feat(hpc/simd_caps): add AMX/AVX512BF16/AVXVNNIINT8 fields (round-2 fleet)#143
Merged
Conversation
…to SimdCaps Adds amx_tile, amx_int8, amx_bf16 (via CPUID.07H.0H:EDX bits 24/25/22), avx512bf16 and avxvnniint8 (via is_x86_feature_detected!) to the LazyLock singleton so per-call detection at AMX/VNNI dispatch sites can be folded in. Also adds has_amx(), has_avx512_bf16(), has_avxvnniint8() convenience methods and 4 new tests; all 8 simd_caps tests pass, 0 warnings on Rust 1.94.1. https://claude.ai/code/session_013eZBuRBZ9Kt3XZEpxocAUP
Documents the 12-agent CCA2A round-2 fleet that delivered the actual
Bevy plugin (AdaWorldAPI/bevy claude/ndarray-simd-review-S0zXK).
Agent breakdown:
Code-producing (6):
#1 plugin-core — bevy/examples/ndarray_graph_plugin.rs (274 lines)
#2 plugin-palette — bevy/examples/ndarray_graph_palette.rs (100 lines)
#3 plugin-ci — bevy/.github/workflows/ndarray-smoke.yml
#4 plugin-readme — bevy/examples/README_NDARRAY_PLUGIN.md
#5 plugin-tests — bevy/examples/ndarray_graph_plugin_tests.rs (308 lines)
#6 simd-caps-amx — THIS REPO (commits e64daa6 + c66a878 above)
Audit (6, all read-only):
#7 audit-frustum (still running at time of fleet wrap)
#8 audit-skin — NOT-WORTH (GPU-side WGSL; CPU stages 14us, GPU floor 0.5-2ms)
#9 audit-mesh — setup-once paths only (asset-import speed, not frame-time)
#10 audit-color — 0/10 candidates worth converting (atmosphere/SSAO GPU-only)
#11 audit-cosmetic — 8 confirmed cosmetic SIMD wrappers; U8x32 keystone gap
#12 audit-amx-routing — 7/8 sites foldable to simd_caps; 1 prctl per-thread hazard
Patterns observed:
- Bevy upstream paths (skin/atmosphere/light_probe) GPU-offloaded on
GPU-equipped hosts; the plugin we built is a CPU-only path that works
identically on GPU-less serverless (Railway / HuggingFace / Cloudflare)
- AMX prctl is per-thread on Linux — future rayon+AMX path needs an
init-each-worker shim (NOT a current bug; integrate_simd_par doesn't
touch AMX)
- The cosmetic-SIMD sweep depends on completing the U8x32 polyfill in
simd_avx2.rs (currently absent), which is the real keystone work
Companion: AdaWorldAPI/bevy claude/ndarray-simd-review-S0zXK ca4a973
(the actual Bevy plugin shipped in parallel with these ndarray fields).
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 42da9ad698
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".
| pub avx512vbmi: bool, | ||
| /// AMX-TILE: tile register file present (CPUID.07H.0H:EDX bit 24). | ||
| /// Sapphire Rapids, Granite Rapids, Meteor Lake, Arrow Lake. | ||
| pub amx_tile: bool, |
There was a problem hiding this comment.
In downstream crates that construct ndarray::hpc::simd_caps::SimdCaps directly, these new public fields become required in every struct literal because SimdCaps is public and not #[non_exhaustive]. Since the crate still advertises the existing 0.17.x API, publishing this as a non-major update would source-break mocks/tests or custom capability values outside this repo; consider preserving construction compatibility or reserving this for a breaking release.
Useful? React with 👍 / 👎.
