feat(agc): store to images selected through image tables by Pipodi · Pull Request #6402 · boykopovar/AnyPS5

Pipodi · 2026-10-15T14:05:16Z

What

An IMAGE_STORE whose T# comes from a runtime-indexed image table used to mark the whole table as an unsupported operation. The dispatch's deferred write-back then failed. This follows #757/#758's table model:

AssignTables now accepts written, non-atomic table images and gives them storage modes. Atomic, packed, BY2/BY4 and dynamic-mip table stores still fault, as before.
A new STORAGE_IMAGE table binding (260). The sampled image table moves to 261 and keeps the variable descriptor count. The runtime ABI goes to v14, so existing shader caches are invalidated.
The driver binds storage views of the tracked storage textures at the same element indices. It marks every bound entry as written, so each dispatch writes back all of the table's storage images. Tables with written entries are never reused.
The descriptor pool of a set that isn't split counts the storage table's descriptors, as the split path does.
A table storage element is checked like a direct one: an element whose minLod clamps above the level it addresses is rejected (it faults when a lane selects it), and the element is looked up as written, so a depth surface under it is retired.
RunStoreTests is now a positive test: a lane stores to the key-1 image, while a null-key store and an EXEC-cleared store write nothing. verifyImageTableRegistration expects the bound written elements, a store-only table binds no sampled elements, and image_store_mip, image_store_pck and image_atomic_add through a table still poison the table with an unsupported operation and bind no storage elements.

Native pipeline replay with storage tables, reuse of table layouts across size classes (which also lets storage-table sets back into chain pools), and rejecting the second writer of a depth plane are follow-ups. @Celegans12 will add them on top (#675).

Technical debt

New entry "Image storage tables" in docs/dev/debt/libSceAgcDriver.md:

every bound storage element is stored back, so two elements over the same bytes race on write-back order (PPSA26344's 0x80001e5f00 binds the stencil plane 0x505f350000 as R8_UINT and R8_UNORM);
the storage table is declared at full capacity, so its sets get dedicated descriptor pools;
the capacity is a heuristic (the smallest of the sampled table capacity, 4096 and a quarter of maxPerStageDescriptorStorageImages), and the sampled capacity isn't reduced by it;
storage elements at or past that absolute index are rejected;
atomic, packed, BY2/BY4 and dynamic-mip table stores still fault;
PresyncSurfaces walks only the sampled table;
untested cases: mixed sampled and storage tables, a T# both read and written, non-uniform storage indexing, non-2D storage modes.

Tested

Windows 11, NVIDIA GeForce RTX 4090, driver 617.42.

ctest on b7e21ca (rebased onto main 6e7ba92), full BUILD_TESTING=ON build: 1364/1366 pass. agc_driver_pixel_lds (vkWaitForFences timeout in the wave64 unbounded case) also fails alone on plain main 6e7ba92; select_tests failed in the parallel run and passes alone.
Ghost of Yōtei (PPSA26344). Base: main d92a8da + fix(agc): bind a read-only V# over the cached address space through the region that holds it #5815 perf(agc): keep Windows host imports of write-watched private memory watched #6060 fix(agc): recycle detiler descriptor pools once their batches complete #6168 fix(agc): sample a depth surface's stencil plane through an R8_UNORM view #5781 fix(libSceNetCtl): answer the DHCP host name, PPPoE name and proxy server as empty strings #6227 + local patches; the runs differ only in this PR. Program 0x80001e5f00 ends with three IMAGE_STOREs through runtime-indexed tables (pc 0x783c/0x785c/0x7868).

Run	This PR	Cache	"unsupported operation" faults	Result
run150	no	warm	yes: `image table: shader 0x80001e5f00 stage 4 pc 0x783c ... unsupported operation`	all workers stop at frame ~828 (~333 s); private memory 32.6 -> 51.8 GB in a minute
run152	yes	cold, 1500 s	0	`0x80001e5f00` compiles (1,457,658 words, 63 s) and dispatches; workers keep running until ~1240 s
run156	yes	warm, 1800 s	0	`0x80001e5f00` v2 dispatched with 150-169 storage images bound (resources 108-343 ms); loading continues to the first in-game scene (~25-30 min)

The game runs predate the review changes, which affect descriptor pool sizing, minLod rejection and depth-surface retirement for table storage elements.

