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 group = 1;
36
layout (constant_id = 8) const int activation_type = 0;
37
layout (constant_id = 9) const float activation_param_0 = 0;
38
layout (constant_id = 10) const float activation_param_1 = 0;
39

40
#define shape_constant_id_offset 11
41
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
42
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
43
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
44
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
45
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
46

47
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
48
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
49
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
50
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
51
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
52

53
#if NCNN_image_shader
54
layout (binding = 0) uniform unfp sampler3D bottom_blob;
55
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
56
layout (binding = 2) uniform unfp sampler3D weight_blob;
57
layout (binding = 3) uniform unfp sampler3D bias_blob;
58
#else
59
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
60
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
61
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
62
layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
63
#endif
64

65
layout (push_constant) uniform parameter
66
{
67
    int dims;
68
    int w;
69
    int h;
70
    int c;
71
    int cstep;
72

73
    int outdims;
74
    int outw;
75
    int outh;
76
    int outc;
77
    int outcstep;
78
} p;
79

80
void main()
81
{
82
    int gx = int(gl_GlobalInvocationID.x);
83
    int gy = int(gl_GlobalInvocationID.y);
84
    int gz = int(gl_GlobalInvocationID.z);
85

86
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
87
        return;
88

89
    afpvec8 sum;
90

91
    if (bias_term == 1)
92
    {
93
#if NCNN_image_shader
94
        sum = image3d_ld8(bias_blob, ivec3(gz, 0, 0));
95
#else
96
        sum = buffer_ld8(bias_data, gz);
97
#endif
98
    }
99
    else
100
    {
101
        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
102
    }
103

104
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
105
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
106

107
    // group convolution
108
    const int channels_g = psc(c) / group;
109
    const int num_output_g = psc(outc) / group;
110

111
    // group id
112
    const int gg = gz / num_output_g;
113

114
#if NCNN_image_shader
115
    for (int y = 0; y < kernel_h; y++)
116
    {
117
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
118
        if (sys < 0 || sys % stride_h != 0)
119
            continue;
120

121
        int sy = sys / stride_h;
122
        if (sy >= psc(h))
123
            continue;
124

125
        for (int x = 0; x < kernel_w; x++)
126
        {
127
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
128
            if (sxs < 0 || sxs % stride_w != 0)
129
                continue;
130

131
            int sx = sxs / stride_w;
132
            if (sx >= psc(w))
133
                continue;
134

135
            int sz = gg * channels_g;
136
            int wx = (y * kernel_w + x) * 8;
137

138
            for (int z = 0; z < channels_g; z++)
139
            {
140
                afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx, sy, sz));
141

142
                afpvec8 k0 = image3d_ld8(weight_blob, ivec3(wx + 0, z, gz));
143
                afpvec8 k1 = image3d_ld8(weight_blob, ivec3(wx + 1, z, gz));
144
                afpvec8 k2 = image3d_ld8(weight_blob, ivec3(wx + 2, z, gz));
145
                afpvec8 k3 = image3d_ld8(weight_blob, ivec3(wx + 3, z, gz));
146
                afpvec8 k4 = image3d_ld8(weight_blob, ivec3(wx + 4, z, gz));
147
                afpvec8 k5 = image3d_ld8(weight_blob, ivec3(wx + 5, z, gz));
148
                afpvec8 k6 = image3d_ld8(weight_blob, ivec3(wx + 6, z, gz));
149
                afpvec8 k7 = image3d_ld8(weight_blob, ivec3(wx + 7, z, gz));
150

151
                // sum += v * k
152
                sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
153
                sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
154
                sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
155
                sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
156
                sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
157
                sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
158
                sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
159
                sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
160

161
                sz += 1;
162
            }
163
        }
164
    }
165
#else
166
    int w_offset_0 = gz * channels_g * kernel_w * kernel_h;
167
    int v_offset_0 = gg * channels_g * psc(cstep);
168

169
    for (int y = 0; y < kernel_h; y++)
170
    {
171
        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
172
        if (sys < 0 || sys % stride_h != 0)
173
            continue;
174

175
        int sy = sys / stride_h;
176
        if (sy >= psc(h))
177
            continue;
178

179
        for (int x = 0; x < kernel_w; x++)
180
        {
181
            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
182
            if (sxs < 0 || sxs % stride_w != 0)
183
                continue;
184

185
            int sx = sxs / stride_w;
186
            if (sx >= psc(w))
187
                continue;
188

189
            int v_offset = v_offset_0 + sy * psc(w) + sx;
190
            int w_offset = w_offset_0 + y * kernel_w + x;
191

192
            for (int z = 0; z < channels_g; z++)
193
            {
194
                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
195

196
                afpvec8 k0 = buffer_ld8(weight_data, w_offset * 8 + 0);
197
                afpvec8 k1 = buffer_ld8(weight_data, w_offset * 8 + 1);
198
                afpvec8 k2 = buffer_ld8(weight_data, w_offset * 8 + 2);
199
                afpvec8 k3 = buffer_ld8(weight_data, w_offset * 8 + 3);
200
                afpvec8 k4 = buffer_ld8(weight_data, w_offset * 8 + 4);
201
                afpvec8 k5 = buffer_ld8(weight_data, w_offset * 8 + 5);
202
                afpvec8 k6 = buffer_ld8(weight_data, w_offset * 8 + 6);
203
                afpvec8 k7 = buffer_ld8(weight_data, w_offset * 8 + 7);
204

205
                // sum += v * k
206
                sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
207
                sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
208
                sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
209
                sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
210
                sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
211
                sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
212
                sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
213
                sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
214

215
                v_offset += psc(cstep);
216
                w_offset += kernel_w * kernel_h;
217
            }
218
        }
219
    }
220
#endif
221

222
    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
223

224
#if NCNN_image_shader
225
    image3d_st8(top_blob, ivec3(gx, gy, gz), sum);
226
#else
227
    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
228

229
    buffer_st8(top_blob_data, gi, sum);
230
#endif
231
}
232