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
#include "flatten_vulkan.h"
16

17
#include "layer_shader_type.h"
18

19
namespace ncnn {
20

21
Flatten_vulkan::Flatten_vulkan()
22
{
23
    support_vulkan = true;
24
    support_image_storage = true;
25

26
    pipeline_flatten = 0;
27
    pipeline_flatten_pack4 = 0;
28
    pipeline_flatten_pack1to4 = 0;
29
    pipeline_flatten_pack8 = 0;
30
    pipeline_flatten_pack1to8 = 0;
31
    pipeline_flatten_pack4to8 = 0;
32
}
33

34
int Flatten_vulkan::create_pipeline(const Option& _opt)
35
{
36
    Option opt = _opt;
37
    const Mat& shape = bottom_shapes.empty() ? Mat() : bottom_shapes[0];
38
    const Mat& out_shape = top_shapes.empty() ? Mat() : top_shapes[0];
39

40
    int elempack = 1;
41
    if (shape.dims == 1) elempack = opt.use_shader_pack8 && shape.w % 8 == 0 ? 8 : shape.w % 4 == 0 ? 4 : 1;
42
    if (shape.dims == 2) elempack = opt.use_shader_pack8 && shape.h % 8 == 0 ? 8 : shape.h % 4 == 0 ? 4 : 1;
43
    if (shape.dims == 3 || shape.dims == 4) elempack = opt.use_shader_pack8 && shape.c % 8 == 0 ? 8 : shape.c % 4 == 0 ? 4 : 1;
44

45
    int out_elempack = 1;
46
    if (out_shape.dims == 1) out_elempack = opt.use_shader_pack8 && out_shape.w % 8 == 0 ? 8 : out_shape.w % 4 == 0 ? 4 : 1;
47

48
    size_t elemsize;
49
    size_t out_elemsize;
50
    if (opt.use_fp16_storage)
51
    {
52
        elemsize = elempack * 2u;
53
        out_elemsize = out_elempack * 2u;
54
    }
55
    else if (opt.use_fp16_packed)
56
    {
57
        elemsize = elempack == 1 ? 4u : elempack * 2u;
58
        out_elemsize = out_elempack == 1 ? 4u : out_elempack * 2u;
59
    }
60
    else
61
    {
62
        elemsize = elempack * 4u;
63
        out_elemsize = out_elempack * 4u;
64
    }
65

66
    Mat shape_packed;
67
    if (shape.dims == 1) shape_packed = Mat(shape.w / elempack, (void*)0, elemsize, elempack);
68
    if (shape.dims == 2) shape_packed = Mat(shape.w, shape.h / elempack, (void*)0, elemsize, elempack);
69
    if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
70
    if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
71

72
    Mat out_shape_packed;
73
    if (out_shape.dims == 1) out_shape_packed = Mat(out_shape.w / out_elempack, (void*)0, out_elemsize, out_elempack);
74

75
    if (!vkdev->shape_support_image_storage(shape_packed) || !vkdev->shape_support_image_storage(out_shape_packed))
76
    {
77
        support_image_storage = false;
78
        opt.use_image_storage = false;
79
    }
80

81
    std::vector<vk_specialization_type> specializations(0 + 10);
82
    specializations[0 + 0].i = std::min(3, shape_packed.dims);
83
    specializations[0 + 1].i = shape_packed.w;
84
    specializations[0 + 2].i = shape_packed.h * shape_packed.d;
85
    specializations[0 + 3].i = shape_packed.c;
86
    specializations[0 + 4].i = shape_packed.cstep;
87
    specializations[0 + 5].i = std::min(3, out_shape_packed.dims);
88
    specializations[0 + 6].i = out_shape_packed.w;
89
    specializations[0 + 7].i = out_shape_packed.h * out_shape_packed.d;
90
    specializations[0 + 8].i = out_shape_packed.c;
91
    specializations[0 + 9].i = out_shape_packed.cstep;
92

