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
#define FLT_MAX 3.402823466e+38
25

26
layout (constant_id = 0) const int pooling_type = 0;
27

28
#define shape_constant_id_offset 1
29
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
30
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
31
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
32
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
33
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
34

35
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
36
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
37
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
38
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
39
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
40

41
#if NCNN_image_shader
42
layout (binding = 0) uniform unfp sampler3D bottom_blob;
43
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
44
#else
45
layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
46
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
47
#endif
48

49
layout (push_constant) uniform parameter
50
{
51
    int dims;
52
    int w;
53
    int h;
54
    int c;
55
    int cstep;
56

57
    int outdims;
58
    int outw;
59
    int outh;
60
    int outc;
61
    int outcstep;
62
} p;
63

64
void main()
65
{
66
    int gx = int(gl_GlobalInvocationID.x);
67
    int gy = int(gl_GlobalInvocationID.y);
68
    int gz = int(gl_GlobalInvocationID.z);
69

70
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
71
        return;
72

73
    afpvec4 res;
74

75
    // calculate adaptive kernel size
76
    const int sx = psc(w) * gx / psc(outw);
77
    const int ex = (psc(w) * (gx + 1) + psc(outw) - 1) / psc(outw);
78
    const int kernel_w = ex - sx;
79
    const int sy = psc(h) * gy / psc(outh);
80
    const int ey = (psc(h) * (gy + 1) + psc(outh) - 1) / psc(outh);
81
    const int kernel_h = ey - sy;
82

83
    if (pooling_type == 0)
84
    {
85
        res = afpvec4(-FLT_MAX);
86

87
#if NCNN_image_shader
88

89
        for (int y = 0; y < kernel_h; y++)
90
        {
91
            for (int x = 0; x < kernel_w; x++)
92
            {
93
                afpvec4 v = image3d_ld4(bottom_blob, ivec3(sx + x, sy + y, gz));
94
                res = max(res, v);
95
            }
96
        }
97
#else
98
        int v_offset = gz * psc(cstep) + sy * psc(w) + sx;
99

100
        for (int y = 0; y < kernel_h; y++)
101
        {
102
            for (int x = 0; x < kernel_w; x++)
103
            {
104
                afpvec4 v = buffer_ld4(bottom_blob_data, v_offset + x);
105
                res = max(res, v);
106
            }
107

108
            v_offset += psc(w);
109
        }
110
#endif
111
    }
112
    else if (pooling_type == 1)
113
    {
114
        vec4 res_fp32 = vec4(0.f);  // force accumulation in fp32
115

116
#if NCNN_image_shader
117
        for (int y = 0; y < kernel_h; y++)
118
        {
119
            for (int x = 0; x < kernel_w; x++)
120
            {
121
                res_fp32 += image3d_ld4(bottom_blob, ivec3(sx + x, sy + y, gz));
122
            }
123
        }
124
#else
125
        int v_offset = gz * psc(cstep) + sy * psc(w) + sx;
126

127
        for (int y = 0; y < kernel_h; y++)
128
        {
129
            for (int x = 0; x < kernel_w; x++)
130
            {
131
                res_fp32 += buffer_ld4(bottom_blob_data, v_offset + x);
132
            }
133

134
            v_offset += psc(w);
135
        }
136
#endif
137

138
        res_fp32 /= float(kernel_h * kernel_w);
139
        res = afpvec4(res_fp32);  // cast to fp16 if possible
140
    }
141

142
#if NCNN_image_shader
143
    image3d_st4(top_blob, ivec3(gx, gy, gz), res);
144
#else
145
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
146

147
    buffer_st4(top_blob_data, gi, res);
148
#endif
149
}
150