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
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 w = 0;
41
layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
42
layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
43
layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
44

45
layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
46
layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
47
layout (constant_id = shape_constant_id_offset + 6) const int outc = 0;
48
layout (constant_id = shape_constant_id_offset + 7) const int outcstep = 0;
49

50
#if NCNN_image_shader
51
layout (binding = 0) uniform unfp sampler3D bottom_blob;
52
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
53
layout (binding = 2) uniform unfp sampler3D weight_blob;
54
layout (binding = 3) uniform unfp sampler3D bias_blob;
55
#else
56
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
57
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
58
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
59
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
60
#endif
61

62
layout (push_constant) uniform parameter
63
{
64
    int w;
65
    int h;
66
    int c;
67
    int cstep;
68

69
    int outw;
70
    int outh;
71
    int outc;
72
    int outcstep;
73
} p;
74

75
void main()
76
{
77
    int gx = int(gl_GlobalInvocationID.x) * 4;
78
    int gy = int(gl_GlobalInvocationID.y);
79

80
    const int outsize = psc(outw) * psc(outh);
81

82
    if (gx >= outsize || gy >= psc(outc))
83
        return;
84

85
    afpvec4 sum0;
86
    afpvec4 sum1;
87
    afpvec4 sum2;
88
    afpvec4 sum3;
89

90
    if (bias_term == 1)
91
    {
92
#if NCNN_image_shader
93
        sum0 = image3d_ld4(bias_blob, ivec3(gy, 0, 0));
94
#else
95
        sum0 = buffer_ld4(bias_data, gy);
96
#endif
97
        sum1 = sum0;
98
        sum2 = sum0;
99
        sum3 = sum0;
100
    }
101
    else
102
    {
103
        sum0 = afpvec4(0.f);
104
        sum1 = afpvec4(0.f);
105
        sum2 = afpvec4(0.f);
106
        sum3 = afpvec4(0.f);
107
    }
108

109
    const int maxk = kernel_w * kernel_h;
110
    const int N = psc(c) * maxk;
111

112
    const ivec4 gx4 = gx + ivec4(0, 1, 2, 3);
113

114
    const ivec4 sy4 = gx4 / psc(outw);
115
    const ivec4 sx4 = gx4 % psc(outw);
116

117
    const ivec4 sxs4 = sx4 * stride_w;
118
    const ivec4 sys4 = sy4 * stride_h;
119

120
#if NCNN_image_shader
121
    for (int z = 0; z < N; z++)
122
    {
123
        const int sz = z / maxk;
124
        const int kk = z % maxk;
125

126
        const int ky = kk / kernel_w;
127
        const int kx = kk % kernel_w;
128

129
        const ivec4 x4 = sxs4 + kx * dilation_w;
130
        const ivec4 y4 = sys4 + ky * dilation_h;
131

132
        afpvec8 v0 = image3d_ld8(bottom_blob, ivec3(x4.r, y4.r, sz));
133
        afpvec8 v1 = image3d_ld8(bottom_blob, ivec3(x4.g, y4.g, sz));
134
        afpvec8 v2 = image3d_ld8(bottom_blob, ivec3(x4.b, y4.b, sz));
135
        afpvec8 v3 = image3d_ld8(bottom_blob, ivec3(x4.a, y4.a, sz));
136

137
        afpvec8 k0 = image3d_ld8(weight_blob, ivec3(z * 4 + 0, gy, 0));
138
        afpvec8 k1 = image3d_ld8(weight_blob, ivec3(z * 4 + 1, gy, 0));
139
        afpvec8 k2 = image3d_ld8(weight_blob, ivec3(z * 4 + 2, gy, 0));
140
        afpvec8 k3 = image3d_ld8(weight_blob, ivec3(z * 4 + 3, gy, 0));
141

142
        // sum += v * k;
143
        sum0.r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
144
        sum0.g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
145
        sum0.b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);
146
        sum0.a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);
147