93
    Mat local_size_xyz(64, 1, 1, (void*)0);
94
    if (out_shape_packed.dims != 0)
95
    {
96
        local_size_xyz.w = std::min(64, out_shape_packed.w);
97
        local_size_xyz.h = 1;
98
        local_size_xyz.c = 1;
99
    }
100

101
    // pack1
102
    if (shape.dims == 0 || (elempack == 1 && out_elempack == 1))
103
    {
104
        pipeline_flatten = new Pipeline(vkdev);
105
        pipeline_flatten->set_optimal_local_size_xyz(local_size_xyz);
106
        pipeline_flatten->create(LayerShaderType::flatten, opt, specializations);
107
    }
108

109
    // pack4
110
    if (shape.dims == 0 || (elempack == 4 && out_elempack == 4))
111
    {
112
        pipeline_flatten_pack4 = new Pipeline(vkdev);
113
        pipeline_flatten_pack4->set_optimal_local_size_xyz(local_size_xyz);
114
        pipeline_flatten_pack4->create(LayerShaderType::flatten_pack4, opt, specializations);
115
    }
116

117
    // pack1to4
118
    if (shape.dims == 0 || (elempack == 1 && out_elempack == 4))
119
    {
120
        pipeline_flatten_pack1to4 = new Pipeline(vkdev);
121
        pipeline_flatten_pack1to4->set_optimal_local_size_xyz(local_size_xyz);
122
        pipeline_flatten_pack1to4->create(LayerShaderType::flatten_pack1to4, opt, specializations);
123
    }
124

125
    // pack8
126
    if ((opt.use_shader_pack8 && shape.dims == 0) || (elempack == 8 && out_elempack == 8))
127
    {
128
        pipeline_flatten_pack8 = new Pipeline(vkdev);
129
        pipeline_flatten_pack8->set_optimal_local_size_xyz(local_size_xyz);
130
        pipeline_flatten_pack8->create(LayerShaderType::flatten_pack8, opt, specializations);
131
    }
132

133
    // pack1to8
134
    if ((opt.use_shader_pack8 && shape.dims == 0) || (elempack == 1 && out_elempack == 8))
135
    {
136
        pipeline_flatten_pack1to8 = new Pipeline(vkdev);
137
        pipeline_flatten_pack1to8->set_optimal_local_size_xyz(local_size_xyz);
138
        pipeline_flatten_pack1to8->create(LayerShaderType::flatten_pack1to8, opt, specializations);
139
    }
140

141
    // pack4to8
142
    if ((opt.use_shader_pack8 && shape.dims == 0) || (elempack == 4 && out_elempack == 8))
143
    {
144
        pipeline_flatten_pack4to8 = new Pipeline(vkdev);
145
        pipeline_flatten_pack4to8->set_optimal_local_size_xyz(local_size_xyz);
146
        pipeline_flatten_pack4to8->create(LayerShaderType::flatten_pack4to8, opt, specializations);
147
    }
148

149
    return 0;
150
}
151

152
int Flatten_vulkan::destroy_pipeline(const Option& /*opt*/)
153
{
154
    delete pipeline_flatten;
155
    pipeline_flatten = 0;
156

157
    delete pipeline_flatten_pack4;
158
    pipeline_flatten_pack4 = 0;
159

160
    delete pipeline_flatten_pack1to4;
161
    pipeline_flatten_pack1to4 = 0;
162

163
    delete pipeline_flatten_pack8;
164
    pipeline_flatten_pack8 = 0;
165

166
    delete pipeline_flatten_pack1to8;
167
    pipeline_flatten_pack1to8 = 0;
168

169
    delete pipeline_flatten_pack4to8;
170
    pipeline_flatten_pack4to8 = 0;
171

172
    return 0;
173
}
174

175
int Flatten_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const
176
{
177
    int dims = bottom_blob.dims;
178

179
    if (dims == 1)
180
    {
181
        top_blob = bottom_blob;
182
        return 0;
183
    }
184

185
    int w = bottom_blob.w;
186
    int h = bottom_blob.h;
187
    int d = bottom_blob.d;
188
    int channels = bottom_blob.c;
189
    size_t elemsize = bottom_blob.elemsize;
190
    int elempack = bottom_blob.elempack;
191

