** SIMD for Humans Easy, powerful, portable, absurdly fast numerical calculations. Includes static dispatch with inlining based on your platform and vector types, zero-allocation iteration, vectorized loading/storing, and support for uneven collections.
It looks something like this:
let lotsof3s = (&[-123.456f32; 128][..]).simditer() .simdmap(f32s(0.0), |v| { f32s(9.0) * v.abs().sqrt().rsqrt().ceil().sqrt() - f32s(4.0) - f32s(2.0) }) .scalar_collect();
Which is analogous to this scalar code:
let lotsof3s = (&[-123.456f32; 128][..]).iter()
.map(|v| {
9.0 * v.abs().sqrt().sqrt().recip().ceil().sqrt() - 4.0 - 2.0
})
.collect::
The vector size is entirely determined by the machine you're compiling for - it attempts to use the largest vector size supported by your machine, and works on any platform or architecture (see below for details).
Compare this to traditional explicit SIMD:
use std::mem::transmute; use stdsimd::{f32x4, f32x8};
let lotsof3s = &mut [-123.456f32; 128][..];
if cfg!(all(not(targetfeature = "avx"), targetfeature = "sse")) {
for ch in init.chunksmut(4) {
let v = f32x4::load(ch, 0);
let scalarabsmask = unsafe { transmute::
Even with all of that boilerplate, this still only supports x86-64 machines with SSE or AVX - and you have to look up each intrinsic to ensure it's usable for your compilation target. * Upcoming Features A rewrite of the iterator API is upcoming, as well as internal changes to better match the direction Rust is taking with explicit SIMD. * Compatibility Faster currently supports any architecture with floating point support, although hardware acceleration is only enabled on machines with x86's vector extensions. ** Performance Here are some extremely unscientific benchmarks which, at least, prove that this isn't any worse than scalar iterators. Even on ancient CPUs, a lot of performance can be extracted out of SIMD. Surprisingly, using SIMD iterators performs better than scalar iterators even on the SSE-less Pentium.
However, intentionally worsening your program's locality for SIMD (as seen in ~tests::benchdeterminmant2~ and ~tests::benchdeterminant3~) is not a worthwhile tradeoff unless you are doing a significant amount of work per vector.
$ RUSTFLAGS="-C target-cpu=ivybridge" cargo bench # host is ivybridge; target has AVX test tests::benchdeterminant2scalar ... bench: 391 ns/iter (+/- 2) test tests::benchdeterminant2simd ... bench: 375 ns/iter (+/- 1) test tests::benchdeterminant3scalar ... bench: 350 ns/iter (+/- 1) test tests::benchdeterminant3simd ... bench: 470 ns/iter (+/- 2) test tests::benchmapscalar ... bench: 6,952 ns/iter (+/- 26) test tests::benchmapsimd ... bench: 875 ns/iter (+/- 3) test tests::benchmapunevensimd ... bench: 880 ns/iter (+/- 3) test tests::benchnopscalar ... bench: 37 ns/iter (+/- 0) test tests::benchnopsimd ... bench: 34 ns/iter (+/- 0) test tests::benchreducescalar ... bench: 6,876 ns/iter (+/- 16) test tests::benchreducesimd ... bench: 835 ns/iter (+/- 2) test tests::benchreduceunevensimd ... bench: 836 ns/iter (+/- 2) test tests::benchzipnopscalar ... bench: 624 ns/iter (+/- 2) test tests::benchzipnopsimd ... bench: 361 ns/iter (+/- 1) test tests::benchzipscalar ... bench: 862 ns/iter (+/- 4) test tests::benchzipsimd ... bench: 771 ns/iter (+/- 2)
RUSTFLAGS="-C target-cpu=x86-64" cargo bench # host is ivybridge; target has SSE2 test tests::benchdeterminant2scalar ... bench: 426 ns/iter (+/- 2) test tests::benchdeterminant2simd ... bench: 376 ns/iter (+/- 2) test tests::benchdeterminant3scalar ... bench: 355 ns/iter (+/- 2) test tests::benchdeterminant3simd ... bench: 486 ns/iter (+/- 3) test tests::benchmapscalar ... bench: 7,157 ns/iter (+/- 59) test tests::benchmapsimd ... bench: 1,886 ns/iter (+/- 10) test tests::benchmapunevensimd ... bench: 1,889 ns/iter (+/- 11) test tests::benchnopscalar ... bench: 38 ns/iter (+/- 0) test tests::benchnopsimd ... bench: 34 ns/iter (+/- 0) test tests::benchreducescalar ... bench: 7,002 ns/iter (+/- 29) test tests::benchreducesimd ... bench: 1,865 ns/iter (+/- 10) test tests::benchreduceunevensimd ... bench: 1,937 ns/iter (+/- 7) test tests::benchzipnopscalar ... bench: 623 ns/iter (+/- 1) test tests::benchzipnopsimd ... bench: 333 ns/iter (+/- 3) test tests::benchzipscalar ... bench: 971 ns/iter (+/- 5) test tests::benchzipsimd ... bench: 525 ns/iter (+/- 3)
$ RUSTFLAGS="-C target-cpu=pentium" cargo bench # host is ivybridge; this only runs the polyfills! test tests::benchdeterminant2scalar ... bench: 427 ns/iter (+/- 2) test tests::benchdeterminant2simd ... bench: 402 ns/iter (+/- 1) test tests::benchdeterminant3scalar ... bench: 354 ns/iter (+/- 1) test tests::benchdeterminant3simd ... bench: 593 ns/iter (+/- 1) test tests::benchmapscalar ... bench: 7,195 ns/iter (+/- 28) test tests::benchmapsimd ... bench: 6,271 ns/iter (+/- 22) test tests::benchmapunevensimd ... bench: 6,288 ns/iter (+/- 22) test tests::benchnopscalar ... bench: 38 ns/iter (+/- 0) test tests::benchnopsimd ... bench: 69 ns/iter (+/- 0) test tests::benchreducescalar ... bench: 7,004 ns/iter (+/- 17) test tests::benchreducesimd ... bench: 6,063 ns/iter (+/- 17) test tests::benchreduceunevensimd ... bench: 6,107 ns/iter (+/- 11) test tests::benchzipnopscalar ... bench: 623 ns/iter (+/- 2) test tests::benchzipnopsimd ... bench: 289 ns/iter (+/- 1) test tests::benchzipscalar ... bench: 972 ns/iter (+/- 3) test tests::benchzipsimd ... bench: 621 ns/iter (+/- 3)