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_max_sh(V512 * dest, V512 op1, V512 op2,
      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(op2, &v3, &v4);
      19      unpack_ph_2twops(*dest, &v7, &v8);
      20  
      21      if ((k&1) || !k)
      22        v5.f32[i] = v1.f32[i] > v3.f32[i] ? v1.f32[i] : v3.f32[i];
      23      else if (zero_mask)
      24        v5.f32[0] = 0;
      25      else
      26        v5.f32[0] = v7.f32[0];
      27     
      28      for (i = 1; i < 8; i++)
      29        v5.f32[i] = v1.f32[i];
      30  
      31      *dest = pack_twops_2ph(v5, v6);
      32  }
      33  
      34  void
      35  test_512 (void)
      36  {
      37    V512 res;
      38    V512 exp;
      39  
      40    init_src();
      41  
      42    emulate_max_sh(&exp, src1, src2,  0x1, 0);
      43    res.xmmh[0] = _mm_max_sh(src1.xmmh[0], src2.xmmh[0]);
      44    check_results(&res, &exp, N_ELEMS, "_mm_max_sh");
      45  
      46    init_dest(&res, &exp);
      47    emulate_max_sh(&exp, src1, src2,  0x1, 0);
      48    res.xmmh[0] = _mm_mask_max_sh(res.xmmh[0], 0x1, src1.xmmh[0], src2.xmmh[0]);
      49    check_results(&res, &exp, N_ELEMS, "_mm_mask_max_sh");
      50  
      51    emulate_max_sh(&exp, src1, src2,  0x3, 1);
      52    res.xmmh[0] = _mm_maskz_max_sh(0x3, src1.xmmh[0], src2.xmmh[0]);
      53    check_results(&res, &exp, N_ELEMS, "_mm_maskz_max_sh");
      54  
      55    emulate_max_sh(&exp, src1, src2,  0x1, 0);
      56    res.xmmh[0] = _mm_max_round_sh(src1.xmmh[0], src2.xmmh[0], 8);
      57    check_results(&res, &exp, N_ELEMS, "_mm_max_round_sh");
      58  
      59    init_dest(&res, &exp);
      60    emulate_max_sh(&exp, src1, src2,  0x1, 0);
      61    res.xmmh[0] = _mm_mask_max_round_sh(res.xmmh[0], 0x1, src1.xmmh[0], src2.xmmh[0], 8);
      62    check_results(&res, &exp, N_ELEMS, "_mm_mask_max_round_sh");
      63  
      64    emulate_max_sh(&exp, src1, src2,  0x3, 1);
      65    res.xmmh[0] = _mm_maskz_max_round_sh(0x3, src1.xmmh[0], src2.xmmh[0], 8);
      66    check_results(&res, &exp, N_ELEMS, "_mm_maskz_max_round_sh");
      67  
      68    if (n_errs != 0)
      69        abort ();
      70  }
      71  
      72