[hugodearaujo]

What

Since #8898, ResolveTextureFormat gives format 36 (11_11_10_FLOAT) its storage form, R32_UINT, which the shader packs and unpacks. A sampled format 36 image is still read through its native B10G11R11_UFLOAT view (SampledTextureFormat, and nativeRepacked in ResourceMaterializer). MultisampleTargetTexture and MultisampleTargetFirstSample kept matching the resident multisampled target by ResolveTextureFormat, so a 4x COLOR_10_11_11 float target, which DecodeColorBuffer renders as B10G11R11_UFLOAT (the same bits as format 36), is refused when it is read through a format 36 texture descriptor of the same shape:

compute shader 0x10b4e03d00: AGC graphics: sampling multisampled color target 0x10e0720000 (960x540, 4 samples, vk format 122) as a 960x540 texture of 4 samples and guest format 36 (vk 98) is not implemented

Assassin's Creed Mirage (PPSA13960) stops on this line 86 to 91 s after start on main.

Both functions now match and fill a multisampled texture in SampledTextureFormat. The repacked format 43 keeps ResolveTextureFormat, as before: its sampled view needs repacked words, which the multisampled path does not produce. For every other format the two give the same Vulkan format, except the 8 and 8_8 sRGB formats on a device that cannot sample them, which were refused before and still are.

agc_driver_sample_multisample_target gains a case that draws a 4x COLOR_10_11_11 float target (0.25, 0.5, 0.125) and reads every sample through format 36 Dim2DMsaa and Dim2DMsaaArray descriptors, and sample 0 through DIM 2D image_load, image_sample_lz and image_sample_l.

Tested

Linux 7.2.11, GCC 16.2.1, Ninja, Release, AMD Radeon RX 9070 XT (RADV, Mesa 26.2.5). The tests ran on main abed035 and this branch on top of it; MultisampleTarget.cpp and the test file are the same on main a120df3, where the title runs below were made.

• ctest -V -R '^agc_driver_sample_multisample_target$' with VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation, VK_LOADER_DEBUG=layer, ANYPS5_NO_SHADER_CACHE=1, ANYPS5_REQUIRE_VULKAN=1:

	result	VUID- lines	validation layer inserted
main	passed	0	yes
main + the new case only	failed: AGC graphics: sampling multisampled color target 0x55eec24a0000 (16x16, 4 samples, vk format 122) as a 16x16 texture of 4 samples and guest format 36 (vk 98) is not implemented	0	yes
this branch	passed	0	yes

No run printed was not found but was requested by env var VK_INSTANCE_LAYERS.

• The same settings over 34 multisample, FMASK and converted-format tests (agc_driver_msaa_layout, agc_driver_msaa_dcc_clear_{2,4,8}, agc_driver_msaa_dcc_compared_import_{2,4,8}, agc_driver_multisample_* including the readback variants, agc_driver_native_fmask, agc_driver_native_msaa_vertex_input, agc_driver_native_multisample_depth, agc_driver_fmask_image_reads, agc_driver_cached_image_multisample_guard, agc_driver_image_msaa_address_on_plane, agc_driver_msaa_enable_single_sample, agc_driver_multisampled_pass_order, agc_driver_packed_unorm4444_multisample_partial_write, agc_driver_sample_mask_multisample, agc_driver_sample_multisample_target, agc_driver_image_converted_unorm, agc_driver_image_repacked_float, image_msaa_load): 34 of 34 pass on main and on this branch, 0 VUID- lines in both; the loader inserted the layer in the 32 tests that open a Vulkan device (agc_driver_msaa_layout and image_msaa_load don't).

