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
#extension GL_GOOGLE_include_directive: enable
26
#include "vulkan_activation.comp"
27

28
layout (constant_id = 0) const int kernel_w = 1;
29
layout (constant_id = 1) const int kernel_h = 1;
30
layout (constant_id = 2) const int dilation_w = 1;
31
layout (constant_id = 3) const int dilation_h = 1;
32
layout (constant_id = 4) const int stride_w = 1;
33
layout (constant_id = 5) const int stride_h = 1;
34
layout (constant_id = 6) const int bias_term = 0;
35
layout (constant_id = 7) const int activation_type = 0;
36
layout (constant_id = 8) const float activation_param_0 = 0;
37
layout (constant_id = 9) const float activation_param_1 = 0;
38

39
#define shape_constant_id_offset 10
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, imfmtc4) writeonly uniform unfp image3D top_blob;
55
layout (binding = 2) uniform unfp sampler3D weight_blob;
56
layout (binding = 3) uniform unfp sampler3D bias_blob;
57
#else
58
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
59
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
60
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
61
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
62
#endif
63

64
layout (push_constant) uniform parameter
65
{
66
    int dims;
67
    int w;
68
    int h;
69
    int c;
70
    int cstep;
71

72
    int outdims;
73
    int outw;
74
    int outh;
75
    int outc;
76
    int outcstep;
77
} p;
78

79
void main()
80
{
81
    int gx = int(gl_GlobalInvocationID.x);
82
    int gy = int(gl_GlobalInvocationID.y);
83
    int gz = int(gl_GlobalInvocationID.z);
84

85
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
86
        return;
87

88
    afpvec4 sum;
89

90
    if (bias_term == 1)
91
    {
92
#if NCNN_image_shader
93
        sum = image3d_ld4(bias_blob, ivec3(gz, 0, 0));
94
#else
95
        sum = buffer_ld4(bias_data, gz);
96
#endif
97
    }
98
    else
99
    {
100
        sum = afpvec4(0.f);
101
    }
102

103
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
104
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
105

106
#if NCNN_image_shader
107
    for (int y = 0; y < kernel_h; y++)
108
    {
109
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
110
        if (sys < 0 || sys % stride_h != 0)
111
            continue;
112

113
        int sy = sys / stride_h;
114
        if (sy >= psc(h))
115
            continue;
116

117
        for (int x = 0; x < kernel_w; x++)
118
        {
119
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
120
            if (sxs < 0 || sxs % stride_w != 0)
121
                continue;
122

123
            int sx = sxs / stride_w;
124
            if (sx >= psc(w))
125
                continue;
126

127
            int wx = (y * kernel_w + x) * 4;
128

129
            for (int z = 0; z < psc(c); z++)
130
            {
131
                afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx, sy, z));
132

133
                afpvec8 k0 = image3d_ld8(weight_blob, ivec3(wx + 0, z, gz));
134
                afpvec8 k1 = image3d_ld8(weight_blob, ivec3(wx + 1, z, gz));
135
                afpvec8 k2 = image3d_ld8(weight_blob, ivec3(wx + 2, z, gz));
136
                afpvec8 k3 = image3d_ld8(weight_blob, ivec3(wx + 3, z, gz));
137

138
                // sum += v * k
139
                sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
140
                sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
141
                sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
142
                sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
143
            }
144
        }
145
    }
146
#else
147
    int w_offset_0 = gz * psc(c) * kernel_w * kernel_h;
148

149
    for (int y = 0; y < kernel_h; y++)
150
    {
151
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
152
        if (sys < 0 || sys % stride_h != 0)
153
            continue;
154

155
        int sy = sys / stride_h;
156
        if (sy >= psc(h))
157
            continue;
158

159
        for (int x = 0; x < kernel_w; x++)
160
        {
161
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
162
            if (sxs < 0 || sxs % stride_w != 0)
163
                continue;
164

165
            int sx = sxs / stride_w;
166
            if (sx >= psc(w))
167
                continue;
168

169
            int v_offset = sy * psc(w) + sx;
170
            int w_offset = w_offset_0 + y * kernel_w + x;
171

172
            for (int z = 0; z < psc(c); z++)
173
            {
174
                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
175

176
                afpvec8 k0 = buffer_ld8(weight_data, w_offset * 4 + 0);
177
                afpvec8 k1 = buffer_ld8(weight_data, w_offset * 4 + 1);
178
                afpvec8 k2 = buffer_ld8(weight_data, w_offset * 4 + 2);
179
                afpvec8 k3 = buffer_ld8(weight_data, w_offset * 4 + 3);
180

181
                // sum += v * k
182
                sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
183
                sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
184
                sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
185
                sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
186

187
                v_offset += psc(cstep);
188
                w_offset += kernel_w * kernel_h;
189
            }
190
        }
191
    }
192
#endif
193

194
    sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
195

196
#if NCNN_image_shader
197
    image3d_st4(top_blob, ivec3(gx, gy, gz), sum);
198
#else
199
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
200

201
    buffer_st4(top_blob_data, gi, sum);
202
#endif
203
}
204