radeonsi: implement transfer_map with user_stride

The stride ends up being aligned by AddrLib in ways that are
inconvenient to express clearly, but basically, a stride that
is aligned to both 64 pixels and 256 bytes will go through
unchanged in practice.

1 files changed, 28 insertions, 5 deletions

diff --git a/src/gallium/drivers/radeon/r600_texture.c b/src/gallium/drivers/radeon/r600_texture.c
index c1f55b47bb..71816220f4 100644
--- a/src/gallium/drivers/radeon/r600_texture.c
+++ b/src/gallium/drivers/radeon/r600_texture.c
@@ -178,14 +178,27 @@ static void r600_copy_from_staging_texture(struct pipe_context *ctx, struct r600
 		       src, 0, &sbox);
 }
 
+static unsigned r600_texture_get_stride(struct si_screen *sscreen,
+					struct r600_texture *rtex,
+					unsigned level)
+{
+	if (sscreen->info.chip_class >= GFX9) {
+		return rtex->surface.u.gfx9.surf_pitch * rtex->surface.bpe;
+	} else {
+		return rtex->surface.u.legacy.level[level].nblk_x *
+		       rtex->surface.bpe;
+	}
+}
+
 static unsigned r600_texture_get_offset(struct si_screen *sscreen,
 					struct r600_texture *rtex, unsigned level,
 					const struct pipe_box *box,
 					unsigned *stride,
 					unsigned *layer_stride)
 {
+	*stride = r600_texture_get_stride(sscreen, rtex, level);
+
 	if (sscreen->info.chip_class >= GFX9) {
-		*stride = rtex->surface.u.gfx9.surf_pitch * rtex->surface.bpe;
 		*layer_stride = rtex->surface.u.gfx9.surf_slice_size;
 
 		if (!box)
@@ -199,8 +212,6 @@ static unsigned r600_texture_get_offset(struct si_screen *sscreen,
 			rtex->surface.u.gfx9.surf_pitch +
 			box->x / rtex->surface.blk_w) * rtex->surface.bpe;
 	} else {
-		*stride = rtex->surface.u.legacy.level[level].nblk_x *
-			  rtex->surface.bpe;
 		assert((uint64_t)rtex->surface.u.legacy.level[level].slice_size_dw * 4 <= UINT_MAX);
 		*layer_stride = (uint64_t)rtex->surface.u.legacy.level[level].slice_size_dw * 4;
 
@@ -1671,7 +1682,9 @@ static void *r600_texture_transfer_map(struct pipe_context *ctx,
 		 * Use the staging texture for uploads if the underlying BO
 		 * is busy.
 		 */
-		if (!rtex->surface.is_linear)
+		if (!rtex->surface.is_linear ||
+		    (user_stride &&
+		     user_stride != r600_texture_get_stride(rctx->screen, rtex, level)))
 			use_staging_texture = true;
 		else if (usage & PIPE_TRANSFER_READ)
 			use_staging_texture =
@@ -1763,9 +1776,19 @@ static void *r600_texture_transfer_map(struct pipe_context *ctx,
 	} else if (use_staging_texture) {
 		struct pipe_resource resource;
 		struct r600_texture *staging;
+		struct pipe_box staging_box = *box;
+
+		if (user_stride) {
+			if (user_stride % rtex->surface.bpe != 0)
+				goto fail_trans;
 
-		r600_init_temp_resource_from_box(&resource, texture, box, level,
+			staging_box.width = user_stride / rtex->surface.bpe;
+			assert(staging_box.width >= box->width);
+		}
+
+		r600_init_temp_resource_from_box(&resource, texture, &staging_box, level,
 						 R600_RESOURCE_FLAG_TRANSFER);
+
 		resource.usage = (usage & PIPE_TRANSFER_READ) ?
 			PIPE_USAGE_STAGING : PIPE_USAGE_STREAM;