• Assassin's Creed Mirage (PPSA13960), main a120df3 and main + this change, otherwise the same build, Cross every 30 s: main stops at 87 s on the line above. With this change that line is gone, and the title stops at 83 s on the next refusal: presenting reads the 0x1fe0000-byte surface at 0x10dfd20000, which overlaps the same 4x target:

  AGC graphics: a CPU read of 0x10dfd20000+0x1fe0000 overlaps multisampled color target 0x10e0720000+0x870000 (4 samples), whose samples are only on the host: guest access to it is not implemented
  

  That is the overlap fix(agc): retire a multisampled color target the title wrote over #8223 and fix(agc): retire a multisampled color target a write-only storage image overwrites #10049 work on. With fix(agc): retire a multisampled color target the title wrote over #8223 (head 331399d) merged on top, that read passes too, and the title stops at 87 s on the partial clear described in fix(agc): retire a multisampled color target the title wrote over #8223's thread: compute shader 0x1000810a00: AGC graphics: a cached image access of 0x10dc1d0000+0x480000 overlaps multisampled color target 0x10dc450000+0x480000 (4 samples), whose samples are only on the host: guest access to it is not implemented. Two other runs with this change (one of them with fix(agc): retire a multisampled color target the title wrote over #8223 merged too) segfaulted on the main thread 3 s after start, before any GPU work: the start-up crash fix(linux): enter the guest entry on a main stack at the console's address #7532 addresses.

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 in their library's technical debt file
• No notes, investigation .md files or images added to the repository
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none (regression from fix(shader_recompiler): pack format 11_11_10 float image stores in shader #8898)
• Unblocks: PPSA13960 Assassin's Creed Mirage (PPSA13960) #7534 (Assassin's Creed Mirage), the stop above
• GPUs tested: AMD Radeon RX 9070 XT (RADV, Mesa 26.2.5)
• AI-assisted: yes

[github-actions]

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 6808	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp core/libs/prx/libSceAgcDriver/tests/execution/SampleMultisampleTarget.cpp
PR 6971	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp
PR 8223	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp core/libs/prx/libSceAgcDriver/tests/execution/SampleMultisampleTarget.cpp
PR 9533	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp
PR 9700	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp
PR 9717	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp
PR 9742	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp
PR 10049	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp
PR 10233	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp core/libs/prx/libSceAgcDriver/tests/execution/SampleMultisampleTarget.cpp
PR 11920	core/libs/prx/libSceAgcDriver/Graphics/src/MultisampleTarget.cpp

[hugodearaujo]

I checked the six PRs the overlap comment could not compare against this branch. #8223, #10049, #10194 and #10233 merge with it without conflicts in MultisampleTarget.cpp and SampleMultisampleTarget.cpp. #6808 and #6971 conflict in MultisampleTarget.cpp with code current main changed after they were written (collectBlocks/fmaskSnapshotResource, absorbDccFill, OrderView), not with the lines this PR changes; neither changes the format MultisampleTargetTexture and MultisampleTargetFirstSample match against (#6971 adds the viewType parameter main already has). The title run with #8223 merged on top is in Tested.

[jotitadev]

Linux run on NVIDIA and lavapipe, since the description covers RADV only. Head f4274307ec on main b986c11a34 (patch applies cleanly). Ubuntu 26.04, GCC 15.2, Release. ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1, -j1. lavapipe selected with VK_ICD_FILENAMES=.../lvp_icd.json.

agc_driver_sample_multisample_target:

	RTX 3070 (NVIDIA 595.99.02)	lavapipe (Mesa 26.0.3)
main + the new case only	fails: sampling multisampled color target ... (16x16, 4 samples, vk format 122) as a 16x16 texture of 4 samples and guest format 36 (vk 98) is not implemented	same failure
this head	passes	passes

Full ^agc suite (without agc_display_buffer_dcc, agc_driver_flip, agc_driver_lds_atomics64, --timeout 30) with #11692 and #11691 applied together on main: 1041/1042 on both devices. The one failure on each also fails on unmodified main on this machine, with the same message: agc_driver_pipeline_journal on the 3070 (the oversized record was not counted as a write failure; it passes when only the NVIDIA ICD is visible), and agc_driver_multisample_retirement on lavapipe (MSAA storage promotion did not retain its source image).

GPUs tested: NVIDIA GeForce RTX 3070 (proprietary 595.99.02, Linux), lavapipe (Mesa 26.0.3, LLVM 21).

