[EagleBearerYS]

What

Since #9533, a multisampled D32S8 target whose stencil has its own plane and whose HTILE is depth-only is imported as a native target. #9533 was merged together with #9717, in a4b76b2c0a.

Four titles that ran before that merge stop on main. JJenkx bisected each one to a4b76b2c0a: Curved Space (PPSA02817, #8691), NINJA GAIDEN: Ragebound (PPSA28002, #8671), Blue Prince (PPSA25009) and New Joe & Mac: Caveman Ninja (PPSA02801, #9091). On main they stop with:

AGC graphics: native multisample depth publication cannot distinguish an unwritten byte from a same-value guest store

What reaches the publication. JJenkx traced it in all four (#8671). Each title reuses the memory of a host-written native target for something new:

• draw input, in NINJA GAIDEN and Blue Prince;
• a colour target and a storage texture, in New Joe & Mac;
• another depth target, in Curved Space.

Why it stops. Building the new resource publishes the old target's planes to the guest (#13021).

• Guest writes are tracked per 4 KiB page. Once the title has stored into a page the host drew, every byte of that page that still holds its snapshot value is undecided, and the publication refuses.
• Curved Space and New Joe & Mac also meet a second refusal: a native multisample depth attachment over the planes of a host-written owner of another identity is not implemented.

The change:

• A cause on each depth read. It's DepthReadCause::Access, the default, or DepthReadCause::Reuse.
  • Reuse marks resources built over the memory: Draw.cpp's vertex and index buffer reads; the two storage-image accesses in ShaderResources.cpp, which also cover the colour-target lookup from refreshResidentTarget; and a new depth surface's three planes.
  • Everything else stays Access: CPU reads, DMA copies, texture reads, shader buffer accesses, dispatch-written ranges and the native sampling path.
• The undecided bytes. Where main refuses an undecided byte, a Reuse publication takes the guest's byte. Every byte the publication can decide is decided as before:
  • the host's bytes on pages the guest didn't write;
  • the guest's bytes where it changed them;
  • bytes covered by driver stores, as recorded.
• Retiring the old owner. After a Reuse publication commits, the old owner is retired: its guards and its owner entries are removed. It is then inactive and no longer host-written, so releaseRetired frees it at the next depth-surface creation (fix(agc): free native multisampled depth surfaces after their guest release #15300).
  • A later bind of its identity imports afresh.
  • A new depth attachment over its planes no longer meets the second refusal.
  • An Access publication leaves the owner resident, as before.
• The AGC debt entry for the publication now says which accesses take the guest's undecided bytes, and that those bytes are inferred.

Why the guest's byte. Before #9533 these targets were not native, so a resource built over their memory read the guest's bytes everywhere.

• This change takes the guest's byte only for the bytes the publication can't decide, and only when the memory is reused.
• So it is never further from the pre-fix(agc): import native multisample D32S8 targets with a separate stencil plane and a depth-only HTILE #9533 result than main is, since main stops. It keeps the host's bytes everywhere the publication can decide them.
• JJenkx's experiment at that check (#8671): Blue Prince renders its intro only with the guest's byte, and with the host's its screenshots are black. NINJA GAIDEN runs either way.

What isn't proven. The untouched bytes on a page the title wrote are inferred, not proven. The write tracking can't tell a byte the title never touched from one it stored with the same value.

• Exact store ranges for guarded planes would decide them. That's the exact follow-up, and it's out of scope here.
• CPU reads and DMA copies keep refusing.

Tested

Linux, on a cloud machine: Ubuntu 24.04, GCC 13.3, CI's configure flags, lavapipe (Mesa 25.2.8, LLVM 20.1.2). The tests ran as an unprivileged user with ANYPS5_REQUIRE_VULKAN=1 GALLIUM_OVERRIDE_CPU_CAPS=avx.

New cases in agc_driver_native_depth_publication (CheckReusePublication):

• The fixture: a 4-sample D32S8 target with a separate stencil plane and a depth-only HTILE. The host draws depth and stencil, then one CPU store leaves the rest of the stencil plane's first page undecided.

• Red on main acf7ac208a, green here. On main all four fail with the refusal above.

  • (0) a vertex buffer read over the page, and (1) an index buffer read;
  • (2) a storage-image lookup, the colour-target path;
  • (3) a new depth target of another identity over the planes.

