Slow "=" operation problem

[Gang Chen <gchen at IfA dot Hawaii dot Edu>]

In the program attached in the end, elements in a large(N~0.4M) double
array multiply with elements in another small(n=6) double array, and the
results are stored in a third larger (N*n) array. To make the problem
clear, this process is repeated 1000 times.

It looks that the storing part ("=" operation) takes much longer time than
the multiplying part ("*" operation). You can see this from the running
time with the propram only has "line 1", "line 2", or "line 3". The
corresponding time is attached at the end of every line.

I suspect this may be related with L1 or L2 caches. I am wondering is
there anyway to speed the storing part up?

Thanks!


/*
 *gcc -O2 -o TESTSPD speedtestOut.c
 *
 */

#include <stdio.h>
#include <stdlib.h>

#define N_LONG (444752)
#define N_SHORT 6
#define NXI3 (N_LONG*N_SHORT)


void testtime(){
  long i,j,k,l_arr;
  double tmp;
  double vecL[N_LONG], vecS[N_SHORT];
  double * sum;

  sum = (double *) malloc(sizeof(double)*NXI3);

  for(i = 0; i < N_LONG; i++){
    vecL[i] = 1.0;
  }
  for(i = 0; i < N_SHORT; i++){
    vecS[i] = 2.0;
  }
  for(i = 0; i < NXI3; i++){
    sum[i] = 0.0;
  }


  for(k = 0; k < 1000; k++){

   for(j = 0; j < N_LONG; j++){
      l_arr = j*N_SHORT;

      for(i = 0; i <N_SHORT ; i++){
	/* line 1 *///tmp = vecS[i]*vecL[j];                //2.730u
0.058s 0:02.79 99.6%     0+0k 0+0io 69pf+0w
	/* line 2 */  sum[l_arr+i] += vecS[i]*vecL[j];     //28.652u
0.156s 0:28.88 99.7%    0+0k 0+0io 69pf+0w
	/* line 3 *///sum[l_arr+i] = vecS[i];              //25.613u
0.232s 0:25.84 100.0%   0+0k 0+0io 69pf+0w

      }
    }


  }/*k*/
}

main ()
{

  testtime();

}