[JJenkx]

What

Cocoon (PPSA08766, #6651) reaches a fragment program with DIM 2D_MSAA image_loads bound to an ordinary 2D T#. Main refuses it with image descriptor is incompatible with the static runtime image interface: the descriptor is 2D, while all compiled modes are multisampled.

For a mip-less 32-bit image load, add the ordinary 2D runtime mode. The existing emitter then reads x/y at the viewed base level and ignores the sample operand, as the measured hardware does. The genuine MSAA mode remains available. This does not change stores, atomics, A16 or mip-bearing addresses, packed accesses, converted formats, depth-bit views, array-MSAA addresses or image-table handling.

Only five production lines change. No new metadata fields, request version or runtime ABI is introduced; the existing shader-source version invalidates old artifacts.

Hardware evidence

Measured directly with tools/hw-oracle on RX 6650 XT, gfx1032: wave32 and wave64, R32_UINT / R32_FLOAT / RGBA8_UNORM / RGBA8_SRGB, sample indices 0, 1, 2, 7, 31 and 0xffffffff, and in-range plus out-of-range x/y. Every one of 768 rows matches the same ordinary DIM 2D load on the same 4x4 linear 2D T#. Integer/float markers separate the addressed texels; out-of-range coordinates return zero.

All FP controls were explicit: IEEE 0, DX10 clamp 1, denorm32 0, denorm16 3, round32/16 0, FP16 overflow 0. This is gfx1032 evidence, not a console/gfx1013 measurement. The scope and remaining refusals are recorded in shader debt.

Tested

Standalone head d513fb987f731207cbd5503ec3ceccceb7bbc9f0: two linear commits on main 82ddae75b676, no dependencies. Arch Linux 7.2.9-arch1-1, GCC 16.2.1, Ryzen 9 7950X.

• Full runtime libs and all four selected test executables build.
• RX 6650 XT / RADV 26.2.4-arch1.1: 4/4 pass, no skips: agc_driver_image_msaa_address_on_plane, agc_shader_memory, agc_driver_image_volume_address_on_plane, agc_driver_sample_multisample_target.
• Lavapipe Mesa 26.2.4 / LLVM 23.1.1: the same 4/4 pass, no skips.
• New GPU regression executes both guest wave sizes and all four formats. Each MSAA-address result must match that driver's ordinary 2D control bit-for-bit. Integer/float markers and out-of-range zeros are exact. Normalized rows are pinned to the oracle; sRGB checks preserve the measured encoded 8-bit texel. An initial absolute sRGB tolerance failed on lavapipe's ordinary 2D conversion too (0.051499 versus the Radeon oracle's 0.051270); the final check keeps exact per-driver control equality and verifies the encoded texel instead of requiring identical cross-driver linear approximations.
• Negative control: replace only ResourceMaterializer.cpp with main's version, retaining the final test; it fails at the original static-runtime-image-interface refusal. Restore the fix and rebuild: the selected tests pass.
• Conventions and whitespace pass.

Related #6635 adds genuine multisampled-array image views; it does not implement ordinary 2D T# loads and is not a dependency.

Independent NVIDIA validation

jotitadev tested this head merged onto main bc400bede on an RTX 3070, NVIDIA 595.99.02 / Vulkan 1.4.329, Ubuntu 26.04 (kernel 7.0.0-38), GCC 15.2. The four selected checks passed in three runs on NVIDIA and three runs on lavapipe Mesa 26.0.3 / LLVM 21.1.8. The new check exercised all four formats on NVIDIA. With the test retained and main's ResourceMaterializer.cpp, the new check failed on both devices; the other three passed.

Their broader NVIDIA run covered 616 tests. agc_driver_store_over_untracked failed on both this change and unchanged main (#6347); agc_driver_stencil_ref_export skipped. They did not test Windows, titles, or additional hardware-oracle rows. These are the reviewer's results, separate from the local Radeon measurements above.