Checklist

🤖 Generated with Claude Code

github-actions · 2026-10-15T14:23:26Z

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on `main`)	Same exports	Same new files
PR 6340	`core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` L4236-4253
PR 6454	`core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` L4228-4232, L4236
PR 7223	`core/libs/prx/libSceAgcDriver/tests/execution/ImageTable.cpp` L371

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 1341	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp`
PR 2434	`docs/dev/debt/libSceAgcDriver.md`
PR 4072	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp`
PR 4339	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvEmitterHelpers.cpp`
PR 4448	`docs/dev/debt/libSceAgcDriver.md`
PR 4472	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp`
PR 4650	`core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvEmitterHelpers.cpp` `docs/dev/debt/libSceAgcDriver.md`
PR 5058	`docs/dev/debt/libSceAgcDriver.md`
PR 5371	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp` `core/shader/recompiler/RuntimeAbi.hpp`
PR 5372	`docs/dev/debt/libSceAgcDriver.md`
PR 5499	`docs/dev/debt/libSceAgcDriver.md`
PR 5500	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 5609	`core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 5925	`core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 5958	`core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp` `core/shader/recompiler/Optimization/src/ResourceTracker.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6146	`docs/dev/debt/libSceAgcDriver.md`
PR 6177	`core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6188	`core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6194	`core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6222	`core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6234	`docs/dev/debt/libSceAgcDriver.md`
PR 6235	`core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvEmitterHelpers.cpp`
PR 6273	`docs/dev/debt/libSceAgcDriver.md`
PR 6300	`docs/dev/debt/libSceAgcDriver.md`
PR 6373	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp`
PR 6412	`core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6414	`core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp`
PR 6445	`docs/dev/debt/libSceAgcDriver.md`
PR 6539	`core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Recompiler.hpp`
PR 6548	`docs/dev/debt/libSceAgcDriver.md`
PR 6553	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp`
PR 6563	`docs/dev/debt/libSceAgcDriver.md`
PR 6601	`docs/dev/debt/libSceAgcDriver.md`
PR 6636	`core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6645	`core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp`
PR 6647	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp`
PR 6718	`core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp`
PR 6727	`docs/dev/debt/libSceAgcDriver.md`
PR 6808	`docs/dev/debt/libSceAgcDriver.md`
PR 6887	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6946	`core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6947	`core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderValidation.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/BindingAllocator.cpp` `core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp` `core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/RuntimeAbi.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvDescriptors.cpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvEmitterHelpers.cpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6959	`core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6965	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6973	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6987	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6989	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 6995	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 7002	`core/libs/prx/libSceAgcDriver/Graphics/include/Context.hpp` `core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp` `core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp` `core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp` `core/shader/recompiler/Recompiler.hpp` `docs/dev/debt/libSceAgcDriver.md`
PR 7154	`core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp` `docs/dev/debt/libSceAgcDriver.md`
PR 7178	`docs/dev/debt/libSceAgcDriver.md`
PR 7201	`core/shader/recompiler/Recompiler.hpp` `core/shader/recompiler/SpirvBackend/src/SpirvMemory/SpirvEmitterHelpers.cpp`
PR 7256	`docs/dev/debt/libSceAgcDriver.md`

EagleBearerYS · 2026-10-15T14:45:44Z

Heads-up on an interaction with #6243 (selecting the record of a T# whose offset the CPU evaluates), which is still open. The two merge without conflicts, but on a merge of both, agc_shader_bindless fails: #6243's selected record direct operations case expects an image_store through a T# keyed per lane to stay an unsupported table operation, and this PR makes that store supported.

With that one expectation changed to TableOperation::Supported for image_store (atomics, PCK and BY2/BY4 keyed per lane still fault, as both PRs expect), the merge passes agc_shader_bindless, agc_driver_image_table and the other shader-memory tests 5 runs in a row on GitHub's Linux and Windows runners. ctest -R '^agc' on the merge fails nothing that main doesn't. A store whose record the CPU evaluates stays a direct binding under #6243; this PR's storage table takes the stores keyed per lane.

