ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2022 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 activation_type = 0;
35
layout (constant_id = 8) const float activation_param_0 = 0;
36
layout (constant_id = 9) const float activation_param_1 = 0;
37

38
#define shape_constant_id_offset 10
39
layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
40
layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
41

42
layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
43
layout (constant_id = shape_constant_id_offset + 3) const int outh = 0;
44
layout (constant_id = shape_constant_id_offset + 4) const int outc = 0;
45
layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
46

47
#if NCNN_image_shader
48
layout (binding = 0) uniform unfp sampler3D col_blob;
49
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
50
layout (binding = 2) uniform unfp sampler3D bias_blob;
51
#else
52
layout (binding = 0) readonly buffer col_blob { sfp col_blob_data[]; };
53
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
54
layout (binding = 2) readonly buffer bias_blob { sfp bias_data[]; };
55
#endif
56

57
layout (push_constant) uniform parameter
58
{
59
    int w;
60
    int h;
61

62
    int outw;
63
    int outh;
64
    int outc;
65
    int outcstep;
66
} p;
67

68
void main()
69
{
70
    int gx = int(gl_GlobalInvocationID.x);
71
    int gy = int(gl_GlobalInvocationID.y);
72
    int gz = int(gl_GlobalInvocationID.z);
73

74
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
75
        return;
76

77
    afp sum;
78

79
    if (bias_term == 1)
80
    {
81
#if NCNN_image_shader
82
        sum = image3d_ld1(bias_blob, ivec3(gz, 0, 0));
83
#else
84
        sum = buffer_ld1(bias_data, gz);
85
#endif
86
    }
87
    else
88
    {
89
        sum = afp(0.f);
90
    }
91

92
    const int maxk = kernel_w * kernel_h;
93

94
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
95
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
96

97
    const int sx_start = (gx < kernel_extent_w) ? 0 : (gx - kernel_extent_w) / stride_w + 1;
98
    const int sx_end = min(gx / stride_w + 1, psc(w));
99

100
    const int sy_start = (gy < kernel_extent_h) ? 0 : (gy - kernel_extent_h) / stride_h + 1;
101
    const int sy_end = min(gy / stride_h + 1, psc(h));
102

103
    for (int sy = sy_start; sy < sy_end; sy += 1)
104
    {
105
        for (int sx = sx_start; sx < sx_end; sx += 1)
106
        {
107
            int h_k = (gy - sy * stride_h);
108
            int w_k = (gx - sx * stride_w);
109

110
            if (h_k % dilation_h == 0 && w_k % dilation_w == 0)
111
            {
112
                h_k /= dilation_h;
113
                w_k /= dilation_w;
114

115
#if NCNN_image_shader
116
                sum += image3d_ld1(col_blob, ivec3(sy * psc(w) + sx, gz * maxk + h_k * kernel_w + w_k, 0));
117
#else
118
                const int gi = (gz * maxk + h_k * kernel_w + w_k) * psc(w) * psc(h) + sy * psc(w) + sx;
119

120
                sum += buffer_ld1(col_blob_data, gi);
121
#endif
122
            }
123
        }
124
    }
125

126
    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
127

128
#if NCNN_image_shader
129
    image3d_st1(top_blob, ivec3(gx, gy, gz), sum);
130
#else
131
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
132

133
    buffer_st1(top_blob_data, gi, sum);
134
#endif
135
}
136