C > Transposition de matrice Intrinsics

__m128 X1,X2,X3,X4;
__m128 X1out,X2out,X3out,X4out;


float X1_sse[4] __attribute__((aligned(16))) = {1, 2, 3, 4};
float X2_sse[4] __attribute__((aligned(16))) = {11, 12, 13, 14};
float X3_sse[4] __attribute__((aligned(16))) = {21, 22, 23, 24};
float X4_sse[4] __attribute__((aligned(16))) = {31, 32, 33, 34};

float X1out_sse[4] __attribute__((aligned(16)));
float X2out_sse[4] __attribute__((aligned(16)));
float X3out_sse[4] __attribute__((aligned(16)));
float X4out_sse[4] __attribute__((aligned(16)));


int main(int argc, int** argv)
{
        double t1, t2, dt, dtmoy=0;
        int N;
        int k, l, i, j, m;

        int i;
        t1 = dtime();
        X1 = _mm_load_ps(X1_sse); X2 = _mm_load_ps(X2_sse); X3 = _mm_load_ps(X3_sse); X4 = _mm_load_ps(X4_sse);
       
       
        X1out = _mm_unpacklo_ps(X1,X3);
        X2out = _mm_unpackhi_ps(X1,X3);
        X3out = _mm_unpacklo_ps(X2,X4);
        X4out = _mm_unpackhi_ps(X2,X4);

        X1 = _mm_unpacklo_ps(X1out,X3out);
        X2 = _mm_unpackhi_ps(X1out,X3out);
        X3 = _mm_unpacklo_ps(X2out,X4out);
        X4 = _mm_unpackhi_ps(X2out,X4out);

        _mm_store_ps(X1out_sse, X1);
        _mm_store_ps(X2out_sse, X2);
        _mm_store_ps(X3out_sse, X3);
        _mm_store_ps(X4out_sse, X4);

        t2 = dtime();
                dt = t2 - t1;
}