Whichever of the two merges second needs that one-line test change. If this one lands first, I'll make it on #6243.

Celegans12 · 2026-10-15T14:58:18Z

@Pipodi I went through this against our local store work, and the model fits #757/#758 nicely. Keeping one element domain with the storage element in 260 and the sampled one in 261, faulting only on use for poisoned or rejected elements, and never reusing tables with written entries are all what I'd have done. A few things I think need fixing before it goes in, plus some I'd be fine leaving to a follow-up.

The two I'd treat as blocking:

The descriptor pool has no room for the storage table on the non-split path (compute, so 0x80001e5f00 included). Binding 260 is declared with StorageTableCapacity descriptors (up to 4096), but the pool sizes at ShaderResources.cpp:1802 still only count plannedStorageImages. The split path (:1741) and draw snapshots count the table, and this one doesn't. NVIDIA doesn't enforce per-type pool counts, so our runs won't show it, but a driver that does will fail the allocation. Once the count is right, a 4096-entry set no longer fits a 1024-image chain pool either, so every build would get its own pool. That's why ours sized 260 from the element count in size classes.
No test still checks that the unsupported table stores fault. The old negative store test became the positive one, so atomic, packed, BY2/BY4 and dynamic-mip table stores aren't covered any more.

Smaller ones:

The table path doesn't do the direct path's minLod check (:4072), so a T# whose minLod clamps above the addressed level gets written anyway. A per-element rejectTableElement would match the direct path.
The element is looked up with cachedStorageTexture(..., guestBytes) (so written is false and RetireDepthSurfaces never runs) but pushed with storageWritten = true (:4289 vs :4297). The direct path keeps the two in step. Either pass true, or reject elements over a depth surface explicitly.
Depth/stencil planes through a table. This one we hit in Yōtei. 0x80001e5f00's storage table holds the stencil plane 0x505f350000 twice: as R8_UINT in its own tile mode, and as R8_UNORM in another. With every element marked written, that's two writers of the same bytes, and write-back order decides the result. On our branch we reject the second writer per element, so the dispatch only faults if a lane actually selects it. That's fine as a follow-up from us if you'd rather keep this PR focused, but a TechnicalDebt line here would help.

Happy to send the depth-plane part and our two extras (pipeline replay with storage tables, and 260 sized in classes of the element count so 233-element sets stay in chain pools) on top once this settles. I'll run it on Yōtei as part of our next game build.

Details and notes

Performance, as a design note rather than a regression, since ours does the same: marking every bound element written means each 0x80001e5f00 dispatch dirties all 150-233 storage images of the table, though only the three stores at 0x783c/0x785c/0x7868 write. Any remembered sampled table overlapping one of them then fails TableCurrent and can't be reused, which matters with Yōtei's 33k-entry sampled tables. A profile split of your 108-343 ms resource time (lookups vs write-back) would show whether it bites. Per-element written bits from the GPU would be the exact fix later.
PresyncSurfaces (:4441-4449) walks ImageTable only and applies the sampled rule to every element, so a store-only table isn't considered.
StorageTableCapacity is a heuristic (maxPerStageDescriptorStorageImages / 4, max 4096), and ImageTableCapacity doesn't subtract it, so ValidateDescriptorLayout could throw on devices with tight limits.
Storage elements at absolute index ≥ 4096 are rejected even when only a few are storage (:4281-4283). That's fine for 0x80001e5f00 (about 233 elements), but mixed tables with many sampled elements first would hit it.
Untested: mixed sampled+storage tables, read and write of the same T#, non-uniform storage indexing, non-2D storage modes (newly allowed), and that the null-key run doesn't fault.

