fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-06 20:18:12 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

winsys/amdgpu: call userq wait ioctl only once
Some checks are pending
macOS-CI / macOS-CI (dri) (push) Waiting to run
Details
macOS-CI / macOS-CI (xlib) (push) Waiting to run
Details

Browse source 
Allocate to hold 256 wait fences. Since there is only one queue per
per ip per process, the idea is that there won't be app or windowing system
that would have large number of job dependencies / wait fences.

If there is an app that has wait fences greater than 256, there won't
be corruption issues since kernel will wait for the extra fences.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/40698>
This commit is contained in:
Yogesh Mohan Marimuthu 2026-03-30 12:45:21 +05:30 • committed by Marge Bot
parent 5df6fc5481
commit 21047e61a5
1 changed files with 5 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

16
src/gallium/winsys/amdgpu/drm/amdgpu_cs.cpp
View file

@ -1657,7 +1657,9 @@ static int amdgpu_cs_submit_ib_userq(struct amdgpu_userq *userq,
}
syncobj_signal_list[num_syncobj_to_signal - 1] = ((struct amdgpu_fence*)csc->fence)->syncobj;
struct drm_amdgpu_userq_fence_info *fence_info;
uint16_t num_wait_fences = 256;
struct drm_amdgpu_userq_fence_info *fence_info = (struct drm_amdgpu_userq_fence_info*)
alloca(num_wait_fences * sizeof(struct drm_amdgpu_userq_fence_info));
struct drm_amdgpu_userq_wait userq_wait_data = {
.waitq_id = userq->userq_handle,
.syncobj_handles = (uintptr_t)syncobj_dependencies_list,
@ -1668,11 +1670,11 @@ static int amdgpu_cs_submit_ib_userq(struct amdgpu_userq *userq,
/* Wait for previous writes to complete before reading from these BOs. */
.bo_write_handles = num_shared_buf_read ? (uintptr_t)shared_buf_kms_handles_read : 0,
.num_syncobj_timeline_handles = num_syncobj_timeline_dependencies,
.num_fences = 0,
.num_fences = num_wait_fences,
.num_syncobj_handles = num_syncobj_dependencies,
.num_bo_read_handles = num_shared_buf_write,
.num_bo_write_handles = num_shared_buf_read,
.out_fences = (uintptr_t)NULL,
.out_fences = (uintptr_t)fence_info,
};
/*
@ -1682,14 +1684,6 @@ static int amdgpu_cs_submit_ib_userq(struct amdgpu_userq *userq,
* To implement this strategy, we use amdgpu_userq_wait() before submitting
* a job, and amdgpu_userq_signal() after to indicate completion.
*/
r = ac_drm_userq_wait(aws->dev, &userq_wait_data);
if (r)
mesa_loge("amdgpu: getting wait num_fences failed\n");
fence_info = (struct drm_amdgpu_userq_fence_info*)
alloca(userq_wait_data.num_fences * sizeof(struct drm_amdgpu_userq_fence_info));
userq_wait_data.out_fences = (uintptr_t)fence_info;
r = ac_drm_userq_wait(aws->dev, &userq_wait_data);
if (r)
mesa_loge("amdgpu: getting wait fences failed\n");