perf(agc): copy small draw snapshots into a per-recorder upload chunk

[meh7an]

What

ShaderResources::PrepareDrawBindings snapshots each guest buffer that a draw reads in place. Since #5906, a snapshot that nothing wrote since the last draw is reused from a per-recorder map (Recorder::UnwrittenDrawSnapshot). On a miss, the draw still collects the range's writes, looks for a kept copy (Recorder::ReusableDrawSnapshot), and otherwise creates a buffer for the copy and keeps it. Most of these ranges are small.

On the S2 page of Sonic Origins Plus's main menu, with #5906, PrepareDrawBindings costs the queue worker 5.22 samples per frame (APS5 sampler, 15 s). Of that, 0.78 goes to Buffer::Buffer, 0.56 to ReusableDrawSnapshot, 0.35 to KeepDrawSnapshot and 0.29 to CollectWrites.

A miss of at most 4 KiB is now copied into a per-recorder upload chunk (Recorder::DrawUpload) and bound at its offset. The chunk is 1 MiB, bump-allocated at the storage buffer offset alignment. The copy skips the collect, the lookup and a buffer of its own.

• DrawBindings::Snapshot gains the offset and size of its bytes in buffer. The descriptor write and the capture inputs use them.
• perf(agc): reuse draw snapshots across draws that nothing wrote between #5906's map keeps the offset with the buffer, so a later draw reuses the copied slice the same way.
• The copy is noted in that map only within a collect epoch (CollectEpoch() != 0), where the queue worker's collects are memoized the same way. A thread without one copies again on every draw.
• Larger snapshots and moved words take the existing paths.
• APS5_DRAW_UPLOAD_MAX changes the limit, and 0 turns this off.

With the chunk, PrepareDrawBindings fell from 5.22 to 3.20 samples per frame:

• Buffer::Buffer 0.78 to 0.31
• ReusableDrawSnapshot 0.56 to 0.03
• KeepDrawSnapshot 0.35 to 0.01
• CollectWrites 0.29 to 0.11

Tested

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

• On main d92a8da with this change, the full ctest suite passes (1253 tests).
  • agc_draw_snapshot_cap now also runs drawSnapshotUploadTests. It checks that inputs up to the limit share the chunk at aligned offsets with their guest bytes, and that a larger one gets its own buffer. A copy made after a CPU store gets a new slice without touching the slice an earlier draw reads, and a full chunk is replaced.
  • unwrittenSnapshotTests checks that the map keeps the offset.
  • The snapshot tests of agc_driver_recorder_tests's unregistered default path read a snapshot through its offset now: resourceReadTests, misalignedRegionTests and misalignedSnapshotTests. drawSnapshotReuseTests covers the kept-copy path, so its input is now 8 KiB, above the limit. No ctest entry runs that path; run through a local flag, its snapshot tests pass on that main with this change. On Windows, drawSnapshotReuseTests, drawSnapshotPatchTests and the kept-indices part of drawInputInPlaceTests stop early: host imports are compared there, not watched.
• Sonic Origins Plus, main menu, S2 page (the 3D background), 1233x694 window.
  • Captured with APS5_TRACE_SECONDS=200:15 on the same machine in one session, caches warm. Traces attached.
  • Both captures ran one build. The baseline turns the chunk off with APS5_DRAW_UPLOAD_MAX=0; this change uses the default limit:

Metric	Baseline	This PR	Change
frame_ms.p50	142.300	136.092	-4.4%
frame_ms.p90	148.909	140.703	-5.5%
mean_frame_ms	143.817	137.142	-4.6%
gpu_busy_ms_per_present	83.726	84.797	+1.3%
Driver.Packet (ms per frame)	141.207	134.532	-4.7%
Graphics.Draw (ms per frame)	60.930	54.816	-10.0%

The FPS shown in the title went from 7.35 to 7.84 (median of six readings). The sampled runs agree: 7.28 to 7.72.

Note: measuring base. These numbers come from my fork's dev, on main 0a56aaa, the last main before #5596. On Windows, later mains compare direct imports on every use, which keeps this menu at 3-4 FPS and hides a change of this size. The build is measure/tape-2026-10-14 (b593e36; its summaries say a38ad5e, the same tree with a test-only fix folded in), plus #5906 ported and this change adapted to it. That patch is measuring-build.patch in the attached zip. The base therefore carries, in both captures:

• 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 and perf(shader): compile resource plan walks instead of interpreting them #6095;
• perf(agc): reuse draw snapshots across draws that nothing wrote between #5906 (from the patch);
• seven smaller steps kept in my fork: e832602, 97b1ff6, 84f5bf0, 52704cf (covered on main by perf(libc): walk shared views in order and remember collectable ranges in Windows write collects #5062), c0c3a6b, 467f668 and f7a41ad.

The comparison tool marks two configuration differences: APS5_DRAW_UPLOAD_MAX, which is the change under test, and the window's first present (2593 vs 2471), because 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. Both summaries also say dirty: the patch above was applied, not committed.

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, summary lines and the measuring patch (one build; baseline APS5_DRAW_UPLOAD_MAX=0, this change the default):
snapshot-5906-traces.zip

[github-actions]

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 7484	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp L5085

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 1191	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp
PR 1368	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 1395	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 4072	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp
PR 5021	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 5441	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp
PR 5499	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp
PR 5609	core/libs/prx/libSceAgcDriver/Graphics/include/Recorder.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 5925	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6177	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6185	core/libs/prx/libSceAgcDriver/Graphics/include/Recorder.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6186	core/libs/prx/libSceAgcDriver/Graphics/include/Recorder.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6188	core/libs/prx/libSceAgcDriver/Graphics/include/Recorder.hpp core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6192	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6194	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6222	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6223	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6279	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp
PR 6360	core/libs/prx/libSceAgcDriver/Graphics/include/Recorder.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6412	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6414	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6481	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6490	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6499	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6529	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp
PR 6539	core/libs/prx/libSceAgcDriver/Graphics/include/Recorder.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6658	core/libs/prx/libSceAgcDriver/Graphics/include/Recorder.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6747	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6763	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6868	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6905	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6937	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 6971	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 7002	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 7085	core/libs/prx/libSceAgcDriver/Graphics/src/Draw.cpp
PR 7332	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp docs/dev/ENVIRONMENT.md
PR 7390	core/libs/prx/libSceAgcDriver/Graphics/include/ShaderResources.hpp core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp
PR 7464	core/libs/prx/libSceAgcDriver/Graphics/include/Recorder.hpp core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 7515	core/libs/prx/libSceAgcDriver/Graphics/src/ShaderResources.cpp

[adriadam10]

Checked the diff against main: upload slices are bump-allocated and never rewritten (a full chunk is replaced and the old one stays alive through the snapshots that reference it), KeepBytes counts the chunk once per batch, the #5906 stamp is still checked after the copy, and the capture inputs and descriptor writes use offset + size. Linux build and the libSceAgcDriver ctest scope are green on RADV and lavapipe (only agc_driver_tessellation segfaults, as on main); agc_draw_snapshot_cap runs drawSnapshotUploadTests here. The measured gain is the author's Windows/NVIDIA run, not reproduced here; merging through the full Build.