gema.cve

#define UNROLL_FACTOR 6

#include <ve_kernel.h>

__ve_kernel__
void
gema_op_d(
    const double * A,
    const double * B,
    double * C)
{
    const int ve_lvl = 256;
    const int pg_off = __pg_ix * 256 * 256;

    /* re-computing these adresses everytime is expensive! */
    const double * my_A = A + pg_off;
    const double * my_B = B + pg_off;
    double * my_C = C + pg_off;

    __vr acc_A[UNROLL_FACTOR];
    __vr acc_B[UNROLL_FACTOR];
    __vr acc_C[UNROLL_FACTOR];
    
    #pragma unroll
    for(int i = 0; i < 256; i += UNROLL_FACTOR)
    {
        #pragma unroll (UNROLL_FACTOR)
        for(int ii = 0; ii < UNROLL_FACTOR; ++ii)
        {
            if(i + ii < 256)
            {
                acc_A[ii] = _vel_vld_vssl(8, my_A + (i + ii) * 256, ve_lvl);
                acc_B[ii] = _vel_vld_vssl(8, my_B + (i + ii) * 256, ve_lvl);
            }
        }

        #pragma unroll (UNROLL_FACTOR)
        for(int ii = 0; ii < UNROLL_FACTOR; ++ii)
        {
            if(i + ii < 256)
            {
                _vel_vst_vssl(
                    _vel_vfaddd_vvvl(acc_A[ii], acc_B[ii], ve_lvl), 
                    8, my_C + (i + ii) * 256, ve_lvl);
            }
        }
    }
}