[RINNECODER]

Upstream PR batch 1, U6: shader semantics correctness (ALU, LDS, export)

Ports 14 of the 16 triaged upstream correctness fixes, one commit per upstream PR ("Ported from upstream PR #N"). Frontend and SPIR-V changes feed our Metal bridge; no Vulkan paths are touched.

Ported

Upstream	Change
boykopovar#5081	NGG group info in SGPR2 for vertex-path shaders
boykopovar#5214	compute wave index in TG_SIZE[11:6]
boykopovar#5110	32_AR export takes alpha from channel 1
boykopovar#5117	s_barrier orders buffer, workgroup and image memory (acquire-release)
boykopovar#5426	GFX10 scratch offsets decode as one dword
boykopovar#4996	MIMG sampler tuples ending in VCC decode
boykopovar#5450	diagnostic names the attribute of an out-of-range interpolation reference
boykopovar#5471	high-half result (op_sel[3]) of v_interp_p2_f16 decodes
boykopovar#5162	VALU carry-out and v_div_scale masks clear inactive lanes
boykopovar#5159	v_cvt_f32_f64 applies the f32 output modifier and denormal flush
boykopovar#5160	per-invocation LDS updates without atomics (Function storage)
boykopovar#5492	ds_add_f32 flushes denormals and NaNs
boykopovar#5498	DS src2 lanes run one at a time in lane order (new IR opcodes)
boykopovar#907	color compare emulation: NOT_EQUAL unordered, TRUNC_COORD point sampling refused, array layer rounds half to even

Notes on how they landed:

• fix(shader): update per-invocation LDS without atomics boykopovar/AnyPS5#5160 conflicted in the shared-atomic update helper and in ShaderMemory.cpp; kept main's barrier condition for non-LDS shared atomics and dropped the unrelated SRT-read-patching test hunk that rode along.
• fix(agc): color compare emulation: review follow-up to #668 boykopovar/AnyPS5#907 was hand-applied to main's reworked emitter (three small edits); its upstream test rewrite targets a file that has diverged and is not ported.
• fix(shader): decode the high-half result of v_interp_p2_f16 boykopovar/AnyPS5#5471: the Vop3InterpDecode hunk is not ported (that test file is not on main); a small decode check replaces it.
• ShaderDiskCache::FormatVersion is bumped 34 -> 35 once (separate commit): MemoryInfo::flushDenormalResults is now encoded in the plan and fix(shader): run DS src2 lanes one at a time in lane order boykopovar/AnyPS5#5498 adds IR opcodes.

Skipped

• perf(shader): compute scalar f16 add, sub and mul with native f16 math boykopovar/AnyPS5#3557 (native f16 scalar add/sub/mul): perf only, and gated on a Float16 capability the Metal target does not report.
• perf(shader): guard packed f16 special-value repairs boykopovar/AnyPS5#5057 (guard packed f16 special-value repairs): perf only, same Float16 gate.

Tests (all run on Metal, ctest, GPU for the first group)