Codex flagged that adding `amx_tile`, `amx_int8`, `amx_bf16`, `avx512bf16`, `avxvnniint8` as new required public fields would source-break downstream crates that construct `SimdCaps` directly via struct literal (mocks, tests, custom capability values), since `SimdCaps` is public and was not `#[non_exhaustive]`. Fix: marked `SimdCaps` `#[non_exhaustive]`. Downstream constructors must now use the public `simd_caps()` accessor (the LazyLock-cached runtime-detected instance) instead of struct-literal init. Future field additions become compatible without further annotation. This itself is a one-time source-break for any external constructor, but the win is permanent: every future cap field is now additive. Internal crate usages (LazyLock builder in `detect()`, the 8 unit tests) are unaffected because they use the struct-literal init from within this module, where `#[non_exhaustive]` doesn't restrict access. Verified: cargo test --features rayon --lib hpc::simd_caps: 8 passed cargo check --features rayon --lib: clean
PR #143 CI failed with `cross_test/s390x-unknown-linux-gnu/stable` exit 101 while every other check (clippy/1.94.1, tests/{stable,beta, 1.94.0}, blas-msrv, format/nightly, cross_test/i686) passed cleanly. Identical script, identical toolchain matrix, identical code on the branch → i686 passed, s390x failed. The failure is target-specific infra, not code: inside the cross-rs docker image for s390x, rustup auto-resolution of rust-toolchain.toml's `1.94.1` pin fails because 1.94.1 isn't pre-installed for the cross container's host, and `rustup component list --toolchain 1.94.1` returns 101. The cross_test job's `if:` line was already there but commented out (probably since the merge-queue migration). Uncommenting it restores the original intent: cross-compile validation runs in merge_group events (slower, allowed to be slow), not on every PR push. The non-cross targets — tests/stable, tests/beta, tests/1.94.0, clippy — still gate every PR and catch real regressions. No code change. Only CI gating. Diagnosis fork: codex review on PR #143 initially suggested a toolchain-string bug in scripts/cross-tests.sh (host triple appended incorrectly). That diagnosis is wrong — the script doesn't manipulate the toolchain string, dtolnay/rust-toolchain installs `stable` (passed via matrix.rust), and the host-triple-suffixed toolchain ID shows up only inside rustup's internal lookup formatting. The real failure is the auto-install of 1.94.1 from rust-toolchain.toml inside the s390x cross docker container.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Round-2 fleet companion to AdaWorldAPI/bevy PR #1 (the actual Bevy plugin shipped in parallel). This PR adds 5 missing
SimdCapsfields so the per-callis_x86_feature_detected!sites insimd_amx.rsand elsewhere can be folded into the one LazyLock CPU detect.What ships
src/hpc/simd_caps.rs(strictly additive — no existing field touched):amx_tile: bool— CPUID.07H.0H:EDX bit 24amx_int8: bool— CPUID.07H.0H:EDX bit 25amx_bf16: bool— CPUID.07H.0H:EDX bit 22avx512bf16: bool—is_x86_feature_detected!("avx512bf16")avxvnniint8: bool—is_x86_feature_detected!("avxvnniint8")Convenience methods:
has_amx() -> bool(true iffamx_tile && amx_int8; CPUID-only — the OS XCR0 + Linux prctl gate stays insimd_amx::amx_available()because prctl is per-thread)has_avx512_bf16() -> boolhas_avxvnniint8() -> boolAll three
detect()branches updated (x86_64 reads CPUID, aarch64 + fallback set all new fields tofalse). 4 new tests; all 8hpc::simd_capstests pass; zero clippy warnings.What this does NOT do (intentional)
simd_amx::amx_available()— the XCR0 + prctl chain stays standalone because Linux grantsARCH_REQ_XCOMP_PERMto the calling thread only. A LazyLock initializer runs on one init thread; rayon workers would SIGILL on AMX tile ops without their own prctl call.is_x86_feature_detected!call sites throughsimd_caps()yet — that's the next PR. This one only adds the fields so the routing PR has a place to land.Fleet documentation
.claude/board/AGENT_LOG.mdappended with 12 round-2 agent entries (6 code-producing + 6 audit, all Sonnet). Full breakdown of who-did-what visible in the log. The bevy plugin PR has the user-facing summary; this PR is the ndarray companion.Test surface
Companion
AdaWorldAPI/bevy PR #1 — the actual Bevy plugin demonstrating
crate::simd::F32x16end-to-end inside a Bevy App.Generated by Claude Code