[mrfelek]

What

A buffer V#'s presence (dwords[2] != 0 && (dwords[0] != 0 || (dwords[1] & 0xffff) != 0)) was supplied as specialization constant BufferBase + i*BufferWords + 2 from the dispatch's V# snapshot, and buildSpecializedModule turns supplied specializations into literal constants. The same guest program therefore compiled to two SPIR-V modules, one where the buffer was bound and one where it was still null, with different RecompileResult::Identity() and no VkPipelineCache hit between them; which one a dispatch got depended on timing.

PPSA26344 has eight compute programs of 1.2–1.45M SPIR-V words reached in both states; every new state cost a 54–320 s NVIDIA pipeline compile, and pipeline journal replay of the other state did not help.

descriptor() in SpirvRuntimeBuffer.cpp now computes presence from the V#'s words at run time, as the registration-module path already does; the formatted-buffer helper takes it from its existing valid parameter, and DescriptorBindingBuilder no longer supplies BufferBase + 2. The GuardedMemoryLoads fixture now encodes validity in the V# words instead of the specialization. ShaderDiskCache FormatVersion is bumped because generated SPIR-V changes.

Tested

Ubuntu 26.10, Ryzen 9 5950X, RTX 3080 Ti 12 GB, NVIDIA 615.71.09. PPSA26344 02.002.000 title-scene load (with a separate blend change applied locally to get past its earlier stop).

• Before: captures of the 1,231,429-word program from two runs differed in exactly one specialization literal (BufferBase+2: 1 vs 0) and 20 OpLogicalAnd folding it. Warm run: pipelines over 100k words took 1,221 s in the first 15 min because 5 journal-replayed giants were rebuilt at 54–70 s each.
• After, cold run: one capture per giant across 8 programs (+~2% SPIR-V each).
• After, warm run: identical capture hashes, no giant rebuilt (journal replay 48 ms–1.26 s, live 18–74 ms); pipelines over 100k words took 21.8 s in the first 15 min.
• Focused tests (tools/select_tests.py, 1143 selected): 1140 pass; the same 3 failures (agc_driver_sample_order, agc_driver_sample_order_grid_reference, agc_driver_sample_order_recipe_depth_locations) occur on main 895f47f on this machine.
• Same tests with ANYPS5_REQUIRE_VULKAN=1 ANYPS5_NO_SHADER_CACHE=1 VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation on main and on this change: the same 3 failures and identical VUID sets (four pre-existing kinds), no new validation errors.

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: PPSA26344 Ghost of Yōtei (PPSA26344) #675
• GPUs tested: NVIDIA RTX 3080 Ti, driver 615.71.09, Ubuntu 26.10
• AI-assisted: yes

Related: #639 added registration-time modules, which already read presence from the V# at run time; this brings the dispatch-time specialization path in line.
Conflicts with #7484 only in the GuardedMemoryLoads test fixture (emitLoad/AccessTrace signatures); whichever lands second needs a small fixture rebase.

AI-assisted. Tested by me on my own machine.

[MarcZX10]

I reviewed the exact head 5106dae. The change removes dispatch-time buffer-presence specialization from the cache identity and derives it from the descriptor words at runtime. The specialized GCN path masks the high-base word before testing its low 16 bits, and the formatted-buffer helper now receives validity from its existing runtime parameter. The GuardedMemoryLoads fixture encodes descriptor validity in the V# words, and the shader disk-cache format is bumped. I found no correctness issue in those paths.\n\nLocal checks run: and All conventions checks passed. both pass. I did not run the local GPU test suite because no usable build tree is available and the root filesystem is full. I inspected the reported same-machine NVIDIA baseline/change results: the author reports 1,140/1,143 selected tests passing on both, the same three existing sample-order failures and identical pre-existing VUID sets, plus matching warm capture hashes across eight large programs. Hosted relinker/product build jobs are policy-skipped for this PR, so I am relying on the author-provided runtime report for GPU validation.

[MarcZX10]

Correction to the review text: the local commands were git diff --check upstream/main...refs/review/9957 and python3 tools/check_conventions.py --base upstream/main --head refs/review/9957; both passed. The shell removed the inline command text from my initial review body. The test and hosted-check attribution in that review is otherwise unchanged.

[MarcZX10]

Reviewed the four-file change and the reported same-machine main/change Vulkan validation runs. The dispatch-time path now derives buffer presence from V# words at run time; the formatted-buffer path uses the existing runtime parameter, and the disk-cache version is bumped for the SPIR-V change. The fixture now encodes validity in descriptor words rather than specialization, matching the production behavior. The author reports the same three known test failures and VUID set on main and the change, with no new validation errors; required intake/conventions checks pass. Approved at head 5106dae.

[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 5371	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp
PR 5609	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp
PR 6280	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp
PR 6412	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp core/shader/recompiler/SpirvBackend/src/SpirvRuntimeBuffer.cpp
PR 6539	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp
PR 7484	core/libs/prx/libSceAgcDriver/tests/GuardedMemoryLoads.cpp core/shader/recompiler/SpirvBackend/src/SpirvRuntimeBuffer.cpp
PR 7642	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp core/shader/recompiler/SpirvBackend/src/SpirvRuntimeBuffer.cpp
PR 7689	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp core/shader/recompiler/SpirvBackend/src/SpirvRuntimeBuffer.cpp
PR 9897	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp
PR 10049	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp
PR 10193	core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp core/shader/recompiler/SpirvBackend/src/SpirvRuntimeBuffer.cpp

[klosixgaming-arch]

Hardware & Environment

• OS: Windows 11 Version 24H2 (OS Build 26100.3194)
• CPU: AMD Ryzen 9 9950X3D (16 Cores, 32 Threads)
• GPU: NVIDIA GeForce RTX 5090 32 GB (Driver: 617.42)
• Toolchain: MinGW-w64 GCC 15.2.0, Ninja 1.13.0, CMake 3.31.6

---

Test Methodology & Verification

1. Security Audit:

   • Inspected commit 5106dae4ce37c9217addcf6d63bf744953d236d3 across all 4 touched files:
     • core/shader/recompiler/SpirvBackend/src/SpirvRuntimeBuffer.cpp
     • core/shader/recompiler/Optimization/src/DescriptorBindingBuilder.cpp
     • core/shader/recompiler/ShaderDiskCache.hpp
     • core/libs/prx/libSceAgcDriver/tests/GuardedMemoryLoads.cpp
   • Verified no untrusted network calls, credential exfiltration, obfuscated binaries, or external downloads.

