[maximilianfv]

What

agc_driver_image_table_reuse and agc_driver_image_table_reuse_draw_pipeline assume that the image table's descriptor set comes from a dedicated pool. That assumption does not hold on every device. The kept-table write counters (writesKept, writesMade), added with kept descriptor sets in #6454 by EagleBearerYS (commit 1d66d85), count writes only for a set that DescriptorCache::Allocate takes from a dedicated pool. ShaderResources.cpp records table writes only when sampled && set < tableWrites.size() && allocation.layout != VK_NULL_HANDLE, and allocation.layout is set only in the dedicated-pool branch. A set gets a dedicated pool only when one of its sizes exceeds ChainPoolSizes (4096 storage buffers, 1024 sampled images, 1024 storage images, 512 samplers). A set from the shared pool chain is freed back to the chain, so the driver has nothing to keep and correctly records nothing.

The only size that can exceed the chain is the sampler table. SamplerTableCapacityFor sizes it as min(1024, Room(stage samplers, 64), Room(set samplers / 4, 64)), first from the plain limits and, when that gives 0, from the update-after-bind limits. MoltenVK on the M1 Pro reports (vulkaninfo) maxPerStageDescriptorSamplers 16, maxPerStageDescriptorUpdateAfterBindSamplers 1024 and maxDescriptorSetUpdateAfterBindSamplers 1024. The plain limits give 0 (16 is below the 64-sampler heap), so the update-after-bind limits apply: min(1024, 960, 192), rounded down to a multiple of 4, is 192, below the chain's 512 samplers. The test's own counters confirm it: dedicatedPools + reusedSets does not grow across the first two dispatches, which is why both tests print the chain line in the runs below. The test then fails at its first write check: three dispatches of one table in one epoch: 0 table descriptor writes kept and 0 made, expected 3 and 0.

This change reads descriptorCache->Counters() before and after the first dispatches. If dedicatedPools + reusedSets did not grow, the table's set came from the shared pool chain. The test then prints the table's descriptor set fits the shared pool chain: kept table sets are not tested, and every ExpectWrites requires 0 kept and 0 made table writes. On devices whose sampler table reaches 1024, the set still comes from a dedicated pool and the original expectations run unchanged. With APS5_NO_POOL_CACHE the descriptor cache is bypassed, so the test also takes the 0/0 path; the expectation is still correct, only the printed reason is then imprecise. The driver is not changed.

Prior art: of the whole-port PRs, #7201 and #8658 (and the PRs #8658 was split into, such as #10419) do not touch this test, but the closed #7332 (feat: macOS on Apple Silicon host support, by Ycaro Gabriel) has the same diagnosis and a test fallback in commits fba033d and 4fab753 (2026-10-19). This PR reimplements that fallback with a different check: #7332 made ExpectWrites take the device and require 0 kept and 0 made writes whenever Counters().dedicatedPools was 0 for the device as a whole, reaching the context through a new VulkanDeviceTestAccess struct; this PR compares the counters around this test's own first dispatches and leaves the ExpectWrites signature as it is. Ycaro Gabriel is credited as co-author. Testers on #8658 reported agc_driver_image_table_reuse failing on an M1 with the same message and passing on an M4, so this expectation depends on the GPU's limits. Open #6340 (MarcZX10) and #6223 (Thunderhawkk) also edit this test, but they add element-reuse checks and change neither ExpectWrites nor the first dispatches.

Tested

Apple M1 Pro, macOS 27.0.1, x86_64 build under Rosetta 2, Vulkan through MoltenVK 1.4.2 from the LunarG Vulkan SDK 1.4.363.

main alone does not build on macOS. I tested in a tree made of mugurc's macOS series (ci/macos-series-2, 27e9689) merged with main at 522e32e, plus these local changes, all identical in every run: the macOS build parts of #10419, #10540, #10562, #10641 and #10642, and two of my unopened fixes (the posix_spawnp child in LazyResolvedAbi.cpp, an OpMemoryBarrier before 32-bit image atomics in SpirvImageEmitter.cpp). Only this PR's change was toggled. All runs used ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1 VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation and ctest --test-dir build -R '^agc_driver_image_table_reuse' -E 'audio_in|voice' -V.

• Without this change, both agc_driver_image_table_reuse and agc_driver_image_table_reuse_draw_pipeline fail (0 of 2) with three dispatches of one table in one epoch: 0 table descriptor writes kept and 0 made, expected 3 and 0.
• With this change, both pass (2 of 2), and each prints the table's descriptor set fits the shared pool chain: kept table sets are not tested.
• With this change and APS5_KEPT_DESCRIPTOR_SETS=0, both pass (2 of 2) and each prints the same chain line.
• Validation output is the same in all runs: 2 Validation Error lines, both VUID-vkCreateDescriptorSetLayout-support-09582, and no was not found but was requested by env var VK_INSTANCE_LAYERS line.
• Not tested on a device whose sampler table reaches 1024, where the dedicated-pool expectations still run. Not tested on Linux or Windows.

