[AdaWorldAPI]

Summary

Three new facade methods, each realized on every backend: AVX-512, AVX2, AVX-only, scalar, NEON, wasm and nightly.

• I32x16::dot4_u8_i8_add(a: U8x64, b: I8x64) adds a[4j..4j+4] · b[4j..4j+4] (u8 × signed i8) to each i32 lane j.
  • With AVX-512 VNNI compiled in, this is one vpdpbusd.
  • Everywhere else, even and odd bytes are widened into separate 16-bit lanes (zero-extended for a, sign-extended for b), then combined with a 16-bit multiply-add. Every product and partial sum is exact. Only the final addition to the accumulator wraps, as vpdpbusd does.
  • Not the vpmaddubsw emulation: that one saturates each pair sum at i16, and 2 * 255 * 127 = 64 770 clamps to 32 767.
• U8x64::low_nibbles returns x & 0x0F.
• U8x64::high_nibbles returns x >> 4. With GFNI compiled in it is a single gf2p8affine; otherwise a 16-bit shift and a mask.

Why: turbovec's VNNI "permute-dot" kernels write this dot product by hand twice, as vpdpbusd on x86 and SDOT on aarch64. They also keep a separate macro variant just for the GFNI high nibble. With these methods in the facade, those kernels can move onto ndarray::simd.

Also fixed: U16x32::to_f32x16_lo/hi on aarch64 and wasm, which came from #385.

• On those targets U16x32 comes from the scalar file while F32x16 is native.
• The method therefore returned a type that is not ndarray::simd::F32x16, and the aarch64 lib tests did not compile at all.
• CI only builds the NEON and wasm parity programs, never the lib tests on those targets, which is how it slipped through.

Tests

simd::dot4_nibble_facade_parity checks four things:

• results against a plain-array reference, over more than 10 000 cases with byte extremes and i32::MIN/i32::MAX accumulators;
• the saturation case;
• that each lane reads only its own four bytes;
• both nibble methods.

The NEON and wasm parity programs gain the same checks (codes 0x500..), so CI covers those backends too.

Build	Path exercised	Result
native (AVX-512 + VNNI + GFNI)	vpdpbusd, gf2p8affine	4/4
config-v4 (AVX-512, no VNNI/GFNI)	widen + vpmaddwd	4/4
config-v3 (AVX2)	256-bit	4/4
config-avx under qemu -cpu SandyBridge	SSE2 quarters	4/4
config-v2 under qemu -cpu Nehalem	scalar	4/4
aarch64 under qemu	NEON	4/4, plus u16x32_facade_parity
nightly-simd	core::simd	4/4
scripts/neon-parity.sh, scripts/wasm-parity.sh	NEON, wasm	pass

Gates:

• Full lib suite (native): 2657 passed.
• cargo clippy --config <tier> -- -D warnings: clean on native, v4, v3, v2 and aarch64. The AVX tier's 8 errors already exist on master; none is on an added line.
• cargo fmt --check: clean.
• The facade-stride guard is clean on the diff.

Disable runs, each red as expected:

• an AVX2 vpmaddubsw realization fails the saturation test and the reference test;
• a wrong GFNI matrix fails the nibble test;
• vpdpbusds in place of vpdpbusd fails the reference test on the i32::MAX accumulator.

Not done

• AVX-VNNI: the 256-bit vpdpbusd isn't used on the AVX2 tier.
• NEON USDOT: not used under i8mm.
• Signed i8 × i8 form: turbovec's xor 0x80 plus correction has no facade word yet.

None of these is measured.

---

Generated by Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

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

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 2768534d-e378-48e9-8f3c-42cc100912c8

• 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 Summary

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

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-29T14:30:11.551435Z	7c4dfe4	Draft marked ready

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.