[Mesa-dev] [Mesa3d-dev] Softpipe and OpenMP

[Jose Fonseca]

On 09/06/15 03:00, Roland Scheidegger wrote:
> Am 08.06.2015 um 09:33 schrieb Chih-Sheng Lin:
>> Hi,
>>
>> I am working on parallelizing softpipe of Mesa-9.1 by using OpenMP.
>>
>> My idea is trying to do sp_setup_tri in parallel with multiple threads.
>>
>> So first in the function sp_vbuf_draw_elements, I duplicate
>> setup_context for avoiding race condition
>>
>> And then sp_setup_tri is set to OpenMP parallel sections with the
>> separate auguments
>>
>> The code is as follows:
>>
>> sp_vbuf_draw_elements(struct vbuf_render *vbr, const ushort *indices,
>> uint nr)
>> {
>>     struct softpipe_vbuf_render *cvbr = softpipe_vbuf_render(vbr);
>>     struct softpipe_context *softpipe = cvbr->softpipe;
>>     const unsigned stride = softpipe->vertex_info_vbuf.size * sizeof(float);
>>     const void *vertex_buffer = cvbr->vertex_buffer;
>>     struct setup_context *setup = cvbr->setup;
>>
>> /*duplicate setup context*/
>> struct setup_context *setup_0, *setup_1;
>> if(cvbr->prim == PIPE_PRIM_TRIANGLES)
>> {
>>      setup_0 = sp_setup_create_context(cvbr->setup);
>>      setup_1 = sp_setup_create_context(cvbr->setup);
>> }
>>
>> ...
>>
>>     case PIPE_PRIM_TRIANGLES:
>> /*
>>        for (i = 2; i < nr; i += 3) {
>>           sp_setup_tri( setup,
>>                         get_vert(vertex_buffer, indices[i-2], stride),
>>                         get_vert(vertex_buffer, indices[i-1], stride),
>>                         get_vert(vertex_buffer, indices[i-0], stride) );
>>        }
>> */
>>
>> /*parallelizing sp_setup_tri by OpenMP parallel sections*/
>> if(nr==6)
>> {
>>      #pragma omp parallel sections
>>      {
>>          #pragma omp section
>>          sp_setup_tri(setup_0, get_vert(vertex_buffer, indices[0],
>> stride), get_vert(vertex_buffer, indices[1],
>> stride),get_vert(vertex_buffer, indices[2], stride));
>>          #pragma omp section
>>          sp_setup_tri(setup_1, get_vert(vertex_buffer, indices[3],
>> stride), get_vert(vertex_buffer, indices[4],
>> stride),get_vert(vertex_buffer, indices[5], stride));
>>      }
>> }
>>      sp_setup_destroy_context(setup_0);
>>      sp_setup_destroy_context(setup_1);
>>
>>        break;
>>
>> And I rebuild mesa and run one test called CPUOverheadTest_onscreen from
>> GFXBench 2.7.5
>>
>> I set OMP_NUM_THREADS=2 and the resolution is 1280x900, it gains 10%
>> performance improvements comparing to OMP_NUM_THREADS=1
>>
>> However, the test result is not correct, but it works very well
>> comparing to my first attempt on parallelization
>>
>> Could someone give me a guide or suggestions for this work?
>>
>> Thanks in advance!
>>
>>
>> Patrick Lin
>
> I'm not sure if that was an issue with what you tested, but note that
> tris can (and often will) overlap and you need to guarantee strict ordering.
> Probably easier if you'd try to parallelize the things inside
> sp_setup_tri, that is the pixel shader load (each shader is ultimately
> run for 2x2 pixels, so you could for instance have 4 threads, and try to
> run 4 2x2 pixel shaders in parallel).
> That said, trying to make softpipe faster by using OpenMP might be a
> nice theoretical excercise but is probably futile. If you want somewhat
> useful performance with a sw renderer try llvmpipe - which already does
> have parallel pixel shader threads (though in practice the scaling is
> not quite optimal once you go past a couple threads, one reason probably
> being that only things after triangle setup can be run in parallel,
> there's also some "unnecessary" waits).

Yes, the point of softpipe existing is to be a nice and simple to 
understand gallium reference driver.  Whereas llvmpipe is its 
multi-threaded JIT-compiled, SIMD-vectorized, hairy brother.  So if you 
make softpipe a hairy multithreaded beast, not only it will be not 
enough to catch up with llvmpipe, but it would deprive us from our 
simple reference driver.

In short, OpenMP softpipe only makes sense to me as a fun experiment / 
exercise.

If you want to learn about these things, I recommend you do

   gitk src/gallium/drivers/llvmpipe

got to the very beginning, and see particularly Keith Whitwell commits, 
and you'll be able to see what things should be done to run softpipe 
with multiple threads.  After all llvmpipe started as nothing other than 
a softpipe fork!

Jose