drm/amdgpu: use scheduler load balancing for SDMA CS
authorChristian König <christian.koenig@amd.com>
Fri, 13 Jul 2018 07:12:44 +0000 (09:12 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Mon, 27 Aug 2018 16:10:45 +0000 (11:10 -0500)
Start to use the scheduler load balancing for userspace SDMA
command submissions.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c

index 02d563c..3ff8042 100644 (file)
@@ -48,7 +48,8 @@ static int amdgpu_ctx_init(struct amdgpu_device *adev,
                           struct drm_file *filp,
                           struct amdgpu_ctx *ctx)
 {
-       unsigned i, j;
+       struct drm_sched_rq *sdma_rqs[AMDGPU_MAX_RINGS];
+       unsigned i, j, num_sdma_rqs;
        int r;
 
        if (priority < 0 || priority >= DRM_SCHED_PRIORITY_MAX)
@@ -80,18 +81,34 @@ static int amdgpu_ctx_init(struct amdgpu_device *adev,
        ctx->init_priority = priority;
        ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
 
-       /* create context entity for each ring */
+       num_sdma_rqs = 0;
        for (i = 0; i < adev->num_rings; i++) {
                struct amdgpu_ring *ring = adev->rings[i];
                struct drm_sched_rq *rq;
 
                rq = &ring->sched.sched_rq[priority];
+               if (ring->funcs->type == AMDGPU_RING_TYPE_SDMA)
+                       sdma_rqs[num_sdma_rqs++] = rq;
+       }
+
+       /* create context entity for each ring */
+       for (i = 0; i < adev->num_rings; i++) {
+               struct amdgpu_ring *ring = adev->rings[i];
 
                if (ring == &adev->gfx.kiq.ring)
                        continue;
 
-               r = drm_sched_entity_init(&ctx->rings[i].entity,
-                                         &rq, 1, &ctx->guilty);
+               if (ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
+                       r = drm_sched_entity_init(&ctx->rings[i].entity,
+                                                 sdma_rqs, num_sdma_rqs,
+                                                 &ctx->guilty);
+               } else {
+                       struct drm_sched_rq *rq;
+
+                       rq = &ring->sched.sched_rq[priority];
+                       r = drm_sched_entity_init(&ctx->rings[i].entity,
+                                                 &rq, 1, &ctx->guilty);
+               }
                if (r)
                        goto failed;
        }