#include <array.hfa>

// traditional "naiive" loops
forall( [M], [N], [P] )
void matmul( array(float, M, P) & src1,
             array(float, P, N) & src2, 
             array(float, M, N) & tgt ) {
    for (i; M) for (j; N) {
        tgt[i][j] = 0;
        for (k; P)
            tgt[i][j] += src1[i][k] * src2[k][j];
    }
}