ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2022 THL A29 Limited, a Tencent company. 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 "cast_mips.h"
16

17
#if __mips_msa
18
#include <msa.h>
19
#endif // __mips_msa
20

21
namespace ncnn {
22

23
Cast_mips::Cast_mips()
24
{
25
    support_packing = true;
26
}
27

28
int Cast_mips::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
29
{
30
    if (type_from == type_to)
31
    {
32
        top_blob = bottom_blob;
33
        return 0;
34
    }
35

36
    int w = bottom_blob.w;
37
    int h = bottom_blob.h;
38
    int d = bottom_blob.d;
39
    int channels = bottom_blob.c;
40
    int dims = bottom_blob.dims;
41
    size_t elemsize = bottom_blob.elemsize;
42
    int elempack = bottom_blob.elempack;
43

44
    size_t out_elemsize = elemsize;
45
    if (type_to == 1)
46
    {
47
        if (type_from == 3)
48
        {
49
            Cast::forward(bottom_blob, top_blob, opt);
50
        }
51

52
        // float32
53
        out_elemsize = 4 * elempack;
54
    }
55
    else if (type_to == 2)
56
    {
57
        // float16
58
        out_elemsize = 2 * elempack;
59
    }
60
    else if (type_to == 3)
61
    {
62
        // int8
63
        out_elemsize = elempack;
64
    }
65
    else if (type_to == 4)
66
    {
67
        // bfloat16
68
        out_elemsize = 2 * elempack;
69
    }
70

71
    if (dims == 1)
72
    {
73
        top_blob.create(w, out_elemsize, elempack, opt.blob_allocator);
74
    }
75
    else if (dims == 2)
76
    {
77
        top_blob.create(w, h, out_elemsize, elempack, opt.blob_allocator);
78
    }
79
    else if (dims == 3)
80
    {
81
        top_blob.create(w, h, channels, out_elemsize, elempack, opt.blob_allocator);
82
    }
83
    else if (dims == 4)
84
    {
85
        top_blob.create(w, h, d, channels, out_elemsize, elempack, opt.blob_allocator);
86
    }
87
    if (top_blob.empty())
88
        return -100;
89

90
    int size = w * h * d * elempack;
91

92
    if (type_from == 1 && type_to == 2)
93
    {
94
        #pragma omp parallel for num_threads(opt.num_threads)
95
        for (int q = 0; q < channels; q++)
96
        {
97
            const float* ptr = bottom_blob.channel(q);
98
            unsigned short* outptr = top_blob.channel(q);
99

100
            int i = 0;
101
#if __mips_msa
102
            for (; i + 7 < size; i += 8)
103
            {
104
                __builtin_prefetch(ptr + 16);
105
                v4f32 _p0 = (v4f32)__msa_ld_w(ptr, 0);
106
                v4f32 _p1 = (v4f32)__msa_ld_w(ptr + 4, 0);
107
                v8i16 _p = __msa_fexdo_h(_p1, _p0);
108
                __msa_st_h(_p, outptr, 0);
109

110
                ptr += 8;
111
                outptr += 8;
112
            }
113
#endif // __mips_msa
114
            for (; i < size; i++)
115
            {
116
                *outptr = float32_to_float16(*ptr);
117
                outptr++;
118
                ptr++;
119
            }
120
        }
121
    }
122

123
    if (type_from == 2 && type_to == 1)
124
    {
125
        #pragma omp parallel for num_threads(opt.num_threads)
126
        for (int q = 0; q < channels; q++)
127
        {
128
            const unsigned short* ptr = bottom_blob.channel(q);
129
            float* outptr = top_blob.channel(q);
130

131
            int i = 0;
132
#if __mips_msa
133
            for (; i + 7 < size; i += 8)
134
            {
135
                __builtin_prefetch(ptr + 16);
136
                v8i16 _p = __msa_ld_h(ptr, 0);
137
                v4f32 _p0 = __msa_fexupr_w(_p);
138
                v4f32 _p1 = __msa_fexupl_w(_p);
139
                __msa_st_w((v4i32)_p0, outptr, 0);
140
                __msa_st_w((v4i32)_p1, outptr + 4, 0);
141

142
                ptr += 8;
143
                outptr += 8;
144
            }
145
#endif // __mips_msa
146
            for (; i < size; i++)
147
            {
148
                *outptr = float16_to_float32(*ptr);
149
                outptr++;
150
                ptr++;
151
            }
152
        }
153
    }
154

155
    if (type_from == 3 && type_to == 1)
156
    {
157
        #pragma omp parallel for num_threads(opt.num_threads)
158
        for (int q = 0; q < channels; q++)
159
        {
160
            const signed char* ptr = bottom_blob.channel(q);
161
            float* outptr = top_blob.channel(q);
162

163
            for (int i = 0; i < size; i++)
164
            {
165
                outptr[i] = (float)ptr[i];
166
            }
167
        }
168
    }
169

170
    if (type_from == 4 && type_to == 1)
171
    {
172
        #pragma omp parallel for num_threads(opt.num_threads)
173
        for (int q = 0; q < channels; q++)
174
        {
175
            const unsigned short* ptr = bottom_blob.channel(q);
176
            float* outptr = top_blob.channel(q);
177

178
            int i = 0;
179
            for (; i < size; i++)
180
            {
181
                *outptr = bfloat16_to_float32(*ptr);
182
                outptr++;
183
                ptr++;
184
            }
185
        }
186
    }
187

188
    if (type_from == 1 && type_to == 4)
189
    {
190
        #pragma omp parallel for num_threads(opt.num_threads)
191
        for (int q = 0; q < channels; q++)
192
        {
193
            const float* ptr = bottom_blob.channel(q);
194
            unsigned short* outptr = top_blob.channel(q);
195

196
            int i = 0;
197
            for (; i < size; i++)
198
            {
199
                *outptr = float32_to_bfloat16(*ptr);
200
                outptr++;
201
                ptr++;
202
            }
203
        }
204
    }
205

206
    return 0;
207
}
208

209
} // namespace ncnn
210