ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2019 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
#endif
20
#if NCNN_fp16_arithmetic
21
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
22
#endif
23

24
layout (constant_id = 0) const int ndim = 0;
25

26
#define shape_constant_id_offset 1
27
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
28
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
29
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
30
layout (constant_id = shape_constant_id_offset + 3) const int d = 0;
31
layout (constant_id = shape_constant_id_offset + 4) const int c = 0;
32
layout (constant_id = shape_constant_id_offset + 5) const int cstep = 0;
33

34
layout (constant_id = shape_constant_id_offset + 6) const int outdims = 0;
35
layout (constant_id = shape_constant_id_offset + 7) const int outw = 0;
36
layout (constant_id = shape_constant_id_offset + 8) const int outh = 0;
37
layout (constant_id = shape_constant_id_offset + 9) const int outd = 0;
38
layout (constant_id = shape_constant_id_offset + 10) const int outc = 0;
39
layout (constant_id = shape_constant_id_offset + 11) const int outcstep = 0;
40

41
#if NCNN_image_shader
42
layout (binding = 0) uniform unfp sampler3D bottom_blob_3d;
43
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
44
#else
45
#if NCNN_fp16_packed
46
layout (binding = 0) readonly buffer bottom_blob { sfpvec2 bottom_blob_data[]; };
47
#else
48
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
49
#endif
50
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
51
#endif
52

53
layout (push_constant) uniform parameter
54
{
55
    int dims;
56
    int w;
57
    int h;
58
    int d;
59
    int c;
60
    int cstep;
61

62
    int outdims;
63
    int outw;
64
    int outh;
65
    int outd;
66
    int outc;
67
    int outcstep;
68
} p;
69

70
void main()
71
{
72
    int gx = int(gl_GlobalInvocationID.x);
73
    int gy = int(gl_GlobalInvocationID.y);
74
    int gz = int(gl_GlobalInvocationID.z);
75

76
    if (gx >= psc(outw) || gy >= psc(outh) * psc(outd) || gz >= psc(outc))
77
        return;
78

79
    ivec4 i4;
80
    if (ndim == 1) i4 = gx * 4 + ivec4(0, 1, 2, 3);
81
    if (ndim == 2) i4 = (gy * 4) * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outw);
82
    if (ndim == 3) i4 = (gz * 4) * psc(outh) * psc(outw) + gy * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outh) * psc(outw);
83
    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);
84

85
    ivec4 x4;
86
    ivec4 y4;
87
    ivec4 z4;
88

89
#if NCNN_image_shader
90
    ivec4 lane4;
91
#else
92
    ivec4 v_offset;
93
#if NCNN_fp16_packed
94
    ivec4 lane2;
95
#endif
96
#endif
97

98
    if (psc(dims) == 1)
99
    {
100
        z4 = ivec4(0);
101
        y4 = ivec4(0);
102
        x4 = i4;
103

104
#if NCNN_image_shader
105
        lane4 = x4 % 4;
106
        x4 = x4 / 4;
107
#else
108
#if NCNN_fp16_packed
109
        v_offset = i4 / 2;
110
        lane2 = i4 % 2;
111
#else
112
        v_offset = i4;
113
#endif
114
#endif
115
    }
116
    else if (psc(dims) == 2)
117
    {
118
        z4 = ivec4(0);
119
        y4 = i4 / psc(w);
120
        x4 = i4 % psc(w);
121

122
#if NCNN_image_shader
123
        lane4 = y4 % 4;
124
        y4 = y4 / 4;
125
#else
126
#if NCNN_fp16_packed
127
        v_offset = ((y4 / 4) * psc(w) + x4) * 2 + (y4 % 4) / 2;
128
        lane2 = y4 % 2;
129
#else
130
        v_offset = ((y4 / 4) * psc(w) + x4) * 4 + y4 % 4;
131
#endif
132
#endif
133
    }
134
    else if (psc(dims) == 3)
135
    {
136
        int size = psc(w) * psc(h);
137