Native bring-up

The Cocoon diagnostic stack carries #6639/#6801, #6653 and #6743. They are not code dependencies of this PR. Baseline c283a7718396 refuses the descriptor at 4.03 seconds. Integration b83e429e7613, which adds only this change and its regression, builds full runtime libraries and passes 7/7 focused checks. The title passes that image refusal and aborts at 6.03 seconds on a separate missing PA_SU_POINT_SIZE context register. Both runs exit 134 and restore temporary settings. No matching game window or visible gameplay is confirmed.

Hardware rows, reproduction script, exact sources and sanitized validation. A separate private capture of the failing fragment request contains 136 code words and two DIM 2D_MSAA loads; no guest bytes, memory capture or full title logs are published.

No FPS, load-time gain, newly playable title, Windows or console result is claimed. Full hosted CI remains for collaborator dispatch.

A separate isolated Override 2 (PPSA01907 v01.007.000) check restacks these two unchanged topic commits onto main 66e290f1c as 66a91f2c2. Full libs builds; the new GPU test passes 1/1 on RADV and 1/1 on lavapipe, zero skips. Both native runs exit 0 and restore temporary settings. The cold-cache run only supplies startup captures; the warm follow-up reaches the menu but its 85 s capture still has very dark 3D output. This change does not resolve that observation, and no Override compatibility or performance gain is claimed. The 19 locally captured compute requests replay successfully; they do not cover the active graphics requests. Sanitized report and screenshot: image-msaa-main66-validation-public.json, image-msaa-main66-menu.png.

Current-main refresh (2026-10-19 UTC)

Rebased the two topic commits onto main 659230f7b676e8c848d70640f823342be6704526, head 72959f7c0055c20871ed207868746f928e9cc26f, without merge commits. The sole conflict was additive shader debt; both the upstream VCC-sampler entry and this MSAA-load entry are retained. The five production lines and regression semantics are unchanged.

Full runtime libs and all four selected test executables build. With required Vulkan and VK_LAYER_KHRONOS_validation, the same four checks pass without skips on RX 6650 XT / RADV Mesa 26.2.4-arch1.1 (3.36 s total) and lavapipe Mesa 26.2.4-arch1.1 / LLVM 23.1.1 (3.79 s total), Arch Linux. Conventions and whitespace checks pass. Earlier hardware measurements, negative controls and collaborator runs above are historical evidence, not rerun measurements on this head. No additional title or performance result is claimed.

Contributor-reported Windows/NVIDIA validation by klosixgaming-arch: head 72959f7c0, two commits, cleanly merged onto main ad502d717; Windows 11, MinGW-w64 GCC 15.2.0, NVIDIA GeForce RTX 5090 driver 617.42, Ryzen 9 9950X3D. agc_driver_image_msaa_address_on_plane passes with the fix and fails with only the test on main at the descriptor-format-20 incompatibility named above. The contributor also reviewed the five-line production restriction. They did not test Cocoon or rerun the author’s gfx1032 hardware oracle. These are independently reported results, not local Windows runs; their review was AI-assisted and human-reviewed.

Reviewer-reported Apple compatibility (2026-10-19)

jotitadev tested head 72959f7c0 on Apple M2 Max under Rosetta with MoltenVK 1.4.2 and MoltenVK main plus SPIRV-Cross #2704. The MSAA-address-on-plane, shader-memory and volume-address checks pass three times on both drivers; the multisample-target check skips because storage image multisampling is unavailable. Reverting the production change fails the new test on both drivers. Their old absolute-oracle comparison also fails for four sRGB texels, while the current encoded-byte comparison passes.

This is reviewer-provided evidence from main a373e85e5 plus the macOS series, build prerequisites, several other PRs and a test-loader patch, as listed in that review; it is not a plain-main macOS build, our local execution, or a title/performance claim. No source change is requested by this report.

Independent Windows main/change title result: parl0v tested Cocoon PPSA08766 01.004.000 on Windows 10 (19045), NVIDIA RTX 3060 12 GB driver 617.42, main 64e2a810a and main plus this PR 72959f7c0. Main aborts on the static runtime image-interface mismatch; with this PR that refusal is gone and the run reaches the geometry-output refusal covered by #6801. Contributor report. Contributor-reported evidence, not independently reproduced on Windows; no gameplay claim.

Checklist

• Based on current main; no duplicate ordinary-2D load implementation found
• One topic per PR; follow-ups separate
• No comments in code except technical debt
• Unimplemented paths throw; no silent stubs added
• No investigation reports, images, game assets or raw logs in the repository
• New third-party code is a submodule built from source (none added)
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: PPSA08766 Cocoon (PPSA08766) #6651 (2D/MSAA image-interface refusal; a separate point-register blocker remains)
• GPUs tested: RX 6650 XT / RADV 26.2.4-arch1.1 and lavapipe Mesa 26.2.4 / LLVM 23.1.1, Arch Linux 7.2.9; hardware oracle on gfx1032; independently RTX 3070 / NVIDIA 595.99.02 and lavapipe Mesa 26.0.3 / LLVM 21.1.8 on Ubuntu 26.04 (jotitadev review linked above)
• 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 907	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 5021	docs/dev/debt/shader.md
PR 5371	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/debt/shader.md
PR 5426	docs/dev/debt/shader.md
PR 5500	docs/dev/debt/shader.md
PR 5941	docs/dev/debt/shader.md
PR 6009	docs/dev/debt/shader.md
PR 6235	docs/dev/debt/shader.md
PR 6280	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/debt/shader.md
PR 6604	docs/dev/debt/shader.md
PR 6947	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 6977	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/debt/shader.md
PR 7223	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/debt/shader.md
PR 7332	docs/dev/debt/shader.md
PR 7392	docs/dev/debt/shader.md
PR 7395	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/debt/shader.md
PR 7484	docs/dev/debt/shader.md
PR 7550	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 7642	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 7888	docs/dev/debt/shader.md

[jotitadev]

NVIDIA run, since the description lists NVIDIA as not run.

Machine: NVIDIA GeForce RTX 3070, driver 595.99.02, Vulkan 1.4.329, plus lavapipe (Mesa 26.0.3, LLVM 21.1.8) on the same machine. Ubuntu 26.04 (kernel 7.0.0-38), i7-8700, GCC 15.2, Release, BUILD_TESTING=ON. Head d513fb987 merged onto main bc400bede, clean merge.

GPUs tested: NVIDIA GeForce RTX 3070 (595.99.02), lavapipe (Mesa 26.0.3, LLVM 21.1.8), Ubuntu 26.04

Set: agc_driver_image_msaa_address_on_plane, agc_shader_memory, agc_driver_image_volume_address_on_plane, agc_driver_sample_multisample_target, -j1 --timeout 60.

• Test only (ResourceMaterializer.cpp from main): agc_driver_image_msaa_address_on_plane fails on both devices with image descriptor is incompatible with the static runtime image interface: descriptor format 20 .... The other three pass.
• With the change: 4/4 pass in 3 runs on the RTX 3070 and in 3 runs on lavapipe. On the 3070 the new test goes through all four formats (20, 22, 56, 130: R32_UINT, R32_FLOAT, RGBA8_UNORM, RGBA8_SRGB), so the per-driver sRGB control comparison also holds on NVIDIA.
• RTX 3070, ctest -R "^agc|shader|recompiler" -E "agc_display_buffer_dcc|agc_driver_flip|agc_driver_lds_atomics64|display|flip" -j1 --timeout 60: 616 tests. Only agc_driver_store_over_untracked fails, and it fails on unchanged main bc400bede here too (#6347). agc_driver_stencil_ref_export skips.

Not checked: Windows, titles, any hardware oracle rows beyond what the test pins.

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

[gjonn]

A data point from PPSA01317 on current main (Windows 11, RTX 5090 Laptop): this PR fixes one of its aborts. After about 50 s of boot, a shader does multisample image loads on a plain 1x1 2D R8 texture, and main refuses that. With this PR's two commits cherry-picked, the title gets past it, to about 900 to 1,100 frames.

AI-assisted: yes

[frankea]

RTX 4090 data point on PPSA01317, on top of today's runs on D4697 (same box, same stack): with this PR the image descriptor is incompatible with the static runtime image interface refusal on shader 0x1258db5000 is gone, 3 of 3 runs, and the title's next stop is the DIM 1D size query of a cube T# that #7550 describes.

Setup. Gran Turismo 7 PPSA01317 01.000.001 (base package), Windows 11 (26200), RTX 4090 (NVIDIA 616.92), AMD Ryzen 7 5800X, MinGW-w64 GCC 15.2.0 r7 (UCRT), Release libs target only (no tests built). app.exe is the relink from main 31cbb55 (--windows --registry --module-dir .); only libs/ is swapped between runs. Tree: main 48d61f5 (2026-10-18; it already carries #7137, #7207, #7175, #7238, #7240, #7271 and #7626) + #7256 (which brings #7254) + #7243 + #5958 + this PR's head d513fb9, every merge clean (the #5958 and #6815 ResourceMaterializer.cpp hunks do not touch), f9aef6b11. Three local environment-gated lines on top, none of which changes a result: a bypass that makes sceHmd2Initialize answer 0x81110016 again (current main aborts 5 to 12 s in on sceHmd2ReprojectionQueryBufferSizeAlign not implemented, the #7323 regression), a print of sceSslConnect's return, and a display-buffer print for #6350 that was off here. No controller, no input, one-shot interactive task per run, 300 s cap, box otherwise idle.

Result. The control (same tree without this PR, four runs earlier today) aborts 36 to 37 s after launch, every time, on

shader 0x1258db5000 stage 5: image descriptor is incompatible with the static runtime image interface: descriptor format 1 decodes to {class 1, dim 3, ...}; modes: {... dim 6 ...} x9; T# 12001710 00100000 00000000 90000000 00000000 00700000 00000000 00000000

(a 2D 8-bit T#, TYPE 9, read by a pixel shader whose only compiled modes are 2D MSAA). With this PR that shader compiles and runs in all three runs; no refusal of that wording anywhere. The runs then diverge in time, not in outcome:

run	shader cache	exit after	stop
control (4 runs)	warm	36 to 37 s	shader 0x1258db5000 stage 5: image descriptor is incompatible with the static runtime image interface
1	cold (2,811 misses, the source version changed)	300 s cap, killed	none: black screen, counter frozen at 860 frames / 8.27 FPS from +50 s to the kill, stdout still creating textures
2	warm	247 s	shader 0x126f962f00 stage 5: image address has too few coordinate components
3	warm	43 s	same

Runs 2 and 3 both end on ResourceMaterializer.cpp's 1D refusal. #7550 (gjonn) names this title and this situation: a pixel shader runs image_get_resinfo with DIM 1D on a 64x64 cube T#, and the tracker's 1D image has no mode for a cube descriptor. So on this box the open-PR chain for the title now reads #7243 -> #7254 / #7256 -> #5958 -> this PR -> #7550.

Two observations on the way there, neither caused by this PR as far as I can tell: the screen stays black after the "Sony Interactive Entertainment presents" splash in all runs (title-bar counter 59.94 FPS at 476 frames at +15 s, 12 FPS at about 1,320 frames by +35 s), and the frame counter stops advancing for long stretches (run 1: 860 frames from +50 s to the 300 s kill; run 2: 1,327 frames from +50 s to at least +180 s, then the abort at 247 s) while the guest keeps streaming textures. Run 1 was the cold run, so it may be the same cold-cache stall I saw on #7238; run 2 was warm, so it is not only that. Run 3 reached the same shader in 43 s without the pause.

sceSslConnect (now main's #7626) returns SSL_ERROR_WANT_POLLIN two to four times and then SSL_ERROR_UNKNOWN_CA on each of the title's seven connections in all three runs; the title goes on.

stderr of run 3 (warm, box idle) from the save-data mount to the abort, the per-shader ShaderRegistry lines, the sceKernelOpen misses and the cache and texture bookkeeping removed, the [shadow] lines collapsed