  The checks: the guest's bytes on the undecided page, the host's bytes elsewhere in the plane, every depth sample, and the old owner retired. A rebind of its identity takes a fresh import. In (3) the new surface imports.

• The same on both:

  • (4) a CPU read and (5) a DMA copy still refuse, and keep the owner and the guest's bytes;
  • (6) a rebind of the same identity publishes nothing;
  • (7) a shader buffer access and (8) a texture read, with nothing undecided, publish the host's bytes, keep the owner and retire nothing.

The existing tests, 49 in all: the native depth, write-tracking, guest memory and multisample tests (ctest -R 'native_depth|native_multisample|depth_publication|write_track|guest_memory|multisample' --repeat until-fail:3).

• On main and here: 47 pass and 2 skip, ×3. The skips are agc_driver_multisample_centered_lines and agc_driver_native_depth_bda_clear, which skip themselves on lavapipe.
• No existing test was changed.

Mutations:

• Making every publication Reuse fails (4)-(8).
• Dropping the retirement fails (0)-(3). (3) then stops at the second refusal.

Checks: tools/check_test_registration.py build and tools/check_conventions.py --base origin/main pass.

The red/green is from the cloud run (lavapipe, CI flags). The GitHub ×10 runs the branch only, because main lacks DepthReadCause.

On today's main: the four commits were rebased onto 9ab6a281ea with the same lines. GitHub's runners, ×10 on that tree (run), the same tests: on Linux, 52 tests, 50 pass and 2 skip; on Windows, 53 tests, 8 pass and 45 skip. The libc and libkernel suites pass on both.

With #15300, #15301 and #15855: the branch merged with main 9f35c937da and #15855 (since merged), the same tests, ×10 on GitHub's runners (run): on Linux, 52 tests, 50 pass and 2 skip; on Windows, 53 tests, 8 pass and 45 skip. The libc and libkernel suites pass on both.

GitHub's Linux and Windows runners, ×10, on acf7ac208a (run). The tests matching native_depth|write_track|guest_memory|multisample:

• Linux: 49 tests, 47 pass and 2 skip, ×10. The skips are the same two lavapipe ones. agc_driver_native_depth_publication passed 10 of 10.
• Windows: 50 tests, 8 pass and 42 skip, ×10.
  • The runner has no Vulkan device, so the AGC driver tests skip themselves, the new cases included.
  • The five agc_write_tracking tests and three guest_memory tests run and pass.
• Both: the libc and libkernel suites pass.

Validation layers (CONTRIBUTING's AGC rule; run). VK_LAYER_KHRONOS_validation ran on GitHub's Linux runner, with main's tree and this branch each running the 35 tests matching native_depth|multisample.

• Each had 33 passes and 2 skips. The layer loaded in both.
• The only VUID is VUID-VkDeviceCreateInfo-pNext-pNext, on main and here. This change adds none.
• llvmpipe (LLVM 20.1.2), Mesa 25.2.8, Ubuntu 24.04.5.

The titles, run by JJenkx on this head merged onto main c4469b60da: Linux 7.2.11, AMD Radeon RX 6650 XT (RADV, Mesa 26.2.6), Release, APS5_PLAYGO_UNLISTED_CHUNKS=installed, 125 s with Cross every 15 s from 30 s. On that main each title stops with the refusal above. With this PR each one gets past it and runs to the scheduled close, exit 0:

• NINJA GAIDEN: Ragebound (PPSA28002) (#8671): stops at 50 s on main. Here it runs 122 s, and the story intro plays at 60 s.
• Curved Space (PPSA02817) (#8691): stops at 18 s on main. Here it runs 121 s, and the "Keep Going?" campaign screen is rendered at 60 s.
• New Joe & Mac: Caveman Ninja (PPSA02801) (#9091): stops at 8 s on main. Here it runs 123 s, and at 60 s the "Classic Modes" menu, opened with Cross, is shown.

AMD Radeon RX 570 (RADV POLARIS10, GFX8), run by ryanc2112 (comment): this head merged onto main 7ad9749ed0, Mesa 26.2.5, Release, with VK_LAYER_KHRONOS_validation.

