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_fmadd_sh(V512 * dest, V512 op1, V512 op2,
      12                  __mmask8 k, int zero_mask, int mask3)
      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[0] = v1.f32[0] * v3.f32[0] + v7.f32[0];
      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        if (mask3)
      30          v5.f32[i] = v7.f32[i];
      31        else
      32          v5.f32[i] = v1.f32[i];
      33      }
      34      *dest = pack_twops_2ph(v5, v6);
      35  }
      36  
      37  void
      38  test_512 (void)
      39  {
      40    V512 res;
      41    V512 exp;
      42  
      43    init_src();
      44  
      45    init_dest(&res, &exp);
      46    emulate_fmadd_sh(&exp, src1, src2,  0x1, 0, 0);
      47    res.xmmh[0] = _mm_fmadd_sh(src1.xmmh[0], src2.xmmh[0], res.xmmh[0]);
      48    check_results(&res, &exp, N_ELEMS, "_mm_fmadd_sh");
      49    init_dest(&res, &exp);
      50    emulate_fmadd_sh(&exp, src1, src2,  0x1, 0, 1);
      51    res.xmmh[0] = _mm_mask3_fmadd_sh(src1.xmmh[0], src2.xmmh[0], res.xmmh[0], 
      52                                      0x1);
      53    check_results(&res, &exp, N_ELEMS, "_mm_mask3_fmadd_sh");
      54    init_dest(&res, &exp);
      55    emulate_fmadd_sh(&exp, src1, src2,  0x1, 0, 0);
      56    res.xmmh[0] = _mm_mask_fmadd_sh(src1.xmmh[0], 0x1, src2.xmmh[0], 
      57                                     res.xmmh[0]);
      58    check_results(&res, &exp, N_ELEMS, "_mm_mask_fmadd_sh");
      59    init_dest(&res, &exp);
      60    emulate_fmadd_sh(&exp, src1, src2,  0x3, 1, 0);
      61    res.xmmh[0] = _mm_maskz_fmadd_sh(0x3, src1.xmmh[0], src2.xmmh[0], 
      62                                      res.xmmh[0]);
      63    check_results(&res, &exp, N_ELEMS, "_mm_maskz_fmadd_sh");
      64  
      65    init_dest(&res, &exp);
      66    emulate_fmadd_sh(&exp, src1, src2,  0x1, 0, 0);
      67    res.xmmh[0] = _mm_fmadd_round_sh(src1.xmmh[0], src2.xmmh[0], res.xmmh[0], 
      68                                      _ROUND_NINT);
      69    check_results(&res, &exp, N_ELEMS, "_mm_fmadd_sh");
      70    init_dest(&res, &exp);
      71    emulate_fmadd_sh(&exp, src1, src2,  0x1, 0, 1);
      72    res.xmmh[0] = _mm_mask3_fmadd_round_sh(src1.xmmh[0], src2.xmmh[0], 
      73                                            res.xmmh[0], 0x1, _ROUND_NINT);
      74    check_results(&res, &exp, N_ELEMS, "_mm_mask3_fmadd_sh");
      75    init_dest(&res, &exp);
      76    emulate_fmadd_sh(&exp, src1, src2,  0x1, 0, 0);
      77    res.xmmh[0] = _mm_mask_fmadd_round_sh(src1.xmmh[0], 0x1, src2.xmmh[0], 
      78                                           res.xmmh[0], _ROUND_NINT);
      79    check_results(&res, &exp, N_ELEMS, "_mm_mask_fmadd_sh");
      80    init_dest(&res, &exp);
      81    emulate_fmadd_sh(&exp, src1, src2,  0x3, 1, 0);
      82    res.xmmh[0] = _mm_maskz_fmadd_round_sh(0x3, src1.xmmh[0], src2.xmmh[0], 
      83                                            res.xmmh[0], _ROUND_NINT);
      84    check_results(&res, &exp, N_ELEMS, "_mm_maskz_fmadd_sh");
      85  
      86    if (n_errs != 0) {
      87        abort ();
      88    }
      89  }
      90