ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2019 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
#endif
20
#if NCNN_fp16_arithmetic
21
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
22
#endif
23

24
#extension GL_GOOGLE_include_directive: enable
25
#include "vulkan_activation.comp"
26

27
layout (constant_id = 0) const int bias_term = 0;
28
layout (constant_id = 1) const int activation_type = 0;
29
layout (constant_id = 2) const float activation_param_0 = 0;
30
layout (constant_id = 3) const float activation_param_1 = 0;
31

32
#define shape_constant_id_offset 4
33
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
34
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
35
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
36
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
37
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
38

39
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
40
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
41
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
42
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
43
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
44

45
#if NCNN_image_shader
46
layout (binding = 0) uniform unfp sampler3D bottom_blob;
47
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
48
layout (binding = 2) uniform unfp sampler3D weight_blob;
49
layout (binding = 3) uniform unfp sampler3D bias_blob;
50
#else
51
layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
52
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
53
#if NCNN_fp16_packed || (NCNN_fp16_storage && !NCNN_fp16_arithmetic)
54
// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
55
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
56
#else
57
layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
58
#endif
59
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
60
#endif
61

62
layout (push_constant) uniform parameter
63
{
64
    int dims;
65
    int w;
66
    int h;
67
    int c;
68
    int cstep;
69

70
    int outdims;
71
    int outw;
72
    int outh;
73
    int outc;
74
    int outcstep;
75
} p;
76

77
void main()
78
{
79
    int gx = int(gl_GlobalInvocationID.x);
80
    int gy = int(gl_GlobalInvocationID.y);
81
    int gz = int(gl_GlobalInvocationID.z);
82

83
    if (gx >= psc(outw) || gy >= 1 || gz >= 1)
84
        return;
85

86
    afpvec4 sum;
87

88
    if (bias_term == 1)
89
    {
90
#if NCNN_image_shader
91
        sum = image3d_ld4(bias_blob, ivec3(gx, 0, 0));
92
#else
93
        sum = buffer_ld4(bias_data, gx);
94
#endif
95
    }
96
    else
97
    {
98
        sum = afpvec4(0.f);
99
    }
100

101
#if NCNN_image_shader
102
    int wx = 0;
103

104
    for (int i = 0; i < psc(w); i++)
105
    {
106
        afpvec4 v = image3d_ld4(bottom_blob, ivec3(i, 0, 0));
107
        afpmat4 k = afpmat4(
108
            image3d_ld4(weight_blob, ivec3(wx + 0, gx, 0)),
109
            image3d_ld4(weight_blob, ivec3(wx + 1, gx, 0)),
110
            image3d_ld4(weight_blob, ivec3(wx + 2, gx, 0)),
111
            image3d_ld4(weight_blob, ivec3(wx + 3, gx, 0))
112
        );
113

114
        sum += v * k;
115

116
        wx += 4;
117
    }
118
#else
119
    int w_offset = gx * psc(w);
120

121
    for (int i = 0; i < psc(w); i++)
122
    {
123
        afpvec4 v = buffer_ld4(bottom_blob_data, i);
124

125
#if NCNN_fp16_packed || (NCNN_fp16_storage && !NCNN_fp16_arithmetic)
126
        // GL_EXT_shader_16bit_storage does not define f16mat4 type :(
127
        afpmat4 k = afpmat4(
128
            buffer_ld4(weight_data, (w_offset + i) * 4 + 0),
129
            buffer_ld4(weight_data, (w_offset + i) * 4 + 1),
130
            buffer_ld4(weight_data, (w_offset + i) * 4 + 2),
131
            buffer_ld4(weight_data, (w_offset + i) * 4 + 3)
132
        );
133
#else
134
        afpmat4 k = afpmat4(weight_data[w_offset + i]);
135
#endif
136

137
        sum += v * k;
138
    }
139
#endif
140

141
    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
142

143
#if NCNN_image_shader
144
    image3d_st4(top_blob, ivec3(gx, 0, 0), sum);
145
#else
146
    buffer_st4(top_blob_data, gx, sum);
147
#endif
148
}
149