ncnn

deconvolution_pack4to1.comp
188 строк · 5.7 Кб
Перенос по словам
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 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 dims = 0;
40
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
41
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
42
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
43
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
44

45
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
46
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
47
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
48
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
49
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
50

51
#if NCNN_image_shader
52
layout (binding = 0) uniform unfp sampler3D bottom_blob;
53
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
54
layout (binding = 2) uniform unfp sampler3D weight_blob;
55
layout (binding = 3) uniform unfp sampler3D bias_blob;
56
#else
57
layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
58
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
59
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
60
layout (binding = 3) readonly buffer bias_blob { sfp bias_data[]; };
61
#endif
62

63
layout (push_constant) uniform parameter
64
{
65
    int dims;
66
    int w;
67
    int h;
68
    int c;
69
    int cstep;
70

71
    int outdims;
72
    int outw;
73
    int outh;
74
    int outc;
75
    int outcstep;
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 sum;
88

89
    if (bias_term == 1)
90
    {
91
#if NCNN_image_shader
92
        sum = image3d_ld1(bias_blob, ivec3(gz, 0, 0));
93
#else
94
        sum = buffer_ld1(bias_data, gz);
95
#endif
96
    }
97
    else
98
    {
99
        sum = afp(0.f);
100
    }
101

102
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
103
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
104

105
#if NCNN_image_shader
106
    for (int y = 0; y < kernel_h; y++)
107
    {
108
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
109
        if (sys < 0 || sys % stride_h != 0)
110
            continue;
111

112
        int sy = sys / stride_h;
113
        if (sy >= psc(h))
114
            continue;
115

116
        for (int x = 0; x < kernel_w; x++)
117
        {
118
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
119
            if (sxs < 0 || sxs % stride_w != 0)
120
                continue;
121

122
            int sx = sxs / stride_w;
123
            if (sx >= psc(w))
124
                continue;
125

126
            int wx = y * kernel_w + x;
127

128
            for (int z = 0; z < psc(c); z++)
129
            {
130
                afpvec4 v = image3d_ld4(bottom_blob, ivec3(sx, sy, z));
131

132
                afpvec4 k = image3d_ld4(weight_blob, ivec3(wx, z, gz));
133

134
                sum += dot(v, k);
135
            }
136
        }
137
    }
138
#else
139
    int w_offset_0 = gz * psc(c) * kernel_w * kernel_h;
140

141
    for (int y = 0; y < kernel_h; y++)
142
    {
143
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
144
        if (sys < 0 || sys % stride_h != 0)
145
            continue;
146

147
        int sy = sys / stride_h;
148
        if (sy >= psc(h))
149
            continue;
150

151
        for (int x = 0; x < kernel_w; x++)
152
        {
153
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
154
            if (sxs < 0 || sxs % stride_w != 0)
155
                continue;
156

157
            int sx = sxs / stride_w;
158
            if (sx >= psc(w))
159
                continue;
160

161
            int v_offset = sy * psc(w) + sx;
162
            int w_offset = w_offset_0 + y * kernel_w + x;
163

164
            for (int z = 0; z < psc(c); z++)
165
            {
166
                afpvec4 v = buffer_ld4(bottom_blob_data, v_offset);
167

168
                afpvec4 k = buffer_ld4(weight_data, w_offset);
169

170
                sum += dot(v, k);
171

172
                v_offset += psc(cstep);
173
                w_offset += kernel_w * kernel_h;
174
            }
175
        }
176
    }
177
#endif
178

179
    sum = activation_afp(sum, activation_type, activation_param_0, activation_param_1);
180

181
#if NCNN_image_shader
182
    image3d_st1(top_blob, ivec3(gx, gy, gz), sum);
183
#else
184
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
185

186
    buffer_st1(top_blob_data, gi, sum);
187
#endif
188
}
189
ncnn

Использование cookies
