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 { sfpvec8 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
    ivec4 ii4;
82
    if (ndim == 1)
83
    {
84
        i4 = gx * 8 + ivec4(0, 1, 2, 3);
85
        ii4 = i4 + 4;
86
    }
87
    if (ndim == 2)
88
    {
89
        i4 = (gy * 8) * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outw);
90
        ii4 = i4 + 4 * psc(outw);
91
    }
92
    if (ndim == 3)
93
    {
94
        i4 = (gz * 8) * psc(outh) * psc(outw) + gy * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outh) * psc(outw);
95
        ii4 = i4 + 4 * psc(outh) * psc(outw);
96
    }
97
    if (ndim == 4)
98
    {
99
        i4 = (gz * 8) * psc(outd) * psc(outh) * psc(outw) + gy * psc(outw) + gx + ivec4(0, 1, 2, 3) * psc(outd) * psc(outh) * psc(outw);
100
        ii4 = i4 + 4 * psc(outd) * psc(outh) * psc(outw);
101
    }
102

103
    ivec4 x4;
104
    ivec4 xx4;
105
    ivec4 y4;
106
    ivec4 yy4;
107
    ivec4 z4;
108
    ivec4 zz4;
109

110
#if NCNN_image_shader
111
    ivec4 lane2;
112
    ivec4 lane2_1;
113
    ivec4 lane4;
114
    ivec4 lane4_1;
115
#else
116
    ivec4 v_offset;
117
    ivec4 vv_offset;
118
#if NCNN_fp16_packed
119
    ivec4 lane2;
120
    ivec4 lane4;
121
#endif
122
#endif
123

124
    if (psc(dims) == 1)
125
    {
126
        z4 = ivec4(0);
127
        y4 = ivec4(0);
128
        x4 = i4;
129
        zz4 = ivec4(0);
130
        yy4 = ivec4(0);
131
        xx4 = ii4;
132

133
#if NCNN_image_shader
134
        lane2 = (x4 % 8) / 4;
135
        lane2_1 = (xx4 % 8) / 4;
136
        lane4 = x4 % 4;
137
        lane4_1 = xx4 % 4;
138
        x4 = x4 / 8;
139
        xx4 = xx4 / 8;
140
#else
141
#if NCNN_fp16_packed
142
        v_offset = i4 / 2;
143
        lane2 = i4 % 2;
144
        vv_offset = ii4 / 2;
145
        lane4 = ii4 % 2;
146
#else
147
        v_offset = i4;
148
        vv_offset = ii4;
149
#endif
150
#endif
151
    }
152
    else if (psc(dims) == 2)
153
    {
154
        z4 = ivec4(0);
155
        y4 = i4 / psc(w);
156
        x4 = i4 % psc(w);
157
        zz4 = ivec4(0);
158
        yy4 = ii4 / psc(w);
159
        xx4 = ii4 % psc(w);
160

161
#if NCNN_image_shader
162
        lane2 = (y4 % 8) / 4;
163
        lane2_1 = (yy4 % 8) / 4;
164
        lane4 = y4 % 4;
165
        lane4_1 = yy4 % 4;
166
        y4 = y4 / 8;
167
        yy4 = yy4 / 8;
168
#else
169
#if NCNN_fp16_packed
170
        v_offset = ((y4 / 8) * psc(w) + x4) * 4 + (y4 % 8) / 2;
171
        lane2 = y4 % 2;
172
        vv_offset = ((yy4 / 8) * psc(w) + xx4) * 4 + (yy4 % 8) / 2;
173
        lane4 = yy4 % 2;
174
#else
175
        v_offset = ((y4 / 8) * psc(w) + x4) * 8 + y4 % 8;
176
        vv_offset = ((yy4 / 8) * psc(w) + xx4) * 8 + yy4 % 8;
177
#endif
178
#endif
179
    }
180
    else if (psc(dims) == 3)
181
    {
182
        int size = psc(w) * psc(h);
183

184
        z4 = i4 / size;
185
        y4 = i4 % size / psc(w);
186
        x4 = i4 % size % psc(w);
187
        zz4 = ii4 / size;
188
        yy4 = ii4 % size / psc(w);
189
        xx4 = ii4 % size % psc(w);
190

191
#if NCNN_image_shader
192
        lane2 = (z4 % 8) / 4;
193
        lane2_1 = (zz4 % 8) / 4;
194
        lane4 = z4 % 4;
195
        lane4_1 = zz4 % 4;
196
        z4 = z4 / 8;
197
        zz4 = zz4 / 8;
198
#else
199
#if NCNN_fp16_packed
200
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 4 + (z4 % 8) / 2;
201
        lane2 = z4 % 2;
202
        vv_offset = ((zz4 / 8) * psc(cstep) + yy4 * psc(w) + xx4) * 4 + (zz4 % 8) / 2;
203
        lane4 = zz4 % 2;
204
#else
205
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 8 + z4 % 8;
206
        vv_offset = ((zz4 / 8) * psc(cstep) + yy4 * psc(w) + xx4) * 8 + zz4 % 8;
207
#endif
208
#endif
209
    }
210
    else // if (psc(dims) == 4)
211
    {
212
        int size = psc(w) * psc(h) * psc(d);
213
        int dsize = psc(w) * psc(h);
214

