1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
5
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
6
// in compliance with the License. You may obtain a copy of the License at
8
// https://opensource.org/licenses/BSD-3-Clause
10
// Unless required by applicable law or agreed to in writing, software distributed
11
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
12
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
13
// specific language governing permissions and limitations under the License.
15
#include "hardswish_mips.h"
21
#include "mips_usability.h"
25
HardSwish_mips::HardSwish_mips()
28
support_packing = true;
32
int HardSwish_mips::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
34
int w = bottom_top_blob.w;
35
int h = bottom_top_blob.h;
36
int d = bottom_top_blob.d;
37
int channels = bottom_top_blob.c;
38
int elempack = bottom_top_blob.elempack;
39
int size = w * h * d * elempack;
41
#pragma omp parallel for num_threads(opt.num_threads)
42
for (int q = 0; q < channels; q++)
44
float* ptr = bottom_top_blob.channel(q);
48
v4f32 _zero = (v4f32)__msa_fill_w(0);
49
v4f32 _one = (v4f32)__msa_fill_w_f32(1.f);
50
v4f32 _alpha = (v4f32)__msa_fill_w_f32(alpha);
51
v4f32 _beta = (v4f32)__msa_fill_w_f32(beta);
52
for (; i + 3 < size; i += 4)
54
__builtin_prefetch(ptr + 16);
55
v4f32 _p = (v4f32)__msa_ld_w(ptr, 0);
56
v4f32 _outp = __msa_fmadd_w(_beta, _p, _alpha);
57
_outp = __msa_fmax_w(_outp, _zero);
58
_outp = __msa_fmin_w(_outp, _one);
59
_outp = __msa_fmul_w(_outp, _p);
60
__msa_st_w((v4i32)_outp, ptr, 0);
69
else if (*ptr > upper)
72
*ptr = *ptr * (*ptr * alpha + beta);