#include <x86intrin.h>
#include <stdio.h>
void stores(void *start, unsigned long size, unsigned long stride, unsigned long offset)
{
  __m256i z = _mm256_set1_epi32(1);
  void *addr;
  void *end = start+size-32;
  unsigned long r=0, i;
  unsigned long incr = (size-32)/stride+1;
  for (i=0; i<1000000000; i+=incr) {
    addr=start;
    do {
      _mm256_storeu_si256((__m256i *)addr,z);
      addr+=stride;
      r++;
    } while (addr<=end);
  }
  /*printf("%ld stores, incr=%ld\n",r,incr);*/
  return;
}