Pipodi · 2026-10-15T15:39:49Z

@Celegans12 thanks for the careful pass. Pushed 085fa565b on top of a rebase onto current main:

Pool sizes on the non-split path: the set's pool now counts the storage table binding's descriptors, as the split path does. As you said, a set declared with the full StorageTableCapacity no longer fits a 1024-image chain pool, so those sets get a dedicated pool for now. I noted that in the debt entry. Your size classes of the element count are the real fix, and I'd be glad to take them as your follow-up together with the native replay.
Negative coverage: verifyImageTableRegistration again checks that image_store_mip (dynamic mip), image_store_pck and image_atomic_add through a table stay TableOperation::Unsupported, poison the table with UnsupportedOperation so it faults on use, and bind no storage elements. I didn't add BY2/BY4: I don't have a hand-encoded instruction for them that I trust, so if your branch has one, I'm happy to add it.
minLod: a storage table element whose minLod clamps above the level it addresses is now rejected per element with rejectTableElement, the same rule the direct path enforces with Require.
written vs storageWritten: the element is now looked up with written = true, so RetireDepthSurfaces runs as it does on the direct path.
Stencil plane through two elements: left for your follow-up as you offered. There's a TechnicalDebt entry under libSceAgcDriver that names 0x80001e5f00 and 0x505f350000 and the write-back order problem.

The limits in your details block are in the same debt entry now too, along with the ones I should have listed in the first place: the storage capacity heuristic and that ImageTableCapacity doesn't subtract it, the absolute-index rejection, PresyncSurfaces walking only the sampled table, the dedicated pools, and the untested cases (mixed tables, read+write of one T#, non-uniform storage indexing, non-2D modes). I'd keep the code changes for those out of this PR unless you think one of them blocks.

Tested on 9161151 (rebased onto main 92e5041): Windows 11, MinGW-w64 GCC 15.2.0, RTX 4090. Full BUILD_TESTING=ON build, then ctest: 1294/1294 pass.

One thing the new check caught: image_store_mip through a table is still planned as Supported with ImageMipMode::DynamicStorage, and only ApplyStaticInterface refuses it. So the test accepts either the plan's Unsupported or a dynamic mip, and then requires the poisoned map and no storage elements in the compiled result.

Dextroid17 · 2026-10-15T19:30:17Z

Tested on Linux with lavapipe, the CI device: llvmpipe (Mesa 26.2.4, LLVM 23.1.1), CachyOS, GCC 16.2.1, CMake 4.4.3, default build type, full build with BUILD_TESTING=ON. Each run used a clean environment (env -i with HOME, PATH, VK_ICD_FILENAMES and GALLIUM_OVERRIDE_CPU_CAPS=avx).

This PR (9161151ef merged into main 44df21632, no conflicts):

agc_driver_image_table and agc_shader_bindless: pass, 3 runs out of 3.
I could not run a check with the fix reverted: ImageTable.cpp uses DescriptorRole::ImageStorageTable, which this PR adds, so the test does not build against main's sources.

Merged together with #6416 (on main 44df21632, no conflicts): full ctest: 2 of 1269 fail, 57 skipped (the same skips as main). The two failures are not from either PR:

agc_driver_mesh_index_cache segfaults on main too (and 5 runs out of 5 on the merge).
agc_frame_dump_io (from fix(agc): disable optional BMP dumps after host I/O failures #5800) also fails on plain main, in 13 of 200 runs on an idle machine: closed pipe pixel write was accepted. Reading the test, its 128×128×4 = 65,536 pixel bytes are exactly Linux's default pipe buffer, so if the reader drains the 54-byte header before the pixel write starts, the empty pipe takes the whole write and it succeeds before the reader closes. Not measured beyond that.

Apart from those, the failed and skipped lists match main's full run.

Not tested: NVIDIA, AMD, Intel, PPSA26344.

AI-assisted: yes

SleglJan · 2026-10-16T21:12:19Z

Data from 007 First Light (PPSA11386 v01.018.000), a title that stores to images selected through image tables in compute.

Setup: Linux 7.2.10, RTX 3080 10 GB, NVIDIA 615.78.08, GCC 16.2.1, Release, bubblewrap sandbox. Stack: main 4186f06 plus open PRs (#5987, #6553, #6465, #6868 and the image-table PRs #6440 and #6567 among them), unpublished local fixes the title needs, local logging and scripted input, and the title's two feedback-wait limits widened in its executable as described in #6172 (local only). This head's three commits (9161151) cherry-picked cleanly onto it. Compared with a run of the same stack without this PR, same save and switches (APS5_HOST_IMPORT_MIB=12288, APS5_RESOURCE_CACHE_ENTRIES=64, APS5_NO_DRAW_SRT_ENTRIES=1, a guest fault dump).

