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 ndim = 0;
26

27
#define shape_constant_id_offset 1
28
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
29
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
30
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
31
layout (constant_id = shape_constant_id_offset + 3) const int d = 0;
32
layout (constant_id = shape_constant_id_offset + 4) const int c = 0;
33
layout (constant_id = shape_constant_id_offset + 5) const int cstep = 0;
34

35
layout (constant_id = shape_constant_id_offset + 6) const int outdims = 0;
36
layout (constant_id = shape_constant_id_offset + 7) const int outw = 0;
37
layout (constant_id = shape_constant_id_offset + 8) const int outh = 0;
38
layout (constant_id = shape_constant_id_offset + 9) const int outd = 0;
39
layout (constant_id = shape_constant_id_offset + 10) const int outc = 0;
40
layout (constant_id = shape_constant_id_offset + 11) const int outcstep = 0;
41

42
#if NCNN_image_shader
43
layout (binding = 0) uniform unfp sampler3D bottom_blob_3d;
44
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
45
#else
46
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
47
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
48
#endif
49

50
layout (push_constant) uniform parameter
51
{
52
    int dims;
53
    int w;
54
    int h;
55
    int d;
56
    int c;
57
    int cstep;
58

59
    int outdims;
60
    int outw;
61
    int outh;
62
    int outd;
63
    int outc;
64
    int outcstep;
65
} p;
66

67
void main()
68
{
69
    int gx = int(gl_GlobalInvocationID.x);
70
    int gy = int(gl_GlobalInvocationID.y);
71
    int gz = int(gl_GlobalInvocationID.z);
72

73
    if (gx >= psc(outw) || gy >= psc(outh) * psc(outd) || gz >= psc(outc))
74
        return;
75

76
    ivec4 i4;
77
    ivec4 ii4;
78
    if (ndim == 1)
79
    {
80
        i4 = gx * 8 + ivec4(0, 1, 2, 3);
81
        ii4 = i4 + 4;
82
    }
83
    if (ndim == 2)
84
    {
85
        i4 = (gy * 8) * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outw);
86
        ii4 = i4 + 4 * psc(outw);
87
    }
88
    if (ndim == 3)
89
    {
90
        i4 = (gz * 8) * psc(outh) * psc(outw) + gy * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outh) * psc(outw);
91
        ii4 = i4 + 4 * psc(outh) * psc(outw);
92
    }
93
    if (ndim == 4)
94
    {
95
        i4 = (gz * 8) * psc(outd) * psc(outh) * psc(outw) + gy * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outd) * psc(outh) * psc(outw);
96
        ii4 = i4 + 4 * psc(outd) * psc(outh) * psc(outw);
97
    }
98

99
    ivec4 x4;
100
    ivec4 xx4;
101
    ivec4 y4;
102
    ivec4 yy4;
103
    ivec4 z4;
104
    ivec4 zz4;
105

106
    if (psc(dims) == 1)
107
    {
108
        z4 = ivec4(0);
109
        y4 = ivec4(0);
110
        x4 = i4;
111
        zz4 = ivec4(0);
112
        yy4 = ivec4(0);
113
        xx4 = ii4;
114
    }
115
    else if (psc(dims) == 2)
116
    {
117
        z4 = ivec4(0);
118
        y4 = i4 / psc(w);
119
        x4 = i4 % psc(w);
120
        zz4 = ivec4(0);
121
        yy4 = ii4 / psc(w);
122
        xx4 = ii4 % psc(w);
123
    }
124
    else if (psc(dims) == 3)
125
    {
126
        int size = psc(w) * psc(h);
127

128
        z4 = i4 / size;
129
        y4 = i4 % size / psc(w);
130
        x4 = i4 % size % psc(w);
131
        zz4 = ii4 / size;
132
        yy4 = ii4 % size / psc(w);
133
        xx4 = ii4 % size % psc(w);
134
    }
135
    else // if (psc(dims) == 4)
136
    {
137
        int size = psc(w) * psc(h) * psc(d);
138
        int dsize = psc(w) * psc(h);
139

140
        z4 = i4 / size;
141
        ivec4 yd4 = i4 % size / dsize;
142
        ivec4 yh4 = i4 % size % dsize / psc(w);
143
        x4 = i4 % size % dsize % psc(w);
144

145
        zz4 = ii4 / size;
146
        ivec4 yyd4 = ii4 % size / dsize;
147
        ivec4 yyh4 = ii4 % size % dsize / psc(w);
148
        xx4 = ii4 % size % dsize % psc(w);
149

150
        y4 = yd4 * psc(h) + yh4;
151
        yy4 = yyd4 * psc(h) + yyh4;
152
    }
153

154
#if NCNN_image_shader
155
    afpvec8 v;
156
    v[0].r = image3d_ld1(bottom_blob_3d, ivec3(x4.r, y4.r, z4.r));
157
    v[0].g = image3d_ld1(bottom_blob_3d, ivec3(x4.g, y4.g, z4.g));
158
    v[0].b = image3d_ld1(bottom_blob_3d, ivec3(x4.b, y4.b, z4.b));
159
    v[0].a = image3d_ld1(bottom_blob_3d, ivec3(x4.a, y4.a, z4.a));
160
    v[1].r = image3d_ld1(bottom_blob_3d, ivec3(xx4.r, yy4.r, zz4.r));
161
    v[1].g = image3d_ld1(bottom_blob_3d, ivec3(xx4.g, yy4.g, zz4.g));
162
    v[1].b = image3d_ld1(bottom_blob_3d, ivec3(xx4.b, yy4.b, zz4.b));
163
    v[1].a = image3d_ld1(bottom_blob_3d, ivec3(xx4.a, yy4.a, zz4.a));
164

165
    if (ndim == 1) image3d_st8(top_blob_3d, ivec3(gx, 0, 0), v);
166
    if (ndim == 2) image3d_st8(top_blob_3d, ivec3(gx, gy, 0), v);
167
    if (ndim == 3 || ndim == 4) image3d_st8(top_blob_3d, ivec3(gx, gy, gz), v);
168
#else
169
    ivec4 v_offset = z4 * psc(cstep) + y4 * psc(w) + x4;
170
    ivec4 vv_offset = zz4 * psc(cstep) + yy4 * psc(w) + xx4;
171

172
    int gi;
173
    if (ndim == 1) gi = gx;
174
    if (ndim == 2) gi = gy * psc(outw) + gx;
175
    if (ndim == 3 || ndim == 4) gi = gz * psc(outcstep) + gy * psc(outw) + gx;
176

177
    buffer_cp1to8(top_blob_data, gi, bottom_blob_data, v_offset, vv_offset);
178
#endif
179
}
180