1  /* { dg-do run { target avx512fp16 } } */
       2  /* { dg-options "-O2 -mavx512fp16 -mavx512dq" } */
       3  
       4  #define AVX512FP16
       5  #include "avx512fp16-helper.h"
       6  
       7  #define N_ELEMS (AVX512F_LEN / 16)
       8  
       9  void NOINLINE
      10  EMULATE(add_ph) (V512 * dest, V512 op1, V512 op2,
      11  		 __mmask32 k, int zero_mask)
      12  {
      13      V512 v1, v2, v3, v4, v5, v6, v7, v8;
      14      int i;
      15      __mmask16 m1, m2;
      16  
      17      m1 = k & 0xffff;
      18      m2 = (k >> 16) & 0xffff;
      19  
      20      unpack_ph_2twops(op1, &v1, &v2);
      21      unpack_ph_2twops(op2, &v3, &v4);
      22      unpack_ph_2twops(*dest, &v7, &v8);
      23  
      24      for (i = 0; i < 16; i++) {
      25          if (((1 << i) & m1) == 0) {
      26              if (zero_mask) {
      27                 v5.f32[i] = 0;
      28              }
      29              else {
      30                 v5.u32[i] = v7.u32[i];
      31              }
      32          }
      33          else {
      34             v5.f32[i] = v1.f32[i] + v3.f32[i];
      35          }
      36  
      37          if (((1 << i) & m2) == 0) {
      38              if (zero_mask) {
      39                 v6.f32[i] = 0;
      40              }
      41              else {
      42                 v6.u32[i] = v8.u32[i];
      43              }
      44          }
      45          else {
      46              v6.f32[i] = v2.f32[i] + v4.f32[i];
      47          }
      48  
      49      }
      50      *dest = pack_twops_2ph(v5, v6);
      51  }
      52  
      53  void
      54  TEST (void)
      55  {
      56    V512 res;
      57    V512 exp;
      58  
      59    init_src();
      60    
      61    EMULATE(add_ph) (&exp, src1, src2, NET_MASK, 0);
      62    HF(res) = INTRINSIC (_add_ph) (HF(src1), HF(src2));
      63    CHECK_RESULT (&res, &exp, N_ELEMS, _add_ph);
      64  
      65    init_dest(&res, &exp);
      66    EMULATE(add_ph) (&exp, src1, src2, MASK_VALUE, 0);
      67    HF(res) = INTRINSIC (_mask_add_ph) (HF(res), MASK_VALUE, HF(src1), HF(src2));
      68    CHECK_RESULT (&res, &exp, N_ELEMS, _mask_add_ph);
      69  
      70    EMULATE(add_ph) (&exp, src1, src2, ZMASK_VALUE, 1);
      71    HF(res) = INTRINSIC (_maskz_add_ph) (ZMASK_VALUE, HF(src1), HF(src2));
      72    CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_add_ph);
      73  
      74  #if AVX512F_LEN == 512
      75    EMULATE(add_ph) (&exp, src1, src2, NET_MASK, 0);
      76    HF(res) = INTRINSIC (_add_round_ph) (HF(src1), HF(src2), _ROUND_NINT);
      77    CHECK_RESULT (&res, &exp, N_ELEMS, _add_round_ph);
      78  
      79    init_dest(&res, &exp);
      80    EMULATE(add_ph) (&exp, src1, src2, MASK_VALUE, 0);
      81    HF(res) = INTRINSIC (_mask_add_round_ph) (HF(res), MASK_VALUE, HF(src1), HF(src2), _ROUND_NINT);
      82    CHECK_RESULT (&res, &exp, N_ELEMS, _mask_add_round_ph);
      83  
      84    EMULATE(add_ph) (&exp, src1, src2, ZMASK_VALUE, 1);
      85    HF(res) = INTRINSIC (_maskz_add_round_ph) (ZMASK_VALUE, HF(src1), HF(src2), _ROUND_NINT);
      86    CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_add_round_ph);
      87  #endif
      88  
      89    if (n_errs != 0) {
      90        abort ();
      91    }
      92  }