tpt-simd

Portable, zero-cost SIMD abstractions for Rust DSP and codec development.

tpt-simd is a workspace of small, focused crates that wrap low-level SIMD intrinsics into the building blocks every codec and DSP project ends up rewriting: complex multiply, fixed-point arithmetic, FFT butterflies, horizontal reductions, transposes, saturating arithmetic and more.

Status: pre-release (v0.1.0). The crates below are implemented with tests and benchmarks; the API may still change before publishing. See spec.txt for the design and todo.md for the roadmap.

Goals

• 40 focused crates plus a tpt-simd umbrella (and the unpublished testutil), each solving one common SIMD problem
• no_std + alloc (std optional)
• Targets: x86_64 (SSE2/AVX2, AVX-512 where a kernel has a path), aarch64 (via LLVM lowering of the portable backend), and a portable fallback everywhere else
• Stable backend via tpt-simd-vector; the nightly feature is reserved for a future core::simd backend and is currently a no-op
• Direct integration path into tpt-kinetix and tpt-cadence

Crates

tpt-simd re-exports most of the crates below (not vector, dct, ntt, audio, bits, geom, transpose, pixel, distance or half); depend on individual crates to keep the dependency graph small.

Area	Crates
Foundation	core (traits, feature detection), vector (stable Simd/SimdMask backend), aligned, testutil (unpublished)
Arithmetic	mul, fixed, saturate, rounding, shift, math (exp/ln/sin/cos/tanh/erf)
Complex / transforms	complex, butterfly, window, convolve, interpolate
Data movement	permute, gather, scatter, select, blend, compare
Reductions	horizontal, reduce, dot
Linear algebra	matrix, blas, sparse
Random numbers	rng (xoshiro256++, Philox4x32-10)
Transforms and signal	dct (DCT-II/III, MDCT), ntt (modular NTT), audio (mix, biquad, resampler), window
Bits and integers	bits (popcount, bitset, delta, rank/select), hash (CRC-32, Adler-32, FNV-1a)
Text and bytes	text (hex, UTF-8, UTF-16 transcoding), search (byte and substring search), half (f16/bf16)
Geometry and images	geom (SoA distance, AABB, quaternions), transpose, pixel (RGB/HSV, histograms)
Similarity and sorting	distance (dot, L2, cosine, Hamming, top-k), sort (sort, select_nth, argsort)

All crate names are prefixed tpt-simd-.

Which crate do I need?

Task	Crate
FIR filter, convolution	convolve (f32, i16)
FFT butterflies, complex math	complex, butterfly
Dot product, matrix multiply	dot, blas (f32/f64 gemm, gemv, LAPACK-style LU/Cholesky)
Sum, mean, variance	reduce
Sparse CG / BiCGSTAB	sparse
Checksums	hash
Hex or UTF-8 validation	text
Substring or byte search	search
Sort or top-k	sort
Nearest-neighbour distances	distance
Random numbers	rng
Fixed-point arithmetic	fixed

hash, sort, search, text and pixel have no runtime dependencies by default; audio needs only libm. All of them build for no_std targets without features.

Getting full performance

Three opt-ins. The first two are off by default so no_std users are unaffected; runtime-dispatch is on by default in the tpt-simd umbrella (use default-features = false to drop it):

• std feature (e.g. tpt-simd = { version = "0.1", features = ["std"] }): float mul_add/round/floor/ceil/trunc/sqrt use the std intrinsics instead of per-lane libm calls. Results are bit-identical; the portable complex multiply is ~5x faster. Without it, those ops fall back to libm.
• -C target-cpu=native (or specific -C target-feature=+avx2,+fma): dispatch is compile-time (see ADR 0001), so the AVX2/FMA paths are only used when the target features are enabled.
• runtime-dispatch feature (implies std; affects tpt-simd-blas gemm and the hash, text and search kernels): detects AVX2+FMA at runtime so no -C target-cpu is needed (gemm f32 ~1.8x faster than the plain build; still below a fully native build). See ADR 0003.

Measured numbers and caveats: docs/benchmarks.md.

Performance against incumbents

Head-to-head benchmarks against established crates (crc32fast, simd-adler32, memchr, faster-hex, simdutf8, glam, gemm, faer, std sort). Ratio = tpt-simd time ÷ incumbent time, measured in the same run; below 1.0 is faster. Measured on an Intel Core i5-13500 (2026-10-12), two same-run passes, shown as a range. Summary:

Area	Operation	Size	Ratio	Notes
hash	CRC-32	4 KB	1.00	parity with crc32fast
search	find_byte	1 MB	0.92–0.93	faster than memchr
search	count_byte	1 MB	0.43	faster than memchr_iter
text	UTF-8 validate	4 KB / 256 KB	1.04–1.09	close to simdutf8
text	hex decode	4 KB	1.64–1.65	slower than faster-hex (lookup-table validator)
sort	radix sort_u32	1 M	0.40–0.41	faster than std
sort	generic sort_unstable	1 M	1.19–1.47	slower than std
matrix	mat3 inverse	3×3	1.74–1.77	slower than glam
gemm f32	512	1.03–1.05	parity with gemm / faer

Full tables, the method, and the caveats (noisy rows, and what is not yet compared): docs/incumbent-benchmarks.md.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.