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
• Unblocks: none
• GPUs tested: Apple M1 Pro, MoltenVK 1.4.2 (Vulkan SDK 1.4.363), macOS 27.0.1
• AI-assisted: yes

🤖 Generated with Claude Code

[MarcZX10]

Reviewed exact head 3beed54422b3bc90e84670e4dc08d291f63ca343. The test classifies the descriptor set from the first two dispatches: dedicated allocations and reuse of a kept dedicated set increment the counters, while shared-chain allocations do not. The production code also records table writes only for dedicated allocations with a non-null layout, so requiring zero writes for the shared-chain case is consistent with the implementation. The test preserves the original kept-write assertions when those counters show a dedicated allocation. Exact-head diff-check, conventions, and merge-tree checks pass; hosted eligibility, conventions, intake, labels, and report pass, while product builds are skipped. I did not run the GPU test locally; I reviewed the author’s M1 Pro run and control results. No blocking issue found.

[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 6223	core/libs/prx/libSceAgcDriver/tests/execution/ImageTableReuse.cpp

[MarcZX10]

A new overlap check for #6223 prompted a manual three-way merge comparison. Both PRs edit ImageTableReuse.cpp; the current two-head merge-tree reports a shared-hunk conflict around the write-count helper and the RunTests dispatch section. The code additions appear conceptually compatible—#6223 adds per-element reuse assertions while this PR handles shared-pool write counters—but the branches need an agreed integration order/rebase so neither assertion set is dropped. The PR body currently says they change neither the write helper nor the first dispatches, while the overlap notice shows the same region and the checklist says Depends on: none. Please coordinate with #6223’s author and record the landing order (or explain how the current heads merge without losing either test) before merge. The prior exact-head functional review remains valid; this request is only for the newly surfaced coordination requirement.

[maximilianfv]

Landing order with #6223, checked against main 6c7a3256f2, this head 3beed54422 and #6223's head 8729fc234f:

• The conflict is between perf(agc): reuse image-table elements whose descriptors repeat #6223 and main, not with this PR. perf(agc): reuse image-table elements whose descriptors repeat #6223's change to ImageTableReuse.cpp does not apply to main on its own (git apply --check fails at the same hunk). It adds ExpectElements(before, 143, 49, "two tables that differ in one record") on the line where main has since added ExpectWrites(before, 0, 4, "two tables that differ in one record, alternating"). The fix is to keep both lines. Its ShaderResources.{hpp,cpp} and debt-file conflicts are also with main; this PR does not touch those files.
• After that rebase, the two PRs do not overlap. I applied perf(agc): reuse image-table elements whose descriptors repeat #6223's test change to main with both lines kept, then applied this PR's diff on top. It applied cleanly, with no hunk touching perf(agc): reuse image-table elements whose descriptors repeat #6223's lines: this PR changes only ExpectWrites's body and the code around the first two dispatches, and perf(agc): reuse image-table elements whose descriptors repeat #6223 adds a new helper and one call after the "two tables that differ in one record" check. Neither assertion set is lost in either order.
• Order: this PR first. perf(agc): reuse image-table elements whose descriptors repeat #6223 needs a rebase onto main either way, and this PR merges cleanly now. I've left Depends on: none, since this PR does not need perf(agc): reuse image-table elements whose descriptors repeat #6223. @Thunderhawkk, if perf(agc): reuse image-table elements whose descriptors repeat #6223 lands first, I'll rebase this one onto it.

I didn't build or run the combined tree. This was a patch-level check only.

[ycbuzi]

Windows check of the dedicated-pool path on three non-Apple GPUs. 3beed54422 merged onto main 2742004ed2 (78a5221a22, clean), Windows 11 Home 10.0.26200, WinLibs MinGW-w64 GCC 15.2.0 (ucrt-posix-seh r7), Release, ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1, one device at a time with ANYPS5_GPU:

GPU	this PR	main 2742004ed2
NVIDIA GeForce RTX 4070 SUPER (0x2783)	both pass	both pass
NVIDIA GeForce GTX 1650 SUPER (0x2187)	both pass	both pass
AMD Radeon(TM) Graphics, RDNA2 iGPU (0x164e)	both pass	both pass

"Both" is agc_driver_image_table_reuse and agc_driver_image_table_reuse_draw_pipeline. On none of the three devices did the PR print the table's descriptor set fits the shared pool chain. So the set came from a dedicated pool there, and the original kept-table expectations ran unchanged and passed, as the description expects for devices whose sampler table reaches 1024. Not run: the Khronos validation layer, MoltenVK. AI-assisted: yes (Claude Code)