| 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(c_fmadd_pch) (V512 * dest, V512 op1, V512 op2, |
| 12 |
__mmask16 k, int zero_mask, int c_flag, |
| 13 |
int is_mask3) |
| 14 |
{ |
| 15 |
V512 v1, v2, v3, v4, v5, v6, v7, v8; |
| 16 |
int i; |
| 17 |
int invert = 1; |
| 18 |
if (c_flag == 1) |
| 19 |
invert = -1; |
| 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 / 2)) & k) == 0) { |
| 27 |
if (zero_mask) { |
| 28 |
v5.f32[i] = 0; |
| 29 |
} |
| 30 |
else { |
| 31 |
v5.u32[i] = is_mask3 ? v3.u32[i] : v7.u32[i]; |
| 32 |
} |
| 33 |
} |
| 34 |
else { |
| 35 |
if ((i % 2) == 0) { |
| 36 |
v5.f32[i] = v1.f32[i] * v7.f32[i] |
| 37 |
- invert * (v1.f32[i+1] * v7.f32[i+1]) + v3.f32[i]; |
| 38 |
} |
| 39 |
else { |
| 40 |
v5.f32[i] = v1.f32[i-1] * v7.f32[i] |
| 41 |
+ invert * (v1.f32[i] * v7.f32[i-1]) + v3.f32[i]; |
| 42 |
|
| 43 |
} |
| 44 |
} |
| 45 |
if (((1 << (i / 2 + 8)) & k) == 0) { |
| 46 |
if (zero_mask) { |
| 47 |
v6.f32[i] = 0; |
| 48 |
} |
| 49 |
else { |
| 50 |
v6.u32[i] = is_mask3 ? v4.u32[i] : v8.u32[i]; |
| 51 |
} |
| 52 |
} |
| 53 |
else { |
| 54 |
if ((i % 2) == 0) { |
| 55 |
v6.f32[i] = v2.f32[i] * v8.f32[i] |
| 56 |
- invert * (v2.f32[i+1] * v8.f32[i+1]) + v4.f32[i]; |
| 57 |
} |
| 58 |
else { |
| 59 |
v6.f32[i] = v2.f32[i-1] * v8.f32[i] |
| 60 |
+ invert * (v2.f32[i] * v8.f32[i-1]) + v4.f32[i]; |
| 61 |
} |
| 62 |
|
| 63 |
} |
| 64 |
} |
| 65 |
|
| 66 |
*dest = pack_twops_2ph(v5, v6); |
| 67 |
} |
| 68 |
|
| 69 |
void |
| 70 |
TEST (void) |
| 71 |
{ |
| 72 |
V512 res; |
| 73 |
V512 exp; |
| 74 |
|
| 75 |
init_src(); |
| 76 |
|
| 77 |
init_dest(&res, &exp); |
| 78 |
EMULATE(c_fmadd_pch)(&exp, src1, src2, NET_CMASK, 0, 1, 0); |
| 79 |
HF(res) = INTRINSIC (_fcmadd_pch) (HF(res), HF(src1), |
| 80 |
HF(src2)); |
| 81 |
CHECK_RESULT (&res, &exp, N_ELEMS, _fcmadd_pch); |
| 82 |
|
| 83 |
init_dest(&res, &exp); |
| 84 |
EMULATE(c_fmadd_pch)(&exp, src1, src2, HALF_MASK, 0, 1, 0); |
| 85 |
HF(res) = INTRINSIC (_mask_fcmadd_pch) (HF(res) ,HALF_MASK, HF(src1), |
| 86 |
HF(src2)); |
| 87 |
CHECK_RESULT (&res, &exp, N_ELEMS, _mask_fcmadd_pch); |
| 88 |
|
| 89 |
init_dest(&res, &exp); |
| 90 |
EMULATE(c_fmadd_pch)(&exp, src1, src2, HALF_MASK, 0, 1, 1); |
| 91 |
HF(res) = INTRINSIC (_mask3_fcmadd_pch) (HF(res), HF(src1), |
| 92 |
HF(src2), HALF_MASK); |
| 93 |
CHECK_RESULT (&res, &exp, N_ELEMS, _mask3_fcmadd_pch); |
| 94 |
|
| 95 |
|
| 96 |
init_dest(&res, &exp); |
| 97 |
EMULATE(c_fmadd_pch)(&exp, src1, src2, HALF_MASK, 1, 1, 0); |
| 98 |
HF(res) = INTRINSIC (_maskz_fcmadd_pch) (HALF_MASK, HF(res), |
| 99 |
HF(src1), HF(src2)); |
| 100 |
CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_fcmadd_pch); |
| 101 |
|
| 102 |
#if AVX512F_LEN == 512 |
| 103 |
init_dest(&res, &exp); |
| 104 |
EMULATE(c_fmadd_pch)(&exp, src1, src2, NET_CMASK, 0, 1, 0); |
| 105 |
HF(res) = INTRINSIC (_fcmadd_round_pch) (HF(res), HF(src1), |
| 106 |
HF(src2), _ROUND_NINT); |
| 107 |
CHECK_RESULT (&res, &exp, N_ELEMS, _fcmadd_pch); |
| 108 |
|
| 109 |
init_dest(&res, &exp); |
| 110 |
EMULATE(c_fmadd_pch)(&exp, src1, src2, HALF_MASK, 0, 1, 0); |
| 111 |
HF(res) = INTRINSIC (_mask_fcmadd_round_pch) (HF(res) ,HALF_MASK, HF(src1), |
| 112 |
HF(src2), _ROUND_NINT); |
| 113 |
CHECK_RESULT (&res, &exp, N_ELEMS, _mask_fcmadd_pch); |
| 114 |
|
| 115 |
init_dest(&res, &exp); |
| 116 |
EMULATE(c_fmadd_pch)(&exp, src1, src2, HALF_MASK, 0, 1, 1); |
| 117 |
HF(res) = INTRINSIC (_mask3_fcmadd_round_pch) (HF(res), HF(src1), |
| 118 |
HF(src2), HALF_MASK, _ROUND_NINT); |
| 119 |
CHECK_RESULT (&res, &exp, N_ELEMS, _mask3_fcmadd_pch); |
| 120 |
|
| 121 |
|
| 122 |
init_dest(&res, &exp); |
| 123 |
EMULATE(c_fmadd_pch)(&exp, src1, src2, HALF_MASK, 1, 1, 0); |
| 124 |
HF(res) = INTRINSIC (_maskz_fcmadd_round_pch) (HALF_MASK, HF(res), |
| 125 |
HF(src1), HF(src2), _ROUND_NINT); |
| 126 |
CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_fcmadd_pch); |
| 127 |
#endif |
| 128 |
|
| 129 |
if (n_errs != 0) { |
| 130 |
abort (); |
| 131 |
} |
| 132 |
} |
| 133 |
|