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
#define shape_constant_id_offset 0
26
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
27
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
28
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
29
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
30
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
31

32
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
33
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
34
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
35
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
36
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
37

38
#if NCNN_image_shader
39
layout (binding = 0) uniform unfp sampler3D bottom_blob_3d;
40
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
41
#else
42
#if NCNN_fp16_packed
43
layout (binding = 0) readonly buffer bottom_blob { sfpvec2 bottom_blob_data[]; };
44
#else
45
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
46
#endif
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 c;
56
    int cstep;
57

58
    int outdims;
59
    int outw;
60
    int outh;
61
    int outc;
62
    int outcstep;
63
} p;
64

65
void main()
66
{
67
    int gx = int(gl_GlobalInvocationID.x);
68
    int gy = int(gl_GlobalInvocationID.y);
69
    int gz = int(gl_GlobalInvocationID.z);
70

71
    if (gx >= psc(outw) || gy >= 1 || gz >= 1)
72
        return;
73

74
    ivec4 i4 = gx * 8 + ivec4(0, 1, 2, 3);
75
    ivec4 ii4 = i4 + 4;
76

77
#if NCNN_image_shader
78
    afpvec8 v;
79

80
    if (psc(dims) == 2)
81
    {
82
        ivec4 y4 = i4 / psc(w);
83
        ivec4 x4 = i4 % psc(w);
84
        ivec4 yy4 = ii4 / psc(w);
85
        ivec4 xx4 = ii4 % psc(w);
86

87
        afpvec8 v0 = image3d_ld8(bottom_blob_3d, ivec3(x4.r, y4.r / 8, 0));
88
        afpvec8 v1 = image3d_ld8(bottom_blob_3d, ivec3(x4.g, y4.g / 8, 0));
89
        afpvec8 v2 = image3d_ld8(bottom_blob_3d, ivec3(x4.b, y4.b / 8, 0));
90
        afpvec8 v3 = image3d_ld8(bottom_blob_3d, ivec3(x4.a, y4.a / 8, 0));
91
        afpvec8 v4 = image3d_ld8(bottom_blob_3d, ivec3(xx4.r, yy4.r / 8, 0));
92
        afpvec8 v5 = image3d_ld8(bottom_blob_3d, ivec3(xx4.g, yy4.g / 8, 0));
93
        afpvec8 v6 = image3d_ld8(bottom_blob_3d, ivec3(xx4.b, yy4.b / 8, 0));
94
        afpvec8 v7 = image3d_ld8(bottom_blob_3d, ivec3(xx4.a, yy4.a / 8, 0));
95

96
        v[0].r = v0[(y4.r % 8) / 4][y4.r % 4];
97
        v[0].g = v1[(y4.g % 8) / 4][y4.g % 4];
98
        v[0].b = v2[(y4.b % 8) / 4][y4.b % 4];
99
        v[0].a = v3[(y4.a % 8) / 4][y4.a % 4];
100
        v[1].r = v4[(yy4.r % 8) / 4][yy4.r % 4];
101
        v[1].g = v5[(yy4.g % 8) / 4][yy4.g % 4];
102
        v[1].b = v6[(yy4.b % 8) / 4][yy4.b % 4];
103
        v[1].a = v7[(yy4.a % 8) / 4][yy4.a % 4];
104
    }
105
    else // if (psc(dims) == 3)
106
    {
107
        int size = psc(w) * psc(h);
108

109
        ivec4 z4 = i4 / size;
110
        ivec4 y4 = i4 % size / psc(w);
111
        ivec4 x4 = i4 % size % psc(w);
112
        ivec4 zz4 = ii4 / size;
113
        ivec4 yy4 = ii4 % size / psc(w);
114
        ivec4 xx4 = ii4 % size % psc(w);
115

116
        afpvec8 v0 = image3d_ld8(bottom_blob_3d, ivec3(x4.r, y4.r, z4.r / 8));
117
        afpvec8 v1 = image3d_ld8(bottom_blob_3d, ivec3(x4.g, y4.g, z4.g / 8));
118
        afpvec8 v2 = image3d_ld8(bottom_blob_3d, ivec3(x4.b, y4.b, z4.b / 8));
119
        afpvec8 v3 = image3d_ld8(bottom_blob_3d, ivec3(x4.a, y4.a, z4.a / 8));
120
        afpvec8 v4 = image3d_ld8(bottom_blob_3d, ivec3(xx4.r, yy4.r, zz4.r / 8));
121
        afpvec8 v5 = image3d_ld8(bottom_blob_3d, ivec3(xx4.g, yy4.g, zz4.g / 8));
122
        afpvec8 v6 = image3d_ld8(bottom_blob_3d, ivec3(xx4.b, yy4.b, zz4.b / 8));
123
        afpvec8 v7 = image3d_ld8(bottom_blob_3d, ivec3(xx4.a, yy4.a, zz4.a / 8));
124

