A collection of AVX-512 accelerated mathematical functions for Rust.
This is very similar to a Rust implementation of the Intel MKL VM Mathematical Functions
There are four public interfaces for each function func:
func(in: &[f64], out: &mut [f64]) -> ()func_par(in: &[f64], out: &mut [f64]) -> ()unsafe func_intr(in: &__m512d, out: &mut __m512d) -> ()unsafe _m512_func_pd(in: __m512d) -> __m512d
The first one is used to compute an array of inputs as quickly as possible for the given function. The benchmarks section features plenty of examples of how this is done. Say you want to calculate x inputs you want to calc on, prealloate a y return array, and call lit_math::exp(&x, &mut y). The second will automatically use a Rayon parallelization scheme to go faster (or slower), depending on your inputs and CPU.
The other two can be used as building blocks to make more complicated functions.
The unroll_fn macro can be used to turn any function with the signature unsafe func_intr(in: &__m512d, out: &mut __m512d) -> () into optimized unrolled and parallel functions like those mentioned in 1 and 2 above. This macro is the one used to generate 1 and 2, so examples for how to use it are in the source for exp.
TODO: expand the interface to handle an arbitrarily large number of input values.
On my Ryzen 7950x, the speedup is roughly 8x with no loss of precision for ln, log2, exp and pow2 when calculating on 2048 sized arrays.