138
        z4 = i4 / size;
139
        y4 = i4 % size / psc(w);
140
        x4 = i4 % size % psc(w);
141

142
#if NCNN_image_shader
143
        lane4 = z4 % 4;
144
        z4 = z4 / 4;
145
#else
146
#if NCNN_fp16_packed
147
        v_offset = ((z4 / 4) * psc(cstep) + y4 * psc(w) + x4) * 2 + (z4 % 4) / 2;
148
        lane2 = z4 % 2;
149
#else
150
        v_offset = ((z4 / 4) * psc(cstep) + y4 * psc(w) + x4) * 4 + z4 % 4;
151
#endif
152
#endif
153
    }
154
    else // if (psc(dims) == 4)
155
    {
156
        int size = psc(w) * psc(h) * psc(d);
157
        int dsize = psc(w) * psc(h);
158

159
        z4 = i4 / size;
160
        ivec4 yd4 = i4 % size / dsize;
161
        ivec4 yh4 = i4 % size % dsize / psc(w);
162
        x4 = i4 % size % dsize % psc(w);
163

164
        y4 = yd4 * psc(h) + yh4;
165

166
#if NCNN_image_shader
167
        lane4 = z4 % 4;
168
        z4 = z4 / 4;
169
#else
170
#if NCNN_fp16_packed
171
        v_offset = ((z4 / 4) * psc(cstep) + y4 * psc(w) + x4) * 2 + (z4 % 4) / 2;
172
        lane2 = z4 % 2;
173
#else
174
        v_offset = ((z4 / 4) * psc(cstep) + y4 * psc(w) + x4) * 4 + z4 % 4;
175
#endif
176
#endif
177
    }
178

179
#if NCNN_image_shader
180
    afpvec4 vr = image3d_ld4(bottom_blob_3d, ivec3(x4.r, y4.r, z4.r));
181
    afpvec4 vg = image3d_ld4(bottom_blob_3d, ivec3(x4.g, y4.g, z4.g));
182
    afpvec4 vb = image3d_ld4(bottom_blob_3d, ivec3(x4.b, y4.b, z4.b));
183
    afpvec4 va = image3d_ld4(bottom_blob_3d, ivec3(x4.a, y4.a, z4.a));
184

185
    afpvec4 v = afpvec4(vr[lane4.r], vg[lane4.g], vb[lane4.b], va[lane4.a]);
186

187
    if (ndim == 1) image3d_st4(top_blob_3d, ivec3(gx, 0, 0), v);
188
    if (ndim == 2) image3d_st4(top_blob_3d, ivec3(gx, gy, 0), v);
189
    if (ndim == 3 || ndim == 4) image3d_st4(top_blob_3d, ivec3(gx, gy, gz), v);
190
#else
191
    int gi;
192
    if (ndim == 1) gi = gx;
193
    if (ndim == 2) gi = gy * psc(outw) + gx;
194
    if (ndim == 3 || ndim == 4) gi = gz * psc(outcstep) + gy * psc(outw) + gx;
195

196
#if NCNN_fp16_packed
197
    afpvec2 vr = buffer_ld2(bottom_blob_data, v_offset.r);
198
    afpvec2 vg = buffer_ld2(bottom_blob_data, v_offset.g);
199
    afpvec2 vb = buffer_ld2(bottom_blob_data, v_offset.b);
200
    afpvec2 va = buffer_ld2(bottom_blob_data, v_offset.a);
201

202
    afpvec4 v = afpvec4(vr[lane2.r], vg[lane2.g], vb[lane2.b], va[lane2.a]);
203

204
    buffer_st4(top_blob_data, gi, v);
205
#else
206
    buffer_cp1to4(top_blob_data, gi, bottom_blob_data, v_offset);
207
#endif
208
#endif
209
}
210