fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-06 02:58:05 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

ac/gpu_info: reduce the tess offchip ring size and compute it proportionately

Browse source 
.. to the CU count. We allocated too much.

This reduces the tess offchip ring size as follows (examples):
- GFX11-12:
  - Navi31, Navi33, and Navi48 get 75% decrease.
  - Navi32 gets 68.75% decrease.
  - Phoenix gets 81.25% decrease.
  - Phoenix2 gets 93.75% decrease.
- GFX10.3:
  - Navi21 and Navi22 get 37.5% decrease.
  - Navi23 and Navi24 get 50% decrease.
  - Rembrandt gets 62.5% decrease.
  - VanGogh gets 75% decrease.
  - Raphael gets 93.75% decrease.
- GFX8-9:
  - Vega10 gets 0% decrease.
  - Vega20 gets 49.6% decrease.
  - Raven gets 65.3% decrease.
  - Raven2 gets 93.7% decrease.
  - Stoney gets 81% decrease.

No difference in performance was measured.

Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34544>
This commit is contained in:
Marek Olšák 2025-04-15 23:11:36 -04:00
parent 9333c0a1ed
commit 6d9e708642
1 changed files with 11 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

16
src/amd/common/ac_gpu_info.c
View file

@ -1741,7 +1741,7 @@ bool ac_query_gpu_info(int fd, void *dev_p, struct radeon_info *info,
*/
unsigned wg_size_in_dwords = info->family == CHIP_HAWAII ? 4096 : 8192;
unsigned wg_size_enum;
unsigned num_workgroups_per_se;
unsigned max_workgroups_per_se;
switch (wg_size_in_dwords) {
case 8192:
@ -1765,15 +1765,21 @@ bool ac_query_gpu_info(int fd, void *dev_p, struct radeon_info *info,
* Gfx6 should limit num_workgroups to 126 (63 per SE)
*/
if (info->gfx_level >= GFX11) {
num_workgroups_per_se = 256;
max_workgroups_per_se = 256;
} else if (info->gfx_level >= GFX10) {
num_workgroups_per_se = 128;
max_workgroups_per_se = 128;
} else if (info->family == CHIP_VEGA12 || info->family == CHIP_VEGA20) {
num_workgroups_per_se = double_offchip_wg ? 128 : 64;
max_workgroups_per_se = double_offchip_wg ? 128 : 64;
} else {
num_workgroups_per_se = double_offchip_wg ? 127 : 63;
max_workgroups_per_se = double_offchip_wg ? 127 : 63;
}
/* Limit to 4 workgroups per CU for TCS, which exhausts LDS if each workgroup occupies 16KB.
* Note that the offchip allocation isn't deallocated until the corresponding TES waves finish.
*/
unsigned num_offchip_wg_per_cu = 4;
unsigned num_workgroups_per_se = MIN2(num_offchip_wg_per_cu * info->max_good_cu_per_sa *
info->max_sa_per_se, max_workgroups_per_se);
unsigned num_workgroups = num_workgroups_per_se * info->max_se;
if (info->gfx_level >= GFX11) {