v0.4.0

What's Changed

• feat: add DIT FFT algorithm & bit reversal control by @smu160 in #37
• Fix build by @Shnatsel in #39
• Require gfni target feature for the AVX-512 codepath by @Shnatsel in #41
• Use a fast, non-secure PRNG in benchmarking harness by @Shnatsel in #43
• Drop lto=true from Cargo.toml by @Shnatsel in #44
• don't multiversion on ARM by @Shnatsel in #45
• perf: add initial cache-blocked dit fft impl by @smu160 in #48
• Use .as_chunks_mut() instead of .chunks_exact_mut() for better performance by @Shnatsel in #51
• Benchmarks: Lower criterion sample size to 20 from its default of 100 by @Shnatsel in #52
• Report throughput in bytes in addition to Melem/s by @Shnatsel in #53
• multithreading via Rayon by @Shnatsel in #50
• Move DIF-only function to DIF file by @Shnatsel in #54
• No explicit inlining by @Shnatsel in #56
• Use mul_add with a float inversion instead of mul_neg_add by @Shnatsel in #57
• Revert "Use mul_add with a float inversion instead of mul_neg_add" by @Shnatsel in #59
• Port to fearless_simd by @Shnatsel in #58
• fearless_simd + BRAVO by @Shnatsel in #64
• Turn x << 1 into x * 2 and add comments on all the other uses of << by @Shnatsel in #66
• Refactor dit planner by @Shnatsel in #67
• Determine SIMD level in the planner by @Shnatsel in #69
• Update examples/bench.rs by @Shnatsel in #71
• Less dynamic dispatch by @Shnatsel in #72
• Switch to released fearless_simd 0.4.0 by @Shnatsel in #77
• targeted conversion to a for loop for just enough inlining by @Shnatsel in #79
• chore: Remove DIF and COBRA implementations by @smu160 in #82
• Use RustFFT more efficiently in benchmarks for a fair comparison by @Shnatsel in #81
• Simplify and optimize BRAVO by @Shnatsel in #83
• Improve standalone benchmarks by @Shnatsel in #84
• Extend large-size benchmarks to 32 bits by @Shnatsel in #85
• Double the chunk size for f64 by @Shnatsel in #86
• README updates by @Shnatsel in #70
• Options: default to 16k parallelism threshold by @Shnatsel in #87
• Make planner fields pub(crate) by @Shnatsel in #88
• Move 'How is it so fast?' down in the README by @Shnatsel in #91
• Remove wide and multiversion dependencies by @Shnatsel in #90
• add a benchmark for planner by @Shnatsel in #76
• Initial work on interleaving performance by @Shnatsel in #94
• Make BRAVO operating on chunks more explicit by @Shnatsel in #96
• perf: add a FFT 32 codelet to fuse first 5 stages by @smu160 in #101
• perf: add tiling to bravo to turn it into co-bravo by @smu160 in #106
• Speed up random signal generation by removing per-element sqrt and sin_cos by @Shnatsel in #109
• Use std::hint::black_box for more accurate benchmarks by @Shnatsel in #110
• Use std::hint::black_box for more accurate small-size benchmarks by @Shnatsel in #111
• perf: remove extra pass used for inverse transforms by @smu160 in #114
• PoC: Make the codelets operate entirely in registers by @Shnatsel in #113
• fix: remove old C FFTW benches, improve methodology, and plots by @smu160 in #117
• feat: add real valued FFT/IFFT by @smu160 in #105
• Use interleave() instead of zip_low()/zip_high() for better perf on AVX2 by @Shnatsel in #116
• release/v0.4.0 by @smu160 in #119

Full Changelog: v0.3.0...v0.4.0