Hugging Face's logo

camenduru
/
ncnn

Model card Files
xet
Community
ncnn
File size: 4,399 Bytes
be903e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
 
// yala is pleased to support the open source community by making ncnn available.
//
//
// Copyright (C) 2022 yala <zhaojunchao@loongson.cn>;<junchao82@qq.com>. All rights reserved.
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

#include "batchnorm_loongarch.h"

#if __loongarch_sx
#include <lsxintrin.h>
#endif // __loongarch_sx

#include "loongarch_usability.h"

namespace ncnn {

BatchNorm_loongarch::BatchNorm_loongarch()
{
#if __loongarch_sx
    support_packing = true;
#endif // __loongarch_sx
}

int BatchNorm_loongarch::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
{
    int dims = bottom_top_blob.dims;
    int elempack = bottom_top_blob.elempack;

    if (dims == 1)
    {
        int w = bottom_top_blob.w * elempack;

#if __loongarch_sx
        int nn_w = w / 4;
        int remain_w_start = nn_w * 4;
#else
        int remain_w_start = 0;
#endif // __loongarch_sx

        float* ptr = bottom_top_blob;

#if __loongarch_sx
        #pragma omp parallel for num_threads(opt.num_threads)
        for (int i = 0; i < nn_w; i++)
        {
            float* ptr0 = ptr + i * 4;

            __m128 _p = (__m128)__lsx_vld(ptr0, 0);
            __m128 _a = (__m128)__lsx_vld((const float*)a_data + i * 4, 0);
            __m128 _b = (__m128)__lsx_vld((const float*)b_data + i * 4, 0);
            _p = __lsx_vfmadd_s(_b, _p, _a);
            __lsx_vst(_p, ptr0, 0);
        }
#endif // __loongarch_sx

        #pragma omp parallel for num_threads(opt.num_threads)
        for (int i = remain_w_start; i < w; i++)
        {
            ptr[i] = b_data[i] * ptr[i] + a_data[i];
        }
    }

    if (dims == 2)
    {
        int w = bottom_top_blob.w * elempack;
        int h = bottom_top_blob.h;

        #pragma omp parallel for num_threads(opt.num_threads)
        for (int i = 0; i < h; i++)
        {
            float* ptr = bottom_top_blob.row(i);
            float a = a_data[i];
            float b = b_data[i];

            int j = 0;
#if __loongarch_sx
            __m128 _a = elempack == 4 ? (__m128)__lsx_vld((const float*)a_data + i * 4, 0) : (__m128)__lsx_vreplfr2vr_s(a);
            __m128 _b = elempack == 4 ? (__m128)__lsx_vld((const float*)b_data + i * 4, 0) : (__m128)__lsx_vreplfr2vr_s(b);
            for (; j + 3 < w; j += 4)
            {
                __builtin_prefetch(ptr + 16);
                __m128 _p = (__m128)__lsx_vld(ptr, 0);
                _p = __lsx_vfmadd_s(_b, _p, _a);
                __lsx_vst(_p, ptr, 0);

                ptr += 4;
            }
#endif // __loongarch_sx
            for (; j < w; j++)
            {
                *ptr = b * *ptr + a;
                ptr++;
            }
        }
    }

    if (dims == 3 || dims == 4)
    {
        int w = bottom_top_blob.w;
        int h = bottom_top_blob.h;
        int d = bottom_top_blob.d;
        int c = bottom_top_blob.c;
        int size = w * h * d * elempack;

        #pragma omp parallel for num_threads(opt.num_threads)
        for (int q = 0; q < c; q++)
        {
            float* ptr = bottom_top_blob.channel(q);
            float a = a_data[q];
            float b = b_data[q];

            int i = 0;
#if __loongarch_sx
            __m128 _a = elempack == 4 ? (__m128)__lsx_vld((const float*)a_data + q * 4, 0) : (__m128)__lsx_vreplfr2vr_s(a);
            __m128 _b = elempack == 4 ? (__m128)__lsx_vld((const float*)b_data + q * 4, 0) : (__m128)__lsx_vreplfr2vr_s(b);
            for (; i + 3 < size; i += 4)
            {
                __builtin_prefetch(ptr + 16);
                __m128 _p = (__m128)__lsx_vld(ptr, 0);
                _p = __lsx_vfmadd_s(_b, _p, _a);
                __lsx_vst(_p, ptr, 0);

                ptr += 4;
            }
#endif // __loongarch_sx
            for (; i < size; i++)
            {
                *ptr = b * *ptr + a;
                ptr++;
            }
        }
    }

    return 0;
}

} // namespace ncnn