[Mesa-dev] [PATCH] anv/pipeline: Enable only one dispatch width in case of per sample shading

[Jason Ekstrand]

On Jul 26, 2016 12:54 PM, "Anuj Phogat" <anuj.phogat at gmail.com> wrote:
>
> Fixes ~45 DEQP sample shading tests:
> ./deqp-vk --deqp-case=dEQP-VK.pipeline.multisample.min_sample_shading*
>
> Many tests exited with VK_ERROR_OUT_OF_DEVICE_MEMORY without this patch.
>
> Cc: Jason Ekstrand <jason at jlekstrand.net>
> Signed-off-by: Anuj Phogat <anuj.phogat at gmail.com>
>
> ---
> Another patch enabling the sample shading is required to test this patch.
> I'll send out the enabling patch once we pass all the sample shading
tests.
> Use https://github.com/aphogat/mesa, branch: review to test the patch.
> ---
>  src/intel/vulkan/gen7_pipeline.c |  9 ++++++++-
>  src/intel/vulkan/gen8_pipeline.c | 12 ++++++++----
>  2 files changed, 16 insertions(+), 5 deletions(-)
>
> diff --git a/src/intel/vulkan/gen7_pipeline.c
b/src/intel/vulkan/gen7_pipeline.c
> index 8ce50be..23535f5 100644
> --- a/src/intel/vulkan/gen7_pipeline.c
> +++ b/src/intel/vulkan/gen7_pipeline.c
> @@ -249,6 +249,8 @@ genX(graphics_pipeline_create)(
>           anv_finishme("primitive_id needs sbe swizzling setup");
>
>        emit_3dstate_sbe(pipeline);
> +      bool per_sample_ps = pCreateInfo->pMultisampleState &&
> +
 pCreateInfo->pMultisampleState->sampleShadingEnable;
>
>        anv_batch_emit(&pipeline->batch, GENX(3DSTATE_PS), ps) {
>           ps.KernelStartPointer0           = pipeline->ps_ksp0;
> @@ -274,7 +276,12 @@ genX(graphics_pipeline_create)(
>
>           ps._32PixelDispatchEnable        = false;
>           ps._16PixelDispatchEnable        = wm_prog_data->dispatch_16;
> -         ps._8PixelDispatchEnable         = wm_prog_data->dispatch_8;
> +         /* On all hardware generations, the only configurations
supporting
> +          * persample dispatch are in which only one dispatch width is
enabled.
> +          */
> +         ps._8PixelDispatchEnable         = wm_prog_data->dispatch_8 &&
> +                                            (!per_sample_ps ||
> +                                             !wm_prog_data->dispatch_16);

I don't think we need to do this.  brw_compile_fs in brw_fs.cpp should
handle this for us based on the shader key.  We should be able to just set
the shader key bits correctly and then trust brw_compile_fs to give us only
one dispatch width.

--Jason

>
>           ps.DispatchGRFStartRegisterforConstantSetupData0 =
>              wm_prog_data->base.dispatch_grf_start_reg,
> diff --git a/src/intel/vulkan/gen8_pipeline.c
b/src/intel/vulkan/gen8_pipeline.c
> index cc10d3a..bde7660 100644
> --- a/src/intel/vulkan/gen8_pipeline.c
> +++ b/src/intel/vulkan/gen8_pipeline.c
> @@ -333,12 +333,19 @@ genX(graphics_pipeline_create)(
>        }
>     } else {
>        emit_3dstate_sbe(pipeline);
> +      bool per_sample_ps = pCreateInfo->pMultisampleState &&
> +
 pCreateInfo->pMultisampleState->sampleShadingEnable;
>
>        anv_batch_emit(&pipeline->batch, GENX(3DSTATE_PS), ps) {
>           ps.KernelStartPointer0     = pipeline->ps_ksp0;
>           ps.KernelStartPointer1     = 0;
>           ps.KernelStartPointer2     = pipeline->ps_ksp0 +
wm_prog_data->prog_offset_2;
> -         ps._8PixelDispatchEnable   = wm_prog_data->dispatch_8;
> +         /* On all hardware generations, the only configurations
supporting
> +          * persample dispatch are in which only one dispatch width is
enabled.
> +          */
> +         ps._8PixelDispatchEnable   = wm_prog_data->dispatch_8 &&
> +                                      (!per_sample_ps ||
> +                                       !wm_prog_data->dispatch_16);
>           ps._16PixelDispatchEnable  = wm_prog_data->dispatch_16;
>           ps._32PixelDispatchEnable  = false;
>           ps.SingleProgramFlow       = false;
> @@ -365,9 +372,6 @@ genX(graphics_pipeline_create)(
>              wm_prog_data->dispatch_grf_start_reg_2;
>        }
>
> -      bool per_sample_ps = pCreateInfo->pMultisampleState &&
> -
 pCreateInfo->pMultisampleState->sampleShadingEnable;
> -
>        anv_batch_emit(&pipeline->batch, GENX(3DSTATE_PS_EXTRA), ps) {
>           ps.PixelShaderValid              = true;
>           ps.PixelShaderKillsPixel         = wm_prog_data->uses_kill;
> --
> 2.5.5
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160726/8305f783/attachment-0001.html>