Mesa (main): freedreno/a6xx: Handle fb_read in sysmem path
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Mon Jun 21 19:02:25 UTC 2021
Module: Mesa
Branch: main
Commit: 5b4f670c1cf0361a7913d256e12544acf7536665
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=5b4f670c1cf0361a7913d256e12544acf7536665
Author: Rob Clark <robdclark at chromium.org>
Date: Sun Jun 20 08:58:42 2021 -0700
freedreno/a6xx: Handle fb_read in sysmem path
Signed-off-by: Rob Clark <robdclark at chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11487>
---
.../ci/deqp-freedreno-a630-bypass-fails.txt | 10 ---
src/gallium/drivers/freedreno/a6xx/fd6_emit.c | 20 +----
src/gallium/drivers/freedreno/a6xx/fd6_gmem.c | 87 ++++++++++++++++++++--
3 files changed, 85 insertions(+), 32 deletions(-)
diff --git a/src/freedreno/ci/deqp-freedreno-a630-bypass-fails.txt b/src/freedreno/ci/deqp-freedreno-a630-bypass-fails.txt
index a317dd6a947..030b7ee4bad 100644
--- a/src/freedreno/ci/deqp-freedreno-a630-bypass-fails.txt
+++ b/src/freedreno/ci/deqp-freedreno-a630-bypass-fails.txt
@@ -1,16 +1,6 @@
-dEQP-GLES31.functional.blend_equation_advanced.barrier.colordodge,Fail
-dEQP-GLES31.functional.blend_equation_advanced.barrier.exclusion,Fail
-dEQP-GLES31.functional.blend_equation_advanced.barrier.multiply,Fail
-dEQP-GLES31.functional.blend_equation_advanced.basic.colordodge,Fail
-dEQP-GLES31.functional.blend_equation_advanced.basic.exclusion,Fail
-dEQP-GLES31.functional.blend_equation_advanced.basic.multiply,Fail
dEQP-GLES31.functional.blend_equation_advanced.msaa.colordodge,Fail
dEQP-GLES31.functional.blend_equation_advanced.msaa.exclusion,Fail
dEQP-GLES31.functional.blend_equation_advanced.msaa.multiply,Fail
-dEQP-GLES31.functional.blend_equation_advanced.srgb.colordodge,Fail
-dEQP-GLES31.functional.blend_equation_advanced.srgb.exclusion,Fail
-dEQP-GLES31.functional.blend_equation_advanced.srgb.multiply,Fail
-dEQP-GLES31.functional.draw_buffers_indexed.overwrite_common.common_blend_eq_buffer_advanced_blend_eq,Fail
dEQP-VK.renderpass.dedicated_allocation.attachment_allocation.input_output.7,Fail
dEQP-VK.renderpass.suballocation.attachment_allocation.input_output.7,Fail
dEQP-VK.renderpass.suballocation.subpass_dependencies.implicit_dependencies.render_passes_5,Fail
diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_emit.c b/src/gallium/drivers/freedreno/a6xx/fd6_emit.c
index 88aef02c77d..1ca7af5d7fc 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_emit.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_emit.c
@@ -249,25 +249,13 @@ fd6_emit_fb_tex(struct fd_ringbuffer *state, struct fd_context *ctx) assert_dt
struct pipe_surface *psurf = pfb->cbufs[0];
struct fd_resource *rsc = fd_resource(psurf->texture);
- uint32_t texconst0 = fd6_tex_const_0(
- psurf->texture, psurf->u.tex.level, psurf->format, PIPE_SWIZZLE_X,
- PIPE_SWIZZLE_Y, PIPE_SWIZZLE_Z, PIPE_SWIZZLE_W);
-
- /* always TILE6_2 mode in GMEM.. which also means no swap: */
- texconst0 &=
- ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
- texconst0 |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
-
- OUT_RING(state, texconst0);
+ OUT_RINGP(state, 0, &ctx->batch->fb_read_patches); /* texconst0, patched in gmem emit */
OUT_RING(state, A6XX_TEX_CONST_1_WIDTH(pfb->width) |
A6XX_TEX_CONST_1_HEIGHT(pfb->height));
- OUT_RINGP(state, A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D),
- &ctx->batch->fb_read_patches);
+ OUT_RING(state, 0); /* texconst2, patched in gmem emit */
OUT_RING(state, A6XX_TEX_CONST_3_ARRAY_PITCH(rsc->layout.layer_size));
-
- OUT_RING(state, A6XX_TEX_CONST_4_BASE_LO(ctx->screen->gmem_base));
- OUT_RING(state, A6XX_TEX_CONST_5_BASE_HI(ctx->screen->gmem_base >> 32) |
- A6XX_TEX_CONST_5_DEPTH(1));
+ OUT_RING(state, 0); /* BASE_LO, patched in gmem emit */
+ OUT_RING(state, 0); /* BASE_HI, patched in gmem emit */
OUT_RING(state, 0); /* texconst6 */
OUT_RING(state, 0); /* texconst7 */
OUT_RING(state, 0); /* texconst8 */
diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c b/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
index 250b5ed7d8e..e5ee771b7c8 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_gmem.c
@@ -243,15 +243,89 @@ use_hw_binning(struct fd_batch *batch)
}
static void
-patch_fb_read(struct fd_batch *batch)
