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 stride = 0;
26
layout (constant_id = 1) const int mode = 0;
27

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

35
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
36
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
37
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
38
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
39
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
40

41
#if NCNN_image_shader
42
layout (binding = 0) uniform unfp sampler3D bottom_blob;
43
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
44
#else
45
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
46
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_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

57
    int outdims;
58
    int outw;
59
    int outh;
60
    int outc;
61
    int outcstep;
62
} p;
63

64
void main()
65
{
66
    int gx = int(gl_GlobalInvocationID.x);
67
    int gy = int(gl_GlobalInvocationID.y);
68
    int gz = int(gl_GlobalInvocationID.z);
69

70
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
71
        return;
72

73
    ivec4 gz4 = gz * 8 + ivec4(0, 1, 2, 3);
74
    ivec4 gzz4 = gz4 + 4;
75

76
    ivec4 z4;
77
    ivec4 zi4;
78
    ivec4 zz4;
79
    ivec4 zii4;
80
    if (mode == 0)
81
    {
82
        z4 = gz4 / (stride * stride);
83
        zi4 = gz4 % (stride * stride);
84
        zz4 = gzz4 / (stride * stride);
85
        zii4 = gzz4 % (stride * stride);
86
    }
87
    else // if (mode == 1)
88
    {
89
        z4 = gz4 % psc(c);
90
        zi4 = gz4 / psc(c);
91
        zz4 = gzz4 % psc(c);
92
        zii4 = gzz4 / psc(c);
93
    }
94
    ivec4 y4 = gy * stride + zi4 / stride;
95
    ivec4 x4 = gx * stride + zi4 % stride;
96
    ivec4 yy4 = gy * stride + zii4 / stride;
97
    ivec4 xx4 = gx * stride + zii4 % stride;
98

99
#if NCNN_image_shader
100
    afpvec8 v;
101
    v[0].r = image3d_ld1(bottom_blob, ivec3(x4.r, y4.r, z4.r));
102
    v[0].g = image3d_ld1(bottom_blob, ivec3(x4.g, y4.g, z4.g));
103
    v[0].b = image3d_ld1(bottom_blob, ivec3(x4.b, y4.b, z4.b));
104
    v[0].a = image3d_ld1(bottom_blob, ivec3(x4.a, y4.a, z4.a));
105
    v[1].r = image3d_ld1(bottom_blob, ivec3(xx4.r, yy4.r, zz4.r));
106
    v[1].g = image3d_ld1(bottom_blob, ivec3(xx4.g, yy4.g, zz4.g));
107
    v[1].b = image3d_ld1(bottom_blob, ivec3(xx4.b, yy4.b, zz4.b));
108
    v[1].a = image3d_ld1(bottom_blob, ivec3(xx4.a, yy4.a, zz4.a));
109

110
    image3d_st8(top_blob, ivec3(gx, gy, gz), v);
111
#else
112
    ivec4 v_offset = z4 * psc(cstep) + y4 * psc(w) + x4;
113
    ivec4 vv_offset = zz4 * psc(cstep) + yy4 * psc(w) + xx4;
114

115
    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
116

117
    buffer_cp1to8(top_blob_data, gi, bottom_blob_data, v_offset, vv_offset);
118
#endif
119
}
120