[meh7an]

What

ResourceCache holds the shader resource objects of draws and dispatches. Its keys are about a hundred words: the device, the stages and their bindings' guest descriptors. Every draw and every dispatch looks one up (Find, or Touch on a recipe hit). The cache indexes its LRU list with an std::unordered_map<Key, iterator, KeyHash>:

• KeyHash is FNV-1a over every word: one multiply per word, each waiting on the one before.
• KeyHash is noexcept, so libstdc++ keeps no hash code in the nodes. A lookup's bucket walk hashes the key of every node it passes again, to find where its bucket ends.

The index is now a HashSlotMap (Execution/include/HashSlotMap.hpp) keyed by the key's hash:

• It uses open addressing in a power-of-two table kept at most half full. A lookup masks the hash, scans a compact array of hashes, and compares the one key it finds.
• The hash is FNV-1a in four lanes over pairs of words, so their multiplies overlap. The lanes are then folded so that the low bits, which pick the slot, depend on all of them. The hash is taken before the mutex.
• When two keys have the same 64-bit hash, the newer entry replaces the older one, counted as an eviction.
• An entry keeps its hash for the evictions (the entry bound and the table element budget) and for RemoveImage.
• Erasing shifts the entries after it back, so the table has no tombstones.
• hashSlotMapTests (graphics tests) covers:
  • inserts, a repeated insert, and growth to 1003 keys;
  • erases, including the backward shift of an erase whose probe run wraps past the end of the table;
  • key 0, which is stored as 1;
  • clearing.

On the S2 page of Sonic Origins Plus's main menu, the queue worker spent, per frame (APS5 sampler, 15 s):

• ResourceCache::Find: 0.52 to 0.27. In a profile of an earlier build, 45 of its 70 samples were its own code, mostly the inlined FNV over the key, and 23 the bucket walk.
• std::_Hashtable::_M_find_before_node, the bucket walk: 0.26 to 0.06. What remains belongs to other hash tables.

That is about 1.5% of the worker's frame (31 samples).

#6492 adds another walk over the cache's entries that reads it->first and it->second. Whichever of the two lands second changes them to it->hash and it->resources.

Tested

Windows 11 (10.0.26300), WinLibs GCC 15.2.0 (UCRT), Release, NVIDIA GeForce RTX 2080 Ti, driver 617.14.

• On main 7d728b1 with this change, the full ctest suite passes 1337 of 1341 tests.
  • agc_driver_gpu_indirect_arguments_misaligned_writer_ahead is not run: main itself does not compile it with MinGW (alignas(65536)).
  • select_tests fails in its setup on Windows without the symlink privilege (WinError 1314 from symlink_to), as on main.
  • agc_driver_multisample_fill_host_only failed once ("the sentinel fill did not land"). That is main's own race, agc_driver_multisample_fill_host_only fails intermittently: the sentinel fill after an absorbed multisampled clear does not land #6512, which fix(agc): record a multisampled target's commands on its own device #6522 fixes. With fix(agc): record a multisampled target's commands on its own device #6522's line added it failed 0 of 200 runs (8 at a time).
  • guest_positioned_close (libkernel and libc only) ended with 0xc0000409 once in the parallel run. Alone it passes 5 of 5.
• Sonic Origins Plus, main menu, S2 page (the 3D background), 1233x694 window.
  • A sampled pair (APS5 sampler, 15 s) took the profile numbers above. The FPS shown in the title went from 10.56 to 10.84 (median of six readings).
  • Captured with APS5_TRACE_SECONDS=200:15 on the same machine in one session, caches warm. Traces attached.
  • Baseline 245e355 (perf(shader): keep a binding's element flags inline #6970), this change 557d4da (the summaries name the commits before them, see the note):

Metric	Baseline	This PR	Change
frame_ms.p50	100.704	98.072	-2.6%
frame_ms.p90	106.216	101.147	-4.8%
mean_frame_ms	103.202	98.211	-4.8%
gpu_busy_ms_per_present	83.660	83.570	-0.1%
Driver.Packet (ms per frame)	101.155	96.211	-4.9%
Driver.Draw (ms per frame)	89.173	83.863	-6.0%

This pair overstates the change: its target is about 1.5% of the worker's frame, and windows of this page vary by a few percent at the median. The title FPS during the two captures was 10.82 and 11.12. The sampled pair, 10.56 to 10.84 FPS, is the better estimate.