Without this PR, the title's compute dispatches fail their deferred write-back at five program sites with deferred write-back failed: image table: … unsupported operation. The failed fault check also skips that dispatch's CPU write-back of copied buffer results; buffers the GPU writes in place are not affected. That path is the same on main and this PR doesn't change it. Counts per site, in the same order in every cell; this stack prints the line once per failure:

	boot, menus and loading	to present 1430	end of run
without this PR	877 / 2 / 2 / 4 / 2	1349 / 190 / 190 / 116 / 11	1369 / 978 / 978 / 121 / 11 (present 2218)
with this PR	0 / 0 / 0 / 0 / 2	0 / 0 / 0 / 0 / 11	0 / 0 / 0 / 0 / 11 (present 1586)

Present numbers differ between the runs from the loading screen on. The frame rate drops at the end of loading at present 1243 without this PR and at 1219 with it. The first column ends at that drop in each run, and it is the like-for-like one.
Four of the five sites no longer fail.
The fifth fails 11 times in both runs, at the same point: twice just before that drop, then once per present for the next nine presents.
- With this PR its message is rejected by the driver: the image storage table exceeds its capacity of 4096 images.
- That message is printed both when the set has no storage-table binding and when the element's index is at or above the capacity. I haven't split the two cases yet.
No stop and no guest fault in either run. The run without this PR used its full 7,200 s budget; I ended the run with this PR at present 1586.
Not comparable between the runs after loading: the player's movement.
- The run with this PR used a newer input script whose stick input never reached the title, so the player stood still. Without this PR the player walked until he got stuck at about present 1430.
- So the later columns and the window-title FPS over presents 1244–1586 (mean 0.15 without this PR, 0.18 with it) cover different scenes.
The run with this PR started from a cold shader cache, because this PR changes recompiler sources and with them the cache's source version.
Not run: ctest on this machine; Dextroid17's comment above has lavapipe.

AI-assisted: yes (Claude Code)

Pipodi · 2026-10-16T21:23:50Z

Thanks, that's the first run on a title that stores through image tables: four of the five sites going from hundreds of failures to zero is what this PR is for.

You're right about the fifth site's message. It comes from one check in ShaderResources.cpp that covers two different failures:

The set has no storage-table binding, although the shader writes through that table element. The storage-table binding is added only for a set where some table is marked as a store table at recompile time. So this case would mean the recompiler's record and the element's written flag disagree.
The element's index is at or above the storage capacity, min(image table capacity, 4096, maxPerStageDescriptorStorageImages / 4), so 4096 on your 3080. The index is the element's position in the whole set: every table element before it in the same set counts, sampled ones included. So a small storage table can still land past 4096 if a large sampled table comes first in the set.

The two need different fixes. Case 2 means giving storage tables their own index space; case 1 is a recompiler bug. 0229b5f (pushed here) gives each case its own message:

set N has no image storage table, but the shader stores to its table element I
the stored table element I of set N is past the image storage table's capacity of C images

It changes only those messages and corrects the debt entry, which said the index counts within the table rather than the set. It cherry-picks on its own, so a run with it would show which case your eleven failures are.

AI-assisted: yes (Claude Code)

oneandonlydean · 2026-10-17T06:31:38Z

Reviewed at 0229b5f. Looks good. Two small things for your next rebase:

main is already at RuntimeAbi v13 (f7b2ebd), so the 12 -> 13 bump here merges silently and the binding move (storage table 260, sampled table 261) would ship under the same v13. Could you bump it to v14 when you rebase?
perf(shader): select the record of a T# whose offset the CPU evaluates #6243 also touches image-table stores. Whichever of feat(agc): store to images selected through image tables #6402 and perf(shader): select the record of a T# whose offset the CPU evaluates #6243 lands second needs to flip one agc_shader_bindless expectation to Supported.

