Re: vectorization in icc

Thanks!


Now it runs fast!


Some performance measures (lm is of type char):


Case 1: This runs in time 1.44s:


                                #pragma ivdep
                                #pragma vector aligned
                                for( j = 0; j < 16; j++ )
                                {
                                                d[ j ] = d[ j ] + d[ j ];
                                                d[ j ] = d[ j ] | B[ j ];


                                                dm[ j ] = d[ j ] & mm[ j ];
                                }
                                lm = 0;
                                for( j = 0; j < 16; j++ ) if( !dm[ j ] ) lm++;
                                m += lm;


Case 2: If I add #pragma novector in the second loop, it runs in time 3.57s.
Case 3: If I add #pragma novector in the first loop, it runs in time 7.67s.
Case 4: If I add #pragma novector in both loops, it runs in time 3.48s.


All in all, the performance boost is nice, almost 2.5X, but what I still
don't get is why Case 3 is so slow? Especially, why it is slower than
Case 4??? Doesn't matter, however.


Thanks again,


Kimmo.

Return to the comp.compilers page.
Search the comp.compilers archives again.