/* Transpose matrix B to both:
*
* - increase cache hits
* - simd GCC vector extensions which is made possible.
* by the transposition, to increase likelyhood of SIMDs.
*
* Note that GCC 6 O=3 is smart enough to use SIMD
* even for the naive CPU method. However this was still way faster.
* */
void mat_mul_cpu_trans_vec(const F *A, const F *B, F *C, size_t n, Cache *cache) {
F tmpf;
size_t i, j, k, k_max, ai, bi;
Vec tmp, a, b;
UNUSED(cache);
mat_trans((F*)B, n);
k_max = (n / VECTOR_NELEMS) * VECTOR_NELEMS;
for (i = 0; i < n; ++i) {
for (j = 0; j < n; ++j) {
vec_zero(&tmp, VECTOR_NELEMS);
for (k = 0; k < k_max; k += VECTOR_NELEMS) {
ai = i * n + k;
bi = j * n + k;
vec_load(&a, VECTOR_NELEMS, A, ai);
vec_load(&b, VECTOR_NELEMS, B, bi);
tmp += a * b;
}
tmpf = 0.0;
for (; k < n; ++k) {
tmpf += A[i*n+k] * B[j*n+k];
}
C[i*n+j] = vec_sum(tmp, VECTOR_NELEMS) + tmpf;
}
}
mat_trans((F*)B, n);
}
main ()
{
i = vec_load (a1);
j = vec_load (a2);
k = vec_add_int4 (i, j);
vec_store (a3, k);
compare_int4 (a3, addi);
k = vec_average_int4 (i, j);
vec_store (a3, k);
compare_int4 (a3, avgi);
f = (float4) vec_load (f1);
g = (float4) vec_load (f2);
h = vec_add_float4 (f, g);
vec_store (f3, h);
compare_float4 (f3, addf);
exit (0);
}