215
        z4 = i4 / size;
216
        ivec4 yd4 = i4 % size / dsize;
217
        ivec4 yh4 = i4 % size % dsize / psc(w);
218
        x4 = i4 % size % dsize % psc(w);
219
        zz4 = ii4 / size;
220
        ivec4 yyd4 = ii4 % size / dsize;
221
        ivec4 yyh4 = ii4 % size % dsize / psc(w);
222
        xx4 = ii4 % size % dsize % psc(w);
223

224
        y4 = yd4 * psc(h) + yh4;
225
        yy4 = yyd4 * psc(h) + yyh4;
226

227
#if NCNN_image_shader
228
        lane2 = (z4 % 8) / 4;
229
        lane2_1 = (zz4 % 8) / 4;
230
        lane4 = z4 % 4;
231
        lane4_1 = zz4 % 4;
232
        z4 = z4 / 8;
233
        zz4 = zz4 / 8;
234
#else
235
#if NCNN_fp16_packed
236
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 4 + (z4 % 8) / 2;
237
        lane2 = z4 % 2;
238
        vv_offset = ((zz4 / 8) * psc(cstep) + yy4 * psc(w) + xx4) * 4 + (zz4 % 8) / 2;
239
        lane4 = zz4 % 2;
240
#else
241
        v_offset = ((z4 / 8) * psc(cstep) + y4 * psc(w) + x4) * 8 + z4 % 8;
242
        vv_offset = ((zz4 / 8) * psc(cstep) + yy4 * psc(w) + xx4) * 8 + zz4 % 8;
243
#endif
244
#endif
245
    }
246

247
#if NCNN_image_shader
248
    afpvec8 v0 = image3d_ld8(bottom_blob_3d, ivec3(x4.r, y4.r, z4.r));
249
    afpvec8 v1 = image3d_ld8(bottom_blob_3d, ivec3(x4.g, y4.g, z4.g));
250
    afpvec8 v2 = image3d_ld8(bottom_blob_3d, ivec3(x4.b, y4.b, z4.b));
251
    afpvec8 v3 = image3d_ld8(bottom_blob_3d, ivec3(x4.a, y4.a, z4.a));
252
    afpvec8 v4 = image3d_ld8(bottom_blob_3d, ivec3(xx4.r, yy4.r, zz4.r));
253
    afpvec8 v5 = image3d_ld8(bottom_blob_3d, ivec3(xx4.g, yy4.g, zz4.g));
254
    afpvec8 v6 = image3d_ld8(bottom_blob_3d, ivec3(xx4.b, yy4.b, zz4.b));
255
    afpvec8 v7 = image3d_ld8(bottom_blob_3d, ivec3(xx4.a, yy4.a, zz4.a));
256

257
    afpvec8 v;
258
    v[0].r = v0[lane2.r][lane4.r];
259
    v[0].g = v1[lane2.g][lane4.g];
260
    v[0].b = v2[lane2.b][lane4.b];
261
    v[0].a = v3[lane2.a][lane4.a];
262
    v[1].r = v4[lane2_1.r][lane4_1.r];
263
    v[1].g = v5[lane2_1.g][lane4_1.g];
264
    v[1].b = v6[lane2_1.b][lane4_1.b];
265
    v[1].a = v7[lane2_1.a][lane4_1.a];
266

267
    if (ndim == 1) image3d_st8(top_blob_3d, ivec3(gx, 0, 0), v);
268
    if (ndim == 2) image3d_st8(top_blob_3d, ivec3(gx, gy, 0), v);
269
    if (ndim == 3 || ndim == 4) image3d_st8(top_blob_3d, ivec3(gx, gy, gz), v);
270
#else
271
    int gi;
272
    if (ndim == 1) gi = gx;
273
    if (ndim == 2) gi = gy * psc(outw) + gx;
274
    if (ndim == 3 || ndim == 4) gi = gz * psc(outcstep) + gy * psc(outw) + gx;
275

276
#if NCNN_fp16_packed
277
    afpvec2 vr = buffer_ld2(bottom_blob_data, v_offset.r);
278
    afpvec2 vg = buffer_ld2(bottom_blob_data, v_offset.g);
279
    afpvec2 vb = buffer_ld2(bottom_blob_data, v_offset.b);
280
    afpvec2 va = buffer_ld2(bottom_blob_data, v_offset.a);
281

282
    afpvec2 vvr = buffer_ld2(bottom_blob_data, vv_offset.r);
283
    afpvec2 vvg = buffer_ld2(bottom_blob_data, vv_offset.g);
284
    afpvec2 vvb = buffer_ld2(bottom_blob_data, vv_offset.b);
285
    afpvec2 vva = buffer_ld2(bottom_blob_data, vv_offset.a);
286

287
    afpvec8 v = afpvec8(vr[lane2.r], vg[lane2.g], vb[lane2.b], va[lane2.a], vvr[lane4.r], vvg[lane4.g], vvb[lane4.b], vva[lane4.a]);
288

289
    buffer_st8(top_blob_data, gi, v);
290
#else
291
    buffer_cp1to8(top_blob_data, gi, bottom_blob_data, v_offset, vv_offset);
292
#endif
293
#endif
294
}
295