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 bias_term = 0;
29
layout (constant_id = 1) const int activation_type = 0;
30
layout (constant_id = 2) const float activation_param_0 = 0;
31
layout (constant_id = 3) const float activation_param_1 = 0;
32

33
#define shape_constant_id_offset 4
34
layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
35
layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
36
layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
37
layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
38

39
layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
40
layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
41
layout (constant_id = shape_constant_id_offset + 6) const int outc = 0;
42
layout (constant_id = shape_constant_id_offset + 7) const int outcstep = 0;
43

44
#if NCNN_image_shader
45
layout (binding = 0) uniform unfp sampler3D bottom_blob;
46
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
47
layout (binding = 2) uniform unfp sampler3D weight_blob;
48
layout (binding = 3) uniform unfp sampler3D bias_blob;
49
#else
50
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
51
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
52
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
53
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
54
#endif
55

56
layout (push_constant) uniform parameter
57
{
58
    int w;
59
    int h;
60
    int c;
61
    int cstep;
62

63
    int outw;
64
    int outh;
65
    int outc;
66
    int outcstep;
67
} p;
68

69
void main()
70
{
71
    int gx = int(gl_GlobalInvocationID.x) * 4;
72
    int gy = int(gl_GlobalInvocationID.y);
73

74
#if NCNN_image_shader
75
    if (gx >= psc(outw) * psc(outh) || gy >= psc(outc))
76
        return;
77
#else
78
    if (gx >= psc(outcstep) || gy >= psc(outc))
79
        return;
80
#endif
81

82
    afpvec4 sum0;
83
    afpvec4 sum1;
84
    afpvec4 sum2;
85
    afpvec4 sum3;
86

87
    if (bias_term == 1)
88
    {
89
#if NCNN_image_shader
90
        afpvec4 b = image3d_ld4(bias_blob, ivec3(gy, 0, 0));
91
#else
92
        afpvec4 b = buffer_ld4(bias_data, gy);
93
#endif
94
        sum0 = b;
95
        sum1 = b;
96
        sum2 = b;
97
        sum3 = b;
98
    }
99
    else
100
    {
101
        sum0 = afpvec4(0.f);
102
        sum1 = afpvec4(0.f);
103
        sum2 = afpvec4(0.f);
104
        sum3 = afpvec4(0.f);
105
    }
106

107
#if NCNN_image_shader
108
    ivec4 gx4 = gx + ivec4(0, 1, 2, 3);
109

110
    ivec4 sy4 = gx4 / psc(w);
111
    ivec4 sx4 = gx4 % psc(w);
112

113
    for (int z = 0; z < psc(c); z++)
114
    {
115
        afpvec8 v0 = image3d_ld8(bottom_blob, ivec3(sx4.r, sy4.r, z));
116
        afpvec8 v1 = image3d_ld8(bottom_blob, ivec3(sx4.g, sy4.g, z));
117
        afpvec8 v2 = image3d_ld8(bottom_blob, ivec3(sx4.b, sy4.b, z));
118
        afpvec8 v3 = image3d_ld8(bottom_blob, ivec3(sx4.a, sy4.a, z));
119

120
        afpvec8 k0 = image3d_ld8(weight_blob, ivec3(0, z, gy));
121
        afpvec8 k1 = image3d_ld8(weight_blob, ivec3(1, z, gy));
122
        afpvec8 k2 = image3d_ld8(weight_blob, ivec3(2, z, gy));
123
        afpvec8 k3 = image3d_ld8(weight_blob, ivec3(3, z, gy));
124

125
        // sum += v * k;
126
        sum0.r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
127
        sum0.g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
128
        sum0.b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);
129
        sum0.a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);
130

131
        sum1.r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
132
        sum1.g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
133
        sum1.b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);
134
        sum1.a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);
135

136
        sum2.r += dot(v2[0], k0[0]) + dot(v2[1], k0[1]);
137
        sum2.g += dot(v2[0], k1[0]) + dot(v2[1], k1[1]);
138
        sum2.b += dot(v2[0], k2[0]) + dot(v2[1], k2[1]);
139
        sum2.a += dot(v2[0], k3[0]) + dot(v2[1], k3[1]);
140

141
        sum3.r += dot(v3[0], k0[0]) + dot(v3[1], k0[1]);
142
        sum3.g += dot(v3[0], k1[0]) + dot(v3[1], k1[1]);
143
        sum3.b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);
144
        sum3.a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);
145
    }
146
#else
147
    int w_offset = gy * psc(c) * 4;
148
    int v_offset = gx;
149

150
    for (int z = 0; z < psc(c); z++)
151
    {
152
        afpvec8 v0 = buffer_ld8(bottom_blob_data, v_offset + 0);
153
        afpvec8 v1 = buffer_ld8(bottom_blob_data, v_offset + 1);
154
        afpvec8 v2 = buffer_ld8(bottom_blob_data, v_offset + 2);
155
        afpvec8 v3 = buffer_ld8(bottom_blob_data, v_offset + 3);
156

157
        afpvec8 k0 = buffer_ld8(weight_data, w_offset + 0);
158
        afpvec8 k1 = buffer_ld8(weight_data, w_offset + 1);
159
        afpvec8 k2 = buffer_ld8(weight_data, w_offset + 2);
160
        afpvec8 k3 = buffer_ld8(weight_data, w_offset + 3);
161

162
        // sum += v * k;
163
        sum0.r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
164
        sum0.g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
165
        sum0.b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);
166
        sum0.a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);
167

168
        sum1.r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
169
        sum1.g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
170
        sum1.b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);
171
        sum1.a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);
172

173
        sum2.r += dot(v2[0], k0[0]) + dot(v2[1], k0[1]);
174
        sum2.g += dot(v2[0], k1[0]) + dot(v2[1], k1[1]);
175
        sum2.b += dot(v2[0], k2[0]) + dot(v2[1], k2[1]);
176
        sum2.a += dot(v2[0], k3[0]) + dot(v2[1], k3[1]);
177

178
        sum3.r += dot(v3[0], k0[0]) + dot(v3[1], k0[1]);
179
        sum3.g += dot(v3[0], k1[0]) + dot(v3[1], k1[1]);
180
        sum3.b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);
181
        sum3.a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);
182

183
        w_offset += 4;
184
        v_offset += psc(cstep);
185
    }
186
#endif
187

188
    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
189
    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
190
    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
191
    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
192

193
#if NCNN_image_shader
194
    image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);
195
    image3d_st4(top_blob, ivec3(sx4.g, sy4.g, gy), sum1);
196
    image3d_st4(top_blob, ivec3(sx4.b, sy4.b, gy), sum2);
197
    image3d_st4(top_blob, ivec3(sx4.a, sy4.a, gy), sum3);
198
#else
199
    int gi = gy * psc(outcstep) + gx;
200

201
    buffer_st4(top_blob_data, gi + 0, sum0);
202
    if (gx + 1 < psc(outcstep)) buffer_st4(top_blob_data, gi + 1, sum1);
203
    if (gx + 2 < psc(outcstep)) buffer_st4(top_blob_data, gi + 2, sum2);
204
    if (gx + 3 < psc(outcstep)) buffer_st4(top_blob_data, gi + 3, sum3);
205
#endif
206
}
207