[Mesa-dev] [PATCH 42/43] i965: Use blorp for CopyImageSubData on gen4-5
Jason Ekstrand
jason at jlekstrand.net
Tue May 16 22:45:36 UTC 2017
We keep the blit path because it's probably faster when it works.
However, now that we can use blorp, we can delete that nasty CPU
fall-back path.
---
src/mesa/drivers/dri/i965/intel_copy_image.c | 140 ++++-----------------------
1 file changed, 17 insertions(+), 123 deletions(-)
diff --git a/src/mesa/drivers/dri/i965/intel_copy_image.c b/src/mesa/drivers/dri/i965/intel_copy_image.c
index 85585c7..b593c3c 100644
--- a/src/mesa/drivers/dri/i965/intel_copy_image.c
+++ b/src/mesa/drivers/dri/i965/intel_copy_image.c
@@ -35,82 +35,6 @@
#include "drivers/common/meta.h"
static void
-copy_image_with_memcpy(struct brw_context *brw,
- struct intel_mipmap_tree *src_mt, int src_level,
- int src_x, int src_y, int src_z,
- struct intel_mipmap_tree *dst_mt, int dst_level,
- int dst_x, int dst_y, int dst_z,
- int src_width, int src_height)
-{
- bool same_slice;
- void *mapped, *src_mapped, *dst_mapped;
- ptrdiff_t src_stride, dst_stride, cpp;
- int map_x1, map_y1, map_x2, map_y2;
- GLuint src_bw, src_bh;
-
- cpp = _mesa_get_format_bytes(src_mt->format);
- _mesa_get_format_block_size(src_mt->format, &src_bw, &src_bh);
-
- assert(src_width % src_bw == 0);
- assert(src_height % src_bh == 0);
- assert(src_x % src_bw == 0);
- assert(src_y % src_bh == 0);
-
- /* If we are on the same miptree, same level, and same slice, then
- * intel_miptree_map won't let us map it twice. We have to do things a
- * bit differently. In particular, we do a single map large enough for
- * both portions and in read-write mode.
- */
- same_slice = src_mt == dst_mt && src_level == dst_level && src_z == dst_z;
-
- if (same_slice) {
- assert(dst_x % src_bw == 0);
- assert(dst_y % src_bh == 0);
-
- map_x1 = MIN2(src_x, dst_x);
- map_y1 = MIN2(src_y, dst_y);
- map_x2 = MAX2(src_x, dst_x) + src_width;
- map_y2 = MAX2(src_y, dst_y) + src_height;
-
- intel_miptree_map(brw, src_mt, src_level, src_z,
- map_x1, map_y1, map_x2 - map_x1, map_y2 - map_y1,
- GL_MAP_READ_BIT | GL_MAP_WRITE_BIT,
- &mapped, &src_stride);
-
- dst_stride = src_stride;
-
- /* Set the offsets here so we don't have to think about while looping */
- src_mapped = mapped + ((src_y - map_y1) / src_bh) * src_stride +
- ((src_x - map_x1) / src_bw) * cpp;
- dst_mapped = mapped + ((dst_y - map_y1) / src_bh) * dst_stride +
- ((dst_x - map_x1) / src_bw) * cpp;
- } else {
- intel_miptree_map(brw, src_mt, src_level, src_z,
- src_x, src_y, src_width, src_height,
- GL_MAP_READ_BIT, &src_mapped, &src_stride);
- intel_miptree_map(brw, dst_mt, dst_level, dst_z,
- dst_x, dst_y, src_width, src_height,
- GL_MAP_WRITE_BIT, &dst_mapped, &dst_stride);
- }
-
- src_width /= (int)src_bw;
- src_height /= (int)src_bh;
-
- for (int i = 0; i < src_height; ++i) {
- memcpy(dst_mapped, src_mapped, src_width * cpp);
- src_mapped += src_stride;
- dst_mapped += dst_stride;
- }
-
- if (same_slice) {
- intel_miptree_unmap(brw, src_mt, src_level, src_z);
- } else {
- intel_miptree_unmap(brw, dst_mt, dst_level, dst_z);
- intel_miptree_unmap(brw, src_mt, src_level, src_z);
- }
-}
-
-static void
copy_miptrees(struct brw_context *brw,
struct intel_mipmap_tree *src_mt,
int src_x, int src_y, int src_z, unsigned src_level,
@@ -118,55 +42,25 @@ copy_miptrees(struct brw_context *brw,
int dst_x, int dst_y, int dst_z, unsigned dst_level,
int src_width, int src_height)
{
- unsigned bw, bh;
-
- if (brw->gen >= 6) {
- brw_blorp_copy_miptrees(brw,
- src_mt, src_level, src_z,
- dst_mt, dst_level, dst_z,
- src_x, src_y, dst_x, dst_y,
- src_width, src_height);
- return;
+ if (brw->gen < 5) {
+ /* On gen4-5, try BLT first.
+ *
+ * Gen4-5 have a single ring for both 3D and BLT operations, so there's
+ * no inter-ring synchronization issues like on Gen6+. It is apparently
+ * faster than using the 3D pipeline. Original Gen4 also has to rebase
+ * and copy miptree slices in order to render to unaligned locations.
+ */
+ if (intel_miptree_copy(brw, src_mt, src_level, src_z, src_x, src_y,
+ dst_mt, dst_level, dst_z, dst_x, dst_y,
+ src_width, src_height))
+ return;
}
- /* We are now going to try and copy the texture using the blitter. If
- * that fails, we will fall back mapping the texture and using memcpy.
- * In either case, we need to do a full resolve.
- */
- intel_miptree_all_slices_resolve_hiz(brw, src_mt);
- intel_miptree_all_slices_resolve_depth(brw, src_mt);
- intel_miptree_all_slices_resolve_color(brw, src_mt, 0);
-
- intel_miptree_all_slices_resolve_hiz(brw, dst_mt);
- intel_miptree_all_slices_resolve_depth(brw, dst_mt);
- intel_miptree_all_slices_resolve_color(brw, dst_mt, 0);
-
- _mesa_get_format_block_size(src_mt->format, &bw, &bh);
-
- /* It's legal to have a WxH that's smaller than a compressed block. This
- * happens for example when you are using a higher level LOD. For this case,
- * we still want to copy the entire block, or else the decompression will be
- * incorrect.
- */
- if (src_width < bw)
- src_width = ALIGN_NPOT(src_width, bw);
-
- if (src_height < bh)
- src_height = ALIGN_NPOT(src_height, bh);
-
- if (intel_miptree_copy(brw, src_mt, src_level, src_z, src_x, src_y,
- dst_mt, dst_level, dst_z, dst_x, dst_y,
- src_width, src_height))
- return;
-
- /* This is a worst-case scenario software fallback that maps the two
- * textures and does a memcpy between them.
- */
- copy_image_with_memcpy(brw, src_mt, src_level,
- src_x, src_y, src_z,
- dst_mt, dst_level,
- dst_x, dst_y, dst_z,
- src_width, src_height);
+ brw_blorp_copy_miptrees(brw,
+ src_mt, src_level, src_z,
+ dst_mt, dst_level, dst_z,
+ src_x, src_y, dst_x, dst_y,
+ src_width, src_height);
}
static void
--
2.5.0.400.gff86faf
More information about the mesa-dev
mailing list