radv: set FORCE_SIMD_DIST(1) for compute when profitable
Ported from RadeonSI. Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This commit is contained in:
@@ -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);
|
||||
|
||||
|
||||
Reference in New Issue
Block a user