
Alias: None
Product: Mesa
Classification: Unclassified
Component: glsl-compiler (show other bugs)
Version: 8.0
Hardware: x86-64 (AMD64) Linux (All)
: medium normal
Assignee: Ian Romanick
QA Contact:
URL:
Whiteboard:
Keywords:
Depends on:
Blocks: i965-perf
  Show dependency treegraph
 
Reported: 2012-09-20 15:11 UTC by Samuel Rødal
Modified: 2019-09-18 19:44 UTC (History)
0 users

See Also:
i915 platform:
i915 features:


Attachments

Description Samuel Rødal 2012-09-20 15:11:37 UTC
When the fragment shader contains code like this:

varying vec4 pos;

void main(void){
 const float factor=0.1;

 vec3 nf1=texture2D( tex2, pos.xy*factor).xyz;
 vec3 nf2=texture2D( tex2, pos.xz*factor).xyz;
 vec3 nf3=texture2D( tex2, pos.zy*factor).xyz;

 ...
}

and this is the only use of pos in the fragment shader, it's more optimal to multiply the pos varying by 0.1 in the vertex shader. This gives a noticeable performance improvement at least on r600.

Is this something that could be automatically done by the GLSL compiler / optimizer?
Comment 1 Samuel Rødal 2012-09-20 15:23:49 UTC
Example change from real-world code implementing this change (although doing the multiplication with a temporary variable in the fragment shader instead of doing it in the vertex shader, maybe Mesa should catch this case too):

     1.5  
     1.6    fragment+='void main(void){';
     1.7 -  fragment+='const float factor=0.1;';
     1.8 +  if saVolume in attr then fragment+='vec3 fpos=pos.xyz*0.1;';
     1.9  
    1.10    // möglichst früh discarden
    1.11    if saBuild in attr then fragment+='if(texture2D(tex3, vec2(gl_TexCoord[0])).r<extra) discard;';
    1.12 @@ -144,9 +144,9 @@
    1.13     if (saWater in attr) then begin
    1.14       fragment+='vec3 _nrm = texture2D(tex2,gl_TexCoord[1].xy+vec2(sin(0.00001*time), cos(0.00001*time)), -4.0).xyz*2.0-1.0;';
    1.15     end else if saVolume in attr then begin
    1.16 -    fragment+='vec3 nf1=texture2D( tex2, pos.xy*factor).xyz;';
    1.17 -    fragment+='vec3 nf2=texture2D( tex2, pos.xz*factor).xyz;';
    1.18 -    fragment+='vec3 nf3=texture2D( tex2, pos.zy*factor).xyz;';
    1.19 +    fragment+='vec3 nf1=texture2D( tex2, fpos.xy).xyz;';
    1.20 +    fragment+='vec3 nf2=texture2D( tex2, fpos.xz).xyz;';
    1.21 +    fragment+='vec3 nf3=texture2D( tex2, fpos.zy).xyz;';
    1.22      fragment+='vec3 _nrm = normalize(vec3(norm2.z*nf1+norm2.y*nf2+norm2.x*nf3)*2.0-1.0);';
    1.23     end else fragment+='vec3 _nrm = texture2D(tex2,vec2(gl_TexCoord[0])).xyz*2.0-1.0;';
    1.24    end else if saDiffuse in attr then fragment+='vec3 normal = normalize(norm);';
    1.25 @@ -184,9 +184,9 @@
    1.26      '+vec2(0.002*sin(time*0.007+watercoord.y*30.0), -0.002+0.001*sin(time*0.002+watercoord.x*100.0)), 1.0/wh.y, 1.0-1.0/wh.y));';
    1.27  
    1.28    end else if (saVolume in attr) and (saDiffuse in attr) and (saTexture in attr) then begin
    1.29 -   fragment+='vec3 f1=texture2D( tex, pos.xy*factor).xyz;';
    1.30 -   fragment+='vec3 f2=texture2D( tex, pos.xz*factor).xyz;';
    1.31 -   fragment+='vec3 f3=texture2D( tex, pos.zy*factor).xyz;';
    1.32 +   fragment+='vec3 f1=texture2D( tex, fpos.xy).xyz;';
    1.33 +   fragment+='vec3 f2=texture2D( tex, fpos.xz).xyz;';
    1.34 +   fragment+='vec3 f3=texture2D( tex, fpos.zy).xyz;';
    1.35     fragment+='vec4 col = vec4(norm2.z*f1+norm2.y*f2+norm2.x*f3, gl_TexCoord[0].x);';
    1.36
Comment 2 GitLab Migration User 2019-09-18 19:44:56 UTC
-- GitLab Migration Automatic Message --

This bug has been migrated to freedesktop.org's GitLab instance and has been closed from further activity.

You can subscribe and participate further through the new bug through this link to our GitLab instance: https://gitlab.freedesktop.org/mesa/mesa/issues/797.


Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.