drm/amdgpu/gfx8: move SET_RESOURCES into the same command stream
authorAlex Deucher <alexander.deucher@amd.com>
Wed, 29 Mar 2017 19:31:22 +0000 (15:31 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 24 May 2017 21:39:50 +0000 (17:39 -0400)
As the KCQ setup.  This way we only have to wait once for the
entire MEC.

Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

index 23bd9a61161246029b9abe554d51d0d2925b27fc..47c417f2c9db38c58253e29b44899f8869ea5c7d 100644 (file)
@@ -4639,57 +4639,7 @@ static void gfx_v8_0_kiq_setting(struct amdgpu_ring *ring)
        WREG32(mmRLC_CP_SCHEDULERS, tmp);
 }
 
-static int gfx_v8_0_kiq_enable(struct amdgpu_ring *ring)
-{
-       struct amdgpu_device *adev = ring->adev;
-       uint32_t scratch, tmp = 0;
-       int r, i;
-
-       r = amdgpu_gfx_scratch_get(adev, &scratch);
-       if (r) {
-               DRM_ERROR("Failed to get scratch reg (%d).\n", r);
-               return r;
-       }
-       WREG32(scratch, 0xCAFEDEAD);
-
-       r = amdgpu_ring_alloc(ring, 11);
-       if (r) {
-               DRM_ERROR("Failed to lock KIQ (%d).\n", r);
-               amdgpu_gfx_scratch_free(adev, scratch);
-               return r;
-       }
-       /* set resources */
-       amdgpu_ring_write(ring, PACKET3(PACKET3_SET_RESOURCES, 6));
-       amdgpu_ring_write(ring, 0);     /* vmid_mask:0 queue_type:0 (KIQ) */
-       amdgpu_ring_write(ring, 0x000000FF);    /* queue mask lo */
-       amdgpu_ring_write(ring, 0);     /* queue mask hi */
-       amdgpu_ring_write(ring, 0);     /* gws mask lo */
-       amdgpu_ring_write(ring, 0);     /* gws mask hi */
-       amdgpu_ring_write(ring, 0);     /* oac mask */
-       amdgpu_ring_write(ring, 0);     /* gds heap base:0, gds heap size:0 */
-       /* write to scratch for completion */
-       amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1));
-       amdgpu_ring_write(ring, (scratch - PACKET3_SET_UCONFIG_REG_START));
-       amdgpu_ring_write(ring, 0xDEADBEEF);
-       amdgpu_ring_commit(ring);
-
-       for (i = 0; i < adev->usec_timeout; i++) {
-               tmp = RREG32(scratch);
-               if (tmp == 0xDEADBEEF)
-                       break;
-               DRM_UDELAY(1);
-       }
-       if (i >= adev->usec_timeout) {
-               DRM_ERROR("KIQ enable failed (scratch(0x%04X)=0x%08X)\n",
-                         scratch, tmp);
-               r = -EINVAL;
-       }
-       amdgpu_gfx_scratch_free(adev, scratch);
-
-       return r;
-}
-
-static int gfx_v8_0_map_queues_enable(struct amdgpu_device *adev)
+static int gfx_v8_0_kiq_kcq_enable(struct amdgpu_device *adev)
 {
        struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
        uint32_t scratch, tmp = 0;
@@ -4702,12 +4652,21 @@ static int gfx_v8_0_map_queues_enable(struct amdgpu_device *adev)
        }
        WREG32(scratch, 0xCAFEDEAD);
 
-       r = amdgpu_ring_alloc(kiq_ring, (8 * adev->gfx.num_compute_rings) + 3);
+       r = amdgpu_ring_alloc(kiq_ring, (8 * adev->gfx.num_compute_rings) + 11);
        if (r) {
                DRM_ERROR("Failed to lock KIQ (%d).\n", r);
                amdgpu_gfx_scratch_free(adev, scratch);
                return r;
        }
+       /* set resources */
+       amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_SET_RESOURCES, 6));
+       amdgpu_ring_write(kiq_ring, 0); /* vmid_mask:0 queue_type:0 (KIQ) */
+       amdgpu_ring_write(kiq_ring, 0x000000FF);        /* queue mask lo */
+       amdgpu_ring_write(kiq_ring, 0); /* queue mask hi */
+       amdgpu_ring_write(kiq_ring, 0); /* gws mask lo */
+       amdgpu_ring_write(kiq_ring, 0); /* gws mask hi */
+       amdgpu_ring_write(kiq_ring, 0); /* oac mask */
+       amdgpu_ring_write(kiq_ring, 0); /* gds heap base:0, gds heap size:0 */
        for (i = 0; i < adev->gfx.num_compute_rings; i++) {
                struct amdgpu_ring *ring = &adev->gfx.compute_ring[i];
                uint64_t mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
@@ -4969,7 +4928,6 @@ static int gfx_v8_0_kiq_init_queue(struct amdgpu_ring *ring)
        struct amdgpu_device *adev = ring->adev;
        struct vi_mqd *mqd = ring->mqd_ptr;
        int mqd_idx = AMDGPU_MAX_COMPUTE_RINGS;
-       int r;
 
        gfx_v8_0_kiq_setting(ring);
 
@@ -5000,9 +4958,7 @@ static int gfx_v8_0_kiq_init_queue(struct amdgpu_ring *ring)
                mutex_unlock(&adev->srbm_mutex);
        }
 
-       r = gfx_v8_0_kiq_enable(ring);
-
-       return r;
+       return 0;
 }
 
 static int gfx_v8_0_kcq_init_queue(struct amdgpu_ring *ring)
@@ -5057,13 +5013,6 @@ static int gfx_v8_0_kiq_resume(struct amdgpu_device *adev)
        if (r)
                goto done;
 
-       ring->ready = true;
-       r = amdgpu_ring_test_ring(ring);
-       if (r) {
-               ring->ready = false;
-               goto done;
-       }
-
        for (i = 0; i < adev->gfx.num_compute_rings; i++) {
                ring = &adev->gfx.compute_ring[i];
 
@@ -5081,10 +5030,20 @@ static int gfx_v8_0_kiq_resume(struct amdgpu_device *adev)
                        goto done;
        }
 
-       r = gfx_v8_0_map_queues_enable(adev);
+       r = gfx_v8_0_kiq_kcq_enable(adev);
        if (r)
                goto done;
 
+       /* Test KIQ */
+       ring = &adev->gfx.kiq.ring;
+       ring->ready = true;
+       r = amdgpu_ring_test_ring(ring);
+       if (r) {
+               ring->ready = false;
+               goto done;
+       }
+
+       /* Test KCQs */
        for (i = 0; i < adev->gfx.num_compute_rings; i++) {
                ring = &adev->gfx.compute_ring[i];
                ring->ready = true;