SIMD pixel.cpp

#include <stdint.h>
#include <math.h>
#include <time.h>
#include <iostream>
#include <immintrin.h>

using namespace std;

#define FENC_STRIDE 64
typedef unsigned char uint8_t;
typedef uint8_t pixel;

// /* Original */
// template <int lx, int ly>
// void sad_x4(const pixel *pix1, const pixel *pix2, const pixel *pix3, const pixel *pix4, const pixel *pix5, intptr_t frefstride, int32_t *res)
// {
//     res[0] = 0;
//     res[1] = 0;
//     res[2] = 0;
//     res[3] = 0;
//     for (int y = 0; y < ly; y++)
//     {
//         for (int x = 0; x < lx; x++)
//         {
//             res[0] += abs(pix1[x] - pix2[x]);
//             res[1] += abs(pix1[x] - pix3[x]);
//             res[2] += abs(pix1[x] - pix4[x]);
//             res[3] += abs(pix1[x] - pix5[x]);
//         }

//         pix1 += FENC_STRIDE;
//         pix2 += frefstride;
//         pix3 += frefstride;
//         pix4 += frefstride;
//         pix5 += frefstride;
//     }
// }

/* SIMD */
template <int lx, int ly>
void sad_x4(const pixel *pix1, const pixel *pix2, const pixel *pix3, const pixel *pix4, const pixel *pix5, intptr_t frefstride, int32_t *res)
{
    // __mm(256)_(sad_x4)_(epu8)
    res[0] = 0;
    res[1] = 0;
    res[2] = 0;
    res[3] = 0;
    for (int y = 0; y < ly; y++)
    {
        __m256i_u _pix1 = _mm256_loadu_si256((__m256i_u *)pix1); // TODO: Doesn't work
        int *_test = (int *)&_pix1;
        for (int x = 0; x < lx; x++)
        {
            cout << "\n"
                 << _test[x];
        }
        for (int x = 0; x < lx; x++)
        {
            res[0] += abs(pix1[x] - pix2[x]);
        }
        for (int x = 0; x < lx; x++)
        {
            res[1] += abs(pix1[x] - pix3[x]);
        }
        for (int x = 0; x < lx; x++)
        {
            res[2] += abs(pix1[x] - pix4[x]);
        }
        for (int x = 0; x < lx; x++)
        {
            res[3] += abs(pix1[x] - pix5[x]);
        }

        pix1 += FENC_STRIDE;
        pix2 += frefstride;
        pix3 += frefstride;
        pix4 += frefstride;
        pix5 += frefstride;
    }
}

int main()
{
    int32_t res[4];
    pixel a[] = {1, 2, 3, 4, 5, 6, 7, 8};

    clock_t start, end;
    start = clock();
    sad_x4<8, 8>(a, a, a, a, a, 0, res);
    end = clock();
    double time_taken = double(end - start) / double(CLOCKS_PER_SEC);
    cout << "\nTime taken: " << time_taken << "s\n";
}