Full CI is running on this head. It'll also go through our game gates before merge.

An IMAGE_STORE whose T# comes from a runtime-indexed image table used to poison the whole table as an unsupported operation. Written, non-atomic table images now get storage modes, a STORAGE_IMAGE table binding (260, the sampled image table moves to 261 and keeps the variable count), and storage views of tracked storage textures at the same element indices. Every bound entry is marked written, so each dispatch writes back all of the table's storage images. Atomic, packed, BY2/BY4 and dynamic-mip table stores still fault. Tables with written entries are not reused. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… elements The shader memory registration test pinned the old refusal. It now checks that the store's table is supported, does not fault, and binds every element as written on the storage table binding, and that a store-only table binds no sampled elements. The driver no longer requires a written flag per table element, so bindings that carry none read as unwritten. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…lements like direct ones The non-split descriptor set's pool now counts the storage table binding's descriptors, as the split path already did. A table storage element whose minimum LOD clamps above the level it addresses is rejected per element, as the direct path refuses it, and the element is looked up as written so a depth surface under it is retired. The registration test checks again that image_store_mip, image_store_pck and image_atomic_add through a table stay unsupported operations that fault on use. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…st its capacity Both cases printed that the image storage table exceeds its capacity. They need different fixes, so each now has its own message with the set, the element's index in the set and the capacity. The debt entry now says that the index counts every table element before it in the set. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Pipodi · 2026-10-17T07:02:11Z

@oneandonlydean thanks. Rebased onto main 6e7ba92 and bumped the runtime ABI to v14 in the commit that moves the bindings, so no commit ships the 260/261 layout under v13. The description says v14 now.

On #6243: agreed. EagleBearerYS offered to make the change on #6243 if this one lands first. If #6243 lands first, I'll flip the image_store expectation in agc_shader_bindless to Supported here.

Tested on b7e21ca: full BUILD_TESTING=ON build, then ctest: 1364/1366. agc_driver_pixel_lds also fails on plain main 6e7ba92 on this machine (a fence timeout in the wave64 unbounded case). select_tests failed only in the parallel run and passes alone.

AI-assisted: yes (Claude Code)

Main's #6402 runs image stores through an image table, so a T# keyed per lane keeps its table with a supported image_store and an unsupported mark for the other operations.

Main's #6402 stores to image tables: a table with a written entry takes the fresh path, as it does on main, and so does one without a sampled table binding, whose entries the fresh path rejects. The shared place step uses the binding as a pointer. Main's #6614 frees descriptor sets from shared pools one by one, while this branch keeps only dedicated pools, so both stay. #7101's bulk key read counts its unmapped-page skips as unmapped, and its key words sit next to this branch's write stamp and compared runs.

#6243 now expects a per-lane table image store to be supported, since main's #6402 runs image stores through image tables. This branch carries #6243's test, so it takes the same expectation.

Brings combo19 (async lane, gpuind, chain-A, 1a f32 mix, boykopovar#7489, S1, bindless images, staging diff, HTILE work, epoch 5) up to upstream main of 2026-10-19. 118 conflicted files, resolved by area; notes in work/research/structural/merge19/notes-*.md. - Versions: FormatVersion 48, RuntimeAbi 18, request serialization 37 (upstream's 31..36 as upstream writes them, our single-workgroup flag is 37, gpuind's trailing block last; includes b75c4e6's fix). - Image tables: upstream's boykopovar#6454 reuse model, scratch indexes, selected records and pointer chains, with our bindless columns and resolve memo in front. - Pairs kept with ours as the default: boykopovar#6210 presenter epoch only with APS5_PRESENT_EPOCH=1; boykopovar#6402 table stores (our path); pending HTILE fills only with APS5_PENDING_HTILE_FILLS=1. boykopovar#6208 takes upstream's final code with DMA pipelined by default (APS5_NO_PIPELINE_DMA=1 off). AI-assisted: yes

Celegans12 mentioned this pull request Oct 15, 2026

fix(shader): keep faulting invocations in their wave and mask their later writes #6416

