Skip to content

v0.16.0

Choose a tag to compare

@tannevaled tannevaled released this 06 Oct 10:10
· 98 commits to main since this release
596c360

Single precision: real-FFT untangle and batched N-D columns (#51, BENCHMARKS.md Round 25).

  • Untangle/retangle kernels for float32 RFFT/IRFFT (AVX2 and NEON, Untangle32/Retangle32).
    • Zen 3: RFFT32 1.30–1.71× and IRFFT32 1.67–2.39× faster.
    • float32 real transforms now take 0.55–0.83× the float64 time (they were 0.79–1.51×).
  • Batched column passes for float32 N-D plans (AVX2 and NEON, StockhamBatchPass32), replacing the gather.
    • Zen 3: PlanN32 64² 2.26×, 256² 1.80×, 1024² 1.72×, 32³ 2.88×; RealPlan2_32 256² 2.05×.
    • float32 N-D now takes 0.46–0.77× the float64 time.
  • Guards: every new entry point bound-checks before any kernel runs, and StockhamBatchPass32 refuses inconsistent layouts.

Checked independently: float32 results are bit-identical to v0.15.0 on arm64 (63/63 hashed cases, special inputs included). On amd64 with AVX2, every value is identical; only NaN bits differ, on 19,339 of 146,328 values from infinite inputs, where x86 propagates the first operand's NaN. The full suites pass on Zen 3.

Not measured: NEON speed on Apple M4 (the machine stayed loaded) and Haswell (load 29).