gallium/radeon: use a common function for DMA blit preparation

this is more robust and probably fixes some bugs already

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

src/gallium/drivers/r600/evergreen_state.c

@@ -3497,14 +3497,10 @@ static void evergreen_dma_copy(struct pipe_context *ctx,
 		return;
 	}
 
-	if (src->format != dst->format || src_box->depth > 1 ||
-	    (rdst->dirty_level_mask | rdst->stencil_dirty_level_mask) & (1 << dst_level)) {
+	if (src_box->depth > 1 ||
+	    !r600_prepare_for_dma_blit(&rctx->b, rdst, dst_level, dstx, dsty,
+					dstz, rsrc, src_level, src_box))
 		goto fallback;
-	}
-
-	if (rsrc->dirty_level_mask & (1 << src_level)) {
-		ctx->flush_resource(ctx, src);
-	}
 
 	src_x = util_format_get_nblocksx(src->format, src_box->x);
 	dst_x = util_format_get_nblocksx(src->format, dst_x);

src/gallium/drivers/r600/r600_state.c

@@ -2972,9 +2972,10 @@ static void r600_dma_copy(struct pipe_context *ctx,
 		return;
 	}
 
-	if (src->format != dst->format || src_box->depth > 1) {
+	if (src_box->depth > 1 ||
+	    !r600_prepare_for_dma_blit(&rctx->b, rdst, dst_level, dstx, dsty,
+					dstz, rsrc, src_level, src_box))
 		goto fallback;
-	}
 
 	src_x = util_format_get_nblocksx(src->format, src_box->x);
 	dst_x = util_format_get_nblocksx(src->format, dst_x);

src/gallium/drivers/radeon/r600_pipe_common.h

@@ -623,6 +623,13 @@ void r600_update_prims_generated_query_state(struct r600_common_context *rctx,
 void r600_streamout_init(struct r600_common_context *rctx);
 
 /* r600_texture.c */
+bool r600_prepare_for_dma_blit(struct r600_common_context *rctx,
+				struct r600_texture *rdst,
+				unsigned dst_level, unsigned dstx,
+				unsigned dsty, unsigned dstz,
+				struct r600_texture *rsrc,
+				unsigned src_level,
+				const struct pipe_box *src_box);
 void r600_texture_get_fmask_info(struct r600_common_screen *rscreen,
 				 struct r600_texture *rtex,
 				 unsigned nr_samples,

src/gallium/drivers/radeon/r600_texture.c

@@ -32,6 +32,61 @@
 #include <errno.h>
 #include <inttypes.h>
 
+bool r600_prepare_for_dma_blit(struct r600_common_context *rctx,
+			       struct r600_texture *rdst,
+			       unsigned dst_level, unsigned dstx,
+			       unsigned dsty, unsigned dstz,
+			       struct r600_texture *rsrc,
+			       unsigned src_level,
+			       const struct pipe_box *src_box)
+{
+	if (!rctx->dma.cs)
+		return false;
+
+	if (util_format_get_blocksizebits(rdst->resource.b.b.format) !=
+	    util_format_get_blocksizebits(rsrc->resource.b.b.format))
+		return false;
+
+	/* MSAA: Blits don't exist in the real world. */
+	if (rsrc->resource.b.b.nr_samples > 1 ||
+	    rdst->resource.b.b.nr_samples > 1)
+		return false;
+
+	/* Depth-stencil surfaces:
+	 *   When dst is linear, the DB->CB copy preserves HTILE.
+	 *   When dst is tiled, the 3D path must be used to update HTILE.
+	 */
+	if (rsrc->is_depth || rdst->is_depth)
+		return false;
+
+	/* DCC as:
+	 *   src: Use the 3D path. DCC decompression is expensive.
+	 *   dst: If overwriting the whole texture, disable DCC and use SDMA.
+	 *        Otherwise, use the 3D path.
+	 * TODO: handle the case when the dst box covers the whole texture
+	 */
+	if (rsrc->dcc_offset || rdst->dcc_offset)
+		return false;
+
+	/* CMASK as:
+	 *   src: Both texture and SDMA paths need decompression. Use SDMA.
+	 *   dst: If overwriting the whole texture, deallocate CMASK and use
+	 *        SDMA. Otherwise, use the 3D path.
+	 * TODO: handle the case when the dst box covers the whole texture
+	 */
+	if (rdst->cmask.size && rdst->dirty_level_mask & (1 << dst_level))
+		return false;
+
+	/* All requirements are met. Prepare textures for SDMA. */
+	if (rsrc->cmask.size && rsrc->dirty_level_mask & (1 << src_level))
+		rctx->b.flush_resource(&rctx->b, &rsrc->resource.b.b);
+
+	assert(!(rsrc->dirty_level_mask & (1 << src_level)));
+	assert(!(rdst->dirty_level_mask & (1 << dst_level)));
+
+	return true;
+}
+
 /* Same as resource_copy_region, except that both upsampling and downsampling are allowed. */
 static void r600_copy_region_with_blit(struct pipe_context *pipe,
 				       struct pipe_resource *dst,

src/gallium/drivers/radeonsi/cik_sdma.c

@@ -226,19 +226,9 @@ void cik_sdma_copy(struct pipe_context *ctx,
 	 */
 	goto fallback;
 
-	if (src->format != dst->format ||
-	    rdst->surface.nsamples > 1 || rsrc->surface.nsamples > 1 ||
-	    (rdst->dirty_level_mask | rdst->stencil_dirty_level_mask) & (1 << dst_level) ||
-	    rdst->dcc_offset || rsrc->dcc_offset) {
+	if (!r600_prepare_for_dma_blit(&sctx->b, rdst, dst_level, dstx, dsty,
+					dstz, rsrc, src_level, src_box))
 		goto fallback;
-	}
-
-	if (rsrc->dirty_level_mask & (1 << src_level)) {
-		if (rsrc->htile_buffer)
-			goto fallback;
-
-		ctx->flush_resource(ctx, src);
-	}
 
 	src_x = util_format_get_nblocksx(src->format, src_box->x);
 	dst_x = util_format_get_nblocksx(src->format, dst_x);

src/gallium/drivers/radeonsi/si_dma.c

@@ -230,17 +230,10 @@ void si_dma_copy(struct pipe_context *ctx,
 	 */
 	goto fallback;
 
-	if (src->format != dst->format || src_box->depth > 1 ||
-	    (rdst->dirty_level_mask | rdst->stencil_dirty_level_mask) & (1 << dst_level) ||
-	    rdst->cmask.size || rdst->fmask.size ||
-	    rsrc->cmask.size || rsrc->fmask.size ||
-	    rdst->dcc_offset || rsrc->dcc_offset) {
+	if (src_box->depth > 1 ||
+	    !r600_prepare_for_dma_blit(&sctx->b, rdst, dst_level, dstx, dsty,
+					dstz, rsrc, src_level, src_box))
 		goto fallback;
-	}
-
-	if (rsrc->dirty_level_mask & (1 << src_level)) {
-		ctx->flush_resource(ctx, src);
-	}
 
 	src_x = util_format_get_nblocksx(src->format, src_box->x);
 	dst_x = util_format_get_nblocksx(src->format, dst_x);