AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ac: adjust maximum HS workgroup size

Browse Source 
This has no effect on triangles because max 64 patches implied max 192
threads, but it improves performance for cases when the number of threads
per patch is > 3.

This improves the score for gfxbench5 "gl_tess_off" (offscreen) by 11%
on Navi48.

Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34863>
This commit is contained in:
Marek Olšák
2025-04-30 22:31:22 -04:00
committed by Marge Bot
parent b960137ebf
commit 870d17012a
1 changed files with 4 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/common/ac_shader_util.c
+4 -7
View File
@@ -942,15 +942,12 @@ uint32_t ac_compute_num_tess_patches(const struct radeon_info *info, uint32_t nu
if (has_primid_instancing_bug && tess_uses_primid)
return 1;
/* 256 threads per workgroup is the hw limit. */
/* 256 threads per workgroup is the hw limit, but 192 performs better. */
const unsigned num_threads_per_patch = MAX2(num_tcs_input_cp, num_tcs_output_cp);
unsigned num_patches = 256 / num_threads_per_patch;
unsigned num_patches = 192 / num_threads_per_patch;
/* Not necessary for correctness, but higher numbers are slower.
* The hardware can do more, but we prefer fully occupied waves.
* eg. 64 triangle patches means 3 fully occupied Wave64 waves.
*/
num_patches = MIN2(num_patches, 64);
/* 127 is the maximum value that fits in tcs_offchip_layout. */
num_patches = MIN2(num_patches, 127);
/* When distributed tessellation is unsupported, switch between SEs
* at a higher frequency to manually balance the workload between SEs.