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 kernel_w = 1;
28
layout (constant_id = 1) const int kernel_h = 1;
29
layout (constant_id = 2) const int dilation_w = 1;
30
layout (constant_id = 3) const int dilation_h = 1;
31
layout (constant_id = 4) const int stride_w = 1;
32
layout (constant_id = 5) const int stride_h = 1;
33
layout (constant_id = 6) const int bias_term = 0;
34
layout (constant_id = 7) const int group = 1;
35
layout (constant_id = 8) const int activation_type = 0;
36
layout (constant_id = 9) const float activation_param_0 = 0;
37
layout (constant_id = 10) const float activation_param_1 = 0;
38

39
#define shape_constant_id_offset 11
40
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
41
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
42
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
43
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
44
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
45

46
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
47
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
48
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
49
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
50
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
51

52
#if NCNN_image_shader
53
layout (binding = 0) uniform unfp sampler3D bottom_blob;
54
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
55
layout (binding = 2) uniform unfp sampler3D weight_blob;
56
layout (binding = 3) uniform unfp sampler3D bias_blob;
57
#else
58
layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
59
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
60
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
61
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
62
#endif
63

64
layout (push_constant) uniform parameter
65
{
66
    int dims;
67
    int w;
68
    int h;
69
    int c;
70
    int cstep;
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
    afpvec4 sum;
89

90
    if (bias_term == 1)
91
    {
92
#if NCNN_image_shader
93
        sum = image3d_ld4(bias_blob, ivec3(gz, 0, 0));
94
#else
95
        sum = buffer_ld4(bias_data, gz);
96
#endif
97
    }
98
    else
99
    {
100
        sum = afpvec4(0.f);
101
    }
102

103
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
104
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
105

106
    // depth-wise deconvolution
107
#if NCNN_image_shader
108
    for (int y = 0; y < kernel_h; y++)
109
    {
110
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
111
        if (sys < 0 || sys % stride_h != 0)
112
            continue;
113

114
        int sy = sys / stride_h;
115
        if (sy >= psc(h))
116
            continue;
117

118
        for (int x = 0; x < kernel_w; x++)
119
        {
120
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
121
            if (sxs < 0 || sxs % stride_w != 0)
122
                continue;
123

124
            int sx = sxs / stride_w;
125
            if (sx >= psc(w))
126
                continue;
127

128
            int wx = y * kernel_w + x;
129

130
            afpvec4 v = image3d_ld4(bottom_blob, ivec3(sx, sy, gz));
131

132
            afpvec4 k = image3d_ld4(weight_blob, ivec3(wx, gz, 0));
133

134
            sum += v * k;
135
        }
136
    }
137
#else
138
    int v_offset_0 = gz * psc(cstep);
139
    int w_offset_0 = gz * kernel_w * kernel_h;
140

141
    for (int y = 0; y < kernel_h; y++)
142
    {
143
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
144
        if (sys < 0 || sys % stride_h != 0)
145
            continue;
146

147
        int sy = sys / stride_h;
148
        if (sy >= psc(h))
149
            continue;
150

151
        for (int x = 0; x < kernel_w; x++)
152
        {
153
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
154
            if (sxs < 0 || sxs % stride_w != 0)
155
                continue;
156

157
            int sx = sxs / stride_w;
158
            if (sx >= psc(w))
159
                continue;
160

161
            int v_offset = v_offset_0 + sy * psc(w) + sx;
162
            int w_offset = w_offset_0 + y * kernel_w + x;
163

164
            afpvec4 v = buffer_ld4(bottom_blob_data, v_offset);
165

166
            afpvec4 k = buffer_ld4(weight_data, w_offset);
167

168
            sum += v * k;
169
        }
170
    }
171
#endif
172

173
    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
174

175
#if NCNN_image_shader
176
    image3d_st4(top_blob, ivec3(gx, gy, gz), sum);
177
#else
178
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
179

180
    buffer_st4(top_blob_data, gi, sum);
181
#endif
182
}
183