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

winsys/amdgpu: do not iterate twice when adding fence dependencies 

The perf difference is very small, 3.25->2.84% in amdgpu_cs_flush()
in the DXMD benchmark.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This commit is contained in:
Samuel Pitoiset 2017-01-27 14:35:23 +01:00
parent 5a6b1aadea
commit 86eb52adad
1 changed files with 32 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

63
src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
View File

@ -941,22 +941,6 @@ static void amdgpu_add_fence_dependency(struct amdgpu_cs *acs,
bo->num_fences = new_num_fences;
}
/* Since the kernel driver doesn't synchronize execution between different
* rings automatically, we have to add fence dependencies manually.
*/
static void amdgpu_add_fence_dependencies(struct amdgpu_cs *acs)
{
struct amdgpu_cs_context *cs = acs->csc;
int i;
cs->request.number_of_dependencies = 0;
for (i = 0; i < cs->num_real_buffers; i++)
amdgpu_add_fence_dependency(acs, &cs->real_buffers[i]);
for (i = 0; i < cs->num_slab_buffers; i++)
amdgpu_add_fence_dependency(acs, &cs->slab_buffers[i]);
}
static void amdgpu_add_fence(struct amdgpu_winsys_bo *bo,
struct pipe_fence_handle *fence)
{
@ -984,6 +968,38 @@ static void amdgpu_add_fence(struct amdgpu_winsys_bo *bo,
bo->num_fences++;
}
/* Since the kernel driver doesn't synchronize execution between different
* rings automatically, we have to add fence dependencies manually.
*/
static void amdgpu_add_fence_dependencies(struct amdgpu_cs *acs)
{
struct amdgpu_cs_context *cs = acs->csc;
unsigned num_buffers;
int i;
cs->request.number_of_dependencies = 0;
num_buffers = cs->num_real_buffers;
for (i = 0; i < num_buffers; i++) {
struct amdgpu_cs_buffer *buffer = &cs->real_buffers[i];
struct amdgpu_winsys_bo *bo = buffer->bo;
amdgpu_add_fence_dependency(acs, buffer);
p_atomic_inc(&bo->num_active_ioctls);
amdgpu_add_fence(bo, cs->fence);
}
num_buffers = cs->num_slab_buffers;
for (i = 0; i < num_buffers; i++) {
struct amdgpu_cs_buffer *buffer = &cs->slab_buffers[i];
struct amdgpu_winsys_bo *bo = buffer->bo;
amdgpu_add_fence_dependency(acs, buffer);
p_atomic_inc(&bo->num_active_ioctls);
amdgpu_add_fence(bo, cs->fence);
}
}
void amdgpu_cs_submit_ib(void *job, int thread_index)
{
struct amdgpu_cs *acs = (struct amdgpu_cs*)job;
@ -1146,7 +1162,6 @@ static int amdgpu_cs_flush(struct radeon_winsys_cs *rcs,
cs->main.base.current.cdw <= cs->main.base.current.max_dw &&
!debug_get_option_noop())) {
struct amdgpu_cs_context *cur = cs->csc;
unsigned i, num_buffers;
/* Set IB sizes. */
amdgpu_ib_finalize(&cs->main);
@ -1183,20 +1198,6 @@ static int amdgpu_cs_flush(struct radeon_winsys_cs *rcs,
pipe_mutex_lock(ws->bo_fence_lock);
amdgpu_add_fence_dependencies(cs);
num_buffers = cur->num_real_buffers;
for (i = 0; i < num_buffers; i++) {
struct amdgpu_winsys_bo *bo = cur->real_buffers[i].bo;
p_atomic_inc(&bo->num_active_ioctls);
amdgpu_add_fence(bo, cur->fence);
}
num_buffers = cur->num_slab_buffers;
for (i = 0; i < num_buffers; i++) {
struct amdgpu_winsys_bo *bo = cur->slab_buffers[i].bo;
p_atomic_inc(&bo->num_active_ioctls);
amdgpu_add_fence(bo, cur->fence);
}
/* Swap command streams. "cst" is going to be submitted. */
cs->csc = cs->cst;
cs->cst = cur;