ArrayFire Matrix Multiplication Vectorization

Question

I am using ArrayFire library for signal processing. I am just curious about how to make more efficient my code. I read vectorization guide in docs, but i just ended up using gfor construct. Is it possible to improve efficiency ? Is there a better way for vectorization ? ( I hope there is :)

Note : I am aiming CUDA performance.

Here is the code which i am trying to improve :

#include <arrayfire.h>
#include <stdio.h>
#include <af/util.h>

static int proc_size = 1024;
static int fft_size  = proc_size * 4;
static int staves    = 288;
static int beams     = 256;

static af::array S;
static af::array B;
static af::array R;

void fn()
{
    gfor ( af::seq i, fft_size )
       R( i , af::span ) = matmul( S( i , af::span ) , B( af::span , af::span , i ) );
}

int main(int, char **)
{
    S = af::randn( fft_size , staves , c32 );

    gfor ( af::seq i, fft_size )
        S( i , af::span ) = af::randn( 1 , staves , c32 );

    B = af::randn( staves , beams , fft_size , af::dtype::c32 );
    R = af::constant( af::cfloat { 0 , 0 } , fft_size , beams );

    try
    {
        af::setDevice( 0 );
        af::info();

        double time = af::timeit(fn);

        printf( "Took %f secs.\n" , time );
    }
    catch (const af::exception &ex)
    {
        fprintf(stderr, "%s\n", ex.what());
        throw;
    }

    return 0;
}