Upstream sync 9 (slice C): AGC driver onto Metal

[RINNECODER]

Upstream sync 9, slice C: AGC driver onto Metal. Range 18c438b10..upstream/main, scopes agc, agc-driver, libSceAgc, libSceAgcDriver (65 commits). 26 upstream commits are cherry-picked with -x, one each. Four fork commits follow: a debt-doc move, a Metal test, and two Metal adaptations. Commit messages that named games or title IDs were reworded.

Ported

On the Metal/Rosetta path (shared State.cpp, QueueState.hpp, Pm4.cpp, libSceAgc, or adapted in Metal code):

• 3960519b1 make host imports over writable pages on macOS (the must-have). Resolved onto the fork's ImportWatched, which does not have upstream's later staysWatched/WindowCopy path, so only the GuestArena::HostWrite around import() is added. No Metal change is needed: no Metal-path source uses write watching (GuestArenaWriteWatched/BeginHostWrite appear only in the Vulkan driver and its tests), so the Metal no-copy aliases never cover read-only watched pages.
• 15955c089 the sample location registers start at their clear-state value 0 (shared QueueState.hpp). Test list conflict resolved.
• acbc57321 skip CMASK decode on single-sample views with viewMip > 0 (shared State.cpp). Applied onto the fork's !multisampled check. The debt note goes into TechnicalDebt.md. The Vulkan test uses cmaskAddress, because the fork has no cmaskFastClear field.
• dd3c13ffd execute INDEX_ATTRIBUTES_INDIRECT (PM4 0x91) (shared Pm4.cpp). On Metal the packet takes the default path in MetalPacketExecution.mm: it is not register-only, so prior GPU work completes before Pm4::Execute reads the record. MetalPacketExecution.mm is unchanged. The debt note moves to TechnicalDebt.md (ea6808d62).
• 7370aaf6d decode GS_FAST_LAUNCH geometry with one point per subgroup (shared State.cpp). Metal adaptation 05a477abd: the Metal mesh path refuses indexed fast-launch draws, as upstream's Vulkan Draw does. This is one line in MetalDraw.mm.
• ff3eae6c9 sample non-sRGB textures through FORCE_DEGAMMA samplers as sRGB views. The upstream test conflict is resolved by dropping its ANISO_OVERRIDE line (that helper is not in the fork) and using the 3-argument RequireSame. Metal adaptation 0a8dc74ff in MetalShaderResources.mm: the bound view is re-viewed with the sRGB twin of its pixel format (R8, RG8, RGBA8, BGRA8, BC1/2/3/7), keeping its levels, slices and swizzle. A format without a twin is refused.
• d6f67d1be keep fused geometry programs out of the caller's memory (libSceAgc ShaderFusion.cpp and the Vulkan ShaderRegistry.cpp; the registry helpers are added beside the fork's). The fused code now lives in libSceAgc-owned host memory. On the Rosetta direct path the CPU reads it through the host page check (HostAccessible), not the GPU ranges.
• fbf619ef0 sceAgcDcbEventWrite ignores the address on non-sample-dump events (libSceAgc). docs/dev/debt/libSceAgc.md does not exist in the fork, so the note goes into TechnicalDebt.md with the title name and ID dropped.
• 0ba462f8e, af19dbfab skip an indexed draw with no indices or instances. The Vulkan Draw.cpp keeps the fork's IndexReadCount. The test keeps the fork's rect-list rejection and adds the two zero-count draws. Metal already returns early for these draws (MetalDraw.mm).

Vulkan driver side, carried to stay close to upstream (not built on macOS):

• 3bc3e35a0 fill kernel that reads its pattern from a buffer
• 090058619 APS5_SKIP_SHADER (Vulkan draws and dispatches only; resolved onto DecodeComputeProgramAddress)
• abc997b0d release the cached address space on device teardown (test main: only the teardown case added)
• 12d07b8ed, a141b8eb6 R8/R8G8 sRGB twins
• e2ee9430b, 65eb466dd, a296b9cac, 3f314030b, 0612c181f CPU snapshots on demand and their fixups
• 0adbdcf45 draw key cache
• fce8aa8c7, 9a1ac6a23 queued-write groups and their ctest entry (only agc_queued_write_groups registered; the other upstream entries need tests the fork lacks)
• 71999169a, faef399e8, 7262b10ae layers of an array surface, slice-0 array target sampled as 2D