• The 17 agc_driver_native_depth_* tests, 3 runs each: on GFX8, 11 of them fail on main, agc_driver_native_depth_publication among them, because the sample shadow lacks depth-attachment usage; fix(agc): give a multisampled depth-stencil sample shadow attachment usage #15750 fixes that. This PR alone gives the same 6 passes and 11 failures as main. With fix(agc): give a multisampled depth-stencil sample shadow attachment usage #15750, all 17 pass on main, and with this PR added all 17 still pass, the new CheckReusePublication cases included. None of these runs reports a VUID- error.
• The AGC, shader, VideoOut and Vulkan tests, this PR on main, once each: the same results as main, with 1,313 passing, the same 14 failing and 40 skipped, and the same 26 pAttachments-00605 validation errors on both sides (AGC: pipelines use differing per-attachment blend states while independentBlend is never enabled (VUID-VkPipelineColorBlendStateCreateInfo-pAttachments-00605) #11990). The 14 failures are known GFX8 or driver issues: the 11 above (fix(agc): give a multisampled depth-stencil sample shadow attachment usage #15750), agc_driver_f32_fma_denormal_mode2 (agc_driver_f32_fma_denormal_mode2 fails on RADV before RDNA2: Mesa splits GLSL Fma, even with NoContraction #15316), agc_driver_tessellation (agc_driver_tessellation segfaults on RADV with Mesa 26.2+: nir_opt_varyings corrupts the TCS input offset's use list (Mesa bug) #5444) and agc_driver_sample_order_recipe_depth_locations (AGC: agc_driver_sample_order_recipe_depth_locations fails on RADV Polaris: the recorded 2x depth draw is mostly at the standard locations, its recipe replay at the custom ones #9759).

Pending: JJenkx's run of Blue Prince (PPSA25009) on this branch: how far it gets, any new stop line, and whether its intro renders. Blue Prince is the title whose intro renders only with the guest's byte (above), so its run checks the byte this PR takes. This PR stays a draft until it's in.

