[tv7]

What

ResolveImageTablesShared (077bc20) hands out an earlier resolution of a variant's image tables when the columns match. It keeps at most 4 per variant and 256 in all within ResolutionMemoWords = 8 Mi words (32 MiB), counting the columns the entries keep alive.

A title with large bindless image tables has more live resolutions than fit in 32 MiB, so they push each other out: the next result memo miss of a variant finds no entry for it and resolves its tables again, although its columns did not change.

This raises ResolutionMemoWords to 64 Mi words (256 MiB). The 256-entry and 4-per-variant bounds stay, so the memo still holds at most 256 resolutions. docs/dev/ARCHITECTURE.md gets the new number.

Tested

• Windows 11, MinGW-w64 GCC (WinLibs 15.2), Release, BUILD_TESTING=ON, on main d0852de: ctest -R "agc|shader|spirv|recomp|scalar_call" -j 12, 846 tests. All pass except shader_cache_runtime_sources (no Ninja on this machine), agc_shader_replay_tool and agc_driver_pm4_replay_tool (0xc0000139, the MinGW runtime), which fail the same way on plain main here. agc_driver_image_array_window_past_layer_limit does not build with this compiler on plain main either (requested alignment '65536' exceeds object file maximum 8192).

• Marvel's Wolverine (PPSA03671), RTX 3070 8 GB, i7-11700KF, 1440p, opening in-game scene. I counted, every 10 s, why ResolveImageTablesShared resolves (local counters, not part of this PR):

  Last four 10 s windows of each run, same scene and settings:

  budget	run	shared	resolved	of them: no entry for the variant	resolve time per 10 s	Populate per result memo miss
  8 Mi words (main)	1	245-276	131-139	128-136	1.62-1.68 s	0.48-0.60 ms
  8 Mi words (main)	2 (time only)				1.84-1.94 s	0.59-0.75 ms
  64 Mi words	1	86-258	46-128	0-10	0.27-1.29 s	0.39-0.69 ms
  128 Mi words	3	271-444	23-97	0-1	0.15-1.13 s	0.26-0.50 ms

  With the larger budget almost no resolution misses for lack of an entry; the ones left are real column changes, which come in bursts as the scene moves (the high windows). The live set settles at ~235-245 entries of 0.6-0.9 MiB each, inside 64 Mi words. (The 64 Mi words run had a build running beside it, so its times are less clean than its counts.)
  The frame rate of this title does not change (~0.25 fps in this scene): its frame time is set by the GPU mutex, which draw recording holds ~60 % of the time, and the resolutions run outside it. In the quiet windows the change saves ~1.4 s of CPU per 10 s there.
  Not a stock-main run: the title needs unmerged PRs (fix(shader): give the frontend's call analysis the dispatch's wave size #7816, feat(shader): read samplers through nested and pointer-keyed descriptor tables #7395, feat(shader): resolve nested image tables through runtime pointers #7223, perf(agc): resolve DISPATCH_INDIRECT thread dimensions on the GPU #8066, perf(agc): keep released image memory in the buffer pool for reuse by type and size #7870, fix(agc): reserve only the sampled textures alive from the storage image budget #7862, perf(agc): skip CPU-path indirect draws that draw nothing before buil… #7930) and local changes to reach this scene. So the numbers show the memo behaviour, not a main result.

Checklist

• Based on current main (d0852de); no open PR changes the resolution memo's bounds or eviction
• One topic per PR
• No comments in code except technical debt
• Unimplemented paths throw (no new paths)
• No notes, investigation .md files or images added
• New third-party code is a submodule built from source (none added)
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none
• Unblocks: none
• GPUs tested: NVIDIA GeForce RTX 3070
• AI-assisted: yes

🤖 Generated with Claude Code

[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.

[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 907	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 5371	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 5925	docs/dev/ARCHITECTURE.md
PR 6044	docs/dev/ARCHITECTURE.md
PR 6280	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 6412	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/ARCHITECTURE.md
PR 6488	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 7550	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 7642	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/ARCHITECTURE.md
PR 7684	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp docs/dev/ARCHITECTURE.md
PR 7689	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 7803	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 8077	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 8527	docs/dev/ARCHITECTURE.md
PR 8625	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 8672	docs/dev/ARCHITECTURE.md
PR 8998	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 9020	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 9335	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp
PR 9748	core/shader/recompiler/Optimization/src/ResourceMaterializer.cpp

[adriadam10]

The change itself is small and the ARCHITECTURE.md update matches it. Two things are missing before it can merge:

• Peak memory. The memo is a single global (resolutionMemo() in ResourceMaterializer.cpp) and this raises what it can keep alive from 32 MiB to 256 MiB of columns for every title and every host. docs/dev/CONVENTIONS and the memory rules ask for the bound (given) plus peak RSS before/after on a large workload. Please add the process RSS peak on main and with the change in the scene you measured, and say what frees the entries on a low-memory host.
• Performance report. The PR is labeled perf: please paste the aps5-summary.jsonl line and the tools/trace_digest.py output of a capture on main and one with the change, same title, scene and machine, as the bot asked. The local counters are useful context but are not that report.

Once both are in, this looks mergeable.

[tv7]

Thanks for the review. Closing this one instead of adding the report.

• Memory: nothing frees the memo early on a low-memory host. The cap is a fixed bound, so every title could keep up to 256 MiB of columns, as you said.
• Gain: in the Wolverine scene I measured, the bigger memo cut image table resolution from about 1.65 s to 0.15-0.6 s per 10 s, but the frame rate did not change. The saved time turned into GPU mutex wait, so a main vs PR capture would show the extra memory and no frame time gain. Also, that scene is only reachable with other unmerged changes, not on plain main.

The cost isn't worth it for every title, so I'm dropping it.