[Mesa-dev] [PATCH v3 2/2] radeonsi: enable TGSI support cap for compute shaders

Bas Nieuwenhuizen bas at basnieuwenhuizen.nl
Tue Apr 19 14:11:52 UTC 2016


On Tue, Apr 19, 2016 at 4:03 PM, Alex Deucher <alexdeucher at gmail.com> wrote:
> On Tue, Apr 19, 2016 at 6:56 AM, Marek Olšák <maraeo at gmail.com> wrote:
>> Reviewed-by: Marek Olšák <marek.olsak at amd.com>
>>
>> Marek
>>
>> On Tue, Apr 19, 2016 at 1:39 AM, Bas Nieuwenhuizen
>> <bas at basnieuwenhuizen.nl> wrote:
>>> v2: Use chip_class instead of family.
>>>
>>> v3: Check kernel version for SI.
>>>
>>> Signed-off-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
>>> ---
>>>  docs/GL3.txt                                  |  4 ++--
>>>  docs/relnotes/11.3.0.html                     |  1 +
>>>  src/gallium/drivers/radeon/r600_pipe_common.c | 21 ++++++++++++++++-----
>>>  src/gallium/drivers/radeonsi/si_pipe.c        | 15 +++++++++++++--
>>>  4 files changed, 32 insertions(+), 9 deletions(-)
>>>
>>> diff --git a/docs/GL3.txt b/docs/GL3.txt
>>> index 3febd6e..6214f8d 100644
>>> --- a/docs/GL3.txt
>>> +++ b/docs/GL3.txt
>>> @@ -167,7 +167,7 @@ GL 4.3, GLSL 4.30:
>>>    GL_ARB_arrays_of_arrays                               DONE (all drivers that support GLSL 1.30)
>>>    GL_ARB_ES3_compatibility                              DONE (all drivers that support GLSL 3.30)
>>>    GL_ARB_clear_buffer_object                            DONE (all drivers)
>>> -  GL_ARB_compute_shader                                 DONE (i965)
>>> +  GL_ARB_compute_shader                                 DONE (i965, radeonsi)
>>>    GL_ARB_copy_image                                     DONE (i965, nv50, nvc0, r600, radeonsi)
>>>    GL_KHR_debug                                          DONE (all drivers)
>>>    GL_ARB_explicit_uniform_location                      DONE (all drivers that support GLSL)
>>> @@ -225,7 +225,7 @@ GL 4.5, GLSL 4.50:
>>>  These are the extensions cherry-picked to make GLES 3.1
>>>  GLES3.1, GLSL ES 3.1
>>>    GL_ARB_arrays_of_arrays                               DONE (all drivers that support GLSL 1.30)
>>> -  GL_ARB_compute_shader                                 DONE (i965)
>>> +  GL_ARB_compute_shader                                 DONE (i965, radeonsi)
>>>    GL_ARB_draw_indirect                                  DONE (i965, nvc0, r600, radeonsi, llvmpipe, softpipe)
>>>    GL_ARB_explicit_uniform_location                      DONE (all drivers that support GLSL)
>>>    GL_ARB_framebuffer_no_attachments                     DONE (i965, nvc0, r600, radeonsi, softpipe)
>>> diff --git a/docs/relnotes/11.3.0.html b/docs/relnotes/11.3.0.html
>>> index 0f9aed8..5a7083c 100644
>>> --- a/docs/relnotes/11.3.0.html
>>> +++ b/docs/relnotes/11.3.0.html
>>> @@ -45,6 +45,7 @@ Note: some of the new features are only available with certain drivers.
>>>
>>>  <ul>
>>>  <li>OpenGL 4.2 on radeonsi</li>
>>> +<li>GL_ARB_compute_shader on radeonsi</li>
>>>  <li>GL_ARB_framebuffer_no_attachments on nvc0, r600, radeonsi, softpipe</li>
>>>  <li>GL_ARB_internalformat_query2 on all drivers</li>
>>>  <li>GL_ARB_robust_buffer_access_behavior on radeonsi</li>
>>> diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
>>> index a7477ab..64da62f 100644
>>> --- a/src/gallium/drivers/radeon/r600_pipe_common.c
>>> +++ b/src/gallium/drivers/radeon/r600_pipe_common.c
>>> @@ -645,23 +645,34 @@ static int r600_get_compute_param(struct pipe_screen *screen,
>>>                         uint64_t *grid_size = ret;
>>>                         grid_size[0] = 65535;
>>>                         grid_size[1] = 65535;
>>> -                       grid_size[2] = 1;
>>> +                       grid_size[2] = 65535;
>>>                 }
>>>                 return 3 * sizeof(uint64_t) ;
>>>