Note: measuring base. These numbers come from my fork's dev, on main 0a56aaa. The resource cache this change indexes is the same on main 7d728b1, which adds the table element budget and RemoveImage to it.

• The baseline is measure/flags-2026-10-17 (245e355, perf(shader): keep a binding's element flags inline #6970).
• This change is measure/keyhash-2026-10-17 (557d4da).
• Each build was made before its commit, so the summaries name the commit before it, dirty: b1909f4 for the baseline and 245e355 for this change.

Both carry these commits from my fork:

• the fixes a71073c, c0d8756, e3c6896, 604f0f3 and 10ff693 (fix(agc): bind a V# past the range limit by the memory behind it #5778);
• make-exe and the stage dumps (522230d, deee71f);
• the crash-report sampler (8d9c310..f641e1e), which took the profiles;
• perf(agc): present 8-bit integer resident images without a readback #6037, perf(agc): skip the register read log while no log is open #6066, perf(agc): skip the binds a continued draw repeats #6077, feat(libs): let relinked Windows games opt into mimalloc #6067, perf(agc): copy small draw snapshots into a per-recorder upload chunk #6082 (its first version, before perf(agc): reuse draw snapshots across draws that nothing wrote between #5906), perf(shader): compile resource plan walks instead of interpreting them #6095, perf(memory): skip 4 MiB runs no write stamped since a range check #6266, perf(agc): replace winpthreads locks and TLS on the draw path #6292, perf(agc): cut per-draw heap allocations #6312, perf(agc): rule out pending write overlaps from the published union #6609, perf(shader): validate vertex stage inputs without building them #6611, perf(agc): move per-draw thread-locals to TEB slots #6615, perf(shader): keep memoized results in a set-associative table #6733 and perf(shader): keep a binding's element flags inline #6970;
• smaller steps kept in my fork:
  • e832602, 97b1ff6, 84f5bf0, ce567bc, a79e58a, a0a8916, 63f7a8e, b048061, 719fb71 and a21a3ce;
  • 3b47f77 (covered on main by perf(libc): walk shared views in order and remember collectable ranges in Windows write collects #5062);
  • e3b8d1c (which perf(agc): fewer heap allocations per recorded draw #6566 covers);
  • 390f624 (a draw decode cache; main reuses decodes through its draw data hits);
  • 40a9184 (the pipeline key outside the store's lock, as perf(agc): journal the pipelines of a run and build them ahead on the next one #5875 does on main, and the thread's last pipeline);
  • b1909f4 (the vertex phase's lists inline, held until perf(libSceAgcDriver): optimize draw descriptor counts and reuse vertex planning vectors #6661 lands).

The comparison tool marks the window's first present as a configuration difference: the window opens 200 s into each run. Both summaries set trace_dropped, because the worker filled its event buffer at about 2,700 draws per frame. The phase totals above stay complete.

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
• Unblocks: none
• GPUs tested: NVIDIA GeForce RTX 2080 Ti, driver 617.14, Windows 11
• AI-assisted: yes

Traces and summary lines (baseline 245e355, whose summary says b1909f4; this change 557d4da, whose summary says 245e355):
keyhash-traces.zip

[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/tests/Graphics.cpp
PR 4040	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5441	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5499	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5500	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 5925	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6146	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6177	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6188	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6192	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6194	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6202	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6223	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6235	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6280	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6412	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6481	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6498	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6548	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6636	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6683	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6808	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 6946	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6947	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 7002	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 7085	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7154	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 7256	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7484	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 7642	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 7668	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 7689	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 7800	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 7948	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Graphics.cpp
PR 8047	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp

[keravyn-ops]

AMD / Windows run of this PR at 5b3c1ea: Radeon 780M (RDNA 3, 0x15bf), AMD driver 25.10.20.01 (32.0.21020.1007), Windows 11 Home 10.0.26100, WinLibs GCC 15.2.0 r7, Release.