Written with AI assistance (Claude Code). I ran the builds and tests above on my machine.

[MarcZX10]

Reviewed exact head f4274307ecd56b995e43983780601bf6f70f2670; approve.

The format selection is consistent with the resource materializer: non-repacked descriptors use SampledTextureFormat, including format 36's native B10G11R11_UFLOAT view, while repacked formats retain ResolveTextureFormat. Both target matching paths (multisample sampling and first-sample reads) use the same selection. The regression covers 4x COLOR_10_11_11 rendered storage read through format 36 using both Dim2DMsaa and Dim2DMsaaArray, plus sample-0 image load/sample operations. I found no conflicting change to these comparisons in the manually checked overlap cases.

Validation evidence: the author reports the focused main/new-case/change comparison and 34 related MSAA tests on RADV with Khronos validation and zero VUIDs. An independent reviewer reports the focused failing-main/passing-change reproduction and an AGC suite run on both RTX 3070 and lavapipe; the sole suite failures also reproduce on unchanged main. Hosted product jobs are skipped by path policy. I could not run a fresh local GPU test here because this environment exposes no GPU device.

[maskym]

Windows + NVIDIA result for head f4274307ec, merged into main f759e165d9 (the PR's GPUs tested line names no NVIDIA run on Windows).

This PR's own tests: agc_driver_sample_multisample_target (1.18 s), all passed.

Host: Windows 11 26300, Intel i7-12700K, NVIDIA GeForce RTX 2070 SUPER (vendor 0x10de, device 0x1e84; Vulkan loader 1.4.357 from the NVIDIA driver), Khronos Validation Layers nightly 1.4.366 (b2b44c0d37) through VK_ADD_LAYER_PATH, WinLibs GCC 15.2.0 (x86_64-ucrt-posix-seh, r6), Release, BUILD_TESTING=ON, Ninja. Build clean.

Same focused run on main f759e165d9 and on this tree, same machine and GPU: VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation, ANYPS5_NO_SHADER_CACHE=1, ANYPS5_REQUIRE_VULKAN=1, VK_LOADER_DEBUG=error,warn,layer, ctest --test-dir build -R '^(agc|guest_video_out|video_out)' --parallel 4 -V.

• main: 1127/1128 passed (9 skipped), the one failure being agc_driver_mesh_barycentric_emulation (AGC: agc_driver_mesh_barycentric_emulation fails since #9193 on devices with mesh shaders and VK_KHR_fragment_shader_barycentric #12121, pre-existing).
• This tree: every pre-existing test gives the same result; the loader inserted VK_LAYER_KHRONOS_validation in every test (998 and 1012 insertions), neither log contains was not found but was requested by env var VK_INSTANCE_LAYERS, and both report exactly one VUID, VUID-VkPipelineColorBlendStateCreateInfo-pAttachments-00605 18 times (pre-existing on main, AGC: pipelines use differing per-attachment blend states while independentBlend is never enabled (VUID-VkPipelineColorBlendStateCreateInfo-pAttachments-00605) #11990 / fix(agc): enable independentBlend for color targets with different blend states #6782): this PR adds no new VUID- error.

The tree also carried the open PRs #11951, #11711, #11704, #11692, #11690, #11794, #11949, all merged cleanly, so the run is a joint one; of those, #11692's new agc_driver_depth_plane_cpu_write_after_page_unwatch fails on this host, reported on that PR. Not run on a PS5.

AI-assisted: yes (Claude Code), the runs and the numbers above are from this machine.

[adriadam10]

Checked the format selection against SampledTextureFormat/ResolveTextureFormat in TextureFormat.cpp: format 36 resolves to its native B10G11R11_UFLOAT view for sampling, repacked format 43 keeps ResolveTextureFormat, so both multisample paths now match the materializer. The new case reads the 4x COLOR_10_11_11 target through format 36 (Dim2DMsaa, Dim2DMsaaArray, sample 0) and fails without the change, as shown by the author on RADV and by @jotitadev on NVIDIA and lavapipe; merged.