Overlap

• feat(agc): publish a written native multisample D32S8 surface to its guest planes on a guest read #13021 (JAB-dev, merged) is the publication this extends. Its structure and its refusals for plain reads are unchanged.
• fix(agc): import a native multisample D32S8 target whose stencil plane the guest stored its clear value into #13023 (JAB-dev) has adjacent edits in DepthSurface.hpp, DepthSurface.cpp (prepareDepthSampling) and NativeMultisampleDepth.cpp, and merges clean (git merge-tree, both on 9ab6a281ea).
  • With fix(agc): import a native multisample D32S8 target whose stencil plane the guest stored its clear value into #13023, a store during the import no longer counts as written, so fewer bytes are undecided. This PR's rule is unchanged.
• fix(agc): publish a written native multisample D32S8 owner without HTILE to its depth and stencil planes #15688, fix(agc): attach a smaller native multisample depth extent with the same Z_X layout to its owner #15690 and fix(agc): release a published native multisample depth owner whose plane a written storage view covers #15707 (hugodearaujo) are a stack for D32S8 owners without HTILE, from Assassin's Creed Mirage (PPSA13960). This PR's targets have a depth-only HTILE, but its rule is in the publication they share, so it reaches their owners too.
  • fix(agc): publish a written native multisample D32S8 owner without HTILE to its depth and stencil planes #15688 publishes a written two-plane owner to its guest planes. The two compose: its publication fills the planes, and this PR changes what a reuse publication does with undecided bytes and what happens to the owner afterwards.
  • fix(agc): attach a smaller native multisample depth extent with the same Z_X layout to its owner #15690 attaches a smaller extent with the same Z_X layout to its owner. That bind then creates no new depth surface, so this PR's reuse publication doesn't run for it. This PR still applies to the binds of another identity that fix(agc): attach a smaller native multisample depth extent with the same Z_X layout to its owner #15690 keeps refusing, such as a larger extent, another layout or an owner with HTILE.
  • fix(agc): release a published native multisample depth owner whose plane a written storage view covers #15707 releases a published owner whose plane a written storage view covers, after cachedStorageTexture publishes it. With this PR, that storage access is a reuse, so the publication has already retired the owner and removed its guards, and fix(agc): release a published native multisample depth owner whose plane a written storage view covers #15707's release skips it as retired. They differ where a storage view overlaps an owner without covering a plane: fix(agc): release a published native multisample depth owner whose plane a written storage view covers #15707 keeps that owner, and this PR retires it after the reuse publication.
  • Their new cases and this PR's sit at the same place in NativeMultisampleDepth.cpp, and keeping both sides resolves that file. The rest merges clean.
• fix(agc): reclaim published native depth owners safely #10469 (Ondra97) publishes D32 owners through the same PublishToGuest, and reclaims a D32 owner once its publication has completed and its sampled views are released.
  • It conflicts with this PR in PublishToGuest's signature, where this PR adds the cause and fix(agc): reclaim published native depth owners safely #10469 adds the D32 check, and in the debt entry, which both rewrite. Keeping both sides resolves them.
  • With both, a reuse publication of a D32 owner would also take the guest's undecided bytes and retire the owner. That combination hasn't been run.
• fix(agc): import a depth plane aliased by a cleared R32F target #15443 and fix(agc): accept a stored refill of a pending HTILE clear #15446 (edubox) add cases at the same place in NativeMultisampleDepth.cpp, and keeping both sides resolves it.
• fix(agc): free native multisampled depth surfaces after their guest release #15300 (Prophet069, merged) lets releaseRetired free a native multisample surface once it's inactive and no longer host-written. Before it, a retired multisample owner stayed listed. After this PR's reuse publication the old owner is both, so fix(agc): free native multisampled depth surfaces after their guest release #15300 frees it. fix(agc): let a sampled depth shadow reclaim memory or fall back to the host #15301 (Prophet069, merged) changes the sampled depth shadow's allocation in DepthSurface.cpp, away from this PR's hunks. The branch merges clean with both and with fix(agc): allocate the sampled depth shadow through the image allocator #15855, and with main acac7c8541.
• fix(agc): retire a depth surface once a surface of another extent takes its planes #15404 (JJenkx) retires displaced single-sample surfaces, while this PR's retirement is for multisample native owners after a reuse publication. They merge clean.
• fix(agc): keep guest writes over a stale stencil plane #14969 (JJenkx) keeps guest writes over a stale stencil plane, in DepthSurface.cpp and the AGC debt file. Its stencil baseline runs only for surfaces whose stencil plane can be mirrored, which are single-sample, so this PR's multisample reuse path doesn't reach it. It leaves retireStencil unchanged. They merge clean.
• I#15600 (oneandonlydean): feat(agc): publish a written native multisample D32S8 surface to its guest planes on a guest read #13021's commit-time window applies here unchanged.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none (the four regressions were reported in discussions Curved Space (PPSA02817) #8691, NINJA GAIDEN: Ragebound (PPSA28002) #8671 and New Joe & Mac: Caveman Ninja (PPSA02801) #9091). Related: fix(agc): import native multisample D32S8 targets with a separate stencil plane and a depth-only HTILE #9533, feat(agc): sample the stencil plane of a native multisampled D32S8 target #9717, feat(agc): publish a written native multisample D32S8 surface to its guest planes on a guest read #13021, fix(agc): import a native multisample D32S8 target whose stencil plane the guest stored its clear value into #13023, agc: a guest store of the same value can be lost while a native multisample depth publication commits #15600
• Unblocks: NINJA GAIDEN: Ragebound (PPSA28002), Curved Space (PPSA02817) and New Joe & Mac: Caveman Ninja (PPSA02801) past the publication refusal, each to the close of a 125 s run (JJenkx's runs above); Blue Prince (PPSA25009) pending
• GPUs tested: Mesa llvmpipe (LLVM 20.1.2), Ubuntu 24.04, on a cloud Linux machine and GitHub's Linux runner; GitHub's Windows runner has no Vulkan device; AMD Radeon RX 570 (RADV POLARIS10, GFX8; ryanc2112); AMD Radeon RX 6650 XT (RADV; JJenkx's title runs)
• AI-assisted: yes

[ryanc2112]

A GFX8 data point: AMD Radeon RX 570 (RADV POLARIS10), Mesa 26.2.5, SteamOS 3.8.28 (Arch distrobox), GCC 16.2.1, Release. Head 7d377e40ba merged onto main 7ad9749ed0 (clean merge), with ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1 VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation and ctest -V --no-tests=error. No log reports was not found but was requested by env var VK_INSTANCE_LAYERS.

