[Mesa-dev] [PATCH 3/3] radv: set FORCE_SIMD_DIST(1) for compute when profitable

[Samuel Pitoiset]

On 12/14/2017 08:35 PM, Bas Nieuwenhuizen wrote:
> Reviewed-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
> 
> Would it make sense to move the compute_resource_limits calculation to
> pipeline creation time?

Yeah, possibly.

> 
> On Thu, Dec 14, 2017 at 3:51 PM, Samuel Pitoiset
> <samuel.pitoiset at gmail.com> wrote:
>> Ported from RadeonSI.
>>
>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>> ---
>>   src/amd/vulkan/radv_cmd_buffer.c | 14 ++++++++++++++
>>   1 file changed, 14 insertions(+)
>>
>> diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
>> index d6aaff707b..4a048485c8 100644
>> --- a/src/amd/vulkan/radv_cmd_buffer.c
>> +++ b/src/amd/vulkan/radv_cmd_buffer.c
>> @@ -2561,6 +2561,7 @@ radv_emit_compute_pipeline(struct radv_cmd_buffer *cmd_buffer)
>>   {
>>          struct radv_shader_variant *compute_shader;
>>          struct radv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
>> +       struct radv_device *device = cmd_buffer->device;
>>          unsigned compute_resource_limits;
>>          unsigned waves_per_threadgroup;
>>          uint64_t va;
>> @@ -2602,6 +2603,19 @@ radv_emit_compute_pipeline(struct radv_cmd_buffer *cmd_buffer)
>>          compute_resource_limits =
>>                  S_00B854_SIMD_DEST_CNTL(waves_per_threadgroup % 4 == 0);
>>
>> +       if (device->physical_device->rad_info.chip_class >= CIK) {
>> +               unsigned num_cu_per_se =
>> +                       device->physical_device->rad_info.num_good_compute_units /
>> +                       device->physical_device->rad_info.max_se;
>> +
>> +               /* Force even distribution on all SIMDs in CU if the workgroup
>> +                * size is 64. This has shown some good improvements if # of
>> +                * CUs per SE is not a multiple of 4.
>> +                */
>> +               if (num_cu_per_se % 4 && waves_per_threadgroup == 1)
>> +                       compute_resource_limits |= S_00B854_FORCE_SIMD_DIST(1);
>> +       }
>> +
>>          radeon_set_sh_reg(cmd_buffer->cs, R_00B854_COMPUTE_RESOURCE_LIMITS,
>>                            compute_resource_limits);
>>
>> --
>> 2.15.1
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev