[MarcZX10]

What

Move the unsupported LDS split check to the start of recordDispatch, before dispatch recording allocates scratch or submits GPU work. Preserve main's CPU-resolved-indirect diagnostic context. The regression requires the LDS-specific refusal message and checks both an ordinary oversized Y dispatch and a z = UINT32_MAX case that distinguishes the early guard from the later refusal.

Fixes #7794.

Tested

• Rebased head 82be87f185: git diff --check upstream/main...HEAD passed.
• Rebased head 82be87f185: python3 tools/check_conventions.py --base upstream/main --head HEAD passed.
• Earlier head 6a189b3dfc: configured and built agc_driver_oversized_dispatch_tests on Linux x86-64. CTest could not execute because this container's Intel Xeon E5-1620 v2 lacks AVX2, FMA, BMI1/BMI2, LZCNT and MOVBE; AnyPS5 aborts during host CPU validation before Vulkan setup.
• A collaborator tested a stronger composition of this change on main 7a36de66b9: NVIDIA RTX 5080 and AMD Radeon Graphics showed 3/3 passes each, with a negative control distinguishing the early refusal. That hardware report predates current main 43a2abaf77; this rebased head has not been rerun on those GPUs. See the detailed report in the PR discussion.
• No local GPU test was run on the rebased head.

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
• No notes or investigation files added
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none (#6786 is in current main)
• Fixes: AGC: an unsupported LDS dispatch split allocates scratch before refusing it #7794
• Unblocks: none
• GPUs tested: see earlier composition report above; not retested on current main
• 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 1395	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 2434	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 4040	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 4072	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 4472	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 5021	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 5431	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 5441	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 5499	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 5925	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 6188	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 6202	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 6290	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 6788	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 6947	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 6960	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 7201	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 7275	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 7449	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 7484	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 7642	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 8014	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 8105	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 8278	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp
PR 8847	core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp

[Neithennn]

Windows NVIDIA run, since the host CPU guard stopped the test on your machine.

Windows 11 Pro 26200, NVIDIA GeForce RTX 4080 (driver 617.42), AMD Ryzen 7 5800X, WinLibs GCC 15.2.0 r7, Ninja, Release, bundled dependencies, BUILD_TESTING=ON, ANYPS5_NO_SHADER_CACHE=1, ANYPS5_REQUIRE_VULKAN=1. Head 6a189b3df merged onto main 36f60df9e; full build succeeds.

• agc_driver_oversized_dispatch: passes (targeted, full suite and 5 serial runs); it prints oversized dispatch tests passed.
• The other LDS tests (agc_driver_lds_*, agc_driver_pixel_lds*, agc_lds_lock_budget, agc_spirv_lds_atomic_lock, agc_mubuf_lds_decode): 20/20 pass.
• Full ctest -j 8: 1631/1631. Unchanged main on the same machine: 1631/1631.

Negative control: with only Execution/src/VulkanDevice.cpp reverted to main (the tightened OversizedDispatch.cpp kept), agc_driver_oversized_dispatch still passes on this GPU. So on the RTX 4080 the refusal is already reached without the reordering, and the test does not tell the two versions apart here. I did not look into which earlier failure the reordering avoids on other devices.

AI-assisted: yes (runs done with Claude Code).

[LeandroLP23]

Ran agc_driver_oversized_dispatch on AMD and Windows. AMD Radeon 860M (RDNA 3.5, 1002:1114), driver 32.0.13058.8001, Windows 11, WinLibs GCC 15.2.0, Release, ANYPS5_NO_SHADER_CACHE=1. Head merged into main 118266a (clean merge). "Main's sources" means this PR's test files over main's other files.

It passes 3 of 3, but on this GPU it can't exercise the change. The split case skips:

oversized dispatch X split: skipped, the device reports a 4294967295 workgroup limit, which never splits a 129199x1x1 dispatch

So it also passes with main's sources. AMD's Windows driver reports a 2^32-1 workgroup count limit, so no dispatch is ever split there. The LDS refusal path needs a GPU with the 65535 limit, such as NVIDIA.

AI-assisted: yes (Claude Code)

[Amoenus]

@MarcZX10 #6786 has merged. I validated a rebased composition 6932eafa20 of your 6a189b3dfc on main 7a36de66b9, retaining your authorship and main's cpuArguments / cpuReason annotations in the relocated LDS diagnostic. It also retains the stronger test from the existing #7794 fork proposal: the original Y case plus z=UINT32_MAX, both requiring the intended LDS refusal.

Three fresh processes on each physical GPU	Result
Strict fixture only, main production unchanged	3/3 expected wrong-refusal failures on RTX 5080 and 3/3 on Radeon Graphics
Rebased early refusal + same fixture	3/3 passes on each GPU, no VUID/errors/warnings in oversized dispatch

The larger-Z control distinguishes the versions even when the first case reaches main's late LDS check. Physical X splitting is unavailable at the reported limits; Y splitting and both LDS refusals execute. This addresses the earlier nondiscriminating-control concern without claiming X coverage.

Full default and explicit relinker/library builds pass. Oversized dispatch, valid device-memory LDS, cross-workgroup flag, indirect limit and pending-producer checks run three times per GPU, plus CPU graphics/PM4 three times: 36 behavioral passes (30 GPU, 6 CPU). The 30 common main checks pass too. The pending-producer fixture has 12 validation errors on main and on this composition, isolated in new #7868; no additional VUIDs, warnings, whole-test skips or timeouts. This is not a blanket validation-clean result.

Windows 11 Enterprise 26200 / GCC 15.2 r7; NVIDIA RTX 5080 (10de:2c02), driver 610.62 / 32.0.16.1062; AMD Radeon(TM) Graphics (1002:13c0), driver 32.0.21036.18. Required Vulkan, disabled application shader cache, Khronos 1.4.363 confirmed loaded, verbose serial CTest, 30 s per process. New main bbb45463c7 landed after the freeze and is not covered by these results. Linux, the unchanged old PR head and title impact were not retested.

Cppcheck/clang-tidy ran too: the fixture has no selected findings; 10 untriaged findings point to unedited device-initialization/header locations, so no analyzer-clean claim. The tested composition is available on our fork for your rebase. #7794 updated in place; I am supporting this PR rather than submitting a duplicate.

AI-assisted adaptation, validation and report: yes, Codex.

[MarcZX10]

Rebased onto current main (43a2abaf77), preserving main’s CPU-resolved-indirect diagnostic context while moving the LDS refusal ahead of later dispatch work. I also added the z = UINT32_MAX regression case from the earlier tested composition report; the current head has not been rerun on those GPUs. Current git diff --check and conventions pass. The earlier CMake build was on the pre-rebase head, and local runtime execution remains blocked by this host CPU’s missing AVX2/BMI features.

[LeandroLP23]

Reran the rebased head 82be87f on AMD's Windows driver. The new z = UINT32_MAX case tells the two versions apart on this GPU, which the earlier fixture couldn't. Radeon 860M (RDNA 3.5, 1002:1114), driver 32.0.13058.8001, Windows 11, WinLibs GCC 15.2.0, Release, ANYPS5_NO_SHADER_CACHE=1. Head merged into main 88a6b34 (clean merge).

agc_driver_oversized_dispatch	Result
This head	pass, 3 of 3
main's sources, this PR's test	fails: oversized dispatch: wrong LDS split refusal: Vulkan dispatch: the LDS of 1x65536x4294967295 workgroups of 36872 bytes each is over maxStorageBufferRange 4294967295

The X split case still skips here (the device reports a 4294967295 workgroup limit, which never splits a 129199x1x1 dispatch), as before. The Z case doesn't depend on that limit.

AI-assisted: yes (Claude Code)

[MarcZX10]

Thanks for rerunning the rebased head with the z = UINT32_MAX case. This provides a distinguishing negative control on the Radeon 860M: the PR passes 3/3 and main’s sources fail with the expected oversized-LDS refusal. The X-split case remains unexercised on this device because its reported workgroup limit never triggers that split; I’ll keep that limitation clear.

[Neithennn]

RTX 4080 / Windows run of the rebased head 82be87f18: the new z = UINT32_MAX case tells the two versions apart on NVIDIA too.

Windows 11 Pro 26200, NVIDIA GeForce RTX 4080 (driver 617.42), WinLibs GCC 15.2.0 r7, Ninja, Release, BUILD_TESTING=ON, ANYPS5_NO_SHADER_CACHE=1, ANYPS5_REQUIRE_VULKAN=1. Head merged onto main 27bcd6a29.

Tree	agc_driver_oversized_dispatch
this head	passes 5/5 serial runs
this head with Execution/src/VulkanDevice.cpp from main (the PR's test kept)	fails 5/5: oversized dispatch: wrong LDS split refusal: Vulkan dispatch: the LDS of 1x65536x4294967295 workgroups of 53256 bytes each is over maxStorageBufferRange 4294967295

On my earlier run of 6a189b3df, main's source passed this test too; the new case is what makes the difference here.

The X split is still not exercised on this GPU: oversized dispatch X split: skipped, the device reports a 2147483647 workgroup limit, which never splits a 129199x1x1 dispatch. The Y split runs.

Full ctest -j 8 on this tree: 1854/1858, the same four as unchanged main on this machine: agc_driver_image_array_window_past_layer_limit is not built (its alignas(65536) hits MinGW's 8192 limit on current main), and the three agc_driver_sample_order* tests that @milescoviello reported on #6683 also fail here.

AI-assisted: yes (runs done with Claude Code).

[milescoviello]

@Neithennn, about the three agc_driver_sample_order* failures: they come from the tests, not from this PR. The tests bind the depth surface's own view as a sampled image. When the surface uses #7318's sampled shadow, that image has no SAMPLED usage (VUID-VkWriteDescriptorSet-descriptorType-00337), and the driver returns other depth.

#9056 makes the tests read through a sampled copy, as SampleOrder.cpp itself does. #9057 fixes a separate barrier on the D32S8 copy image that the layer flags (03320).

Results with #9056:

• RTX 4070 Ti (Linux): all three pass.
• RTX 3080 (Windows, gdols): two pass, and _grid_reference skips.
• Radeon 680M: _grid_reference still loses the device, which is a separate problem.

AI-assisted: yes (Claude Code)

[adriadam10]

Re-checked against current main: the early LDS-split refusal in recordDispatch still precedes any allocation (same condition and message, cpuArguments/cpuReason context kept, late check removed). The z = UINT32_MAX case fails on main's sources, as measured by @LeandroLP23 (Radeon 860M) and @Neithennn (RTX 4080); merged.