| 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 |
#ifndef __REDUCEPH__ |
| 11 |
#define __REDUCEPH__ |
| 12 |
V512 borrow_reduce_ps(V512 v, int imm8) |
| 13 |
{ |
| 14 |
V512 temp; |
| 15 |
switch (imm8) |
| 16 |
{ |
| 17 |
case 1: temp.zmm = _mm512_mask_reduce_ps (v.zmm, 0xffff, v.zmm, 1);break; |
| 18 |
case 2: temp.zmm = _mm512_mask_reduce_ps (v.zmm, 0xffff, v.zmm, 2);break; |
| 19 |
case 3: temp.zmm = _mm512_mask_reduce_ps (v.zmm, 0xffff, v.zmm, 3);break; |
| 20 |
case 4: temp.zmm = _mm512_mask_reduce_ps (v.zmm, 0xffff, v.zmm, 4);break; |
| 21 |
case 5: temp.zmm = _mm512_mask_reduce_ps (v.zmm, 0xffff, v.zmm, 5);break; |
| 22 |
case 6: temp.zmm = _mm512_mask_reduce_ps (v.zmm, 0xffff, v.zmm, 6);break; |
| 23 |
case 7: temp.zmm = _mm512_mask_reduce_ps (v.zmm, 0xffff, v.zmm, 7);break; |
| 24 |
case 8: temp.zmm = _mm512_mask_reduce_ps (v.zmm, 0xffff, v.zmm, 8);break; |
| 25 |
} |
| 26 |
return temp; |
| 27 |
} |
| 28 |
#endif |
| 29 |
|
| 30 |
void NOINLINE |
| 31 |
EMULATE(reduce_ph) (V512 * dest, V512 op1, |
| 32 |
__mmask32 k, int imm8, int zero_mask) |
| 33 |
{ |
| 34 |
V512 v1, v2, v3, v4, v5, v6, v7, v8; |
| 35 |
V512 t1,t2; |
| 36 |
int i; |
| 37 |
__mmask16 m1, m2; |
| 38 |
|
| 39 |
m1 = k & 0xffff; |
| 40 |
m2 = (k >> 16) & 0xffff; |
| 41 |
|
| 42 |
unpack_ph_2twops(op1, &v1, &v2); |
| 43 |
unpack_ph_2twops(*dest, &v7, &v8); |
| 44 |
t1 = borrow_reduce_ps(v1, imm8); |
| 45 |
t2 = borrow_reduce_ps(v2, imm8); |
| 46 |
|
| 47 |
for (i = 0; i < 16; i++) { |
| 48 |
if (((1 << i) & m1) == 0) { |
| 49 |
if (zero_mask) { |
| 50 |
v5.f32[i] = 0; |
| 51 |
} |
| 52 |
else { |
| 53 |
v5.u32[i] = v7.u32[i]; |
| 54 |
} |
| 55 |
} |
| 56 |
else { |
| 57 |
v5.f32[i] = t1.f32[i]; |
| 58 |
} |
| 59 |
|
| 60 |
if (((1 << i) & m2) == 0) { |
| 61 |
if (zero_mask) { |
| 62 |
v6.f32[i] = 0; |
| 63 |
} |
| 64 |
else { |
| 65 |
v6.u32[i] = v8.u32[i]; |
| 66 |
} |
| 67 |
} |
| 68 |
else { |
| 69 |
v6.f32[i] = t2.f32[i]; |
| 70 |
} |
| 71 |
|
| 72 |
} |
| 73 |
*dest = pack_twops_2ph(v5, v6); |
| 74 |
} |
| 75 |
|
| 76 |
void |
| 77 |
TEST (void) |
| 78 |
{ |
| 79 |
V512 res; |
| 80 |
V512 exp; |
| 81 |
|
| 82 |
init_src(); |
| 83 |
|
| 84 |
EMULATE(reduce_ph) (&exp, src1, NET_MASK, 6, 0); |
| 85 |
HF(res) = INTRINSIC (_reduce_ph) (HF(src1), 6); |
| 86 |
CHECK_RESULT (&res, &exp, N_ELEMS, _reduce_ph); |
| 87 |
|
| 88 |
init_dest(&res, &exp); |
| 89 |
EMULATE(reduce_ph) (&exp, src1, MASK_VALUE, 5, 0); |
| 90 |
HF(res) = INTRINSIC (_mask_reduce_ph) (HF(res), MASK_VALUE, HF(src1), 5); |
| 91 |
CHECK_RESULT (&res, &exp, N_ELEMS, _mask_reduce_ph); |
| 92 |
|
| 93 |
EMULATE(reduce_ph) (&exp, src1, ZMASK_VALUE, 4, 1); |
| 94 |
HF(res) = INTRINSIC (_maskz_reduce_ph) (ZMASK_VALUE, HF(src1), 4); |
| 95 |
CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_reduce_ph); |
| 96 |
|
| 97 |
#if AVX512F_LEN == 512 |
| 98 |
EMULATE(reduce_ph) (&exp, src1, NET_MASK, 6, 0); |
| 99 |
HF(res) = INTRINSIC (_reduce_round_ph) (HF(src1), 6, _ROUND_CUR); |
| 100 |
CHECK_RESULT (&res, &exp, N_ELEMS, _reduce_round_ph); |
| 101 |
|
| 102 |
init_dest(&res, &exp); |
| 103 |
EMULATE(reduce_ph) (&exp, src1, MASK_VALUE, 5, 0); |
| 104 |
HF(res) = INTRINSIC (_mask_reduce_round_ph) (HF(res), MASK_VALUE, HF(src1), 5, _ROUND_CUR); |
| 105 |
CHECK_RESULT (&res, &exp, N_ELEMS, _mask_reduce_round_ph); |
| 106 |
|
| 107 |
EMULATE(reduce_ph) (&exp, src1, ZMASK_VALUE, 4, 1); |
| 108 |
HF(res) = INTRINSIC (_maskz_reduce_round_ph) (ZMASK_VALUE, HF(src1), 4, _ROUND_CUR); |
| 109 |
CHECK_RESULT (&res, &exp, N_ELEMS, _maskz_reduce_round_ph); |
| 110 |
#endif |
| 111 |
|
| 112 |
if (n_errs != 0) { |
| 113 |
abort (); |
| 114 |
} |
| 115 |
} |
| 116 |
|