libsimd.h

C++ library of fast, approximate math function for Intel AVX.

How to use?

Include the relevant header in your file, e.g.:

#include "libsimdrcp.h"

You need to enable the relevant compiler flags, see below for GCC and Intel compilers.

To compile with gcc, use e.g.

g++ my_source.cpp -o my_exec -O3 -mfma -mavx2 -ffast-math

To compile with icpc, use e.g.

icpc my_source.cpp -o my_exec -O3 -march=core-avx2 -fma

Currently supported functions:

Calculates 1/x, using one Newton-Raphson iterations on the start guess from the approximate AVX intrinsic _mm256_rcp_ps. Accurate to 32-bit precision, but faster than _m256_div_ps(ONE, q).

#include "libsimdrcp.h"
__m256 _mm256_rcp1s_ps(const __m256 &q)

Calculates 1/sqrt(x), using one Newton-Raphson iterations on the start guess from the approximate AVX intrinsic _mm256_rsqrt_ps. Accurate to 32-bit precision, but faster than _m256_div_ps(ONE, _mm256_sqrt_ps(q)).

#include "libsimdrsqrt.h"
__m256 _mm256_rsqrt1s_ps(const __m256 &q)

Calculates exp(x) using a bit shifting technique. Extremely fast, but has an error of about 10%.

#include "libsimdexp.h"
__m256 _mm256_expfaster_ps(const __m256 &q)

Only valid for -126 < x < 0.0. Calculates exp(x) via bit shifting techniques and the Newton-Raphson approximation. Not super fast, but has an error of about 0.0001% or 0.001%, depending on which approximation is used for the reciprocal 1/x. If high accuracy is desired, use _mm256_rcp1s_ps() instead of _mm256_rcp_ps().

#include "libsimdexp.h"
__m256 _mm256_expfastsmallneg_ps(const __m256 &q)

Prints a __m256 vector to std. out.

#include "libsimdtools.h"
void _mm256_print_ps(const __m256 &q)

License:

The license for this software has very little restriction. If you do use libsimd for academic publications, please cite this GitHub repository.

Anders S. Christensen (2015) "libsimd - a C++ library of fast, approximate math functions for Intel AVX." https://github.com/andersx/libsimd

There will be a preprint on arXiv soon, if I have enough time to wrap things up.