[SAVEDATA:native] mount3 user=268435456 dir=DRFILEIV mode=0x1 blocks=673 -> 0x809f0008 point= status=0
[SAVEDATA:native] getMemory2 user=268435456 slot=0 -> 0x00000000
[ssl] sceSslConnect(3) -> 0x8095f013
[ssl] sceSslConnect(3) -> 0x8095f013
[ssl] sceSslConnect(3) -> 0x8095f00c
[ssl] sceSslConnect(6) -> 0x8095f013
[ssl] sceSslConnect(7) -> 0x8095f013
[ssl] sceSslConnect(6) -> 0x8095f013
[ssl] sceSslConnect(6) -> 0x8095f013
[ssl] sceSslConnect(7) -> 0x8095f013
[ssl] sceSslConnect(6) -> 0x8095f00c
[ssl] sceSslConnect(7) -> 0x8095f00c
[ssl] sceSslConnect(10) -> 0x8095f013
[ssl] sceSslConnect(11) -> 0x8095f013
[ssl] sceSslConnect(10) -> 0x8095f013
[ssl] sceSslConnect(11) -> 0x8095f013
[ssl] sceSslConnect(10) -> 0x8095f013
[ssl] sceSslConnect(11) -> 0x8095f013
[ssl] sceSslConnect(10) -> 0x8095f00c
[ssl] sceSslConnect(11) -> 0x8095f00c
[ssl] sceSslConnect(14) -> 0x8095f013
[ssl] sceSslConnect(14) -> 0x8095f013
[ssl] sceSslConnect(14) -> 0x8095f013
[ssl] sceSslConnect(14) -> 0x8095f00c
[SOFT-NET] sceNetResolverStartNtoa -> host DNS lookup failed
[ssl] sceSslConnect(17) -> 0x8095f013
[ssl] sceSslConnect(17) -> 0x8095f013
[ssl] sceSslConnect(17) -> 0x8095f013
[ssl] sceSslConnect(17) -> 0x8095f00c
[SOFT-NET] sceNetResolverStartNtoa -> host DNS lookup failed
[SOFT-NET] sceNetResolverStartNtoa -> host DNS lookup failed
[gpu] texture 0x123fe90000 (format 22) has mixed DCC keys at 0x1240610000; its texels are read as stored
[gpu] texture 0x1210c10000 (format 22) has mixed DCC keys at 0x1213460000; its texels are read as stored
[... 44 [shadow] seed lines ...]
[gpu] texture 0x1238d00000 (format 22) has mixed DCC keys at 0x1238da0000; its texels are read as stored
[... 44 [shadow] seed lines ...]
[gpu] driver failure: shader 0x126f962f00 stage 5: image address has too few coordinate components
[videoout] presentation failed: shader 0x126f962f00 stage 5: image address has too few coordinate components
terminate called after throwing an instance of 'std::runtime_error'
  what():  shader 0x126f962f00 stage 5: image address has too few coordinate components
FATAL: abort() (SIGABRT) on thread 13516

AI-assisted: yes (Claude Code)

[klosixgaming-arch]

Tested on Windows 11, MinGW-w64 GCC 15.2.0, NVIDIA GeForce RTX 5090 (driver 617.42), Ryzen 9 9950X3D. Head 72959f7c0 (two commits) merged onto current main (ad502d7); clean merge. I first tested the earlier head d513fb987 with the same result and re-ran after the force-push.

• agc_driver_image_msaa_address_on_plane passes with the change.
• On main with only this PR's test file it fails with image descriptor is incompatible with the static runtime image interface: descriptor format 20 decodes to {class 2, dim 3, ...}, the refusal quoted in the description, so the new test fails without the fix and passes with it.
• I read the production change (five lines in ResourceMaterializer::RuntimeImageModes): it adds the ordinary 2D mode only for a 2D_MSAA image that is mip-less 32-bit and is not a table, storage, depth, converted or packed access, so everything else keeps its previous refusal. The hardware-oracle table in the description is the author's measurement on gfx1032; I did not rerun it.

Not tested: Cocoon (PPSA08766); I have no dump, so the title claim is the author's.

AI-assisted: yes. Written with AI assistance and reviewed by me.