2. Negative Baseline Verification:

   • Reverting the runtime buffer changes in SpirvRuntimeBuffer.cpp while retaining the updated test fixture reproduces the pipeline specialization mismatch, with agc_guarded_memory_loads failing due to unexpected specialization constant behavior when binding dynamic vs null buffers.

3. Positive Test Results:

   • Recompiled and executed targeted suite:
     • agc_guarded_memory_loads: PASSED (0.03s)
     • agc_driver_graphics: PASSED (7.21s)
     • shader_cache_native_paths: PASSED (0.99s)
     • shader_cache_runtime_sources: PASSED (55.34s)
   • Reading buffer V# presence at runtime directly from descriptor words instead of specializing via specialization constant BufferBase + 2 avoids generating redundant SPIR-V module permutations for bound vs null descriptors and prevents pipeline cache bloat.

4. Conventions:

   • python tools/check_conventions.py --base origin/main passed cleanly with zero warnings or errors.

LGTM!

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

[mrfelek]

Prior art I missed when opening this: #5353 (@Pipodi) moved both buffer stride and presence to run time for the same Yōtei compute programs. It was closed on Oct 17 because, with cold caches, the first frame came later (261 s) and total compile work didn't drop.

How this one differs:

• It covers presence only. Stride, the format word and base misalignment stay specialized. The SPIR-V grows ~2% per giant (e.g. 1,231,429 → 1,255,837 words).
• The gain is on later launches, not the cold one. Before this change, a giant first seen with a buffer still null and later with it bound (or the other way round) got a second identity. The journal-replayed pipeline then missed and was rebuilt live at 54–70 s each. Warm run, first 15 min: 1,221 s → 21.8 s for pipelines over 100k words, with no giant rebuilt.
• Cold cost is about the same, as perf(recompiler): read the buffer stride and presence at run time in compute modules #5353 found. The cold run with this change compiled one module per giant (8 programs, 68–78 s each, 320 s for the 1.45M-word one). I haven't measured cold time-to-first-frame against main.
• I haven't measured per-dispatch GPU time, which @CnRJay asked for on perf(recompiler): read the buffer stride and presence at run time in compute modules #5353 (Pipodi measured no per-dispatch GPU cost for the stride+presence variant there). I can add an A/B if wanted.

The format-word variant Pipodi found on Oct 22 (discussion #675) is a separate key; this PR doesn't touch it.

This post was prepared with AI assistance (Claude); I reviewed the results.

[BKlaasWerkman]

Intel run, since the generated SPIR-V changes and the runs above are NVIDIA only. Head 5106dae was merged into main 15e71db (clean merge) and compared with that main on the same machine.

Setup: Windows 11 Pro 10.0.26300, WinLibs GCC 15.2.0 r7, Release. Environment: VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation (1.4.363), ANYPS5_REQUIRE_VULKAN=1 and ANYPS5_NO_SHADER_CACHE=1. The run covers the 1,250 tests that tools/select_tests.py selects for this diff.

GPU	this PR	main
Intel Arc Graphics (Meteor Lake, 101.9150)	1222 passed, 26 skipped, 2 failed	the same
NVIDIA GeForce RTX 3060 Ti (617.42)	1243 passed, 4 skipped, 3 failed	the same

• No test changes result, and on both GPUs the validation messages are the same as main's.
• agc_guarded_memory_loads, with the updated fixture, prints the same seven passing cases on both GPUs.
• The failures also occur on main:
  • On Intel, agc_storage_write_back and agc_driver_multisample_retirement fail.
  • On NVIDIA, the agc_driver_sample_order trio fails.

AI-assisted: yes (Claude Code)

[oneandonlydean]

@mrfelek this now conflicts with main in ShaderDiskCache.hpp: the branch bumps FormatVersion 44 -> 45, and main is already at 46 (#8564's 4_4_4_4 target change and the one before it). Could you merge main and take the next free number? That's 47 today, but #9852 and #10049 are also moving the cache format, so whichever lands first keeps 47 and the others move up. The rest of the branch merges cleanly, so it's just the version line, then a CI rerun.

[mrfelek]

@oneandonlydean Rebased onto main 887a964 and took FormatVersion 47; full ctest with validation matches main (same 3 sample_order failures). AI-assisted. Tested by me on my own machine.

[oneandonlydean]

@mrfelek Thanks for the rebase. A correction to my earlier ask, sorry for the churn: this PR doesn't need a FormatVersion bump at all. It changes runtime SPIR-V generation, not the encoded cache layout (there's no ShaderDiskCache.cpp change), and each changed runtime source already changes SourceVersion, which keys the cache directory and is checked on read. Could you drop the ShaderDiskCache.hpp FormatVersion edit and keep main's value? Several other open PRs need the next numbers.

[mrfelek]

@oneandonlydean done: dropped the FormatVersion change and kept main's value; the PR now only touches the runtime SPIR-V and its test. AI-assisted. Tested by me on my own machine.