1  /* { dg-do run { target avx512fp16 } } */
       2  /* { dg-options "-O2 -mavx512fp16 -mavx512dq" } */
       3  
       4  
       5  #define AVX512FP16
       6  #include "avx512fp16-helper.h"
       7  
       8  #define N_ELEMS (AVX512F_LEN / 16)
       9  
      10  void NOINLINE
      11  EMULATE(cvtw2_ph) (V512 * dest, V512 op1,
      12  		 __mmask32 k, int zero_mask)
      13  {
      14    V512 v1, v2, v3, v4, v5, v6, v7, v8;
      15    int i;
      16    __mmask16 m1, m2;
      17  
      18    m1 = k & 0xffff;
      19    m2 = (k >> 16) & 0xffff;
      20  
      21    unpack_ph_2twops(*dest, &v7, &v8);
      22  
      23    for (i = 0; i < 16; i++) {
      24        if (((1 << i) & m1) == 0) {
      25  	  if (zero_mask) {
      26  	      v5.f32[i] = 0;
      27  	  }
      28  	  else {
      29  	      v5.f32[i] = v7.f32[i];
      30  	  }
      31        }
      32        else {
      33  	  v5.f32[i] = op1.u16[i];
      34  
      35        }
      36  
      37        if (((1 << i) & m2) == 0) {
      38  	  if (zero_mask) {
      39  	      v6.f32[i] = 0;
      40  	  }
      41  	  else {
      42  	      v6.f32[i] = v8.f32[i];
      43  	  }
      44        }
      45        else {
      46  	  v6.f32[i] = op1.u16[i+16];
      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(cvtw2_ph)(&exp, src3, NET_MASK, 0);
      62    HF(res) = INTRINSIC (_cvtepu16_ph) (SI(src3));
      63    CHECK_RESULT (&res, &exp, N_ELEMS, _cvtepu16_ph);
      64  
      65    init_dest(&res, &exp);
      66    EMULATE(cvtw2_ph)(&exp, src3, MASK_VALUE, 0);
      67    HF(res) = INTRINSIC (_mask_cvtepu16_ph) (HF(res), MASK_VALUE, SI(src3));
      68    CHECK_RESULT (&res, &exp, N_ELEMS, _mask_cvtepu16_ph);
      69  
      70    EMULATE(cvtw2_ph)(&exp, src3, ZMASK_VALUE, 1);
      71    HF(res) = INTRINSIC (_maskz_cvtepu16_ph) (ZMASK_VALUE, SI(src3));
      72    CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_cvtepu16_ph);
      73  
      74  #if AVX512F_LEN == 512
      75    EMULATE(cvtw2_ph)(&exp, src3, NET_MASK, 0);
      76    HF(res) = INTRINSIC (_cvt_roundepu16_ph) (SI(src3), _ROUND_NINT);
      77    CHECK_RESULT (&res, &exp, N_ELEMS, _cvt_roundepu16_ph);
      78  
      79    init_dest(&res, &exp);
      80    EMULATE(cvtw2_ph)(&exp, src3, MASK_VALUE, 0);
      81    HF(res) = INTRINSIC (_mask_cvt_roundepu16_ph) (HF(res), MASK_VALUE, SI(src3), _ROUND_NINT);
      82    CHECK_RESULT (&res, &exp, N_ELEMS, _mask_cvt_roundepu16_ph);
      83  
      84    EMULATE(cvtw2_ph)(&exp, src3, ZMASK_VALUE, 1);
      85    HF(res) = INTRINSIC (_maskz_cvt_roundepu16_ph) (ZMASK_VALUE, SI(src3), _ROUND_NINT);
      86    CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_cvt_roundepu16_ph);
      87  #endif
      88  
      89    if (n_errs != 0) {
      90        abort ();
      91    }
      92  }
      93