fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-03-18 14:30:37 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

broadcom/common: Optimize CSD super-group packing

Browse source 
Return one work group per super group when the work group size
is multiple of 16 (elements per batch) and recalculate max_wgs_per_sg
only when TSY barriers cut the available QPU threads.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37836>
This commit is contained in:
Daivik Bhatia 2025-09-05 19:35:36 +05:30 • committed by Iago Toral Quiroga
parent 099ac5be1a
commit 1326d52d23
1 changed files with 9 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
src/broadcom/common/v3d_util.c
View file

@ -41,6 +41,12 @@ v3d_csd_choose_workgroups_per_supergroup(struct v3d_device_info *devinfo,
if (has_subgroups)
return 1;
/* If the workgroup size is a multiple of 16 (elements per batch),
* the lane occupancy is already maximized.
*/
if (wg_size % 16 == 0)
return 1;
/* Compute maximum number of batches in a supergroup for this workgroup size.
* Each batch is 16 elements, and we can have up to 16 work groups in a
* supergroup:
@ -56,11 +62,13 @@ v3d_csd_choose_workgroups_per_supergroup(struct v3d_device_info *devinfo,
* available, so we can have at least 2 supergroups executing in parallel
* and we don't stall all our QPU threads when a supergroup hits a barrier.
*/
uint32_t max_wgs_per_sg = 16;
if (has_tsy_barrier) {
uint32_t max_qpu_threads = devinfo->qpu_count * threads;
max_batches_per_sg = MIN2(max_batches_per_sg, max_qpu_threads / 2);
max_wgs_per_sg = max_batches_per_sg * 16 / wg_size;
}
uint32_t max_wgs_per_sg = max_batches_per_sg * 16 / wg_size;
uint32_t best_wgs_per_sg = 1;
uint32_t best_unused_lanes = 16;