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
layout (constant_id = 1) const int with_scalar = 0;
27
layout (constant_id = 2) const float const_b = 0;
28

29
#define shape_constant_id_offset 3
30
layout (constant_id = shape_constant_id_offset + 0) const int adims = 0;
31
layout (constant_id = shape_constant_id_offset + 1) const int aw = 0;
32
layout (constant_id = shape_constant_id_offset + 2) const int ah = 0;
33
layout (constant_id = shape_constant_id_offset + 3) const int ac = 0;
34
layout (constant_id = shape_constant_id_offset + 4) const int acstep = 0;
35

36
layout (constant_id = shape_constant_id_offset + 5) const int bdims = 0;
37
layout (constant_id = shape_constant_id_offset + 6) const int bw = 0;
38
layout (constant_id = shape_constant_id_offset + 7) const int bh = 0;
39
layout (constant_id = shape_constant_id_offset + 8) const int bc = 0;
40
layout (constant_id = shape_constant_id_offset + 9) const int bcstep = 0;
41

42
layout (constant_id = shape_constant_id_offset + 10) const int outdims = 0;
43
layout (constant_id = shape_constant_id_offset + 11) const int outw = 0;
44
layout (constant_id = shape_constant_id_offset + 12) const int outh = 0;
45
layout (constant_id = shape_constant_id_offset + 13) const int outc = 0;
46
layout (constant_id = shape_constant_id_offset + 14) const int outcstep = 0;
47

48
#if NCNN_image_shader
49
layout (binding = 0) uniform unfp sampler3D a_blob_3d;
50
layout (binding = 1) uniform unfp sampler3D b_blob_3d;
51
layout (binding = 2, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
52
#else
53
layout (binding = 0) buffer a_blob { sfpvec8 a_blob_data[]; };
54
layout (binding = 1) readonly buffer b_blob { sfpvec8 b_blob_data[]; };
55
layout (binding = 2) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
56
#endif
57

58
layout (push_constant) uniform parameter
59
{
60
    int adims;
61
    int aw;
62
    int ah;
63
    int ac;
64
    int acstep;
65

66
    int bdims;
67
    int bw;
68
    int bh;
69
    int bc;
70
    int bcstep;
71

72
    int outdims;
73
    int outw;
74
    int outh;
75
    int outc;
76
    int outcstep;
77
} p;
78

79
void main()
80
{
81
    int gx = int(gl_GlobalInvocationID.x);
82
    int gy = int(gl_GlobalInvocationID.y);
83
    int gz = int(gl_GlobalInvocationID.z);
84

85
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
86
        return;
87

88
#if NCNN_image_shader
89
    afpvec8 v1 = image3d_ld8(a_blob_3d, ivec3(gx, gy, gz));
90
#else
91
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
92

93
    afpvec8 v1 = buffer_ld8(a_blob_data, gi);
94
#endif
95

96
    afpvec8 v2;
97

98
    if (with_scalar == 1)
99
    {
100
        // scalar
101
        v2[0] = afpvec4(afp(const_b));
102
        v2[1] = afpvec4(afp(const_b));
103
    }
104
    else
105
    {
106
        // no broadcast
107
#if NCNN_image_shader
108
        v2 = image3d_ld8(b_blob_3d, ivec3(gx, gy, gz));
109
#else
110
        v2 = buffer_ld8(b_blob_data, gi);
111
#endif
112
    }
113

114
    afpvec8 res;
115

116
    if (op_type == 0)
117
    {
118
        res[0] = v1[0] + v2[0];
119
        res[1] = v1[1] + v2[1];
120
    }
121
    if (op_type == 1)
122
    {
123
        res[0] = v1[0] - v2[0];
124
        res[1] = v1[1] - v2[1];
125
    }
126
    if (op_type == 2)
127
    {
128
        res[0] = v1[0] * v2[0];
129
        res[1] = v1[1] * v2[1];
130
    }
131
    if (op_type == 3)
132
    {
133
        res[0] = v1[0] / v2[0];
134
        res[1] = v1[1] / v2[1];
135
    }
136
    if (op_type == 4)
137
    {
138
        res[0] = max(v1[0], v2[0]);
139
        res[1] = max(v1[1], v2[1]);
140
    }
141
    if (op_type == 5)
142
    {
143
        res[0] = min(v1[0], v2[0]);
144
        res[1] = min(v1[1], v2[1]);
145
    }
146
    if (op_type == 6)
147
    {
148
        res[0] = pow(v1[0], v2[0]);
149
        res[1] = pow(v1[1], v2[1]);
150
    }
151
    if (op_type == 7)
152
    {
153
        res[0] = v2[0] - v1[0];
154
        res[1] = v2[1] - v1[1];
155
    }
156
    if (op_type == 8)
157
    {
158
        res[0] = v2[0] / v1[0];
159
        res[1] = v2[1] / v1[1];
160
    }
161
    if (op_type == 9)
162
    {
163
        res[0] = pow(v2[0], v1[0]);
164
        res[1] = pow(v2[1], v1[1]);
165
    }
166
    if (op_type == 10)
167
    {
168
#if NCNN_moltenvk
169
        res[0] = afpvec4(atan(vec4(v1[0]), vec4(v2[0])));
170
        res[1] = afpvec4(atan(vec4(v1[1]), vec4(v2[1])));
171
#else
172
        res[0] = atan(v1[0], v2[0]);
173
        res[1] = atan(v1[1], v2[1]);
174
#endif
175
    }
176
    if (op_type == 11)
177
    {
178
#if NCNN_moltenvk
179
        res[0] = afpvec4(atan(vec4(v2[0]), vec4(v1[0])));
180
        res[1] = afpvec4(atan(vec4(v2[1]), vec4(v1[1])));
181
#else
182
        res[0] = atan(v2[0], v1[0]);
183
        res[1] = atan(v2[1], v1[1]);
184
#endif
185
    }
186

187
#if NCNN_image_shader
188
    image3d_st8(top_blob_3d, ivec3(gx, gy, gz), res);
189
#else
190
    buffer_st8(top_blob_data, gi, res);
191
#endif
192
}
193