ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2020 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
#version 450
16

17
#if NCNN_fp16_storage
18
#extension GL_EXT_shader_16bit_storage: require
19
struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
20
#endif
21
#if NCNN_fp16_arithmetic
22
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
23
#endif
24

25
layout (constant_id = 0) const int op_type = 0;
26

27
#define shape_constant_id_offset 1
28
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
29
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
30
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
31
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
32
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
33

34
#if NCNN_image_shader
35
layout (binding = 0) uniform unfp sampler3D bottom_blob_3d;
36
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
37
#else
38
layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
39
#endif
40

41
layout (push_constant) uniform parameter
42
{
43
    int dims;
44
    int w;
45
    int h;
46
    int c;
47
    int cstep;
48
} p;
49

50
void main()
51
{
52
    int gx = int(gl_GlobalInvocationID.x);
53
    int gy = int(gl_GlobalInvocationID.y);
54
    int gz = int(gl_GlobalInvocationID.z);
55

56
    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
57
        return;
58

59
#if NCNN_image_shader
60
    afpvec8 v = image3d_ld8(bottom_blob_3d, ivec3(gx, gy, gz));
61
#else
62
    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
63

64
    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
65
#endif
66

67
    afpvec8 res;
68

69
    if (op_type == 0)
70
    {
71
        res[0] = abs(v[0]);
72
        res[1] = abs(v[1]);
73
    }
74
    if (op_type == 1)
75
    {
76
        res[0] = -v[0];
77
        res[1] = -v[1];
78
    }
79
    if (op_type == 2)
80
    {
81
        res[0] = floor(v[0]);
82
        res[1] = floor(v[1]);
83
    }
84
    if (op_type == 3)
85
    {
86
        res[0] = ceil(v[0]);
87
        res[1] = ceil(v[1]);
88
    }
89
    if (op_type == 4)
90
    {
91
        res[0] = v[0] * v[0];
92
        res[1] = v[1] * v[1];
93
    }
94
    if (op_type == 5)
95
    {
96
        res[0] = sqrt(v[0]);
97
        res[1] = sqrt(v[1]);
98
    }
99
    if (op_type == 6)
100
    {
101
        res[0] = inversesqrt(v[0]);
102
        res[1] = inversesqrt(v[1]);
103
    }
104
    if (op_type == 7)
105
    {
106
        res[0] = exp(v[0]);
107
        res[1] = exp(v[1]);
108
    }
109
    if (op_type == 8)
110
    {
111
        res[0] = log(v[0]);
112
        res[1] = log(v[1]);
113
    }
114
    if (op_type == 9)
115
    {
116
        res[0] = sin(v[0]);
117
        res[1] = sin(v[1]);
118
    }
119
    if (op_type == 10)
120
    {
121
        res[0] = cos(v[0]);
122
        res[1] = cos(v[1]);
123
    }
124
    if (op_type == 11)
125
    {
126
        res[0] = tan(v[0]);
127
        res[1] = tan(v[1]);
128
    }
129
    if (op_type == 12)
130
    {
131
        res[0] = asin(v[0]);
132
        res[1] = asin(v[1]);
133
    }
134
    if (op_type == 13)
135
    {
136
        res[0] = acos(v[0]);
137
        res[1] = acos(v[1]);
138
    }
139
    if (op_type == 14)
140
    {
141
        res[0] = atan(v[0]);
142
        res[1] = atan(v[1]);
143
    }
144
    if (op_type == 15)
145
    {
146
        res[0] = afp(1.f) / v[0];
147
        res[1] = afp(1.f) / v[1];
148
    }
149
    if (op_type == 16)
150
    {
151
#if NCNN_moltenvk
152
        res[0] = afpvec4(tanh(vec4(v[0])));
153
        res[1] = afpvec4(tanh(vec4(v[1])));
154
#else
155
        res[0] = tanh(v[0]);
156
        res[1] = tanh(v[1]);
157
#endif
158
    }
159
    if (op_type == 17)
160
    {
161
        res[0] = log(v[0]) * afp(0.434294481903);
162
        res[1] = log(v[1]) * afp(0.434294481903);
163
    }
164
    if (op_type == 18)
165
    {
166
        res[0] = round(v[0]);
167
        res[1] = round(v[1]);
168
    }
169
    if (op_type == 19)
170
    {
171
        res[0] = trunc(v[0]);
172
        res[1] = trunc(v[1]);
173
    }
174

175
#if NCNN_image_shader
176
    image3d_st8(top_blob_3d, ivec3(gx, gy, gz), res);
177
#else
178
    buffer_st8(bottom_top_blob_data, gi, res);
179
#endif
180
}
181