On GFX8, 11 of the 17 agc_driver_native_depth_* tests fail on main because the sample shadow lacks depth-attachment usage. That includes agc_driver_native_depth_publication, which fails with native D32S8 publication lost live samples, HTILE, padding or guest writes. #15750 fixes it, so I also ran this PR on top of #15750 to let the new CheckReusePublication cases run here. The two merge together cleanly.

agc_driver_native_depth_* (17 tests), 3 runs each:

	Result	VUID- errors
main	6 pass, 11 fail (the GFX8 failures #15750 fixes)	0
This PR	the same 6 pass and 11 fail	0
main + #15750	17 of 17 pass	0
main + #15750 + this PR	17 of 17 pass, publication included	0

The 1,327 AGC, shader, VideoOut and Vulkan tests, this PR on main, once each:

• Same results as main: 1,313 pass, the same 14 fail and 40 skip.
• Validation errors unchanged: 26 pAttachments-00605 (AGC: pipelines use differing per-attachment blend states while independentBlend is never enabled (VUID-VkPipelineColorBlendStateCreateInfo-pAttachments-00605) #11990) on both sides.
• The 14 failures are known GFX8 or driver issues: 11 agc_driver_native_depth_* (fix(agc): give a multisampled depth-stencil sample shadow attachment usage #15750), agc_driver_f32_fma_denormal_mode2 (agc_driver_f32_fma_denormal_mode2 fails on RADV before RDNA2: Mesa splits GLSL Fma, even with NoContraction #15316), agc_driver_tessellation (agc_driver_tessellation segfaults on RADV with Mesa 26.2+: nir_opt_varyings corrupts the TCS input offset's use list (Mesa bug) #5444) and agc_driver_sample_order_recipe_depth_locations (AGC: agc_driver_sample_order_recipe_depth_locations fails on RADV Polaris: the recorded 2x depth draw is mostly at the standard locations, its recipe replay at the custom ones #9759).

AI-assisted: yes (Claude Code)

[EagleBearerYS]

@JJenkx thank you for the three title runs, and @ryanc2112 for the RX 570 run. I've added both to the description.

• NINJA GAIDEN: Ragebound (PPSA28002), Curved Space (PPSA02817) and New Joe & Mac: Caveman Ninja (PPSA02801) stop at the publication refusal on main c4469b60da, at 50 s, 18 s and 8 s. With this PR each one runs to the scheduled close of the 125 s run and exits 0.
• On the RX 570, this PR on top of fix(agc): give a multisampled depth-stencil sample shadow attachment usage #15750 passes all 17 agc_driver_native_depth_* tests, the new reuse cases included, with no VUID- errors, and the wider AGC, shader, VideoOut and Vulkan set matches main.

The PR stays a draft until Blue Prince (PPSA25009) is in, since its intro is the case that needs the guest's byte.

[AlfaRojo]

Windows check of head 7d377e40ba, merged onto main 3cb1cd29ec (cfcaf07ecf, clean merge), on a discrete NVIDIA and an integrated AMD GPU, since the PR's own runs were on lavapipe.

Setup: Windows 11 Pro (10.0.26200), WinLibs GCC 15.2.0 r7 UCRT, Release, ANYPS5_AGC_EXECUTION_TESTS including NativeMultisampleDepth, ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1, ANYPS5_GPU selecting the device (confirmed by Selected GPU); no validation layer on this machine.

GPU	agc_driver_native_depth_publication	ctest -R 'native_depth|native_multisample' (18 tests)
NVIDIA GeForce RTX 4070 (10de:2786, driver 32.0.16.1692)	3 of 3 pass	17 pass, 1 skip
AMD Radeon Graphics 1002:13c0 (RDNA 2 integrated, driver 32.0.21043.5001)	3 of 3 pass	17 pass, 1 skip

The skip is the pre-existing skipped native HTILE write-coverage assertion: the write tracker does not cover the imported metadata range in CheckDepthTarget, which main prints on these GPUs too. No red arm: the new cases use DepthReadCause from this PR's headers, so the test does not build against main's code. No title run.

AI-assisted: yes.