ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2018 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
#define FLT_MAX 3.402823466e+38
25

26
layout (constant_id = 0) const int pooling_type = 0;
27
layout (constant_id = 1) const int kernel_w = 1;
28
layout (constant_id = 2) const int kernel_h = 1;
29
layout (constant_id = 3) const int stride_w = 1;
30
layout (constant_id = 4) const int stride_h = 1;
31
layout (constant_id = 5) const int pad_left = 0;
32
layout (constant_id = 6) const int pad_right = 0;
33
layout (constant_id = 7) const int pad_top = 0;
34
layout (constant_id = 8) const int pad_bottom = 0;
35
layout (constant_id = 9) const int global_pooling = 0;
36
layout (constant_id = 10) const int pad_mode = 0;
37
layout (constant_id = 11) const int avgpool_count_include_pad = 0;
38

39
#define shape_constant_id_offset 12
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, imfmtc1) writeonly uniform unfp image3D top_blob;
55
#else
56
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
57
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
58
#endif
59

60
layout (push_constant) uniform parameter
61
{
62
    int dims;
63
    int w;
64
    int h;
65
    int c;
66
    int cstep;
67

68
    int outdims;
69
    int outw;
70
    int outh;
71
    int outc;
72
    int outcstep;
73

74
    int wtailpad;
75
    int htailpad;
76
} p;
77

78
void main()
79
{
80
    int gx = int(gl_GlobalInvocationID.x);
81
    int gy = int(gl_GlobalInvocationID.y);
82
    int gz = int(gl_GlobalInvocationID.z);
83

84
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
85
        return;
86

87
    afp res;
88

89
    if (pooling_type == 0)
90
    {
91
        res = afp(-FLT_MAX);
92

93
#if NCNN_image_shader
94
        int sx = gx * stride_w;
95
        int sy = gy * stride_h;
96

97
        for (int y = 0; y < kernel_h; y++)
98
        {
99
            for (int x = 0; x < kernel_w; x++)
100
            {
101
                afp v = image3d_ld1(bottom_blob, ivec3(sx + x, sy + y, gz));
102
                res = max(res, v);
103
            }
104
        }
105
#else
106
        int v_offset = gz * psc(cstep) + gy * stride_h * psc(w) + gx * stride_w;
107

108
        for (int y = 0; y < kernel_h; y++)
109
        {
110
            for (int x = 0; x < kernel_w; x++)
111
            {
112
                afp v = buffer_ld1(bottom_blob_data, v_offset + x);
113
                res = max(res, v);
114
            }
115

116
            v_offset += psc(w);
117
        }
118
#endif
119
    }
120
    if (pooling_type == 1 && avgpool_count_include_pad == 0)
121
    {
122
        res = afp(0.f);
123
        int area = 0;
124

125
        int sx = gx * stride_w;
126
        int sy = gy * stride_h;
127

128
#if NCNN_image_shader
129
        for (int y = 0; y < kernel_h; y++)
130
        {
131
            if (sy + y < pad_top)
132
                continue;
133

134
            if (sy + y >= psc(h) - pad_bottom - p.htailpad)
135
                break;
136

137
            for (int x = 0; x < kernel_w; x++)
138
            {
139
                if (sx + x < pad_left)
140
                    continue;
141

142
                if (sx + x >= psc(w) - pad_right - p.wtailpad)
143
                    break;
144

145
                res += image3d_ld1(bottom_blob, ivec3(sx + x, sy + y, gz));
146
                area += 1;
147
            }
148
        }
149
#else
150
        int v_offset = gz * psc(cstep) + sy * psc(w) + sx;
151

152
        for (int y = 0; y < kernel_h; y++)
153
        {
154
            if (sy + y < pad_top)
155
            {
156
                v_offset += psc(w);
157
                continue;
158
            }
159

160
            if (sy + y >= psc(h) - pad_bottom - p.htailpad)
161
                break;
162

163
            for (int x = 0; x < kernel_w; x++)
164
            {
165
                if (sx + x < pad_left)
166
                {
167
                    continue;
168
                }
169

170
                if (sx + x >= psc(w) - pad_right - p.wtailpad)
171
                    break;
172

173
                res += buffer_ld1(bottom_blob_data, v_offset + x);
174
                area += 1;
175
            }
176

177
            v_offset += psc(w);
178
        }
179
#endif
180

181
        res /= afp(area);
182
    }
183
    if (pooling_type == 1 && avgpool_count_include_pad == 1)
184
    {
185
        res = afp(0.f);
186

187
#if NCNN_image_shader
188
        int sx = gx * stride_w;
189
        int sy = gy * stride_h;
190

191
        for (int y = 0; y < kernel_h; y++)
192
        {
193
            for (int x = 0; x < kernel_w; x++)
194
            {
195
                res += image3d_ld1(bottom_blob, ivec3(sx + x, sy + y, gz));
196
            }
197
        }
198
#else
199
        int v_offset = gz * psc(cstep) + gy * stride_h * psc(w) + gx * stride_w;
200

201
        for (int y = 0; y < kernel_h; y++)
202
        {
203
            for (int x = 0; x < kernel_w; x++)
204
            {
205
                res += buffer_ld1(bottom_blob_data, v_offset + x);
206
            }
207

208
            v_offset += psc(w);
209
        }
210
#endif
211

212
        res /= afp(kernel_w * kernel_h);
213
    }
214

215
#if NCNN_image_shader
216
    image3d_st1(top_blob, ivec3(gx, gy, gz), res);
217
#else
218
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
219

220
    buffer_st1(top_blob_data, gi, res);
221
#endif
222
}
223