Fast memory transpose with SSE, AVX, and OpenMP

Question

I need a fast memory transpose algorithm for my Gaussian convolution function in C/C++. What I do now is

convolute_1D
transpose
convolute_1D
transpose
         


        

Answer 1

Consider this 4x4 transpose.

struct MATRIX {
    union {
        float  f[4][4];
        __m128 m[4];
        __m256 n[2];
    };
};
MATRIX myTranspose(MATRIX in) {

    // This takes 15 assembler instructions (compile not inline), 
    // and is faster than XMTranspose
    // Comes in like this  1  2  3  4  5  6  7  8
    //                     9 10 11 12 13 14 15 16
    //
    // Want the result     1  5  9 13  2  6 10 14
    //                     3  7 11 15  4  8 12 16

    __m256 t0, t1, t2, t3, t4, t5, n0, n1;
    MATRIX result;

    n0 = in.n[0];                                               // n0 =  1,  2,  3,  4,  5,  6,  7,  8
    n1 = in.n[1];                                               // n1 =  9, 10, 11, 12, 13, 14, 15, 16
    t0 = _mm256_unpacklo_ps(n0, n1);                            // t0 =  1,  9,  2, 10,  5, 13,  6, 14
    t1 = _mm256_unpackhi_ps(n0, n1);                            // t1 =  3, 11,  4, 12,  7, 15,  8, 16

    t2 = _mm256_permute2f128_ps(t0, t1, 0x20);                  // t2 =  1,  9,  2, 10,  3, 11,  4, 12 
    t3 = _mm256_permute2f128_ps(t0, t1, 0x31);                  // t3 =  5, 13,  6, 14,  7, 15,  8, 16

    t4 = _mm256_unpacklo_ps(t2, t3);                            // t2 =  1,  5,  9, 13,  3,  7, 11, 15
    t5 = _mm256_unpackhi_ps(t2, t3);                            // t3 =  2,  6, 10, 14,  4,  8, 12, 16

    result.n[0] = _mm256_permute2f128_ps(t4, t5, 0x20);         // t6 =  1,  5,  9, 13,  2,  6, 10, 14
    result.n[1] = _mm256_permute2f128_ps(t4, t5, 0x31);         // t7 =  3,  7, 11, 15,  4,  8, 12, 16
    return result;
}