125
        v[0].r = v0[(z4.r % 8) / 4][z4.r % 4];
126
        v[0].g = v1[(z4.g % 8) / 4][z4.g % 4];
127
        v[0].b = v2[(z4.b % 8) / 4][z4.b % 4];
128
        v[0].a = v3[(z4.a % 8) / 4][z4.a % 4];
129
        v[1].r = v4[(zz4.r % 8) / 4][zz4.r % 4];
130
        v[1].g = v5[(zz4.g % 8) / 4][zz4.g % 4];
131
        v[1].b = v6[(zz4.b % 8) / 4][zz4.b % 4];
132
        v[1].a = v7[(zz4.a % 8) / 4][zz4.a % 4];
133
    }
134

135
    image3d_st8(top_blob, ivec3(gx, 0, 0), v);
136
#else
137
#if NCNN_fp16_packed
138
    ivec4 v_offset;
139
    ivec4 lane4;
140
    ivec4 vv_offset;
141
    ivec4 lane8;
142

143
    if (psc(dims) == 2)
144
    {
145
        ivec4 y4 = i4 / psc(w);
146
        ivec4 x4 = i4 % psc(w);
147
        ivec4 yy4 = ii4 / psc(w);
148
        ivec4 xx4 = ii4 % psc(w);
149

150
        v_offset = ((y4 / 8) * psc(w) + x4) * 4 + (y4 % 8) / 2;
151
        lane4 = y4 % 2;
152
        vv_offset = ((yy4 / 8) * psc(w) + xx4) * 4 + (yy4 % 8) / 2;
153
        lane8 = yy4 % 2;
154
    }
155
    else // if (psc(dims) == 3)
156
    {
157
        int size = psc(w) * psc(h);
158

159
        ivec4 z4 = i4 / size;
160
        ivec4 y4 = i4 % size / psc(w);
161
        ivec4 x4 = i4 % size % psc(w);
162
        ivec4 zz4 = ii4 / size;
163
        ivec4 yy4 = ii4 % size / psc(w);
164
        ivec4 xx4 = ii4 % size % psc(w);
165

166
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 4 + (z4 % 8) / 2;
167
        lane4 = z4 % 2;
168
        vv_offset = ((zz4 / 8) * psc(cstep) + yy4 * psc(w) + xx4) * 4 + (zz4 % 8) / 2;
169
        lane8 = zz4 % 2;
170
    }
171

172
    afpvec2 vr = buffer_ld2(bottom_blob_data, v_offset.r);
173
    afpvec2 vg = buffer_ld2(bottom_blob_data, v_offset.g);
174
    afpvec2 vb = buffer_ld2(bottom_blob_data, v_offset.b);
175
    afpvec2 va = buffer_ld2(bottom_blob_data, v_offset.a);
176

177
    afpvec2 vvr = buffer_ld2(bottom_blob_data, vv_offset.r);
178
    afpvec2 vvg = buffer_ld2(bottom_blob_data, vv_offset.g);
179
    afpvec2 vvb = buffer_ld2(bottom_blob_data, vv_offset.b);
180
    afpvec2 vva = buffer_ld2(bottom_blob_data, vv_offset.a);
181

182
    afpvec8 v = afpvec8(vr[lane4.r], vg[lane4.g], vb[lane4.b], va[lane4.a], vvr[lane8.r], vvg[lane8.g], vvb[lane8.b], vva[lane8.a]);
183

184
    buffer_st8(top_blob_data, gx, v);
185
#else
186
    ivec4 v_offset;
187
    ivec4 vv_offset;
188

189
    if (psc(dims) == 2)
190
    {
191
        ivec4 y4 = i4 / psc(w);
192
        ivec4 x4 = i4 % psc(w);
193
        ivec4 yy4 = ii4 / psc(w);
194
        ivec4 xx4 = ii4 % psc(w);
195

196
        v_offset = ((y4 / 8) * psc(w) + x4) * 8 + y4 % 8;
197
        vv_offset = ((yy4 / 8) * psc(w) + xx4) * 8 + yy4 % 8;
198
    }
199
    else // if (psc(dims) == 3)
200
    {
201
        int size = psc(w) * psc(h);
202

203
        ivec4 z4 = i4 / size;
204
        ivec4 y4 = i4 % size / psc(w);
205
        ivec4 x4 = i4 % size % psc(w);
206
        ivec4 zz4 = ii4 / size;
207
        ivec4 yy4 = ii4 % size / psc(w);
208
        ivec4 xx4 = ii4 % size % psc(w);
209

210
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 8 + z4 % 8;
211
        vv_offset = ((zz4 / 8) * psc(cstep) + yy4 * psc(w) + xx4) * 8 + zz4 % 8;
212
    }
213

214
    buffer_cp1to8(top_blob_data, gx, bottom_blob_data, v_offset, vv_offset);
215
#endif
216
#endif
217
}
218