Open

12 tasks done

EagleBearerYS mentioned this pull request Oct 15, 2026

perf(shader): select the record of a T# whose offset the CPU evaluates #6243

Merged

11 tasks done

Pipodi force-pushed the feat/image-table-stores branch from 42ef89c to 9161151 Compare October 15, 2026 15:39

EagleBearerYS mentioned this pull request Oct 15, 2026

perf(agc): share one image table per heap and write epoch through capture, compare, resolution and descriptors #6454

Merged

11 tasks done

fdfawef324 mentioned this pull request Oct 15, 2026

fix(agc): fault image table entries the device or guest memory cannot hold #6481

Open

12 tasks done

This was referenced Oct 15, 2026

fix(agc): drop resource-cache entries that hold released sampled textures #6492

Merged

fix(agc): release the texture caches when a device allocation runs out of memory #6174

Merged

github-actions Bot added shader core/shader docs agc libSceAgc* labels Oct 15, 2026

binarycleric mentioned this pull request Oct 16, 2026

fix(agc): bind image and sampler tables through update-after-bind limits #6647

Merged

11 tasks done

SleglJan mentioned this pull request Oct 16, 2026

fix(agc): place device memory in host memory when the driver refuses a device-local allocation #5987

Merged

12 tasks done

This was referenced Oct 16, 2026

feat(recompiler): run single-wave wave64 compute programs at one lane per invocation on 32-lane hosts #5405

Merged

perf(agc): reserve the storage cache's budget and held textures, not an eighth, from the sampled budget #6783

Closed

SleglJan mentioned this pull request Oct 16, 2026

fix(shader): poison unsupported biased volume table entries #6881

Merged

12 tasks done

oneandonlydean self-assigned this Oct 17, 2026

Pipodi and others added 4 commits October 17, 2026 08:45

Pipodi force-pushed the feat/image-table-stores branch from 0229b5f to b7e21ca Compare October 17, 2026 07:01

SleglJan mentioned this pull request Oct 17, 2026

fix(agc): let the storage image cache budget follow the heap's use like the sampled one #7071

Merged

11 tasks done

yabanana mentioned this pull request Oct 17, 2026

feat(shader): resolve nested image tables through runtime pointers #7223

Merged

12 tasks done

oneandonlydean merged commit 8f0e276 into boykopovar:main Oct 17, 2026
8 of 9 checks passed

SuarezPM mentioned this pull request Oct 21, 2026

feat(shader): support 32-bit image atomics through keyed image tables #8792

Merged

12 tasks done

Navigation Menu

Repository navigation

feat(agc): store to images selected through image tables - #6402

feat(agc): store to images selected through image tables#6402
oneandonlydean merged 4 commits into
boykopovar:mainfrom
Pipodi:feat/image-table-stores

Pipodi commented Oct 15, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 15, 2026 •

edited

Loading

Uh oh!

EagleBearerYS commented Oct 15, 2026 •

edited

Loading

Uh oh!

Celegans12 commented Oct 15, 2026

Uh oh!

Pipodi commented Oct 15, 2026

Uh oh!

Dextroid17 commented Oct 15, 2026

Uh oh!

SleglJan commented Oct 16, 2026

Uh oh!

Pipodi commented Oct 16, 2026

Uh oh!

oneandonlydean commented Oct 17, 2026

Uh oh!

Pipodi commented Oct 17, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Conversation

Pipodi commented Oct 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Technical debt

Tested

Checklist

Uh oh!

github-actions Bot commented Oct 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

EagleBearerYS commented Oct 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Celegans12 commented Oct 15, 2026

Uh oh!

Pipodi commented Oct 15, 2026

Uh oh!

Dextroid17 commented Oct 15, 2026

Uh oh!

SleglJan commented Oct 16, 2026

Uh oh!

Pipodi commented Oct 16, 2026

Uh oh!

oneandonlydean commented Oct 17, 2026

Uh oh!

Pipodi commented Oct 17, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Pipodi commented Oct 15, 2026 •

edited

Loading

github-actions Bot commented Oct 15, 2026 •

edited

Loading

EagleBearerYS commented Oct 15, 2026 •

edited

Loading