Skipped

• Already on main: baaad78e8, 8c554cc42 (Upstream sync 6: shared AGC driver fixes onto Metal (slice A, first chunk) #442); d4b59ce9f (Port upstream AGC driver shared slice (U5) #475); c6a422342 (Upstream PR batch 1 U8b: Metal depth, MSAA and fill-kernel correctness #484); d7818c9e5 (Upstream PR batch 1 U7: tessellation and draw-state keys #483).
• Vulkan presentation chain and resident present: c60b52d2d, 4589edf7f, a062eda97, ed95e6618, 506215722, 3ddfc2395, ccd8591fb.
• Pipeline cache and library: 4307e3cb5.
• Windows only: aec5ba133, 83787e692, c99c89962.
• Upstream MSAA model (MultisampleTarget, not in the fork, which has its own model): f7fd7569f, 5acd68f94, 3ebad9045.
• Depend on Vulkan series not yet ported:
  • draw snapshot cap: 0a1dff2fc, d6698c904
  • BC block copy / PerUseCopy: 126a1e413, cf9dd4f18
  • depth storage reuse: 696d1660c
  • HTILE memory-value fills: 6ffc99bed
  • partially committed snapshots: 955b56a2b, 8d8fe39ee
  • BDA host-import windows and holds: ca95f7409, d275a5881, e9249a0a4, c871b5893, c29e06c69, 972ce1224, e5d65569e, fd2c300e4
  • depth plane uploads: 875ecc139
  • pending storage results on unwatch: e56a7c235, db4fd5f29
• Shader recompiler TessellationLowering conflict, left to the shader slice: 866ed445b.

Evidence

• arm64 guest replay config (-DANYPS5_METAL_UTILITIES_ONLY=ON -DANYPS5_METAL_SHADER_BRIDGE=ON -DANYPS5_BUILD_METAL_SHADER_BRIDGE_TESTS=ON -DANYPS5_METAL_GUEST_REPLAY=ON): full build, then ctest excluding the GUI anyps5_metal_presentation_replay. Before 0a8dc74ff: 57/58 passed. The failure was anyps5_metal_driver_semantics_imagesampleforcedegamma, which runs the ported upstream execution test on the Metal driver: "guest format 56, which is not sRGB, ... not implemented". After 0a8dc74ff, the affected set (driver_semantics*, render_state) passed 10/10.

• anyps5_metal_render_state gains Sync9SharedDecode. It covers the sample-location defaults, CMASK at view mip 1 versus mip 0, and INDEX_ATTRIBUTES_INDIRECT loading and refusal. With the QueueState.hpp line removed, it fails with "a sample location register does not start at its clear-state value 0".

• Rosetta x86_64 (-DANYPS5_ROSETTA_LIBS=ON): rosetta_libs rosetta_engine metal_utility_tests rosetta_direct_graphics_tests built, and ctest metal_utility_tests|rosetta_direct_graphics passed 2/2.

• Vulkan-device targets (tests/Graphics.cpp, tests/Recorder.cpp, tests/Pm4.cpp, most tests/execution/*) are not built on macOS. Their hunks are carried but not compiled or run here.

• Game smokes (tools/smoke.sh <build> 300 --map, eboot lock, private image clone), same build dir:

  head	main menu	leader select	map generation	fault	skipped draws/dispatches
  main 304c53c5d	<= 47 s	<= 56 s	<= 77 s	none	4 lines: 1 draw, 1 dispatch with unresolved DCC metadata; 1 + x10 dispatches with mismatched texture dimensions
  05a477abd (before the Metal degamma commit)	<= 51 s	<= 64 s	<= 83 s	none	same 4 lines
  0a8dc74ff (final)	<= 51 s	<= 63 s	<= 83 s	none	same 4 lines

  All three runs reached the 3D map and ran to the 300 s wall (exit 0 on SIGTERM), with 0 unresolved imports. The only unimplemented call is the same Ajm Opus mode in each. The timings are asset-open times from a shared machine running other agents. They are not a performance comparison.

• No performance claims; nothing was measured for speed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DgW4n3qYgi5jAAohEmXRQj

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T08:15:24.399667Z	0a8dc74	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: d7148766-4b60-4bb8-8fbe-cfa6d7e0db9b

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0a8dc74ff3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid applying the parent texture swizzle twice

When FORCE_DEGAMMA is used with a non-sRGB texture whose guest descriptor has a non-identity channel mapping, view is already a swizzled child created by MetalTexture::createViews, and passing view.swizzle while creating another child composes that same mapping a second time. Metal explicitly combines a child view's swizzle with its parent's (Apple documentation), so mappings such as R/B swaps can be undone and produce incorrect sampled colors; create the sRGB child with an identity-relative swizzle so it preserves the parent's effective mapping.

Useful? React with 👍 / 👎.