[Mesa-dev] r600g: status of my work on the shader optimization

Vadim Girlin vadimgirlin at gmail.com
Fri Mar 1 23:28:40 PST 2013

On 03/02/2013 05:15 AM, Sebastien Caty wrote:
> On February 28, 2013 12:19:30 PM Vadim Girlin wrote:
>> Make sure that you have recent patches, some problems with R7xx chips
>> were fixed yesterday. Then please check if there are any regressions
>> with piglit tests (as compared to the piglit results with R600_SB=0). If
>> there are regressions, send me the shader dumps printed when you run the
>> regressed tests with "R600_SB_DUMP=3 R600_DUMP_SHADERS=2". If you could
>> also figure out the failing shader in each case (as described in [1]),
>> this will help to spot the problem. Also you might want to check if
>> there are any problems with some other more simple and/or open-source
>> applications, so that it will be easier for me to reproduce the issue
>> (if it's not r7xx-specific). One more way to help me with debugging your
>> issue with SS3 is to use apitrace to create the minimal trace that
>> reproduces it (again, if only it's not r7xx-specific, so that it may be
>> reproduced with my evergreen card, otherwise you could try to apply the
>> method [1] to locate the guilty shader and send me the shader dump).
>> Anyway, it's better to start with piglit - it's usually a lot easier to
>> debug any issues with a small and simple apps such as the piglit tests.
>> Vadim
> Got piglit running after figuring out that piglit uses LLVM if it's available
> even if R600_LLVM is not set. I had a lot of GPU lockups and CS errors.

If you are building mesa with R600 LLVM backend enabled, then the driver 
uses LLVM backend by default, so you have to disable it explicitly with 
R600_LLVM=0 if you don't want to use it. LLVM backend has its own 
problems with piglit, so for now it's better to disable it when you are 
testing r600-sb to avoid mixing r600-sb's and LLVM's issues, probably 
it's simpler just to build mesa without LLVM backend for this testing.


> I reduced piglit tests to the shaders set. Most of them passes correctly, 3
> are marked as fixed, 9 of them regressions, no chrases.
> Trying to dig more and found out which shader is causing trouble but I haven't
> found out how to run a specific test with piglit. Documentation isn't to
> friendly...Id appreciate some help with this.
> Test that failed :
> glx-visuals-stencil-pixmap
> glsl-const-builtin-notEqual-06
> glsl-fs-discard-02
> glsl-fs-fogcolor-statechange
> glsl-fs-loop-two-counter-02
> glsl-fs-loop-two-counter-04
> glsl-fs-vec4-indexing-5
> glsl-fs-vec4-indexing-6
> glsl-uniform-non-uniform-array-compare
>> [1]: see section "Regression debugging" at
>>        http://people.freedesktop.org/~vadimg/r600-sb.html
>>>> Vadim
>>> _______________________________________________
>>> mesa-dev mailing list
>>> mesa-dev at lists.freedesktop.org
>>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev

More information about the mesa-dev mailing list