fix(agc): give a depth stencil copy's layout barrier both aspects by milescoviello · Pull Request #9057 · boykopovar/AnyPS5

milescoviello · 2026-10-21T11:09:19Z

What

The depth sample reorder (SampleOrder.cpp, Depth) copies the depth plane into an image of the surface's own format before redrawing it in the guest's order. For a D32_SFLOAT_S8_UINT surface that only reorders depth (SampleOrderMethod::Depth), the copy image is still D32S8. But its first transition to GENERAL named the depth aspect alone. Without separateDepthStencilLayouts, Vulkan requires a depth/stencil format's barrier to name both aspects. The Khronos layer reports this as VUID-VkImageMemoryBarrier-image-03320, 6 times over the sample order tests, from the 2x and 4x "depth format 130 through a depth pass" cases.

The barrier now takes its aspects from the format (HasStencil(target.format)), as transitionLocated in the same file already does. The copy region still moves depth only when only depth is reordered.

Tested

Rebased onto main d2e42c3 as 7333de6aee (commit unchanged, git range-diff). RTX 4070 Ti (615.71.09, Ubuntu 26.04.1, GCC 15.2, Release), 3 runs each: with test(agc): read the sample order tests' depth through a sampled copy #9056 (53949c0775) the 4 sample_order tests pass; alone, 3 of the 4 still fail there, as on main, which test(agc): read the sample order tests' depth through a sampled copy #9056 fixes.

Ubuntu 24.04, GCC 13.3, NVIDIA RTX 4070 Ti, driver 595.99.02, on main 275e896.

Khronos validation layer 1.4.363 over the 4 tests matching sample_order:
- main: VUID-VkImageMemoryBarrier-image-03320 ×6 and VUID-VkWriteDescriptorSet-descriptorType-00337 ×18.
- With this change: 03320 is gone, and nothing is new.
Pass and fail counts are unchanged by this change, under the layer and without it (3 of 3, and under lavapipe). agc_driver_sample_order_planning passes. The other three fail on this device on main and with this change alike: the remaining 00337 reads are a test issue, which test(agc): read the sample order tests' depth through a sampled copy #9056 fixes. That PR passes all 4 here.
AMD, Windows: @DavidSamir ran it with #9056 on a Radeon 680M (driver 32.0.21030.2001) under the 1.4.363 layer. The 6 03320 errors on main 27fabb6 are gone.
NVIDIA, Windows: @MaxLikesCode ran it on an RTX 4090 (driver 617.42, Windows 11) on main 2c4b116, under the 1.4.363 layer, 3 runs each. main + test(agc): read the sample order tests' depth through a sampled copy #9056 still has 6 03320 per run; with this PR added there is no VUID, and the pass and skip counts are unchanged.
@EagleBearerYS (fix(agc): keep the guest's sample order where pixel (0,0)'s locations stand in for a 2x2 grid #6683's author) reviewed it. They confirmed this was the only barrier in SampleOrder.cpp naming fewer aspects than its image's format has, and ran it with test(agc): read the sample order tests' depth through a sampled copy #9056 on GitHub's runners (lavapipe: the sample order tests pass 3 runs in a row).
Older layer, lavapipe: @xXPBSXx ran all 483 built agc_driver_*_tests under the Ubuntu 24.04 layer package (1.3.275) on main 950388b. main reports 03320 only from agc_driver_sample_order; with this change that test passes with no 03320, and only test(agc): read the sample order tests' depth through a sampled copy #9056's 00337 remains. One run.
NVIDIA and AMD, Windows: @wachi35676 ran it with #9056 on an RTX 5070 Ti (driver 617.42) and an AMD Radeon iGPU (gfx1036, driver 32.0.21042.62), Windows 11, 3 runs each: with both PRs sample_order and _recipe_depth_locations pass 3 of 3 on both GPUs (they fail on main). No validation layer on that machine.
lavapipe on the current host (Mesa 26.0.8-1ubuntu0.3, LLVM 21.1.8, Ubuntu 26.04.1), this PR alone merged onto main a6dcea9: the 4 sample_order tests pass 3 of 3. On the RTX 4070 Ti (615.71.09) in the same build, 3 of the 4 still fail, as described above: test(agc): read the sample order tests' depth through a sampled copy #9056 fixes those.

