[Mesa-dev] [PATCH 3/6] winsys/amdgpu: use a separate fence list for syncobjs
Marek Olšák
maraeo at gmail.com
Wed Feb 6 21:20:45 UTC 2019
From: Marek Olšák <marek.olsak at amd.com>
---
src/gallium/winsys/amdgpu/drm/amdgpu_cs.c | 31 ++++++++++-------------
src/gallium/winsys/amdgpu/drm/amdgpu_cs.h | 1 +
2 files changed, 15 insertions(+), 17 deletions(-)
diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
index b202ce21cf9..a059a3958d4 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
@@ -884,38 +884,40 @@ static void amdgpu_cs_context_cleanup(struct amdgpu_cs_context *cs)
}
for (i = 0; i < cs->num_slab_buffers; i++) {
p_atomic_dec(&cs->slab_buffers[i].bo->num_cs_references);
amdgpu_winsys_bo_reference(&cs->slab_buffers[i].bo, NULL);
}
for (i = 0; i < cs->num_sparse_buffers; i++) {
p_atomic_dec(&cs->sparse_buffers[i].bo->num_cs_references);
amdgpu_winsys_bo_reference(&cs->sparse_buffers[i].bo, NULL);
}
cleanup_fence_list(&cs->fence_dependencies);
+ cleanup_fence_list(&cs->syncobj_dependencies);
cleanup_fence_list(&cs->syncobj_to_signal);
cs->num_real_buffers = 0;
cs->num_slab_buffers = 0;
cs->num_sparse_buffers = 0;
amdgpu_fence_reference(&cs->fence, NULL);
memset(cs->buffer_indices_hashlist, -1, sizeof(cs->buffer_indices_hashlist));
cs->last_added_bo = NULL;
}
static void amdgpu_destroy_cs_context(struct amdgpu_cs_context *cs)
{
amdgpu_cs_context_cleanup(cs);
FREE(cs->real_buffers);
FREE(cs->slab_buffers);
FREE(cs->sparse_buffers);
FREE(cs->fence_dependencies.list);
+ FREE(cs->syncobj_dependencies.list);
FREE(cs->syncobj_to_signal.list);
}
static struct radeon_cmdbuf *
amdgpu_cs_create(struct radeon_winsys_ctx *rwctx,
enum ring_type ring_type,
void (*flush)(void *ctx, unsigned flags,
struct pipe_fence_handle **fence),
void *flush_ctx,
@@ -1111,21 +1113,24 @@ static void amdgpu_cs_add_fence_dependency(struct radeon_cmdbuf *rws,
{
struct amdgpu_cs *acs = amdgpu_cs(rws);
struct amdgpu_cs_context *cs = acs->csc;
struct amdgpu_fence *fence = (struct amdgpu_fence*)pfence;
util_queue_fence_wait(&fence->submitted);
if (is_noop_fence_dependency(acs, fence))
return;
- add_fence_to_list(&cs->fence_dependencies, fence);
+ if (amdgpu_fence_is_syncobj(fence))
+ add_fence_to_list(&cs->syncobj_dependencies, fence);
+ else
+ add_fence_to_list(&cs->fence_dependencies, fence);
}
static void amdgpu_add_bo_fence_dependencies(struct amdgpu_cs *acs,
struct amdgpu_cs_buffer *buffer)
{
struct amdgpu_cs_context *cs = acs->csc;
struct amdgpu_winsys_bo *bo = buffer->bo;
unsigned new_num_fences = 0;
for (unsigned j = 0; j < bo->num_fences; ++j) {
@@ -1360,67 +1365,59 @@ void amdgpu_cs_submit_ib(void *job, int thread_index)
num_chunks++;
/* Fence */
if (has_user_fence) {
chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_FENCE;
chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_fence) / 4;
chunks[num_chunks].chunk_data = (uintptr_t)&acs->fence_chunk;
num_chunks++;
}
- /* Dependencies */
+ /* Fence dependencies. */
unsigned num_dependencies = cs->fence_dependencies.num;
- unsigned num_syncobj_dependencies = 0;
-
if (num_dependencies) {
struct drm_amdgpu_cs_chunk_dep *dep_chunk =
alloca(num_dependencies * sizeof(*dep_chunk));
- unsigned num = 0;
for (unsigned i = 0; i < num_dependencies; i++) {
struct amdgpu_fence *fence =
(struct amdgpu_fence*)cs->fence_dependencies.list[i];
- if (amdgpu_fence_is_syncobj(fence)) {
- num_syncobj_dependencies++;
- continue;
- }
-
assert(util_queue_fence_is_signalled(&fence->submitted));
- amdgpu_cs_chunk_fence_to_dep(&fence->fence, &dep_chunk[num++]);
+ amdgpu_cs_chunk_fence_to_dep(&fence->fence, &dep_chunk[i]);
}
chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_DEPENDENCIES;
- chunks[num_chunks].length_dw = sizeof(dep_chunk[0]) / 4 * num;
+ chunks[num_chunks].length_dw = sizeof(dep_chunk[0]) / 4 * num_dependencies;
chunks[num_chunks].chunk_data = (uintptr_t)dep_chunk;
num_chunks++;
}
/* Syncobj dependencies. */
+ unsigned num_syncobj_dependencies = cs->syncobj_dependencies.num;
if (num_syncobj_dependencies) {
struct drm_amdgpu_cs_chunk_sem *sem_chunk =
alloca(num_syncobj_dependencies * sizeof(sem_chunk[0]));
- unsigned num = 0;
- for (unsigned i = 0; i < num_dependencies; i++) {
+ for (unsigned i = 0; i < num_syncobj_dependencies; i++) {
struct amdgpu_fence *fence =
- (struct amdgpu_fence*)cs->fence_dependencies.list[i];
+ (struct amdgpu_fence*)cs->syncobj_dependencies.list[i];
if (!amdgpu_fence_is_syncobj(fence))
continue;
assert(util_queue_fence_is_signalled(&fence->submitted));
- sem_chunk[num++].handle = fence->syncobj;
+ sem_chunk[i].handle = fence->syncobj;
}
chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_SYNCOBJ_IN;
- chunks[num_chunks].length_dw = sizeof(sem_chunk[0]) / 4 * num;
+ chunks[num_chunks].length_dw = sizeof(sem_chunk[0]) / 4 * num_syncobj_dependencies;
chunks[num_chunks].chunk_data = (uintptr_t)sem_chunk;
num_chunks++;
}
/* Syncobj signals. */
unsigned num_syncobj_to_signal = cs->syncobj_to_signal.num;
if (num_syncobj_to_signal) {
struct drm_amdgpu_cs_chunk_sem *sem_chunk =
alloca(num_syncobj_to_signal * sizeof(sem_chunk[0]));
diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.h b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.h
index 60b05910c71..17c0581b6f4 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_cs.h
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_cs.h
@@ -96,20 +96,21 @@ struct amdgpu_cs_context {
struct amdgpu_cs_buffer *sparse_buffers;
int buffer_indices_hashlist[4096];
struct amdgpu_winsys_bo *last_added_bo;
unsigned last_added_bo_index;
unsigned last_added_bo_usage;
uint32_t last_added_bo_priority_usage;
struct amdgpu_fence_list fence_dependencies;
+ struct amdgpu_fence_list syncobj_dependencies;
struct amdgpu_fence_list syncobj_to_signal;
struct pipe_fence_handle *fence;
/* the error returned from cs_flush for non-async submissions */
int error_code;
};
struct amdgpu_cs {
struct amdgpu_ib main; /* must be first because this is inherited */
--
2.17.1
More information about the mesa-dev
mailing list