summaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
authorRob Clark <robdclark@chromium.org>2019-05-03 13:10:22 -0700
committerRob Clark <robdclark@chromium.org>2019-05-04 11:50:44 -0700
commit857d9f3b02c95713d845c54bc6cc1faf877095bb (patch)
tree227801642d6710d0ee6d742c4875bba6300a6bd8 /src
parent6ffb58726b937db1149855ed1fca4690c100158f (diff)
freedreno/a6xx: UBWC fixes
A few fixes that get UBWC working for the games/benchmarks where I noticed problems before (in particular and manhattan, and stk (modulo image support for UBWC when compute shaders are used for post-process effects): + fix the size of the UBWC meta buffer (ie, the offset to color pixel data) that is returned by ->fill_ubwc_buffer_sizes() + correct size/layout for 8 and 16 byte per pixel formats + limit the supported formats.. Note all formats that can be tiled can be compressed. Signed-off-by: Rob Clark <robdclark@chromium.org>
Diffstat (limited to 'src')
-rw-r--r--src/gallium/drivers/freedreno/a6xx/fd6_resource.c87
-rw-r--r--src/gallium/drivers/freedreno/a6xx/fd6_texture.c2
2 files changed, 78 insertions, 11 deletions
diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_resource.c b/src/gallium/drivers/freedreno/a6xx/fd6_resource.c
index e15a7ba44bb..ca475f30ef6 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_resource.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_resource.c
@@ -26,6 +26,7 @@
*/
#include "fd6_resource.h"
+#include "fd6_format.h"
#include "a6xx.xml.h"
@@ -161,12 +162,51 @@ setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format forma
return size;
}
+/* A subset of the valid tiled formats can be compressed. We do
+ * already require tiled in order to be compressed, but just because
+ * it can be tiled doesn't mean it can be compressed.
+ */
+static bool
+ok_ubwc_format(enum a6xx_color_fmt fmt)
+{
+ switch (fmt) {
+ case RB6_R10G10B10A2_UINT:
+ case RB6_R10G10B10A2_UNORM:
+ case RB6_R11G11B10_FLOAT:
+ case RB6_R16_FLOAT:
+ case RB6_R16G16B16A16_FLOAT:
+ case RB6_R16G16B16A16_SINT:
+ case RB6_R16G16B16A16_UINT:
+ case RB6_R16G16_FLOAT:
+ case RB6_R16G16_SINT:
+ case RB6_R16G16_UINT:
+ case RB6_R16_SINT:
+ case RB6_R16_UINT:
+ case RB6_R32G32B32A32_SINT:
+ case RB6_R32G32B32A32_UINT:
+ case RB6_R32G32_SINT:
+ case RB6_R32G32_UINT:
+ case RB6_R32_SINT:
+ case RB6_R32_UINT:
+ case RB6_R5G6B5_UNORM:
+ case RB6_R8G8B8A8_SINT:
+ case RB6_R8G8B8A8_UINT:
+ case RB6_R8G8B8A8_UNORM:
+ case RB6_R8G8B8_UNORM:
+ case RB6_R8G8_SINT:
+ case RB6_R8G8_UINT:
+ case RB6_R8G8_UNORM:
+ case RB6_X8Z24_UNORM:
+ return true;
+ default:
+ return false;
+ }
+}
+
uint32_t
fd6_fill_ubwc_buffer_sizes(struct fd_resource *rsc)
{
-#define RGB_TILE_WIDTH 16
#define RBG_TILE_WIDTH_ALIGNMENT 64
-#define RGB_TILE_HEIGHT 4
#define RGB_TILE_HEIGHT_ALIGNMENT 16
#define UBWC_PLANE_SIZE_ALIGNMENT 4096
@@ -174,23 +214,50 @@ fd6_fill_ubwc_buffer_sizes(struct fd_resource *rsc)
uint32_t width = prsc->width0;
uint32_t height = prsc->height0;
+ if (!ok_ubwc_format(fd6_pipe2color(prsc->format)))
+ return 0;
+
/* limit things to simple single level 2d for now: */
if ((prsc->depth0 != 1) || (prsc->array_size != 1) || (prsc->last_level != 0))
return 0;
+ uint32_t block_width, block_height;
+ switch (rsc->cpp) {
+ case 2:
+ case 4:
+ block_width = 16;
+ block_height = 4;
+ break;
+ case 8:
+ block_width = 8;
+ block_height = 4;
+ break;
+ case 16:
+ block_width = 4;
+ block_height = 4;
+ break;
+ default:
+ return 0;
+ }
+
uint32_t meta_stride =
- ALIGN_POT(DIV_ROUND_UP(width, RGB_TILE_WIDTH), RBG_TILE_WIDTH_ALIGNMENT);
- uint32_t meta_scanlines =
- ALIGN_POT(DIV_ROUND_UP(height, RGB_TILE_HEIGHT), RGB_TILE_HEIGHT_ALIGNMENT);
- uint32_t meta_plane =
- ALIGN_POT(meta_stride * meta_scanlines, UBWC_PLANE_SIZE_ALIGNMENT);
+ ALIGN_POT(DIV_ROUND_UP(width, block_width), RBG_TILE_WIDTH_ALIGNMENT);
+ uint32_t meta_height =
+ ALIGN_POT(DIV_ROUND_UP(height, block_height), RGB_TILE_HEIGHT_ALIGNMENT);
+ uint32_t meta_size =
+ ALIGN_POT(meta_stride * meta_height, UBWC_PLANE_SIZE_ALIGNMENT);
- rsc->offset = meta_plane;
+ /* UBWC goes first, then color data.. this constraint is mainly only
+ * because it is what the kernel expects for scanout. For non-2D we
+ * could just use a separate UBWC buffer..
+ */
+ rsc->ubwc_offset = 0;
+ rsc->offset = meta_size;
rsc->ubwc_pitch = meta_stride;
- rsc->ubwc_size = meta_plane >> 2;
+ rsc->ubwc_size = meta_size >> 2; /* in dwords??? */
rsc->tile_mode = TILE6_3;
- return rsc->ubwc_size;
+ return meta_size;
}
uint32_t
diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_texture.c b/src/gallium/drivers/freedreno/a6xx/fd6_texture.c
index f780ebdf0aa..b2430cc6c3e 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_texture.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_texture.c
@@ -266,7 +266,7 @@ fd6_sampler_view_create(struct pipe_context *pctx, struct pipe_resource *prsc,
format, rsc->slices[lvl].pitch) * rsc->cpp);
so->offset = fd_resource_offset(rsc, lvl, cso->u.tex.first_layer);
- so->ubwc_enabled = rsc->ubwc_size && u_minify(prsc->width0, lvl) >= 16;
+ so->ubwc_enabled = rsc->ubwc_size && !fd_resource_level_linear(prsc, lvl);
}
so->texconst0 |= fd6_tex_const_0(prsc, lvl, cso->format,