Not tested: Intel.

Checklist

🤖 Generated with Claude Code

DavidSamir · 2026-10-21T11:49:00Z

AMD Radeon 680M (AMD 32.0.21030.2001, Windows) result for this PR together with #9056, under the validation layer: #9056 (comment). The VUID-VkImageMemoryBarrier-image-03320 errors (6 on main 27fabb6b7) are gone with both PRs.

AI-assisted: yes

EagleBearerYS · 2026-10-21T11:50:39Z

Read against SampleOrder.cpp on main: this is right, and it fixes a bug in our #6683's code (57ceb9e). Thanks for finding it.

A D32S8 surface whose stencil isn't reordered (SampleOrderMethod::Depth, as when it has no stencil address) still gets a D32S8 copy from copyFor. That copy's first UNDEFINED to GENERAL barrier named the depth aspect alone. Nothing in the driver enables separateDepthStencilLayouts, so VUID-VkImageMemoryBarrier-image-03320 applies and the barrier must name both.
layoutAspects follows the copy's format, as transitionLocated does for the surface. The copy region and the descriptor stay depth-only, so only the layout barrier changes.
It was the only barrier in SampleOrder.cpp that named fewer aspects than its image's format has. The tests' own ToGeneral calls already take both.
Checked here, with this PR and test(agc): read the sample order tests' depth through a sampled copy #9056 merged on main 275e896 (run):
- On GitHub's Linux runner (lavapipe), agc_driver_sample_order, agc_driver_sample_order_planning, agc_driver_pixel_sample_iteration and agc_driver_sample_location_grid_fallback pass 3 runs in a row, and the grid-reference and recipe tests skip.
- On Windows, without a GPU, 4 of those 6 skip and 2 pass.
- The runners have no validation layer, so the 03320 count isn't reproduced here.

milescoviello · 2026-10-21T11:56:13Z

Thanks @EagleBearerYS for the review, and for checking that no other barrier in SampleOrder.cpp names fewer aspects than its image's format. Thanks @DavidSamir for the 680M run under the layer. Both are in the Tested section now.

AI-assisted: yes (Claude Code)

xXPBSXx · 2026-10-21T19:52:12Z

Validation-layer data point for this PR (data only).

Setup: Linux (WSL2 Ubuntu 24.04, GCC 13.3, Release), llvmpipe, vulkan-validationlayers 1.3.275.0-1 from the Ubuntu archive (older than the SDK a driver ships with), VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation. main 950388bfd; this PR's change (taken as git diff upstream/main...refs/pr/9057) applied on top of it.

Before (main): across all 483 built agc_driver_*_tests the layer reports VUID-VkImageMemoryBarrier-image-03320 once, from agc_driver_sample_order: barrier on a depth/stencil image (D32_SFLOAT_S8_UINT) with only the DEPTH aspect and no separateDepthStencilLayouts.

After (this PR): agc_driver_sample_order run again with the layer: 0 reports of that VUID, test passes. VUID-VkWriteDescriptorSet-descriptorType-00337 (a depth view without SAMPLED usage used as a sampled image) is still reported 3 times in that run's log; #9056 is about that one.

One run each, on llvmpipe only.

Not run: other GPUs for the layer run, other platforms, the other tests with the change under the layer.

AI-assisted (Claude Code).

milescoviello · 2026-10-21T19:53:16Z

Thanks @xXPBSXx. So 03320 also comes only from agc_driver_sample_order across the 483 tests, and this PR clears it there. I added it to the Tested section.

AI-assisted: yes (Claude Code)

github-actions · 2026-10-22T13:12:16Z

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 7642	`core/libs/prx/libSceAgcDriver/Graphics/src/SampleOrder.cpp`
PR 9508	`core/libs/prx/libSceAgcDriver/Graphics/src/SampleOrder.cpp`
PR 10193	`core/libs/prx/libSceAgcDriver/Graphics/src/SampleOrder.cpp`
PR 10194	`core/libs/prx/libSceAgcDriver/Graphics/src/SampleOrder.cpp`