• New core/libs/prx/libSceAgcDriver/tests/metal_semantics/: a small shim presents the Vulkan test-device interface on Metal (guest buffer descriptors are mirrored into shared MTLBuffers), so upstream execution tests compile unchanged and run on the GPU: ValuMaskInactiveLanes (fix(shader): clear the bits of inactive lanes in VALU carry-out and v_div_scale masks boykopovar/AnyPS5#5162), VCvtF32F64ResultModes (fix(shader): apply the f32 output modifier and denormal flush to v_cvt_f32_f64 boykopovar/AnyPS5#5159), LdsAddF32 (fix(shader): flush ds_add_f32 denormals and NaNs as the hardware does boykopovar/AnyPS5#5492), LdsSrc2LaneOrder (fix(shader): run DS src2 lanes one at a time in lane order boykopovar/AnyPS5#5498), BarrierGlobalMemory (fix(shader): order buffer and image memory at s_barrier boykopovar/AnyPS5#5117; the barrier check is on the emitted SPIR-V semantics, the exchange itself is not a deterministic race detector).
• Own Metal tests: TgSizeWaveIndex (fix(shader): put the compute wave index in TG_SIZE[11:6] boykopovar/AnyPS5#5214, buffer-store variant of the upstream global-store test), FunctionLdsUpdates (fix(shader): update per-invocation LDS without atomics boykopovar/AnyPS5#5160, standalone copy of upstream's SPIR-V check), ColorCompareReject (fix(agc): color compare emulation: review follow-up to #668 boykopovar/AnyPS5#907 TRUNC_COORD), Vop3InterpHighHalf (fix(shader): decode the high-half result of v_interp_p2_f16 boykopovar/AnyPS5#5471).
• Existing unit tests now built for Metal: ImageSampleGather (fix(shader): accept MIMG sampler tuples ending in VCC boykopovar/AnyPS5#4996), FlatScratch (fix(shader): decode GFX10 scratch offsets as one dword boykopovar/AnyPS5#5426), InterpolationReferenceRange (fix(recompiler): name the attribute of an out-of-range interpolation reference boykopovar/AnyPS5#5450).
• anyps5_metal_guest_draw_replay gains two real draws: 32_AR export (fix(shader): take the 32_AR export alpha from channel 1 as GFX10 does boykopovar/AnyPS5#5110) and an NGG vertex shader reading the group info (fix(recompiler): give vertex-path shaders the NGG group info boykopovar/AnyPS5#5081, wave32 and wave64).
• Each of the tests above was shown to fail with its source fix reverted (mutation check), except: NOT_EQUAL unordered and RoundEven layer from fix(agc): color compare emulation: review follow-up to #668 boykopovar/AnyPS5#907 have no running test (they need a sampled-image execution path; the upstream test for it does not compile on main, see below).
• Not registered: LdsSrc2 and LdsSrc2Wide (use device-memory LDS, which the Metal bridge rejects). The upstream files for fix(shader): take the 32_AR export alpha from channel 1 as GFX10 does boykopovar/AnyPS5#5110, fix(recompiler): give vertex-path shaders the NGG group info boykopovar/AnyPS5#5081, fix(shader): update per-invocation LDS without atomics boykopovar/AnyPS5#5160, fix(shader): put the compute wave index in TG_SIZE[11:6] boykopovar/AnyPS5#5214, fix(shader): decode the high-half result of v_interp_p2_f16 boykopovar/AnyPS5#5471 execution tests are in tree as ported but are Vulkan-only.
• Results: ctest on the shader-bridge/guest-replay configuration, 50/50 passed (before the final rebase onto main, which touched no shader code).

Game smoke (one run per side, 280 s wall, eboot lock, on AC power)

Branch vs main, same title image: both reach guest entry, no fault, no uncaught exception, 0 skipped draws or dispatches, no conversion failures, identical open sequence for the first 2600 log lines (intro movies), same single unimplemented import. The branch run was cold-cache (format 35: 60 shaders converted, 0 failures); the main run used the warm cache. The branch run progressed further into asset loading within the wall time; run-to-run timing varies on the shared machine. Screens were not compared pixel by pixel, and neither run was confirmed to reach a 3D map in 280 s.

Performance

Not measured.

Pre-existing issues seen (not touched)

• tests/execution/EmulatedColorCompare.cpp on main does not compile (expectedVariant, RunSrgbCubeZero undefined).
• MetalReplay/UpstreamCorrectness/CMakeLists.txt is not included by any build file (already tracked as [BT-09] 20 tests (VideoOut, Flip, UpstreamCorrectness) are reachable only through an undocumented diagnostic wrapper #194).
• A 32 KiB LDS compute shader exceeds Metal's threadgroup limit by the bridge's few extra bytes under the debug layer (LdsSrc2LaneOrder runs without the layer for that reason).

🤖 Generated with Claude Code

Docs: the nine upstream docs hunks (5110, 5117, 5426, 5159, 5162, 5160, 5492, 5498, 907) are ported into docs/dev/TechnicalDebt.md (this tree has no docs/dev/debt/shader.md); the stale DS src2, LDS float atomics, f32 denormal, output modifier and color compare entries are updated, and the ShaderDiskCache history now has a 35 line.

[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 2026-10-14T01:37:25.850885Z	c131503	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: a41b8e05-ed21-4307-ad87-aef388af145a

• 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: c13150315d

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]

Skip lane serialization for function-backed LDS src2

When a vertex or fragment shader uses a DS src2 opcode, LDS is backed by per-invocation Function storage, so different lanes cannot alias each other's data. Nevertheless, both new EmitSharedSrc2U32 and EmitSharedSrc2U64 unconditionally enter this loop, electing every active lane and issuing a workgroup-scoped barrier after each update. This turns each otherwise independent src2 operation into 32/64 sequential iterations and reintroduces the workgroup barrier that the function-LDS path avoids elsewhere; bypass the lane-ordered helper and update directly when InFunctionStorage(state, mem.kind) is true.

Useful? React with 👍 / 👎.