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
layout (constant_id = 0) const int across_spatial = 0;
26
layout (constant_id = 1) const int across_channel = 0;
27
layout (constant_id = 2) const int channel_shared = 0;
28
layout (constant_id = 3) const int scale_term = 0;
29
layout (constant_id = 4) const float channel_shared_scale = 1.f;
30

31
#define shape_constant_id_offset 5
32
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
33
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
34
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
35
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
36
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
37

38
#if NCNN_image_shader
39
layout (binding = 0) uniform unfp sampler3D bottom_blob;
40
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
41
layout (binding = 2) uniform unfp sampler3D coeffs_blob;
42
layout (binding = 3) uniform unfp sampler3D scale_blob;
43
#else
44
layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
45
layout (binding = 1) readonly buffer coeffs_blob { sfpvec8 coeffs_blob_data[]; };
46
layout (binding = 2) readonly buffer scale_blob { sfpvec8 scale_blob_data[]; };
47
#endif
48

49
layout (push_constant) uniform parameter
50
{
51
    int dims;
52
    int w;
53
    int h;
54
    int c;
55
    int cstep;
56
} p;
57

58
void main()
59
{
60
    int gx = int(gl_GlobalInvocationID.x);
61
    int gy = int(gl_GlobalInvocationID.y);
62
    int gz = int(gl_GlobalInvocationID.z);
63

64
    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
65
        return;
66

67
#if NCNN_image_shader
68
    afpvec8 v = image3d_ld8(bottom_blob, ivec3(gx, gy, gz));
69
#else
70
    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
71

72
    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
73
#endif
74

75
    afpvec8 a;
76

77
    if (across_spatial == 1 && across_channel == 1)
78
    {
79
#if NCNN_image_shader
80
        a = image3d_ld8(coeffs_blob, ivec3(0, 0, 0));
81
#else
82
        a = buffer_ld8(coeffs_blob_data, 0);
83
#endif
84
    }
85

86
    if (across_spatial == 1 && across_channel == 0)
87
    {
88
#if NCNN_image_shader
89
        a = image3d_ld8(coeffs_blob, ivec3(gz, 0, 0));
90
#else
91
        a = buffer_ld8(coeffs_blob_data, gz);
92
#endif
93
    }
94

95
    if (across_spatial == 0 && across_channel == 1)
96
    {
97
#if NCNN_image_shader
98
        a = image3d_ld8(coeffs_blob, ivec3(gy * psc(w) + gx, 0, 0));
99
#else
100
        a = buffer_ld8(coeffs_blob_data, gy * psc(w) + gx);
101
#endif
102
    }
103

104
#if !NCNN_image_shader && (NCNN_fp16_packed || NCNN_fp16_storage)
105
    // NOTE coeffs may produce (X, undef, X, undef) on nvidia fp16p/fp16s
106
    // TODO only enable this workaround for some nvidia driver
107
    if (across_channel == 1)
108
    {
109
        a = afpvec8(afpvec4(a[0].r), afpvec4(a[0].r));
110
    }
111
#endif
112

113
    v[0] = v[0] * a[0];
114
    v[1] = v[1] * a[1];
115

116
    if (scale_term == 1)
117
    {
118
        if (channel_shared == 1)
119
        {
120
            v[0] = v[0] * afp(channel_shared_scale);
121
            v[1] = v[1] * afp(channel_shared_scale);
122
        }
123
        else
124
        {
125
#if NCNN_image_shader
126
            afpvec8 scale = image3d_ld8(scale_blob, ivec3(gz, 0, 0));
127
#else
128
            afpvec8 scale = buffer_ld8(scale_blob_data, gz);
129
#endif
130

131
            v[0] = v[0] * scale[0];
132
            v[1] = v[1] * scale[1];
133
        }
134
    }
135

136
#if NCNN_image_shader
137
    image3d_st8(top_blob, ivec3(gx, gy, gz), v);
138
#else
139
    buffer_st8(bottom_top_blob_data, gi, v);
140
#endif
141
}
142