ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2023 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
#endif
20
#if NCNN_fp16_arithmetic
21
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
22
#endif
23

24
#extension GL_GOOGLE_include_directive: enable
25
#include "vulkan_activation.comp"
26

27
layout (constant_id = 0) const int kernel_w = 1;
28
layout (constant_id = 1) const int dilation_w = 1;
29
layout (constant_id = 2) const int stride_w = 1;
30
layout (constant_id = 3) const int bias_term = 0;
31
layout (constant_id = 4) const int activation_type = 0;
32
layout (constant_id = 5) const float activation_param_0 = 0;
33
layout (constant_id = 6) const float activation_param_1 = 0;
34

35
#define shape_constant_id_offset 7
36
layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
37
layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
38

39
layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
40
layout (constant_id = shape_constant_id_offset + 3) const int outh = 0;
41

42
#if NCNN_image_shader
43
layout (binding = 0) uniform unfp sampler3D bottom_blob;
44
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
45
layout (binding = 2) uniform unfp sampler3D weight_blob;
46
layout (binding = 3) uniform unfp sampler3D bias_blob;
47
#else
48
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
49
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
50
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
51
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
52
#endif
53

54
layout (push_constant) uniform parameter
55
{
56
    int w;
57
    int h;
58

59
    int outw;
60
    int outh;
61
} p;
62

63
void main()
64
{
65
    int gx = int(gl_GlobalInvocationID.x) * 2;
66
    int gy = int(gl_GlobalInvocationID.y) * 2;
67

68
    if (gx >= psc(outw) || gy >= psc(outh))
69
        return;
70

71
    const ivec2 gx2 = gx + ivec2(0, 1);
72
    const ivec2 gy2 = gy + ivec2(0, 1);
73

74
    afpvec4 sum0 = afpvec4(0.0f);
75
    afpvec4 sum1 = afpvec4(0.0f);
76
    afpvec4 sum2 = afpvec4(0.0f);
77
    afpvec4 sum3 = afpvec4(0.0f);
78

79
    if (bias_term == 1)
80
    {
81
#if NCNN_image_shader
82
        sum0 = image3d_ld4(bias_blob, ivec3(gy2.x, 0, 0));
83
        sum2 = image3d_ld4(bias_blob, ivec3(gy2.y, 0, 0));
84
#else
85
        sum0 = buffer_ld4(bias_data, gy2.x);
86
        sum2 = buffer_ld4(bias_data, gy2.y);
87
#endif
88
        sum1 = sum0;
89
        sum3 = sum2;
90
    }
91

92
#if NCNN_image_shader
93

94
    ivec2 v_offset = gx2 * stride_w;
95

96
    for (int y = 0; y < psc(h); y++)
97
    {
98
        int wx = 0;
99

100
        for (int x = 0; x < kernel_w; x++)
101
        {
102
            afp v0 = image3d_ld1(bottom_blob, ivec3(v_offset.x + x * dilation_w, y, 0));
103
            afp v1 = image3d_ld1(bottom_blob, ivec3(v_offset.y + x * dilation_w, y, 0));
104
            
105
            afpvec4 k0 = image3d_ld4(weight_blob, ivec3(wx, y, gy2.x));
106
            afpvec4 k1 = image3d_ld4(weight_blob, ivec3(wx, y, gy2.y));
107

108
            sum0 += v0 * k0;
109
            sum1 += v1 * k0;
110
            sum2 += v0 * k1;
111
            sum3 += v1 * k1;
112

113
            wx += 1;
114
        }
115
    }
116
    
117
#else
118

119
    ivec2 v_offset = gx2 * stride_w;
120
    ivec2 w_offset = gy2 * psc(h) * kernel_w;
121
    
122
    for (int y = 0; y < psc(h); y++)
123
    {    
124
        for (int x = 0; x < kernel_w; x++)
125
        {
126
            afp v0 = buffer_ld1(bottom_blob_data, v_offset.x + x * dilation_w);
127
            afp v1 = buffer_ld1(bottom_blob_data, v_offset.y + x * dilation_w);
128
            
129
            afpvec4 k0 = buffer_ld4(weight_data, w_offset.x + x);
130
            afpvec4 k1 = buffer_ld4(weight_data, w_offset.y + x);
131

132
            sum0 += v0 * k0;
133
            sum1 += v1 * k0;
134
            sum2 += v0 * k1;
135
            sum3 += v1 * k1;
136
        }       
137
        v_offset += psc(w);
138
        w_offset += kernel_w;
139
    }
140
    
141
#endif
142

143
    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
144
    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
145
    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
146
    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
147
    
148
#if NCNN_image_shader
149

150
    image3d_st4(top_blob, ivec3(gx2.x, gy2.x, 0), sum0);
151
    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, 0), sum1);
152
    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, 0), sum2);
153
    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, 0), sum3);
154
    
155
#else
156

157
    const int gi = gy * psc(outw) + gx;
158

159
    buffer_st4(top_blob_data, gi, sum0);
160
    if (gx + 1 < psc(outw)) buffer_st4(top_blob_data, gi + 1, sum1);
161
    if (gy + 1 < psc(outh)) buffer_st4(top_blob_data, gi + psc(outw), sum2);
162
    if (gy + 1 < psc(outh) && gx + 1 < psc(outw)) buffer_st4(top_blob_data, gi + psc(outw) + 1, sum3);
163
    
164
#endif
165
}
166