ncnn

1
// Xavier Hsinyuan is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2021 Xavier Hsinyuan <thelastlinex@hotmail.com>. All rights reserved.
4
//
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
7
//
8
// https://opensource.org/licenses/BSD-3-Clause
9
//
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.
14

15
#include "relu_riscv.h"
16

17
#if __riscv_vector
18
#include <riscv_vector.h>
19
#endif // __riscv_vector
20

21
namespace ncnn {
22

23
ReLU_riscv::ReLU_riscv()
24
{
25
#if __riscv_vector
26
    support_packing = true;
27
#if __riscv_zfh
28
    support_fp16_storage = true;
29
#endif
30
#endif
31
}
32

33
int ReLU_riscv::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
34
{
35
#if __riscv_vector && __riscv_zfh
36
    int elembits = bottom_top_blob.elembits();
37

38
    if (opt.use_fp16_storage && elembits == 16)
39
    {
40
        return forward_inplace_fp16s(bottom_top_blob, opt);
41
    }
42
#endif
43

44
    int w = bottom_top_blob.w;
45
    int h = bottom_top_blob.h;
46
    int d = bottom_top_blob.d;
47
    int channels = bottom_top_blob.c;
48
    int elempack = bottom_top_blob.elempack;
49
    int size = w * h * d * elempack;
50

51
    #pragma omp parallel for num_threads(opt.num_threads)
52
    for (int q = 0; q < channels; q++)
53
    {
54
        float* ptr = bottom_top_blob.channel(q);
55
        if (slope == 0.f)
56
        {
57
#if __riscv_vector
58
            int n = size;
59
            while (n > 0)
60
            {
61
                size_t vl = vsetvl_e32m8(n);
62

63
                vfloat32m8_t _p = vle32_v_f32m8(ptr, vl);
64
                _p = vfmax_vf_f32m8(_p, 0.f, vl);
65
                vse32_v_f32m8(ptr, _p, vl);
66

67
                ptr += vl;
68
                n -= vl;
69
            }
70
#else  // __riscv_vector
71
            for (int i = 0; i < size; i++)
72
            {
73
                if (*ptr < 0)
74
                    *ptr = 0;
75
                ptr++;
76
            }
77
#endif // __riscv_vector
78
        }
79
        else
80
        {
81
#if __riscv_vector
82
            int n = size;
83
            while (n > 0)
84
            {
85
                size_t vl = vsetvl_e32m8(n);
86

87
                vfloat32m8_t _p = vle32_v_f32m8(ptr, vl);
88
                _p = vfmul_vf_f32m8_m(vmflt_vf_f32m8_b4(_p, .0f, vl), _p, _p, slope, vl); //slope: float(float32_t)
89
                vse32_v_f32m8(ptr, _p, vl);
90

91
                ptr += vl;
92
                n -= vl;
93
            }
94
#else  // __riscv_vector
95
            for (int i = 0; i < size; i++)
96
            {
97
                if (*ptr < 0)
98
                    *ptr *= slope;
99
                ptr++;
100
            }
101
#endif // __riscv_vector
102
        }
103
    }
104

105
    return 0;
106
}
107

108
#if __riscv_vector && __riscv_zfh
109
int ReLU_riscv::forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) const
110
{
111
    int w = bottom_top_blob.w;
112
    int h = bottom_top_blob.h;
113
    int d = bottom_top_blob.d;
114
    int channels = bottom_top_blob.c;
115
    int elempack = bottom_top_blob.elempack;
116
    int size = w * h * d * elempack;
117

118
    #pragma omp parallel for num_threads(opt.num_threads)
119
    for (int q = 0; q < channels; q++)
120
    {
121
        __fp16* ptr = bottom_top_blob.channel(q);
122
        if (slope == 0.f)
123
        {
124
            int n = size;
125
            while (n > 0)
126
            {
127
                size_t vl = vsetvl_e16m8(n);
128

129
                vfloat16m8_t _p = vle16_v_f16m8(ptr, vl);
130
                _p = vfmax_vf_f16m8(_p, (__fp16)0.f, vl);
131
                vse16_v_f16m8(ptr, _p, vl);
132

133
                ptr += vl;
134
                n -= vl;
135
            }
136
        }
137
        else
138
        {
139
            int n = size;
140
            __fp16 _slope = (__fp16)slope;
141
            while (n > 0)
142
            {
143
                size_t vl = vsetvl_e16m8(n);
144

145
                vfloat16m8_t _p = vle16_v_f16m8(ptr, vl);
146
                _p = vfmul_vf_f16m8_m(vmflt_vf_f16m8_b2(_p, .0f, vl), _p, _p, _slope, vl);
147
                vse16_v_f16m8(ptr, _p, vl);
148

149
                ptr += vl;
150
                n -= vl;
151
            }
152
        }
153
    }
154

155
    return 0;
156
}
157

158
#endif
159
} // namespace ncnn
160