radv: set FORCE_SIMD_DIST(1) for compute when profitable
authorSamuel Pitoiset <samuel.pitoiset@gmail.com>
Thu, 14 Dec 2017 14:51:20 +0000 (15:51 +0100)
committerSamuel Pitoiset <samuel.pitoiset@gmail.com>
Thu, 14 Dec 2017 21:20:59 +0000 (22:20 +0100)
Ported from RadeonSI.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
src/amd/vulkan/radv_cmd_buffer.c

index d6aaff7..4a04848 100644 (file)
@@ -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);