[tannevaled]

NEON kernels for the correctly rounded float32 Exp and Log on arm64, which until now ran the Go transcriptions.

exp32CRNEON ports Arm's cr_expf, an AdvSIMD routine to begin with, on 8 floats per iteration.

• FCVTL/FCVTL2 widen. VFMLA/VFMLS do expCR's fused operations in its order.
• The table is read through two general registers per pair, as cr_expf does, because NEON has no gather. FCVTN/FCVTN2 round once.
• A block with a lane outside [−512, 512] or a NaN (FACGE) is handed back to exp32CR before anything is stored.

log32CRNEON is the twin of log32CRAVX2, running logCR's fast path (CORE-MATH's logf).

• The integer part works in 32-bit lanes: m, j, e, x == 1, and the positive-normal test.
• Per pair: tz (UXTL), e (SXTL, SCVTF), tr[j] and tl[j] through general registers, then logCR's unfused VFMUL/VFADD in its order. Both bounds are rounded once.
• A block with a lane that is not a positive normal, or whose bounds round apart, is handed back to logCR.

Instructions Go's assembler does not name (FCVTL/FCVTN and their 2 forms, FACGE, UXTL, SXTL, SCVTF, CMGT, BIC) are emitted as WORDs. go tool objdump decodes each one as intended, with operands in the right roles.

Proof on darwin/arm64 over all 2³² inputs: TestExp32CorrectlyRounded and TestLog32CorrectlyRounded now run the NEON kernels. Both find 0 incorrectly rounded, and logCR equals the independent oracle everywhere.

The tests catch planted defects:

• Widening the wrong half (FCVTL for FCVTL2) fails TestExp32NEONMatchesGo and the exhaustive sample (1602 wrong of 65536).
• Dropping the positive-normal BIC fails TestLog32NEONMatchesGo (log(0) = −88.03) and the sample (33016 wrong).

Both kernels are bit-identical to the Go path on 300 000 inputs, every length 0–70 and in place, and both are fenced. Their constant blocks' layouts are tested.

⚠ Not yet measured. The arm64 hosts are loaded: this Mac is at load 78–103 and cfarm424 at 48 on 64 cores. A gated abbench (v0.35.0 vs this branch, load ≤ 8) is queued. This PR will not be merged before its numbers are in and in docs/perf.md.

Also: s390x (IBM z17) is now proven for float32 Exp and Log over all 2³² inputs (0 incorrectly rounded). That completes the set of 64-bit architectures. docs/perf.md will record it with the numbers.

🤖 Generated with Claude Code

[tannevaled]

On hold, not to be merged here. ndarray's 1-D kernels are moving into go-simd/floats (go-scikit performance campaign, kernel K1), so there is one implementation. These NEON kernels go there with the AVX2 ones already on main (exp32CRAVX2, log32CRAVX2, logAVX2).

The port keeps:

• the bit-identity oracles (expCR, logCR);
• the 2³² exhaustive tests;
• the objdump checks of the WORD-encoded instructions;
• the hand-back of special lanes.

This PR stays open until the go-simd/floats PR is merged, then it will be closed with a pointer to it.

🤖 Generated with Claude Code

[tannevaled]

Superseded by go-simd/floats#29 (merged 2026-11-07). These NEON kernels now live in go-simd/floats as Float32Exp and Float32Log, with the AVX2 kernels from #94 and #96. ndarray will import them from there, so there is one implementation.

🤖 Generated with Claude Code