Results (of ippsSum) depends on the alignment

[Marc_L_]

Hello,

I am observing slightly different results when calling ippsSum_64f with the same numbers but in a differently aligned array; is it a bug or a documented behaviour ?

Thanks !

PS: Here is a simple C program that shows this phenomenon:

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

#include "ipp.h"

#define SIZE 1000
#define NB_TESTS 10

int main (void) {
  int alignment,test,k;
  double input1[SIZE], input2[SIZE+1];

  for (alignment = 0; alignment < 8; alignment ++) {
    int failed = 0;

    for (test = 0; test < NB_TESTS; test ++) {
      double r1, r2;
      double *p = (double *) (alignment + (char *)input2);

      for (k = 0; k < SIZE; k ++) {
        input1 =  (double)rand() / (double)RAND_MAX;
        p = input1;
      }

      ippsSum_64f(input1, SIZE, &r1);
      ippsSum_64f(p, SIZE, &r2);

      if (r1 != r2) {
        printf("[%d %% 8] diff = %e\n", alignment, r2 - r1);
        failed ++;
      }
    }
    printf("[%d %% 8] %d / %d tests failed\n", alignment, failed, NB_TESTS);
  }

  return 0;
}

When the alignment is "0 % 8", ippsSum returns the same result on both inputs but when it's non-zero the test some times succeeds, some times fails (with an error around 1e-13).

[Shaojuan_Z_Intel]

Hi Marc,

For floating point computation, there could be small epsilon difference in result, and that is expected for different memory alignment. Here is an article discusses about the variation https://software.intel.com/en-us/articles/getting-reproducible-results-with-intel-mkl. Although the article talks about MKL, the same rule applies to IPP. Hope this helps. Thanks!

[Marc_L_]

Thank you for your reply !