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
#if NCNN_fp16_packed
47
layout (binding = 0) readonly buffer bottom_blob { sfpvec2 bottom_blob_data[]; };
48
#else
49
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
50
#endif
51
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
52
#endif
53

54
layout (push_constant) uniform parameter
55
{
56
    int dims;
57
    int w;
58
    int h;
59
    int d;
60
    int c;
61
    int cstep;
62

63
    int outdims;
64
    int outw;
65
    int outh;
66
    int outd;
67
    int outc;
68
    int outcstep;
69
} p;
70

71
void main()
72
{
73
    int gx = int(gl_GlobalInvocationID.x);
74
    int gy = int(gl_GlobalInvocationID.y);
75
    int gz = int(gl_GlobalInvocationID.z);
76

77
    if (gx >= psc(outw) || gy >= psc(outh) * psc(outd) || gz >= psc(outc))
78
        return;
79

80
    ivec4 i4;
81
    if (ndim == 1) i4 = gx * 4 + ivec4(0, 1, 2, 3);
82
    if (ndim == 2) i4 = (gy * 4) * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outw);
83
    if (ndim == 3) i4 = (gz * 4) * psc(outh) * psc(outw) + gy * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outh) * psc(outw);
84
    if (ndim == 4) i4 = (gz * 4) * psc(outd) * psc(outh) * psc(outw) + gy * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outd) * psc(outh) * psc(outw);
85

86
    ivec4 x4;
87
    ivec4 y4;
88
    ivec4 z4;
89

90
#if NCNN_image_shader
91
    ivec4 lane2;
92
    ivec4 lane4;
93
#else
94
    ivec4 v_offset;
95
#if NCNN_fp16_packed
96
    ivec4 lane2;
97
#endif
98
#endif
99

100
    if (psc(dims) == 1)
101
    {
102
        z4 = ivec4(0);
103
        y4 = ivec4(0);
104
        x4 = i4;
105

106
#if NCNN_image_shader
107
        lane2 = (x4 % 8) / 4;
108
        lane4 = x4 % 4;
109
        x4 = x4 / 8;
110
#else
111
#if NCNN_fp16_packed
112
        v_offset = i4 / 2;
113
        lane2 = i4 % 2;
114
#else
115
        v_offset = i4;
116
#endif
117
#endif
118
    }
119
    else if (psc(dims) == 2)
120
    {
121
        z4 = ivec4(0);
122
        y4 = i4 / psc(w);
123
        x4 = i4 % psc(w);
124

125
#if NCNN_image_shader
126
        lane2 = (y4 % 8) / 4;
127
        lane4 = y4 % 4;
128
        y4 = y4 / 8;
129
#else
130
#if NCNN_fp16_packed
131

132
        v_offset = ((y4 / 8) * psc(w) + x4) * 4 + (y4 % 8) / 2;
133
        lane2 = y4 % 2;
134
#else
135
        v_offset = ((y4 / 8) * psc(w) + x4) * 8 + y4 % 8;
136
#endif
137
#endif
138
    }
139
    else if (psc(dims) == 3)
140
    {
141
        int size = psc(w) * psc(h);
142

143
        z4 = i4 / size;
144
        y4 = i4 % size / psc(w);
145
        x4 = i4 % size % psc(w);
146

147
#if NCNN_image_shader
148
        lane2 = (z4 % 8) / 4;
149
        lane4 = z4 % 4;
150
        z4 = z4 / 8;
151
#else
152
#if NCNN_fp16_packed
153
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 4 + (z4 % 8) / 2;
154
        lane2 = z4 % 2;
155
#else
156
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 8 + z4 % 8;
157
#endif
158
#endif
159
    }
160
    else // if (psc(dims) == 4)
161
    {
162
        int size = psc(w) * psc(h) * psc(d);
163
        int dsize = psc(w) * psc(h);
164

165
        z4 = i4 / size;
166
        ivec4 yd4 = i4 % size / dsize;
167
        ivec4 yh4 = i4 % size % dsize / psc(w);
168
        x4 = i4 % size % dsize % psc(w);
169

170
        y4 = yd4 * psc(h) + yh4;
171

172
#if NCNN_image_shader
173
        lane2 = (z4 % 8) / 4;
174
        lane4 = z4 % 4;
175
        z4 = z4 / 8;
176
#else
177
#if NCNN_fp16_packed
178
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 4 + (z4 % 8) / 2;
179
        lane2 = z4 % 2;
180
#else
181
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 8 + z4 % 8;
182
#endif
183
#endif
184
    }
185

186
#if NCNN_image_shader
187
    afpvec8 v0 = image3d_ld8(bottom_blob_3d, ivec3(x4.r, y4.r, z4.r));
188
    afpvec8 v1 = image3d_ld8(bottom_blob_3d, ivec3(x4.g, y4.g, z4.g));
189
    afpvec8 v2 = image3d_ld8(bottom_blob_3d, ivec3(x4.b, y4.b, z4.b));
190
    afpvec8 v3 = image3d_ld8(bottom_blob_3d, ivec3(x4.a, y4.a, z4.a));
191

192
    afpvec4 v;
193
    v.r = v0[lane2.r][lane4.r];
194
    v.g = v1[lane2.g][lane4.g];
195
    v.b = v2[lane2.b][lane4.b];
196
    v.a = v3[lane2.a][lane4.a];
197

198
    if (ndim == 1) image3d_st4(top_blob_3d, ivec3(gx, 0, 0), v);
199
    if (ndim == 2) image3d_st4(top_blob_3d, ivec3(gx, gy, 0), v);
200
    if (ndim == 3 || ndim == 4) image3d_st4(top_blob_3d, ivec3(gx, gy, gz), v);
201
#else
202
    int gi;
203
    if (ndim == 1) gi = gx;
204
    if (ndim == 2) gi = gy * psc(outw) + gx;
205
    if (ndim == 3 || ndim == 4) gi = gz * psc(outcstep) + gy * psc(outw) + gx;
206

207
#if NCNN_fp16_packed
208
    afpvec2 vr = buffer_ld2(bottom_blob_data, v_offset.r);
209
    afpvec2 vg = buffer_ld2(bottom_blob_data, v_offset.g);
210
    afpvec2 vb = buffer_ld2(bottom_blob_data, v_offset.b);
211
    afpvec2 va = buffer_ld2(bottom_blob_data, v_offset.a);
212

213
    afpvec4 v = afpvec4(vr[lane2.r], vg[lane2.g], vb[lane2.b], va[lane2.a]);
214

215
    buffer_st4(top_blob_data, gi, v);
216
#else
217
    buffer_cp1to4(top_blob_data, gi, bottom_blob_data, v_offset);
218
#endif
219
#endif
220
}
221