Every ctest entry built from the test targets this PR changes passes (14):

• agc_driver_undithered_alpha_to_coverage_state (0.11 s): 289: Existing undithered alpha-to-coverage states passed
• agc_driver_alpha_to_coverage_state (0.02 s): 290: Alpha-to-coverage state and refusal tests passed
• agc_driver_prepared_graphics_prebuild_unset (0.16 s): 292: prebuild policy native readback passed
• agc_driver_prepared_graphics_prebuild_0 (0.13 s): 293: prebuild policy native readback passed
• agc_driver_prepared_graphics_prebuild_1 (0.13 s): 294: prebuild policy native readback passed
• agc_driver_graphics (4.93 s): 327: Graphics validation tests passed
• agc_driver_prepared_graphics (0.13 s): 588: prepared graphics ABI tests passed
• agc_driver_prepared_unregistered_graphics (0.15 s): 591: prepared unregistered graphics tests passed
• agc_sparse_buffer_holes (0.04 s): 708: sparse buffer hole tests passed
• agc_pipeline_journal (0.07 s): 712: pipeline journal recipe, record and replay tests passed
• agc_descriptor_pool_reset_failure (0.02 s)
• agc_driver_address_space_teardown (0.02 s): 767: address-space device teardown releases cached guest allocations
• agc_compute_descriptor_layout_flags (0.02 s): 798: Descriptor layout native and captured binding flags passed
• agc_driver_covered_centroid_state (0.02 s): 799: Covered-centroid and multisample state checks passed

AI-assisted: yes (runs made on the device above).

[adriadam10]

Thanks, the HashSlotMap itself reads correctly (growth, backward-shift erase including the wrapped run, key 0, and the test pins them). It now conflicts with main, and not mechanically:

• ResourceCache on main has adaptive capacity (capacity, fixedCapacity, baseEntries, shrink window) and new helpers noteEvicted(key), noteReturn(key), noteUse(key), dropReleasedTextures(evicted). Find, Insert and Touch call them, and Insert no longer uses the static capacity lambda your diff edits (ShaderResources.cpp around the Insert hunk).
• noteEvicted/noteUse still take a Key and use KeyHash{}(key) for their own tables, so KeyHash cannot simply be removed. Decide whether they take the new hashKey value (then evictedKeys/usedKeys are keyed by it) or keep the old hash.
• ResourceCache has other places that read entries as pairs (->first/->second) on main; they need the Entry struct too.

Please rebase on current main and re-measure: noteUse adds work per Find/Touch, so the ~1.5% of the worker's frame may differ. Keep the trace-vs-sampler caveat in the body.

[github-actions]

This pull request is labeled perf, so it needs the summary line of a capture on main and one with the change, from the same title, scene and machine, as the Performance report form asks. Please paste both aps5-summary.jsonl lines, attach the two traces and add the output of tools/trace_digest.py.

[meh7an]

Closing this: on current main it shows no gain.

• Rebased onto main 1d41645 (ae33fb5).
  • Main's adaptive-capacity helpers take the four-lane hash now, so KeyHash and its pass over the key on every Find and Touch are gone.
  • The full suite passes 1738 of 1741. The three that don't also fail on plain main: agc_driver_pipelined_dispatch_item failed 1 of 40 runs under load on both.
• Re-measured on the Sonic Origins Plus S2 page, with three interleaved runs per side on main 1d41645:
  • frame p50: main 86.25 / 86.83 / 91.13 ms, this change 85.73 / 87.07 / 104.35 ms (median +0.3%);
  • title FPS: main 11.91 / 12.22 / 12.41, this change 11.70 / 11.80 / 12.33.

That isn't worth a new map in the cache. Thanks @adriadam10 for the review and @keravyn-ops for the AMD run.