drm/amdgpu/gfx9: allocate queues horizontally across pipes
authorAlex Deucher <alexander.deucher@amd.com>
Mon, 5 Jun 2017 14:58:56 +0000 (10:58 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 6 Jun 2017 21:01:37 +0000 (17:01 -0400)
Pipes provide better concurrency than queues, therefore we want to make
sure that apps use queues from different pipes whenever possible.

Optimize for the trivial case where an app will consume rings in order,
therefore we don't want adjacent rings to belong to the same pipe.

gfx9 was missed when these patches were rebased.

Reviewed-by: Tom St Denis <tom.stdenis@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c

index 9502353ec325f02db8fe305bc447b2d70cfee538..0c48f6c1d7f54f53455b7df5e04b68984bc71b43 100644 (file)
@@ -1483,9 +1483,44 @@ static int gfx_v9_0_ngg_en(struct amdgpu_device *adev)
        return 0;
 }
 
+static int gfx_v9_0_compute_ring_init(struct amdgpu_device *adev, int ring_id,
+                                     int mec, int pipe, int queue)
+{
+       int r;
+       unsigned irq_type;
+       struct amdgpu_ring *ring = &adev->gfx.compute_ring[ring_id];
+
+       ring = &adev->gfx.compute_ring[ring_id];
+
+       /* mec0 is me1 */
+       ring->me = mec + 1;
+       ring->pipe = pipe;
+       ring->queue = queue;
+
+       ring->ring_obj = NULL;
+       ring->use_doorbell = true;
+       ring->doorbell_index = AMDGPU_DOORBELL_MEC_RING0 + ring_id;
+       ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr
+                               + (ring_id * GFX9_MEC_HPD_SIZE);
+       sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue);
+
+       irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP
+               + ((ring->me - 1) * adev->gfx.mec.num_pipe_per_mec)
+               + ring->pipe;
+
+       /* type-2 packets are deprecated on MEC, use type-3 instead */
+       r = amdgpu_ring_init(adev, ring, 1024,
+                            &adev->gfx.eop_irq, irq_type);
+       if (r)
+               return r;
+
+
+       return 0;
+}
+
 static int gfx_v9_0_sw_init(void *handle)
 {
-       int i, r, ring_id;
+       int i, j, k, r, ring_id;
        struct amdgpu_ring *ring;
        struct amdgpu_kiq *kiq;
        struct amdgpu_device *adev = (struct amdgpu_device *)handle;
@@ -1547,69 +1582,23 @@ static int gfx_v9_0_sw_init(void *handle)
                        return r;
        }
 
-       /* set up the compute queues */
-       for (i = 0, ring_id = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; i++) {
-               unsigned irq_type;
-
-               if (!test_bit(i, adev->gfx.mec.queue_bitmap))
-                       continue;
-
-               if (WARN_ON(ring_id >= AMDGPU_MAX_COMPUTE_RINGS))
-                       break;
-
-               ring = &adev->gfx.compute_ring[ring_id];
-
-               /* mec0 is me1 */
-               ring->me = ((i / adev->gfx.mec.num_queue_per_pipe)
-                               / adev->gfx.mec.num_pipe_per_mec)
-                               + 1;
-               ring->pipe = (i / adev->gfx.mec.num_queue_per_pipe)
-                               % adev->gfx.mec.num_pipe_per_mec;
-               ring->queue = i % adev->gfx.mec.num_queue_per_pipe;
-
-               ring->ring_obj = NULL;
-               ring->use_doorbell = true;
-               ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr + (ring_id * GFX9_MEC_HPD_SIZE);
-               ring->doorbell_index = AMDGPU_DOORBELL_MEC_RING0 + ring_id;
-               sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue);
-
-               irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP
-                       + ((ring->me - 1) * adev->gfx.mec.num_pipe_per_mec)
-                       + ring->pipe;
-
-               /* type-2 packets are deprecated on MEC, use type-3 instead */
-               r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq,
-                                    irq_type);
-               if (r)
-                       return r;
-
-               ring_id++;
-       }
-
-       /* set up the compute queues */
-       for (i = 0, ring_id = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; i++) {
-               unsigned irq_type;
-
-               /* max 32 queues per MEC */
-               if ((i >= 32) || (i >= AMDGPU_MAX_COMPUTE_RINGS)) {
-                       DRM_ERROR("Too many (%d) compute rings!\n", i);
-                       break;
+       /* set up the compute queues - allocate horizontally across pipes */
+       ring_id = 0;
+       for (i = 0; i < adev->gfx.mec.num_mec; ++i) {
+               for (j = 0; j < adev->gfx.mec.num_queue_per_pipe; j++) {
+                       for (k = 0; k < adev->gfx.mec.num_pipe_per_mec; k++) {
+                               if (!amdgpu_is_mec_queue_enabled(adev, i, k, j))
+                                       continue;
+
+                               r = gfx_v9_0_compute_ring_init(adev,
+                                                              ring_id,
+                                                              i, k, j);
+                               if (r)
+                                       return r;
+
+                               ring_id++;
+                       }
                }
-               ring = &adev->gfx.compute_ring[i];
-               ring->ring_obj = NULL;
-               ring->use_doorbell = true;
-               ring->doorbell_index = (AMDGPU_DOORBELL64_MEC_RING0 + i) << 1;
-               ring->me = 1; /* first MEC */
-               ring->pipe = i / 8;
-               ring->queue = i % 8;
-               ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr + (i * GFX9_MEC_HPD_SIZE);
-               sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue);
-               irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP + ring->pipe;
-               /* type-2 packets are deprecated on MEC, use type-3 instead */
-               r = amdgpu_ring_init(adev, ring, 1024,
-                                    &adev->gfx.eop_irq, irq_type);
-               if (r)
-                       return r;
        }
 
        r = gfx_v9_0_kiq_init(adev);