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
#if NCNN_fp16_packed || (NCNN_fp16_storage && !NCNN_fp16_arithmetic)
61
// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
62
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
63
#else
64
layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
65
#endif
66
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
67
#endif
68

69
layout (push_constant) uniform parameter
70
{
71
    int dims;
72
    int w;
73
    int h;
74
    int c;
75
    int cstep;
76

77
    int outdims;
78
    int outw;
79
    int outh;
80
    int outc;
81
    int outcstep;
82
} p;
83

84
void main()
85
{
86
    int gx = int(gl_GlobalInvocationID.x);
87
    int gy = int(gl_GlobalInvocationID.y);
88
    int gz = int(gl_GlobalInvocationID.z);
89

90
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
91
        return;
92

93
    afpvec4 sum;
94

95
    if (bias_term == 1)
96
    {
97
#if NCNN_image_shader
98
        sum = image3d_ld4(bias_blob, ivec3(gz, 0, 0));
99
#else
100
        sum = buffer_ld4(bias_data, gz);
101
#endif
102
    }
103
    else
104
    {
105
        sum = afpvec4(0.f);
106
    }
107

108
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
109
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
110

111
    // group convolution
112
    const int channels_g = psc(c) / group;
113
    const int num_output_g = psc(outc) / group;
114

115
    // group id
116
    const int gg = gz / num_output_g;
117

118
#if NCNN_image_shader
119
    for (int y = 0; y < kernel_h; y++)
120
    {
121
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
122
        if (sys < 0 || sys % stride_h != 0)
123
            continue;
124

125
        int sy = sys / stride_h;
126
        if (sy >= psc(h))
127
            continue;
128

129
        for (int x = 0; x < kernel_w; x++)
130
        {
131
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
132
            if (sxs < 0 || sxs % stride_w != 0)
133
                continue;
134

135
            int sx = sxs / stride_w;
136
            if (sx >= psc(w))
137
                continue;
138

139
            int sz = gg * channels_g;
140
            int wx = (y * kernel_w + x) * 4;
141

142
            for (int z = 0; z < channels_g; z++)
143
            {
144
                afpvec4 v = image3d_ld4(bottom_blob, ivec3(sx, sy, sz));
145

146
                afpmat4 k = afpmat4(
147
                    image3d_ld4(weight_blob, ivec3(wx + 0, z, gz)),
148
                    image3d_ld4(weight_blob, ivec3(wx + 1, z, gz)),
149
                    image3d_ld4(weight_blob, ivec3(wx + 2, z, gz)),
150
                    image3d_ld4(weight_blob, ivec3(wx + 3, z, gz))
151
                );
152

153
                sum += v * k;
154

155
                sz += 1;
156
            }
157
        }
158
    }
159
#else
160
    int w_offset_0 = gz * channels_g * kernel_w * kernel_h;
161
    int v_offset_0 = gg * channels_g * psc(cstep);
162

163
    for (int y = 0; y < kernel_h; y++)
164
    {
165
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
166
        if (sys < 0 || sys % stride_h != 0)
167
            continue;
168

169
        int sy = sys / stride_h;
170
        if (sy >= psc(h))
171
            continue;
172

173
        for (int x = 0; x < kernel_w; x++)
174
        {
175
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
176
            if (sxs < 0 || sxs % stride_w != 0)
177
                continue;
178

179
            int sx = sxs / stride_w;
180
            if (sx >= psc(w))
181
                continue;
182

183
            int v_offset = v_offset_0 + sy * psc(w) + sx;
184
            int w_offset = w_offset_0 + y * kernel_w + x;
185

186
            for (int z = 0; z < channels_g; z++)
187
            {
188
                afpvec4 v = buffer_ld4(bottom_blob_data, v_offset);
189

190
#if NCNN_fp16_packed || (NCNN_fp16_storage && !NCNN_fp16_arithmetic)
191
                // GL_EXT_shader_16bit_storage does not define f16mat4 type :(
192
                afpmat4 k = afpmat4(
193
                    buffer_ld4(weight_data, w_offset * 4 + 0),
194
                    buffer_ld4(weight_data, w_offset * 4 + 1),
195
                    buffer_ld4(weight_data, w_offset * 4 + 2),
196
                    buffer_ld4(weight_data, w_offset * 4 + 3)
197
                );
198
#else
199
                afpmat4 k = afpmat4(weight_data[w_offset]);
200
#endif
201

202
                sum += v * k;
203

204
                v_offset += psc(cstep);
205
                w_offset += kernel_w * kernel_h;
206
            }
207
        }
208
    }
209
#endif
210

211
    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
212

213
#if NCNN_image_shader
214
    image3d_st4(top_blob, ivec3(gx, gy, gz), sum);
215
#else
216
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
217

218
    buffer_st4(top_blob_data, gi, sum);
219
#endif
220
}
221