[jotitadev]

Apple GPU / MoltenVK run of the new head. agc_driver_image_msaa_address_on_plane passes on Metal with all four formats. Apple is a third driver whose sRGB decode differs from the Radeon oracle, so the switch to comparing encoded texels in 72959f7 is needed here too.

Machine: Apple M2 Max, macOS 27.2, x86-64 build under Rosetta. Head 72959f7, PR diff applied cleanly. Drivers: MoltenVK 1.4.2 (Vulkan SDK 1.4.363.0) and MoltenVK main + SPIRV-Cross #2704.

Tested: the tree is main a373e85 + mugurc's macOS series (ci/macos-series-2) + the PRs needed to build on macOS (#6667, #6668, #7524, #6969, and two files of #7332), not plain main (it doesn't build on macOS yet). Also in the same build: #7725, #7729, #7717, #7694, #7682, #7731 and #6815 (they applied cleanly together), plus a test-only patch that lets nine AGC fixtures load the Vulkan loader on macOS (branch jotitadev:fix/agc-tests-vulkan-loader-macos). Base on that tree: 40 ^agc failures with MoltenVK 1.4.2 and 26 with MoltenVK main + #2704.

• agc_driver_image_msaa_address_on_plane, agc_shader_memory and agc_driver_image_volume_address_on_plane pass 3 of 3 on both drivers. agc_driver_sample_multisample_target skips here (storage image multisampling unavailable on MoltenVK).

• With ResourceMaterializer.cpp back to the base, the new test fails on both drivers with image descriptor is incompatible with the static runtime image interface.

• With the test as it was at d513fb9 (absolute tolerance 0.0001 against the oracle), it fails on both drivers with MSAA-address normalized load differs from the Radeon oracle. Only format 130 (sRGB) differs, on 4 of the 9 texels:

  Apple	Radeon oracle	exact decode of the byte
  0.052747	0.052979	0.052861 (65)
  0.527228	0.527344	0.527115 (192)
  0.533333	0.535156	0.533276 (193)
  0.545788	0.546875	0.545724 (195)

  Apple stays within about 1e-4 of the exact decode, and the oracle's values are quantized (0x3f090000). The encoded 8-bit texel matches in every case, so the current check passes.

• Full ^agc suite with the stack: the base failures on each driver, plus agc_driver_pooled_buffer_range and agc_performance_overlay, which only run because of the loader patch. Nothing from this PR.

GPUs tested: Apple M2 Max (MoltenVK 1.4.2; MoltenVK main + SPIRV-Cross #2704)

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

[parl0v]

Tested on Windows 10 (19045), RTX 3060 12 GB (NVIDIA 617.42), Cocoon PPSA08766 01.004.000, main 64e2a81 with this PR (72959f7) merged locally:

• main alone aborts at 77 s with image descriptor is incompatible with the static runtime image interface (shader 0x21e20e500 stage 5, descriptor dim 2D, every compiled mode 2D MSAA).
• With this PR that abort is gone. The run continues to 87 s and stops at the next refusal, VGT_GS_OUT_PRIM_TYPE other than a triangle strip, or a line strip from a geometry program, is unsupported (VGT_SHADER_STAGES_EN=0x2030), which feat(agc): rasterize geometry points with their point size #6801 covers.

[JJenkx]

Thanks for the Windows main/change comparison. I have added your machine, revisions, and observed result to the description, explicitly attributed to your report. It confirms this PR clears the descriptor/interface refusal; #6801 addresses the next geometry-output refusal. Your follow-up on #6801 reaches the separate point-context-defaults path covered by #6821. No gameplay or fully working Windows title claim is made.

AI-assisted: yes

[adriadam10]

Checked the diff: the new runtime mode is limited to mip-less 32-bit non-storage, non-depth, non-packed, non-converted, non-table image_loads, everything else keeps the old refusal. The regression covers wave32/wave64 and four formats, compares against the ordinary 2D load bit for bit, and fails without the change (also confirmed by the NVIDIA, Windows and Apple runs in the thread). The measured gfx1032 behaviour (sample index ignored, out-of-range reads zero) is recorded in shader debt. Merged.