The depth sample reorder copies a D32S8 surface into an image of the same format. When only depth is reordered, the copy's transition to GENERAL named the depth aspect alone, which Vulkan does not allow for a depth/stencil format without separateDepthStencilLayouts (VUID-VkImageMemoryBarrier-image-03320). The barrier now names both aspects of a format with stencil, as transitionLocated does; the copy itself still moves depth only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

milescoviello · 2026-10-23T18:56:30Z

Rebased onto main d2e42c3 as 7333de6aee, with the commit unchanged (git range-diff), together with #9056 (53949c0775). On the RTX 4070 Ti (615.71.09, Linux), the four sample_order tests with both PRs pass 4 of 4, 3 runs out of 3. This PR alone still fails three of them there, as the Tested section says: #9056 fixes those. The full table is on #9056.

AI-assisted: yes (Claude Code)

adriadam10 · 2026-10-23T23:01:46Z

Checked the barrier in SampleOrder.cpp: the depth-only copy is still a D32_SFLOAT_S8_UINT image, and without separateDepthStencilLayouts its barrier must name both aspects. It now follows HasStencil(target.format) like transitionLocated, while the copy region still moves depth only. Third-party runs under the validation layer (680M, RTX 4090, lavapipe) show VUID-VkImageMemoryBarrier-image-03320 gone; merged.

This was referenced Oct 21, 2026

test(agc): read the sample order tests' depth through a sampled copy #9056

Merged

fix(agc): keep the guest's sample order where pixel (0,0)'s locations stand in for a 2x2 grid #6683

Merged

This was referenced Oct 21, 2026

feat(agc): preserve initial native multisample color contents #6973

Merged

feat(agc): publish native color for read-only shader buffers #7002

Draft

milescoviello mentioned this pull request Oct 21, 2026

fix(agc): refuse LDS split before allocation #7827

Merged

12 tasks done

github-actions Bot added the agc libSceAgc* label Oct 21, 2026

BKlaasWerkman mentioned this pull request Oct 23, 2026

fix(shader,agc): declare ImageMSArray only for multisampled arrayed storage images #9978

Merged

12 tasks done

LeandroLP23 mentioned this pull request Oct 23, 2026

build(cmake): compile the Windows test exit wrapper once for every test #10124

Merged

12 tasks done

milescoviello force-pushed the fix/sample-order-depth-copy-aspects branch from 78254be to 7333de6 Compare October 23, 2026 18:56

adriadam10 self-assigned this Oct 23, 2026

adriadam10 merged commit a23f8fc into boykopovar:main Oct 23, 2026
12 checks passed

Amoenus mentioned this pull request Oct 23, 2026

feat(agc): back extra guest mip levels with storage images #7085

Open

12 tasks done

Navigation Menu

Repository navigation

fix(agc): give a depth stencil copy's layout barrier both aspects - #9057

fix(agc): give a depth stencil copy's layout barrier both aspects#9057
adriadam10 merged 1 commit into
boykopovar:mainfrom
milescoviello:fix/sample-order-depth-copy-aspects

milescoviello commented Oct 21, 2026 •

edited

Loading

Uh oh!

DavidSamir commented Oct 21, 2026 •

edited

Loading

Uh oh!

EagleBearerYS commented Oct 21, 2026

Uh oh!

milescoviello commented Oct 21, 2026

Uh oh!

xXPBSXx commented Oct 21, 2026

Uh oh!

milescoviello commented Oct 21, 2026

Uh oh!

github-actions Bot commented Oct 22, 2026 •

edited

Loading

Uh oh!

milescoviello commented Oct 23, 2026

Uh oh!

Uh oh!

adriadam10 commented Oct 23, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

milescoviello commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Tested

Checklist

Uh oh!

DavidSamir commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

EagleBearerYS commented Oct 21, 2026

Uh oh!

milescoviello commented Oct 21, 2026

Uh oh!

xXPBSXx commented Oct 21, 2026

Uh oh!

milescoviello commented Oct 21, 2026

Uh oh!

github-actions Bot commented Oct 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

milescoviello commented Oct 23, 2026

Uh oh!

Uh oh!

adriadam10 commented Oct 23, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

milescoviello commented Oct 21, 2026 •

edited

Loading

DavidSamir commented Oct 21, 2026 •

edited

Loading

github-actions Bot commented Oct 22, 2026 •

edited

Loading