Why using AVX on array created in heap result in an error?

Question

When I use AVX to perform GEMM, I encountered such problem: when using 2d array created on stack, there won't be any error, but if I use array created on heap, then segmentation fault will happen.

This code works well.

    double A[2048][2048];
    double B[2048][2048];
    double C[2048][2048];
    __m256d vec_res = _mm256_setzero_pd();
    __m256d vec_1 = _mm256_setzero_pd();
    __m256d vec_2 = _mm256_setzero_pd();

    for (int i = 0; i < M; i++){
        for (int j = 0; j < M; j++){
            vec_1 = _mm256_set1_pd(A[i][j]);
            for (int k = 0; k < M; k += 4){
                vec_2 = _mm256_load_pd(&B[j][k]);
                vec_res = _mm256_load_pd(&C[i][k]);
                vec_res = _mm256_add_pd(vec_res ,_mm256_mul_pd(vec_1, vec_2));
                _mm256_store_pd(&C[i][k], vec_res);

            }
        }
    }

But when I create array on heap, then segmentation fault will happen:

void Malloc2D(int height, int width, double ** & Mat)
{
    Mat = new double * [height];
    for(int i = 0; i < height; i++)
    {
        Mat[i] = new double [width];
    }
}

int main()
{
    int M, N, K;
    M = N = K = 1024;
    double ** A = nullptr;
    double ** B = nullptr;
    double ** C = nullptr;
    Malloc2D(M, N, A);
    Malloc2D(N, K, B);
    Malloc2D(M, K, C);
    __m256d vec_res = _mm256_setzero_pd();
    __m256d vec_1 = _mm256_setzero_pd();
    __m256d vec_2 = _mm256_setzero_pd();

    for (int i = 0; i < M; i++){
        for (int j = 0; j < M; j++){
            vec_1 = _mm256_set1_pd(A[i][j]);
            for (int k = 0; k < M; k += 4){
                vec_2 = _mm256_load_pd(&B[j][k]);
                vec_res = _mm256_load_pd(&C[i][k]);
                vec_res = _mm256_add_pd(vec_res ,_mm256_mul_pd(vec_1, vec_2));
                _mm256_store_pd(&C[i][k], vec_res);

            }
        }
    }
    return 0;
}

The error code is

Process finished with exit code 139 (interrupted by signal 11: SIGSEGV)

And IDE detected that error happened with this assembly code

vmovapd %ymm0,(%rdx)

Should I use 1d array instead of 2d?

doenst matter (because values are equal), but you allocate MxN and then loop MxM — 463035818_is_not_an_ai, Sep 08 '21 at 09:21
Yes, you should use a 1D array, or an actual 2D array (array of arrays, not an array of pointers), for performance reasons. — Peter Cordes, Sep 08 '21 at 09:24
But for correctness, if you're going to use alignment-required stores, you need to ask for aligned memory (which requires C++17 if you want that with with new/delete, I think). — Peter Cordes, Sep 08 '21 at 09:25

Why using AVX on array created in heap result in an error?

0 Answers0