output[ __umul24(i, j_dim) + j ] = outval0 ;
output[ __umul24(i, j_dim) + j+1 ] = outval1 ;
output[ __umul24(i, j_dim) + j+2 ] = outval2 ;
- output[ __umul24(i, j_dim) + j+3 ] = outval3;
- output[ __umul24(i, j_dim) + j+4 ] = outval4;
- output[ __umul24(i, j_dim) + j+5 ] = outval5;
+ output[ __umul24(i, j_dim) + j+3 ] = outval3 ;
+ output[ __umul24(i, j_dim) + j+4 ] = outval4 ;
+ output[ __umul24(i, j_dim) + j+5 ] = outval5 ;
output[ __umul24(i, j_dim) + j+6 ] = outval6 ;
output[ __umul24(i, j_dim) + j+7 ] = outval7 ;
}