148
        sum1.r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
149
        sum1.g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
150
        sum1.b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);
151
        sum1.a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);
152

153
        sum2.r += dot(v2[0], k0[0]) + dot(v2[1], k0[1]);
154
        sum2.g += dot(v2[0], k1[0]) + dot(v2[1], k1[1]);
155
        sum2.b += dot(v2[0], k2[0]) + dot(v2[1], k2[1]);
156
        sum2.a += dot(v2[0], k3[0]) + dot(v2[1], k3[1]);
157

158
        sum3.r += dot(v3[0], k0[0]) + dot(v3[1], k0[1]);
159
        sum3.g += dot(v3[0], k1[0]) + dot(v3[1], k1[1]);
160
        sum3.b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);
161
        sum3.a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);
162
    }
163
#else
164
    int w_offset = gy * N * 4;
165

166
    for (int z = 0; z < N; z++)
167
    {
168
        const int sz = z / maxk;
169
        const int kk = z % maxk;
170

171
        const int ky = kk / kernel_w;
172
        const int kx = kk % kernel_w;
173

174
        const ivec4 v_offset = sz * psc(cstep) + (sys4 + ky * dilation_h) * psc(w) + sxs4 + kx * dilation_w;
175

176
        afpvec8 v0 = buffer_ld8(bottom_blob_data, v_offset.r);
177
        afpvec8 v1 = buffer_ld8(bottom_blob_data, v_offset.g);
178
        afpvec8 v2 = buffer_ld8(bottom_blob_data, v_offset.b);
179
        afpvec8 v3 = buffer_ld8(bottom_blob_data, v_offset.a);
180

181
        afpvec8 k0 = buffer_ld8(weight_data, w_offset + 0);
182
        afpvec8 k1 = buffer_ld8(weight_data, w_offset + 1);
183
        afpvec8 k2 = buffer_ld8(weight_data, w_offset + 2);
184
        afpvec8 k3 = buffer_ld8(weight_data, w_offset + 3);
185

186
        // sum += v * k;
187
        sum0.r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
188
        sum0.g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
189
        sum0.b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);
190
        sum0.a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);
191

192
        sum1.r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
193
        sum1.g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
194
        sum1.b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);
195
        sum1.a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);
196

197
        sum2.r += dot(v2[0], k0[0]) + dot(v2[1], k0[1]);
198
        sum2.g += dot(v2[0], k1[0]) + dot(v2[1], k1[1]);
199
        sum2.b += dot(v2[0], k2[0]) + dot(v2[1], k2[1]);
200
        sum2.a += dot(v2[0], k3[0]) + dot(v2[1], k3[1]);
201

202
        sum3.r += dot(v3[0], k0[0]) + dot(v3[1], k0[1]);
203
        sum3.g += dot(v3[0], k1[0]) + dot(v3[1], k1[1]);
204
        sum3.b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);
205
        sum3.a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);
206

207
        w_offset += 4;
208
    }
209
#endif
210

211
    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
212
    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
213
    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
214
    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
215

216
#if NCNN_image_shader
217
    image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);
218
    image3d_st4(top_blob, ivec3(sx4.g, sy4.g, gy), sum1);
219
    image3d_st4(top_blob, ivec3(sx4.b, sy4.b, gy), sum2);
220
    image3d_st4(top_blob, ivec3(sx4.a, sy4.a, gy), sum3);
221
#else
222
    const int gi = gy * psc(outcstep) + gx;
223

224
    buffer_st4(top_blob_data, gi, sum0);
225
    if (gx + 1 < outsize) buffer_st4(top_blob_data, gi + 1, sum1);
226
    if (gx + 2 < outsize) buffer_st4(top_blob_data, gi + 2, sum2);
227
    if (gx + 3 < outsize) buffer_st4(top_blob_data, gi + 3, sum3);
228
#endif
229
}
230