gcc/testsuite/gcc.target/i386/avx512bw-vpermt2w-2.c

   1 /* { dg-do run } */
   2 /* { dg-options "-O2 -mavx512bw" } */
   3 /* { dg-require-effective-target avx512bw } */
   4
   5 #define AVX512BW
   6 #include "avx512f-helper.h"
   7
   8 #define SIZE (AVX512F_LEN / 16)
   9 #include <math.h>
  10 #include <limits.h>
  11 #include <float.h>
  12 #include "avx512f-mask-type.h"
  13
  14 #define NUM 32
  15
  16 void
  17 CALC (short *dst, short *src1, short *ind, short *src2)
  18 {
  19   int i;
  20
  21   for (i = 0; i < SIZE; i++)
  22     {
  23       unsigned long long offset = ind[i] & (SIZE - 1);
  24       unsigned long long cond = ind[i] & SIZE;
  25
  26       dst[i] = cond ? src2[offset] : src1[offset];
  27     }
  28 }
  29
  30 void
  31 TEST (void)
  32 {
  33   int i, j;
  34   UNION_TYPE (AVX512F_LEN, i_w) s1, s2, res1, res2, res3, ind;
  35   short res_ref[SIZE];
  36
  37   MASK_TYPE mask = MASK_VALUE;
  38
  39   for (i = 0; i < NUM; i++)
  40     {
  41       for (j = 0; j < SIZE; j++)
  42         {
  43           ind.a[j] = i * (j << 1);
  44           s1.a[j] = DEFAULT_VALUE;
  45           s2.a[j] = 1.5 * i * 2 * j;
  46
  47           res1.a[j] = DEFAULT_VALUE;
  48           res2.a[j] = DEFAULT_VALUE;
  49           res3.a[j] = DEFAULT_VALUE;
  50         }
  51
  52       CALC (res_ref, s1.a, ind.a, s2.a);
  53
  54       res1.x = INTRINSIC (_permutex2var_epi16) (s1.x, ind.x, s2.x);
  55       res2.x =
  56         INTRINSIC (_mask_permutex2var_epi16) (s1.x, mask, ind.x, s2.x);
  57       res3.x =
  58         INTRINSIC (_maskz_permutex2var_epi16) (mask, s1.x, ind.x,
  59                                                s2.x);
  60
  61       if (UNION_CHECK (AVX512F_LEN, i_w) (res1, res_ref))
  62         abort ();
  63
  64       MASK_MERGE (i_w) (res_ref, mask, SIZE);
  65       if (UNION_CHECK (AVX512F_LEN, i_w) (res2, res_ref))
  66         abort ();
  67
  68       MASK_ZERO (i_w) (res_ref, mask, SIZE);
  69       if (UNION_CHECK (AVX512F_LEN, i_w) (res3, res_ref))
  70         abort ();
  71     }
  72 }