drm/radeon: fix up ring functions for compute rings
authorAlex Deucher <alexander.deucher@amd.com>
Thu, 11 Apr 2013 13:36:17 +0000 (09:36 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Thu, 27 Jun 2013 14:49:09 +0000 (10:49 -0400)
The compute rings use RELEASE_MEM rather then EOP
packets for writing fences and there is no SYNC_PFP_ME
packet on the compute rings.

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/radeon/cik.c

index 19a6b3c31304e6e5979ddd76124da303c73564cd..fa4c9acc2f4b109be5f16ffd7838aea32fb89098 100644 (file)
@@ -1706,7 +1706,7 @@ int cik_ring_test(struct radeon_device *rdev, struct radeon_ring *ring)
 }
 
 /**
- * cik_fence_ring_emit - emit a fence on the gfx ring
+ * cik_fence_gfx_ring_emit - emit a fence on the gfx ring
  *
  * @rdev: radeon_device pointer
  * @fence: radeon fence object
@@ -1714,8 +1714,8 @@ int cik_ring_test(struct radeon_device *rdev, struct radeon_ring *ring)
  * Emits a fence sequnce number on the gfx ring and flushes
  * GPU caches.
  */
-void cik_fence_ring_emit(struct radeon_device *rdev,
-                        struct radeon_fence *fence)
+void cik_fence_gfx_ring_emit(struct radeon_device *rdev,
+                            struct radeon_fence *fence)
 {
        struct radeon_ring *ring = &rdev->ring[fence->ring];
        u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
@@ -1742,6 +1742,44 @@ void cik_fence_ring_emit(struct radeon_device *rdev,
        radeon_ring_write(ring, 0);
 }
 
+/**
+ * cik_fence_compute_ring_emit - emit a fence on the compute ring
+ *
+ * @rdev: radeon_device pointer
+ * @fence: radeon fence object
+ *
+ * Emits a fence sequnce number on the compute ring and flushes
+ * GPU caches.
+ */
+void cik_fence_compute_ring_emit(struct radeon_device *rdev,
+                                struct radeon_fence *fence)
+{
+       struct radeon_ring *ring = &rdev->ring[fence->ring];
+       u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
+
+       /* RELEASE_MEM - flush caches, send int */
+       radeon_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 5));
+       radeon_ring_write(ring, (EOP_TCL1_ACTION_EN |
+                                EOP_TC_ACTION_EN |
+                                EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
+                                EVENT_INDEX(5)));
+       radeon_ring_write(ring, DATA_SEL(1) | INT_SEL(2));
+       radeon_ring_write(ring, addr & 0xfffffffc);
+       radeon_ring_write(ring, upper_32_bits(addr));
+       radeon_ring_write(ring, fence->seq);
+       radeon_ring_write(ring, 0);
+       /* HDP flush */
+       /* We should be using the new WAIT_REG_MEM special op packet here
+        * but it causes the CP to hang
+        */
+       radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
+       radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
+                                WRITE_DATA_DST_SEL(0)));
+       radeon_ring_write(ring, HDP_MEM_COHERENCY_FLUSH_CNTL >> 2);
+       radeon_ring_write(ring, 0);
+       radeon_ring_write(ring, 0);
+}
+
 void cik_semaphore_ring_emit(struct radeon_device *rdev,
                             struct radeon_ring *ring,
                             struct radeon_semaphore *semaphore,
@@ -4051,9 +4089,12 @@ void cik_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
        radeon_ring_write(ring, 0);
        radeon_ring_write(ring, 1 << vm->id);
 
-       /* sync PFP to ME, otherwise we might get invalid PFP reads */
-       radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
-       radeon_ring_write(ring, 0x0);
+       /* compute doesn't have PFP */
+       if (ridx == RADEON_RING_TYPE_GFX_INDEX) {
+               /* sync PFP to ME, otherwise we might get invalid PFP reads */
+               radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
+               radeon_ring_write(ring, 0x0);
+       }
 }
 
 /**