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
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 dilation_w = 1;
30
layout (constant_id = 2) const int stride_w = 1;
31
layout (constant_id = 3) const int bias_term = 0;
32
layout (constant_id = 4) const int activation_type = 0;
33
layout (constant_id = 5) const float activation_param_0 = 0;
34
layout (constant_id = 6) const float activation_param_1 = 0;
35

36
#define shape_constant_id_offset 7
37
layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
38
layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
39

40
layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
41
layout (constant_id = shape_constant_id_offset + 3) const int outh = 0;
42

43
#if NCNN_image_shader
44
layout (binding = 0) uniform unfp sampler3D bottom_blob;
45
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
46
layout (binding = 2) uniform unfp sampler3D weight_blob;
47
layout (binding = 3) uniform unfp sampler3D bias_blob;
48
#else
49
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
50
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
51
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
52
layout (binding = 3) readonly buffer bias_blob { sfp bias_data[]; };
53
#endif
54

55
layout (push_constant) uniform parameter
56
{
57
    int w;
58
    int h;
59

60
    int outw;
61
    int outh;
62
} p;
63

64
void main()
65
{
66
    int gx = int(gl_GlobalInvocationID.x) * 2;
67
    int gy = int(gl_GlobalInvocationID.y) * 2;
68

69
    if (gx >= psc(outw) || gy >= psc(outh))
70
        return;
71

72
    const ivec2 gx2 = gx + ivec2(0, 1);
73
    const ivec2 gy2 = gy + ivec2(0, 1);
74

75
    afp sum0 = afp(0.0f);
76
    afp sum1 = afp(0.0f);
77
    afp sum2 = afp(0.0f);
78
    afp sum3 = afp(0.0f);
79

80
    if (bias_term == 1)
81
    {
82
#if NCNN_image_shader
83
        sum0 = image3d_ld1(bias_blob, ivec3(gy2.x, 0, 0));
84
        sum2 = image3d_ld1(bias_blob, ivec3(gy2.y, 0, 0));
85
#else
86
        sum0 = buffer_ld1(bias_data, gy2.x);
87
        sum2 = buffer_ld1(bias_data, gy2.y);
88
#endif
89
        sum1 = sum0;
90
        sum3 = sum2;
91
    }
92

93
#if NCNN_image_shader
94

95
    ivec2 v_offset = gx2 * stride_w;
96

97
    for (int y = 0; y < psc(h); y++)
98
    {
99
        int wx = 0;
100

101
        for (int x = 0; x < kernel_w; x++)
102
        {
103
            afpvec8 v0 = image3d_ld8(bottom_blob, ivec3(v_offset.x + x * dilation_w, y, 0));
104
            afpvec8 v1 = image3d_ld8(bottom_blob, ivec3(v_offset.y + x * dilation_w, y, 0));
105
            
106
            afpvec8 k0 = image3d_ld8(weight_blob, ivec3(wx, y, gy2.x));
107
            afpvec8 k1 = image3d_ld8(weight_blob, ivec3(wx, y, gy2.y));
108

109
            sum0 += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
110
            sum1 += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
111
            sum2 += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
112
            sum3 += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
113

114
            wx += 1;
115
        }
116
    }
117
    
118
#else
119

120
    ivec2 v_offset = gx2 * stride_w;
121
    ivec2 w_offset = gy2 * psc(h) * kernel_w;
122
    
123
    for (int y = 0; y < psc(h); y++)
124
    {    
125
        for (int x = 0; x < kernel_w; x++)
126
        {
127
            afpvec8 v0 = buffer_ld8(bottom_blob_data, v_offset.x + x * dilation_w);
128
            afpvec8 v1 = buffer_ld8(bottom_blob_data, v_offset.y + x * dilation_w);
129
            
130
            afpvec8 k0 = buffer_ld8(weight_data, w_offset.x + x);
131
            afpvec8 k1 = buffer_ld8(weight_data, w_offset.y + x);
132

133
            sum0 += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
134
            sum1 += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
135
            sum2 += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
136
            sum3 += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
137
        }       
138
        v_offset += psc(w);
139
        w_offset += kernel_w;
140
    }
141
    
142
#endif
143

144
    sum0 = activation_afp(sum0, activation_type, activation_param_0, activation_param_1);
145
    sum1 = activation_afp(sum1, activation_type, activation_param_0, activation_param_1);
146
    sum2 = activation_afp(sum2, activation_type, activation_param_0, activation_param_1);
147
    sum3 = activation_afp(sum3, activation_type, activation_param_0, activation_param_1);
148
    
149
#if NCNN_image_shader
150

151
    image3d_st1(top_blob, ivec3(gx2.x, gy2.x, 0), sum0);
152
    image3d_st1(top_blob, ivec3(gx2.y, gy2.x, 0), sum1);
153
    image3d_st1(top_blob, ivec3(gx2.x, gy2.y, 0), sum2);
154
    image3d_st1(top_blob, ivec3(gx2.y, gy2.y, 0), sum3);
155
    
156
#else
157

158
    const int gi = gy * psc(outw) + gx;
159

160
    buffer_st1(top_blob_data, gi, sum0);
161
    if (gx + 1 < psc(outw)) buffer_st1(top_blob_data, gi + 1, sum1);
162
    if (gy + 1 < psc(outh)) buffer_st1(top_blob_data, gi + psc(outw), sum2);
163
    if (gy + 1 < psc(outh) && gx + 1 < psc(outw)) buffer_st1(top_blob_data, gi + psc(outw) + 1, sum3);
164
    
165
#endif
166
}
167