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
#define FLT_MAX 3.402823466e+38
26

27
layout (constant_id = 0) const int pooling_type = 0;
28

29
#define shape_constant_id_offset 1
30
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
31
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
32
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
33
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
34
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
35

36
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
37
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
38
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
39
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
40
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
41

42
#if NCNN_image_shader
43
layout (binding = 0) uniform unfp sampler3D bottom_blob;
44
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
45
#else
46
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
47
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
48
#endif
49

50
layout (push_constant) uniform parameter
51
{
52
    int dims;
53
    int w;
54
    int h;
55
    int c;
56
    int cstep;
57

58
    int outdims;
59
    int outw;
60
    int outh;
61
    int outc;
62
    int outcstep;
63
} p;
64

65
void main()
66
{
67
    int gx = int(gl_GlobalInvocationID.x);
68
    int gy = int(gl_GlobalInvocationID.y);
69
    int gz = int(gl_GlobalInvocationID.z);
70

71
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
72
        return;
73

74
    afpvec8 res;
75

76
    // calculate adaptive kernel size
77
    const int sx = psc(w) * gx / psc(outw);
78
    const int ex = (psc(w) * (gx + 1) + psc(outw) - 1) / psc(outw);
79
    const int kernel_w = ex - sx;
80
    const int sy = psc(h) * gy / psc(outh);
81
    const int ey = (psc(h) * (gy + 1) + psc(outh) - 1) / psc(outh);
82
    const int kernel_h = ey - sy;
83

84
    if (pooling_type == 0)
85
    {
86
        res = afpvec8(afpvec4(-FLT_MAX), afpvec4(-FLT_MAX));
87

88
#if NCNN_image_shader
89

90
        for (int y = 0; y < kernel_h; y++)
91
        {
92
            for (int x = 0; x < kernel_w; x++)
93
            {
94
                afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx + x, sy + y, gz));
95
                res[0] = max(res[0], v[0]);
96
                res[1] = max(res[1], v[1]);
97
            }
98
        }
99
#else
100
        int v_offset = gz * psc(cstep) + sy * psc(w) + sx;
101

102
        for (int y = 0; y < kernel_h; y++)
103
        {
104
            for (int x = 0; x < kernel_w; x++)
105
            {
106
                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
107
                res[0] = max(res[0], v[0]);
108
                res[1] = max(res[1], v[1]);
109
            }
110

111
            v_offset += psc(w);
112
        }
113
#endif
114
    }
115
    else if (pooling_type == 1)
116
    {
117
        mat2x4 res_fp32 = mat2x4(vec4(0.f), vec4(0.f));  // force accumulation in fp32
118
        
119
#if NCNN_image_shader
120
        for (int y = 0; y < kernel_h; y++)
121
        {
122
            for (int x = 0; x < kernel_w; x++)
123
            {
124
                afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx + x, sy + y, gz));
125
                res_fp32[0] += v[0];
126
                res_fp32[1] += v[1];
127
            }
128
        }
129
#else
130
        int v_offset = gz * psc(cstep) + sy * psc(w) + sx;
131

132
        for (int y = 0; y < kernel_h; y++)
133
        {
134
            for (int x = 0; x < kernel_w; x++)
135
            {
136
                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
137
                res_fp32[0] += v[0];
138
                res_fp32[1] += v[1];
139
            }
140

141
            v_offset += psc(w);
142
        }
143
#endif
144

145
        int area = kernel_h * kernel_w;
146
        res_fp32[0] /= float(area);
147
        res_fp32[1] /= float(area);
148
        res = afpvec8(res_fp32);  // cast to fp16 if possible
149
    }
150

151
#if NCNN_image_shader
152
    image3d_st8(top_blob, ivec3(gx, gy, gz), res);
153
#else
154
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
155

156
    buffer_st8(top_blob_data, gi, res);
157
#endif
158
}
159