gcc/testsuite/gcc.target/i386/avx512f-vpermi2ps-2.c

   1 /* { dg-do run } */
   2 /* { dg-options "-O2 -mavx512f" } */
   3 /* { dg-require-effective-target avx512f } */
   4
   5 #define AVX512F
   6
   7 #include "avx512f-helper.h"
   8
   9 #define SIZE (AVX512F_LEN / 32)
  10 #include "avx512f-mask-type.h"
  11 #include "math.h"
  12 #include "values.h"
  13
  14 static void
  15 CALC (float *dst, float *src1, int *ind, float *src2)
  16 {
  17   int i;
  18
  19   for (i = 0; i < SIZE; i++)
  20     {
  21       unsigned long long offset = ind[i] & (SIZE - 1);
  22       unsigned long long cond = ind[i] & SIZE;
  23
  24       dst[i] = cond ? src2[offset] : src1[offset];
  25     }
  26 }
  27
  28 void static
  29 TEST (void)
  30 {
  31   int i, k;
  32   UNION_TYPE (AVX512F_LEN,) s1, s2, res;
  33   UNION_TYPE (AVX512F_LEN, i_d) ind;
  34   float res_ref[SIZE];
  35
  36   union
  37   {
  38     float f;
  39     int i;
  40   } ind_copy[SIZE];
  41
  42   MASK_TYPE mask = MASK_VALUE;
  43
  44   for (i = 0; i < SIZE; i++)
  45     {
  46       /* Some of the integer indexes may be interpreted as floating point
  47          values in mask-merge mode, that's why we use IND_COPY.  */
  48       ind.a[i] = ind_copy[i].i = 17 * (i << 1);
  49       s1.a[i] = 42.5 * i + 1;
  50       s2.a[i] = 22.5 * i;
  51
  52       res.a[i] = DEFAULT_VALUE;
  53     }
  54
  55   CALC (res_ref, s1.a, ind.a, s2.a);
  56
  57   res.x = INTRINSIC (_mask2_permutex2var_ps) (s1.x, ind.x, mask, s2.x);
  58
  59   /* Standard MASK_MERGE cannot be used since VPERMI2PS in mask-merge mode
  60      merges vectors of two different types (_m512 and __m512i).  */
  61   for (k = 0; k < SIZE; k++)
  62     res_ref[k] = (mask & (1LL << k)) ? res_ref[k] : ind_copy[k].f;
  63
  64   if (UNION_CHECK (AVX512F_LEN,) (res, res_ref))
  65     abort ();
  66 }