+patch_fb_read_gmem(struct fd_batch *batch)
{
+ unsigned num_patches = fd_patch_num_elements(&batch->fb_read_patches);
+ if (!num_patches)
+ return;
+
+ struct fd_screen *screen = batch->ctx->screen;
const struct fd_gmem_stateobj *gmem = batch->gmem_state;
+ struct pipe_framebuffer_state *pfb = &batch->framebuffer;
+ struct pipe_surface *psurf = pfb->cbufs[0];
+ uint32_t texconst0 = fd6_tex_const_0(
+ psurf->texture, psurf->u.tex.level, psurf->format, PIPE_SWIZZLE_X,
+ PIPE_SWIZZLE_Y, PIPE_SWIZZLE_Z, PIPE_SWIZZLE_W);
+
+ /* always TILE6_2 mode in GMEM.. which also means no swap: */
+ texconst0 &=
+ ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
+ texconst0 |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
+
+ for (unsigned i = 0; i < num_patches; i++) {
+ struct fd_cs_patch *patch = fd_patch_element(&batch->fb_read_patches, i);
+ patch->cs[0] = texconst0;
+ patch->cs[2] = A6XX_TEX_CONST_2_PITCH(gmem->bin_w * gmem->cbuf_cpp[0]) |
+ A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D);
+ patch->cs[4] = A6XX_TEX_CONST_4_BASE_LO(screen->gmem_base);
+ patch->cs[5] = A6XX_TEX_CONST_5_BASE_HI(screen->gmem_base >> 32) |
+ A6XX_TEX_CONST_5_DEPTH(1);
+ }
+ util_dynarray_clear(&batch->fb_read_patches);
+}
+
+static void
+patch_fb_read_sysmem(struct fd_batch *batch)
+{
+ unsigned num_patches = fd_patch_num_elements(&batch->fb_read_patches);
+ if (!num_patches)
+ return;
+
+ struct pipe_framebuffer_state *pfb = &batch->framebuffer;
+ struct pipe_surface *psurf = pfb->cbufs[0];
+ if (!psurf)
+ return;
- for (unsigned i = 0; i < fd_patch_num_elements(&batch->fb_read_patches);
- i++) {
+ struct fd_resource *rsc = fd_resource(psurf->texture);
+ unsigned lvl = psurf->u.tex.level;
+ unsigned layer = psurf->u.tex.first_layer;
+ bool ubwc_enabled = fd_resource_ubwc_enabled(rsc, lvl);
+ uint64_t iova = fd_bo_get_iova(rsc->bo) + fd_resource_offset(rsc, lvl, layer);
+ uint64_t ubwc_iova = fd_bo_get_iova(rsc->bo) + fd_resource_ubwc_offset(rsc, lvl, layer);
+ uint32_t texconst0 = fd6_tex_const_0(
+ psurf->texture, psurf->u.tex.level, psurf->format, PIPE_SWIZZLE_X,
+ PIPE_SWIZZLE_Y, PIPE_SWIZZLE_Z, PIPE_SWIZZLE_W);
+ uint32_t block_width, block_height;
+ fdl6_get_ubwc_blockwidth(&rsc->layout, &block_width, &block_height);
+
+ for (unsigned i = 0; i < num_patches; i++) {
struct fd_cs_patch *patch = fd_patch_element(&batch->fb_read_patches, i);
- *patch->cs =
- patch->val | A6XX_TEX_CONST_2_PITCH(gmem->bin_w * gmem->cbuf_cpp[0]);
+ patch->cs[0] = texconst0;
+ patch->cs[2] = A6XX_TEX_CONST_2_PITCH(fd_resource_pitch(rsc, lvl)) |
+ A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D);
+ /* This is cheating a bit, since we can't use OUT_RELOC() here.. but
+ * the render target will already have a reloc emitted for RB_MRT state,
+ * so we can get away with manually patching in the address here:
+ */
+ patch->cs[4] = A6XX_TEX_CONST_4_BASE_LO(iova);
+ patch->cs[5] = A6XX_TEX_CONST_5_BASE_HI(iova >> 32) |
+ A6XX_TEX_CONST_5_DEPTH(1);
+
+ if (!ubwc_enabled)
+ continue;
+
+ patch->cs[3] |= A6XX_TEX_CONST_3_FLAG;
+ patch->cs[7] = A6XX_TEX_CONST_7_FLAG_LO(ubwc_iova);
+ patch->cs[8] = A6XX_TEX_CONST_8_FLAG_HI(ubwc_iova >> 32);
+ patch->cs[9] = A6XX_TEX_CONST_9_FLAG_BUFFER_ARRAY_PITCH(
+ rsc->layout.ubwc_layer_size >> 2);
+ patch->cs[10] =
+ A6XX_TEX_CONST_10_FLAG_BUFFER_PITCH(
+ fdl_ubwc_pitch(&rsc->layout, lvl)) |
+ A6XX_TEX_CONST_10_FLAG_BUFFER_LOGW(util_logbase2_ceil(
+ DIV_ROUND_UP(u_minify(psurf->texture->width0, lvl), block_width))) |
+ A6XX_TEX_CONST_10_FLAG_BUFFER_LOGH(util_logbase2_ceil(
+ DIV_ROUND_UP(u_minify(psurf->texture->height0, lvl), block_height)));
}
util_dynarray_clear(&batch->fb_read_patches);
}
@@ -741,7 +815,7 @@ fd6_emit_tile_init(struct fd_batch *batch) assert_dt
emit_zs(ring, pfb->zsbuf, batch->gmem_state);
emit_mrt(ring, pfb, batch->gmem_state);
emit_msaa(ring, pfb->samples);
- patch_fb_read(batch);
+ patch_fb_read_gmem(batch);
if (use_hw_binning(batch)) {
/* enable stream-out during binning pass: */
@@ -1523,6 +1597,7 @@ fd6_emit_sysmem_prep(struct fd_batch *batch) assert_dt
emit_zs(ring, pfb->zsbuf, NULL);
emit_mrt(ring, pfb, NULL);
emit_msaa(ring, pfb->samples);
+ patch_fb_read_sysmem(batch);
update_render_cntl(batch, pfb, false);
More information about the mesa-commit
mailing list