AVX Programming, how to do within C mex file

Hi everybody, fma operations or more generally using avx programming may be very usefull on some circonstancies, better accuracy within dot product for example, faster execution of unrolled loop as blas routines do,....

Is it possible to use this approach within C mex file? Is it nessessary to add specific compilation options at mex command?

Any help or hints about this subject will be great. I have not the fixed-point designer toolbox which probably answers to my question. However I don't use any fixed-point arithmetic.

Best regards

Alain

Answer 1

[Walter Roberson]

If you are using Visual Studio then see https://devblogs.microsoft.com/cppblog/avx-512-auto-vectorization-in-msvc/

[Alain Barraud]

Following your answer AVX.c contains a fma basic test

__m256d veca = _mm256_setr_pd(6.0, 6.0, 6.0, 6.0);

__m256d vecb = _mm256_setr_pd(2.0, 2.0, 2.0, 2.0);

__m256d vecc = _mm256_setr_pd(7.0, 7.0, 7.0, 7.0);

__m256d resultat = _mm256_fmaddsub_pd(veca, vecb, vecc);

double* res = (double*)&resultat;

wihin an appropriate mex gateway with some includes as follows

#include <immintrin.h>

#include <stdio.h>

#include "mex.h"

#include "matrix.h"

mex COMPFLAGS='$COMPFLAGS /arch:AVX2' AVX.c

Building with 'Microsoft Visual C++ 2019 (C)'.

MEX completed successfully.

I obtain the attended results