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(div_ph) (V512 * dest, V512 op1, V512 op2,
      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(op1, &v1, &v2);
      22      unpack_ph_2twops(op2, &v3, &v4);
      23      unpack_ph_2twops(*dest, &v7, &v8);
      24  
      25      for (i = 0; i < 16; i++) {
      26          if (((1 << i) & m1) == 0) {
      27              if (zero_mask) {
      28                 v5.f32[i] = 0;
      29              }
      30              else {
      31                 v5.u32[i] = v7.u32[i];
      32              }
      33          }
      34          else {
      35             v5.f32[i] = v1.f32[i] / v3.f32[i];
      36          }
      37  
      38          if (((1 << i) & m2) == 0) {
      39              if (zero_mask) {
      40                 v6.f32[i] = 0;
      41              }
      42              else {
      43                 v6.u32[i] = v8.u32[i];
      44              }
      45          }
      46          else {
      47              v6.f32[i] = v2.f32[i] / v4.f32[i];
      48          }
      49  
      50      }
      51      *dest = pack_twops_2ph(v5, v6);
      52  }
      53  
      54  void
      55  TEST (void)
      56  {
      57    V512 res;
      58    V512 exp;
      59  
      60    init_src();
      61  
      62    EMULATE(div_ph) (&exp, src1, src2, NET_MASK, 0);
      63    HF(res) = INTRINSIC (_div_ph) (HF(src1), HF(src2));
      64    CHECK_RESULT (&res, &exp, N_ELEMS, _div_ph);
      65  
      66    init_dest(&res, &exp);
      67    EMULATE(div_ph) (&exp, src1, src2, MASK_VALUE, 0);
      68    HF(res) = INTRINSIC (_mask_div_ph) (HF(res), MASK_VALUE, HF(src1), HF(src2));
      69    CHECK_RESULT (&res, &exp, N_ELEMS, _mask_div_ph);
      70  
      71    EMULATE(div_ph) (&exp, src1, src2, ZMASK_VALUE, 1);
      72    HF(res) = INTRINSIC (_maskz_div_ph) (ZMASK_VALUE, HF(src1), HF(src2));
      73    CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_div_ph);
      74  
      75  #if AVX512F_LEN == 512
      76  #if AVX512F_LEN == 512
      77    EMULATE(div_ph) (&exp, src1, src2, NET_MASK, 0);
      78    HF(res) = INTRINSIC (_div_round_ph) (HF(src1), HF(src2), _ROUND_NINT);
      79    CHECK_RESULT (&res, &exp, N_ELEMS, _div_ph);
      80  
      81    init_dest(&res, &exp);
      82    EMULATE(div_ph) (&exp, src1, src2, MASK_VALUE, 0);
      83    HF(res) = INTRINSIC (_mask_div_round_ph) (HF(res), MASK_VALUE, HF(src1), HF(src2), _ROUND_NINT);
      84    CHECK_RESULT (&res, &exp, N_ELEMS, _mask_div_ph);
      85    
      86    EMULATE(div_ph) (&exp, src1, src2, ZMASK_VALUE, 1);
      87    HF(res) = INTRINSIC (_maskz_div_round_ph) (ZMASK_VALUE, HF(src1), HF(src2), _ROUND_NINT);
      88    CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_div_ph);
      89  #endif
      90  #endif
      91  
      92    if (n_errs != 0) {
      93        abort ();
      94    }
      95  }
      96  
      97