drm/amdgpu: fix dependency issue

The problem is that executing the jobs in the right order doesn't give you the right result
because consecutive jobs executed on the same engine are pipelined.
In other words job B does it buffer read before job A has written it's result.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu.h

@@ -1129,6 +1129,7 @@ struct amdgpu_job {
 	void			*owner;
 	uint64_t		fence_ctx; /* the fence_context this job uses */
 	bool                    vm_needs_flush;
+	bool			need_pipeline_sync;
 	unsigned		vm_id;
 	uint64_t		vm_pd_addr;
 	uint32_t		gds_base, gds_size;

drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c

@@ -160,6 +160,8 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
 		return r;
 	}
+	if (ring->funcs->emit_pipeline_sync && job && job->need_pipeline_sync)
+		amdgpu_ring_emit_pipeline_sync(ring);
 
 	if (vm) {
 		r = amdgpu_vm_flush(ring, job);

drivers/gpu/drm/amd/amdgpu/amdgpu_job.c

@@ -57,6 +57,7 @@ int amdgpu_job_alloc(struct amdgpu_device *adev, unsigned num_ibs,
 	(*job)->vm = vm;
 	(*job)->ibs = (void *)&(*job)[1];
 	(*job)->num_ibs = num_ibs;
+	(*job)->need_pipeline_sync = false;
 
 	amdgpu_sync_create(&(*job)->sync);
 
@@ -152,6 +153,9 @@ static struct dma_fence *amdgpu_job_dependency(struct amd_sched_job *sched_job)
 		fence = amdgpu_sync_get_fence(&job->sync);
 	}
 
+	if (amd_sched_dependency_optimized(fence, sched_job->s_entity))
+		job->need_pipeline_sync = true;
+
 	return fence;
 }
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

@@ -614,7 +614,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
 	if (ring->funcs->init_cond_exec)
 		patch_offset = amdgpu_ring_init_cond_exec(ring);
 
-	if (ring->funcs->emit_pipeline_sync)
+	if (ring->funcs->emit_pipeline_sync && !job->need_pipeline_sync)
 		amdgpu_ring_emit_pipeline_sync(ring);
 
 	if (ring->funcs->emit_vm_flush && vm_flush_needed) {

drivers/gpu/drm/amd/scheduler/gpu_scheduler.c

@@ -236,6 +236,23 @@ static void amd_sched_entity_clear_dep(struct dma_fence *f, struct dma_fence_cb
 	dma_fence_put(f);
 }
 
+bool amd_sched_dependency_optimized(struct dma_fence* fence,
+				    struct amd_sched_entity *entity)
+{
+	struct amd_gpu_scheduler *sched = entity->sched;
+	struct amd_sched_fence *s_fence;
+
+	if (!fence || dma_fence_is_signaled(fence))
+		return false;
+	if (fence->context == entity->fence_context)
+		return true;
+	s_fence = to_amd_sched_fence(fence);
+	if (s_fence && s_fence->sched == sched)
+		return true;
+
+	return false;
+}
+
 static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
 {
 	struct amd_gpu_scheduler *sched = entity->sched;

drivers/gpu/drm/amd/scheduler/gpu_scheduler.h

@@ -158,4 +158,6 @@ int amd_sched_job_init(struct amd_sched_job *job,
 		       void *owner);
 void amd_sched_hw_job_reset(struct amd_gpu_scheduler *sched);
 void amd_sched_job_recovery(struct amd_gpu_scheduler *sched);
+bool amd_sched_dependency_optimized(struct dma_fence* fence,
+				    struct amd_sched_entity *entity);
 #endif