>>>         case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
>>>                 if (ret) {
>>>                         uint64_t *block_size = ret;
>>> -                       block_size[0] = 256;
>>> -                       block_size[1] = 256;
>>> -                       block_size[2] = 256;
>>> +                       if (rscreen->chip_class >= SI && HAVE_LLVM >= 0x309 &&
>>> +                           ir_type == PIPE_SHADER_IR_TGSI) {
>>> +                               block_size[0] = 2048;
>>> +                               block_size[1] = 2048;
>>> +                               block_size[2] = 2048;
>>> +                       } else {
>>> +                               block_size[0] = 256;
>>> +                               block_size[1] = 256;
>>> +                               block_size[2] = 256;
>>> +                       }
>>>                 }
>>>                 return 3 * sizeof(uint64_t);
>>>
>>>         case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
>>>                 if (ret) {
>>>                         uint64_t *max_threads_per_block = ret;
>>> -                       *max_threads_per_block = 256;
>>> +                       if (rscreen->chip_class >= SI && HAVE_LLVM >= 0x309 &&
>>> +                           ir_type == PIPE_SHADER_IR_TGSI)
>>> +                               *max_threads_per_block = 2048;
>>> +                       else
>>> +                               *max_threads_per_block = 256;
>>>                 }
>>>                 return sizeof(uint64_t);
>>>
>>> diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
>>> index f22cd03..7501a8f 100644
>>> --- a/src/gallium/drivers/radeonsi/si_pipe.c
>>> +++ b/src/gallium/drivers/radeonsi/si_pipe.c
>>> @@ -447,6 +447,8 @@ static int si_get_param(struct pipe_screen* pscreen, enum pipe_cap param)
>>>
>>>  static int si_get_shader_param(struct pipe_screen* pscreen, unsigned shader, enum pipe_shader_cap param)
>>>  {
>>> +       struct si_screen *sscreen = (struct si_screen *)pscreen;
>>> +
>>>         switch(shader)
>>>         {
>>>         case PIPE_SHADER_FRAGMENT:
>>> @@ -464,9 +466,18 @@ static int si_get_shader_param(struct pipe_screen* pscreen, unsigned shader, enu
>>>                 case PIPE_SHADER_CAP_PREFERRED_IR:
>>>                         return PIPE_SHADER_IR_NATIVE;
>>>
>>> -               case PIPE_SHADER_CAP_SUPPORTED_IRS:
>>> -                       return 0;
>>> +               case PIPE_SHADER_CAP_SUPPORTED_IRS: {
>>> +                       int ir = 1 << PIPE_SHADER_IR_NATIVE;
>>>
>>> +                       /* Old kernels disallowed shader register writes using
>>> +                        * COPY_DATA packets that are used for indirect dispatches. */
>>> +                       if (HAVE_LLVM >= 0x309 && (sscreen->b.chip_class >= CIK ||
>>> +                                                  (sscreen->b.info.drm_major == 2 &&
>>> +                                                   sscreen->b.info.drm_minor >= 45)))
>
> Can you hide this in the winsys for SI?  If we support SI with amdgpu
> at some point, this will need to be changed.
>
> Alex

You mean only doing this check in the radeon winsys?  Would just
preemptively allowing SI + drm_major==3  in this if statement also be
okay?

I think that should have the same behavior as hiding it in the winsys
and the amdgpu winsys always saying it is supported.

- Bas

>>> +                               ir |= 1 << PIPE_SHADER_IR_TGSI;
>>> +
>>> +                       return ir;
>>> +               }
>>>                 case PIPE_SHADER_CAP_DOUBLES:
>>>                         return HAVE_LLVM >= 0x0307;
>>>
>>> --
>>> 2.8.0
>>>
>>> _______________________________________________
>>> mesa-dev mailing list
>>> mesa-dev at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list