ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2021 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 dims = 0;
35
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
36
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
37
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
38
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
39

40
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
41
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
42
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
43
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
44
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
45

46
#if NCNN_image_shader
47
layout (binding = 0) uniform unfp sampler3D bottom_blob;
48
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
49
layout (binding = 2) uniform unfp sampler3D weight_blob;
50
layout (binding = 3) uniform unfp sampler3D bias_blob;
51
#else
52
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
53
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
54
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
55
layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
56
#endif
57

58
layout (push_constant) uniform parameter
59
{
60
    int dims;
61
    int w;
62
    int h;
63
    int c;
64
    int cstep;
65

66
    int outdims;
67
    int outw;
68
    int outh;
69
    int outc;
70
    int outcstep;
71
} p;
72

73
void main()
74
{
75
    int gx = int(gl_GlobalInvocationID.x);
76
    int gy = int(gl_GlobalInvocationID.y);
77
    int gz = int(gl_GlobalInvocationID.z);
78

79
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= 1)
80
        return;
81

82
    afpvec8 sum;
83

84
    if (bias_term == 1)
85
    {
86
#if NCNN_image_shader
87
        sum = image3d_ld8(bias_blob, ivec3(gx, 0, 0));
88
#else
89
        sum = buffer_ld8(bias_data, gx);
90
#endif
91
    }
92
    else
93
    {
94
        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
95
    }
96

97
#if NCNN_image_shader
98
    for (int i = 0; i < psc(w) / 8; i++)
99
    {
100
        afpvec8 v;
101
        v[0].r = image3d_ld1(bottom_blob, ivec3(i * 8 + 0, gy, 0));
102
        v[0].g = image3d_ld1(bottom_blob, ivec3(i * 8 + 1, gy, 0));
103
        v[0].b = image3d_ld1(bottom_blob, ivec3(i * 8 + 2, gy, 0));
104
        v[0].a = image3d_ld1(bottom_blob, ivec3(i * 8 + 3, gy, 0));
105
        v[1].r = image3d_ld1(bottom_blob, ivec3(i * 8 + 4, gy, 0));
106
        v[1].g = image3d_ld1(bottom_blob, ivec3(i * 8 + 5, gy, 0));
107
        v[1].b = image3d_ld1(bottom_blob, ivec3(i * 8 + 6, gy, 0));
108
        v[1].a = image3d_ld1(bottom_blob, ivec3(i * 8 + 7, gy, 0));
109

110
        afpvec8 k0 = image3d_ld8(weight_blob, ivec3(i * 8 + 0, gx, 0));
111
        afpvec8 k1 = image3d_ld8(weight_blob, ivec3(i * 8 + 1, gx, 0));
112
        afpvec8 k2 = image3d_ld8(weight_blob, ivec3(i * 8 + 2, gx, 0));
113
        afpvec8 k3 = image3d_ld8(weight_blob, ivec3(i * 8 + 3, gx, 0));
114
        afpvec8 k4 = image3d_ld8(weight_blob, ivec3(i * 8 + 4, gx, 0));
115
        afpvec8 k5 = image3d_ld8(weight_blob, ivec3(i * 8 + 5, gx, 0));
116
        afpvec8 k6 = image3d_ld8(weight_blob, ivec3(i * 8 + 6, gx, 0));
117
        afpvec8 k7 = image3d_ld8(weight_blob, ivec3(i * 8 + 7, gx, 0));
118

119
        // sum += v * k
120
        sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
121
        sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
122
        sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
123
        sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
124
        sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
125
        sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
126
        sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
127
        sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
128
    }
129
#else
130
    int v_offset = gy * psc(w);
131
    int w_offset = gx * psc(w) / 8;
132

133
    for (int i = 0; i < psc(w) / 8; i++)
134
    {
135
        afpvec8 v;
136
        v[0].r = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 0);
137
        v[0].g = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 1);
138
        v[0].b = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 2);
139
        v[0].a = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 3);
140
        v[1].r = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 4);
141
        v[1].g = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 5);
142
        v[1].b = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 6);
143
        v[1].a = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 7);
144

145
        afpvec8 k0 = buffer_ld8(weight_data, (w_offset + i) * 8 + 0);
146
        afpvec8 k1 = buffer_ld8(weight_data, (w_offset + i) * 8 + 1);
147
        afpvec8 k2 = buffer_ld8(weight_data, (w_offset + i) * 8 + 2);
148
        afpvec8 k3 = buffer_ld8(weight_data, (w_offset + i) * 8 + 3);
149
        afpvec8 k4 = buffer_ld8(weight_data, (w_offset + i) * 8 + 4);
150
        afpvec8 k5 = buffer_ld8(weight_data, (w_offset + i) * 8 + 5);
151
        afpvec8 k6 = buffer_ld8(weight_data, (w_offset + i) * 8 + 6);
152
        afpvec8 k7 = buffer_ld8(weight_data, (w_offset + i) * 8 + 7);
153

154
        // sum += v * k
155
        sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
156
        sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
157
        sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
158
        sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
159
        sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
160
        sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
161
        sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
162
        sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
163
    }
164
#endif
165

166
    sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
167

168
#if NCNN_image_shader
169
    image3d_st1(top_blob, ivec3(gx * 8 + 0, gy, 0), sum[0].r);
170
    image3d_st1(top_blob, ivec3(gx * 8 + 1, gy, 0), sum[0].g);
171
    image3d_st1(top_blob, ivec3(gx * 8 + 2, gy, 0), sum[0].b);
172
    image3d_st1(top_blob, ivec3(gx * 8 + 3, gy, 0), sum[0].a);
173
    image3d_st1(top_blob, ivec3(gx * 8 + 4, gy, 0), sum[1].r);
174
    image3d_st1(top_blob, ivec3(gx * 8 + 5, gy, 0), sum[1].g);
175
    image3d_st1(top_blob, ivec3(gx * 8 + 6, gy, 0), sum[1].b);
176
    image3d_st1(top_blob, ivec3(gx * 8 + 7, gy, 0), sum[1].a);
177
#else
178
    const int gi = gy * psc(outw) + gx * 8;
179
    buffer_st1(top_blob_data, gi + 0, sum[0].r);
180
    buffer_st1(top_blob_data, gi + 1, sum[0].g);
181
    buffer_st1(top_blob_data, gi + 2, sum[0].b);
182
    buffer_st1(top_blob_data, gi + 3, sum[0].a);
183
    buffer_st1(top_blob_data, gi + 4, sum[1].r);
184
    buffer_st1(top_blob_data, gi + 5, sum[1].g);
185
    buffer_st1(top_blob_data, gi + 6, sum[1].b);
186
    buffer_st1(top_blob_data, gi + 7, sum[1].a);
187
#endif
188
}
189