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 8
       9  
      10  void NOINLINE
      11  emulate_sqrt_sh(V512 * dest, V512 op1,
      12  		__mmask8 k, int zero_mask)
      13  {
      14    V512 v1, v2, v3, v4, v5, v6, v7, v8;
      15    int i;
      16  
      17    unpack_ph_2twops(op1, &v1, &v2);
      18    unpack_ph_2twops(*dest, &v7, &v8);
      19  
      20    if ((k&1) || !k)
      21      v5.f32[0] = sqrtf(v1.f32[0]);
      22    else if (zero_mask)
      23      v5.f32[0] = 0;
      24    else
      25      v5.f32[0] = v7.f32[0];
      26  
      27    for (i = 1; i < 8; i++)
      28      v5.f32[i] = v1.f32[i];
      29  
      30    *dest = pack_twops_2ph(v5, v6);
      31  }
      32  
      33  void
      34  test_512 (void)
      35  {
      36    V512 res;
      37    V512 exp;
      38  
      39    init_src();
      40  
      41    emulate_sqrt_sh(&exp, src1,  0x1, 0);
      42    res.xmmh[0] = _mm_sqrt_round_sh(exp.xmmh[0], src1.xmmh[0], _ROUND_NINT);
      43    check_results(&res, &exp, N_ELEMS, "_mm_sqrt_round_sh");
      44  
      45    init_dest(&res, &exp);
      46    emulate_sqrt_sh(&exp, src1,  0x1, 0);
      47    res.xmmh[0] = _mm_mask_sqrt_round_sh(res.xmmh[0], 0x1, exp.xmmh[0],
      48                                           src1.xmmh[0], _ROUND_NINT);
      49    check_results(&res, &exp, N_ELEMS, "_mm_mask_sqrt_round_sh");
      50  
      51    emulate_sqrt_sh(&exp, src1,  0x1, 1);
      52    res.xmmh[0] = _mm_maskz_sqrt_round_sh(0x1, exp.xmmh[0], src1.xmmh[0], _ROUND_NINT);
      53    check_results(&res, &exp, N_ELEMS, "_mm_maskz_sqrt_round_sh");
      54  
      55    if (n_errs != 0) {
      56        abort ();
      57    }
      58  
      59  }
      60