r600g: implement wait-free buffer transfer for DISCARD_RANGE

Reviewed-by: Alex Deucher <[email protected]>
Reviewed-by: Christian König <[email protected]>

4 files changed, 50 insertions, 16 deletions

diff --git a/src/gallium/drivers/r600/r600_blit.c b/src/gallium/drivers/r600/r600_blit.c
index 829d979d2e4..8cdfae8b3f0 100644
--- a/src/gallium/drivers/r600/r600_blit.c
+++ b/src/gallium/drivers/r600/r600_blit.c
@@ -259,11 +259,9 @@ static void r600_clear_depth_stencil(struct pipe_context *ctx,
 	r600_blitter_end(ctx);
 }
 
-static void r600_copy_buffer(struct pipe_context *ctx,
-			     struct pipe_resource *dst,
-			     unsigned dstx,
-			     struct pipe_resource *src,
-			     const struct pipe_box *src_box)
+void r600_copy_buffer(struct pipe_context *ctx, struct
+		      pipe_resource *dst, unsigned dstx,
+		      struct pipe_resource *src, const struct pipe_box *src_box)
 {
 	struct r600_context *rctx = (struct r600_context*)ctx;
 
diff --git a/src/gallium/drivers/r600/r600_buffer.c b/src/gallium/drivers/r600/r600_buffer.c
index 165427eddf4..27dddcc1cf7 100644
--- a/src/gallium/drivers/r600/r600_buffer.c
+++ b/src/gallium/drivers/r600/r600_buffer.c
@@ -44,20 +44,24 @@ static struct pipe_transfer *r600_get_transfer(struct pipe_context *ctx,
 					       const struct pipe_box *box)
 {
 	struct r600_context *rctx = (struct r600_context*)ctx;
-	struct pipe_transfer *transfer = util_slab_alloc(&rctx->pool_transfers);
+	struct r600_transfer *transfer = util_slab_alloc(&rctx->pool_transfers);
 
-	transfer->resource = resource;
-	transfer->level = level;
-	transfer->usage = usage;
-	transfer->box = *box;
-	transfer->stride = 0;
-	transfer->layer_stride = 0;
-	transfer->data = NULL;
+	assert(box->x + box->width <= resource->width0);
+
+	transfer->transfer.resource = resource;
+	transfer->transfer.level = level;
+	transfer->transfer.usage = usage;
+	transfer->transfer.box = *box;
+	transfer->transfer.stride = 0;
+	transfer->transfer.layer_stride = 0;
+	transfer->transfer.data = NULL;
+	transfer->staging = NULL;
+	transfer->offset = 0;
 
 	/* Note strides are zero, this is ok for buffers, but not for
 	 * textures 2d & higher at least.
 	 */
-	return transfer;
+	return &transfer->transfer;
 }
 
 static void r600_set_constants_dirty_if_bound(struct r600_context *rctx,
@@ -126,6 +130,25 @@ static void *r600_buffer_transfer_map(struct pipe_context *pipe,
 			r600_set_constants_dirty_if_bound(rctx, &rctx->ps_constbuf_state, rbuffer);
 		}
 	}
+	else if ((transfer->usage & PIPE_TRANSFER_DISCARD_RANGE) &&
+		 !(transfer->usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
+		 rctx->screen->has_streamout &&
+		 /* The buffer range must be aligned to 4. */
+		 transfer->box.x % 4 == 0 && transfer->box.width % 4 == 0) {
+		assert(transfer->usage & PIPE_TRANSFER_WRITE);
+
+		/* Check if mapping this buffer would cause waiting for the GPU. */
+		if (rctx->ws->cs_is_buffer_referenced(rctx->cs, rbuffer->cs_buf, RADEON_USAGE_READWRITE) ||
+		    rctx->ws->buffer_is_busy(rbuffer->buf, RADEON_USAGE_READWRITE)) {
+			/* Do a wait-free write-only transfer using a temporary buffer. */
+			struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
+
+			rtransfer->staging = (struct r600_resource*)
+				pipe_buffer_create(pipe->screen, PIPE_BIND_VERTEX_BUFFER,
+						   PIPE_USAGE_STAGING, transfer->box.width);
+			return rctx->ws->buffer_map(rtransfer->staging->cs_buf, rctx->cs, PIPE_TRANSFER_WRITE);
+		}
+	}
 
 	data = rctx->ws->buffer_map(rbuffer->cs_buf, rctx->cs, transfer->usage);
 	if (!data)
@@ -137,7 +160,17 @@ static void *r600_buffer_transfer_map(struct pipe_context *pipe,
 static void r600_buffer_transfer_unmap(struct pipe_context *pipe,
 					struct pipe_transfer *transfer)
 {
-	/* no-op */
+	struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
+
+	if (rtransfer->staging) {
+		struct pipe_box box;
+		u_box_1d(0, transfer->box.width, &box);
+
+		/* Copy the staging buffer into the original one. */
+		r600_copy_buffer(pipe, transfer->resource, transfer->box.x,
+				 &rtransfer->staging->b.b, &box);
+		pipe_resource_reference((struct pipe_resource**)&rtransfer->staging, NULL);
+	}
 }
 
 static void r600_transfer_destroy(struct pipe_context *ctx,
diff --git a/src/gallium/drivers/r600/r600_pipe.c b/src/gallium/drivers/r600/r600_pipe.c
index 2a102621183..52725fc5c67 100644
--- a/src/gallium/drivers/r600/r600_pipe.c
+++ b/src/gallium/drivers/r600/r600_pipe.c
@@ -197,7 +197,7 @@ static struct pipe_context *r600_create_context(struct pipe_screen *screen, void
 		return NULL;
 
 	util_slab_create(&rctx->pool_transfers,
-			 sizeof(struct pipe_transfer), 64,
+			 sizeof(struct r600_transfer), 64,
 			 UTIL_SLAB_SINGLETHREADED);
 
 	rctx->context.screen = screen;
diff --git a/src/gallium/drivers/r600/r600_pipe.h b/src/gallium/drivers/r600/r600_pipe.h
index 2f16ff6e51c..7fdf2f8c444 100644
--- a/src/gallium/drivers/r600/r600_pipe.h
+++ b/src/gallium/drivers/r600/r600_pipe.h
@@ -456,6 +456,9 @@ void evergreen_cb(struct r600_context *rctx, struct r600_pipe_state *rstate,
 void evergreen_update_dual_export_state(struct r600_context * rctx);
 
 /* r600_blit.c */
+void r600_copy_buffer(struct pipe_context *ctx, struct
+		      pipe_resource *dst, unsigned dstx,
+		      struct pipe_resource *src, const struct pipe_box *src_box);
 void r600_init_blit_functions(struct r600_context *rctx);
 void r600_blit_uncompress_depth(struct pipe_context *ctx,
 		struct r600_resource_texture *texture,