有效地使用英特尔编译器SVML`\uu m128\u mm\u sincos\u ps()`

以下是关于 godbolt

#include <stdio.h>
#include <stdlib.h>
#include <mathimf.h>
#include <immintrin.h>

void test (const float * restrict mI, 
           float * restrict mCOmega,
           float * restrict mSOmega,
           float paramOmega,
           int numRows,
           int numCols)
{   
    __m128 x, sin_x, cos_x;
    int ii;
    for (ii = 0; ii < numRows * numCols; ii = ii + 4)
    {   x = _mm_loadu_ps((float*)&mI[ii]);
        x = _mm_mul_ps(x, _mm_set1_ps(paramOmega));
        sin_x = _mm_sincos_ps(&cos_x, x);
        _mm_storeu_ps((float*)&mCOmega[ii], cos_x);
        _mm_storeu_ps((float*)&mSOmega[ii], sin_x);
    }   
}