gcc/testsuite/gcc.target/i386/avx2-vpsadbw-2.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target avx2 } */
   3 /* { dg-options "-O2 -mavx2" } */
   4
   5 #include <string.h>
   6 #include "avx2-check.h"
   7
   8 static void
   9 compute_sadbw256 (unsigned char *s1, unsigned char *s2, unsigned short *r)
  10 {
  11   int i;
  12   unsigned char tmp[32];
  13
  14   for (i = 0; i < 32; i++)
  15     tmp[i] = s1[i] > s2[i] ? s1[i] - s2[i] : s2[i] - s1[i];
  16
  17   memset (r, 0, 32);
  18
  19   for (i = 0; i < 8; i++)
  20     r[0] += tmp[i];
  21
  22   for (i = 8; i < 16; i++)
  23     r[4] += tmp[i];
  24
  25   for (i = 16; i < 24; i++)
  26     r[8] += tmp[i];
  27
  28   for (i = 24; i < 32; i++)
  29     r[12] += tmp[i];
  30 }
  31
  32 void static
  33 avx2_test (void)
  34 {
  35   union256i_b s1, s2;
  36   union256i_w res;
  37   unsigned short res_ref[16];
  38   int i, j;
  39   int fail = 0;
  40
  41   for (i = 0; i < 10; i++)
  42     {
  43       for (j = 0; j < 32; j++)
  44         {
  45           s1.a[j] = j * i;
  46           s2.a[j] = j + 20;
  47         }
  48
  49       res.x = _mm256_sad_epu8 (s1.x, s2.x);;
  50       compute_sadbw256 (s1.a, s2.a, res_ref);
  51
  52       fail += check_union256i_w (res, res_ref);
  53     }
  54
  55   if (fail != 0)
  56     abort ();
  57 }