[Mesa-dev] [PATCH 4/6] gallium/radeon: fix performance of buffer readbacks

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

We want cached GTT for all non-persistent read mappings.
Set level = 0 on purpose.

Use dma_copy, because resource_copy_region causes a failure in the PBO
read of piglit/getteximage-luminance.

If Rocket League used the READ flag, it should get cached GTT.

v2: mask out UNSYNCHRONIZED

Cc: 13.0 17.0 <mesa-stable at lists.freedesktop.org>
---
 src/gallium/drivers/radeon/r600_buffer_common.c | 17 +++++++++--------
 1 file changed, 9 insertions(+), 8 deletions(-)

diff --git a/src/gallium/drivers/radeon/r600_buffer_common.c b/src/gallium/drivers/radeon/r600_buffer_common.c
index da6f020..762181f 100644
--- a/src/gallium/drivers/radeon/r600_buffer_common.c
+++ b/src/gallium/drivers/radeon/r600_buffer_common.c
@@ -374,38 +374,39 @@ static void *r600_buffer_transfer_map(struct pipe_context *ctx,
 			if (staging) {
 				data += box->x % R600_MAP_BUFFER_ALIGNMENT;
 				return r600_buffer_get_transfer(ctx, resource, level, usage, box,
 								ptransfer, data, staging, offset);
 			}
 		} else {
 			/* At this point, the buffer is always idle (we checked it above). */
 			usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
 		}
 	}
-	/* Using a staging buffer in GTT for larger reads is much faster. */
+	/* Use a staging buffer in cached GTT for reads. */
 	else if ((usage & PIPE_TRANSFER_READ) &&
-		 !(usage & (PIPE_TRANSFER_WRITE |
-			    PIPE_TRANSFER_PERSISTENT)) &&
-		 rbuffer->domains & RADEON_DOMAIN_VRAM &&
+		 !(usage & PIPE_TRANSFER_PERSISTENT) &&
+		 (rbuffer->domains & RADEON_DOMAIN_VRAM ||
+		  rbuffer->flags & RADEON_FLAG_GTT_WC) &&
 		 r600_can_dma_copy_buffer(rctx, 0, box->x, box->width)) {
 		struct r600_resource *staging;
 
 		staging = (struct r600_resource*) pipe_buffer_create(
 				ctx->screen, 0, PIPE_USAGE_STAGING,
 				box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT));
 		if (staging) {
 			/* Copy the VRAM buffer to the staging buffer. */
-			ctx->resource_copy_region(ctx, &staging->b.b, 0,
-						  box->x % R600_MAP_BUFFER_ALIGNMENT,
-						  0, 0, resource, level, box);
+			rctx->dma_copy(ctx, &staging->b.b, 0,
+				       box->x % R600_MAP_BUFFER_ALIGNMENT,
+				       0, 0, resource, 0, box);
 
-			data = r600_buffer_map_sync_with_rings(rctx, staging, PIPE_TRANSFER_READ);
+			data = r600_buffer_map_sync_with_rings(rctx, staging,
+							       usage & ~PIPE_TRANSFER_UNSYNCHRONIZED);
 			if (!data) {
 				r600_resource_reference(&staging, NULL);
 				return NULL;
 			}
 			data += box->x % R600_MAP_BUFFER_ALIGNMENT;
 
 			return r600_buffer_get_transfer(ctx, resource, level, usage, box,
 							ptransfer, data, staging, 0);
 		}
 	}
-- 
2.7.4