192
    int total = w * h * d * channels * elempack;
193

194
    int out_elempack = opt.use_shader_pack8 && total % 8 == 0 ? 8 : total % 4 == 0 ? 4 : 1;
195
    size_t out_elemsize = elemsize / elempack * out_elempack;
196

197
    if (opt.use_fp16_packed && !opt.use_fp16_storage)
198
    {
199
        if (out_elempack == 8) out_elemsize = 8 * 2u;
200
        if (out_elempack == 4) out_elemsize = 4 * 2u;
201
        if (out_elempack == 1) out_elemsize = 4u;
202
    }
203

204
    if (dims == 2 && elempack == 1 && !(opt.use_fp16_packed && !opt.use_fp16_storage && out_elempack != 1))
205
    {
206
        top_blob = bottom_blob;
207
        top_blob.dims = 1;
208
        top_blob.w = total / out_elempack;
209
        top_blob.h = 1;
210
        top_blob.cstep = top_blob.w;
211
        top_blob.elemsize = out_elemsize;
212
        top_blob.elempack = out_elempack;
213
        return 0;
214
    }
215

216
    top_blob.create(total / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
217
    if (top_blob.empty())
218
        return -100;
219

220
    std::vector<VkMat> bindings(2);
221
    bindings[0] = bottom_blob;
222
    bindings[1] = top_blob;
223

224
    std::vector<vk_constant_type> constants(10);
225
    constants[0].i = std::min(3, bottom_blob.dims);
226
    constants[1].i = bottom_blob.w;
227
    constants[2].i = bottom_blob.h * bottom_blob.d;
228
    constants[3].i = bottom_blob.c;
229
    constants[4].i = bottom_blob.cstep;
230
    constants[5].i = std::min(3, top_blob.dims);
231
    constants[6].i = top_blob.w;
232
    constants[7].i = top_blob.h * top_blob.d;
233
    constants[8].i = top_blob.c;
234
    constants[9].i = top_blob.cstep;
235

236
    const Pipeline* pipeline = 0;
237
    if (elempack == 1 && out_elempack == 1)
238
    {
239
        pipeline = pipeline_flatten;
240
    }
241
    else if (elempack == 4 && out_elempack == 4)
242
    {
243
        pipeline = pipeline_flatten_pack4;
244
    }
245
    else if (elempack == 1 && out_elempack == 4)
246
    {
247
        pipeline = pipeline_flatten_pack1to4;
248
    }
249
    else if (elempack == 8 /*&& out_elempack == 8*/)
250
    {
251
        pipeline = pipeline_flatten_pack8;
252
    }
253
    else if (elempack == 1 && out_elempack == 8)
254
    {
255
        pipeline = pipeline_flatten_pack1to8;
256
    }
257
    else if (elempack == 4 && out_elempack == 8)
258
    {
259
        pipeline = pipeline_flatten_pack4to8;
260
    }
261

262
    cmd.record_pipeline(pipeline, bindings, constants, top_blob);
263

264
    return 0;
265
}
266

267
int Flatten_vulkan::forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const
268
{
269
    int dims = bottom_blob.dims;
270

271
    if (dims == 1)
272
    {
273
        top_blob = bottom_blob;
274
        return 0;
275
    }
276

277
    int w = bottom_blob.w;
278
    int h = bottom_blob.h;
279
    int d = bottom_blob.d;
280
    int channels = bottom_blob.c;
281
    size_t elemsize = bottom_blob.elemsize;
282
    int elempack = bottom_blob.elempack;
283

284
    int total = w * h * d * channels * elempack;
285

286
    int out_elempack = opt.use_shader_pack8 && total % 8 == 0 ? 8 : total % 4 == 0 ? 4 : 1;
287
    size_t out_elemsize = elemsize / elempack * out_elempack;
288

289
    if (opt.use_fp16_packed && !opt.use_fp16_storage)
290
    {
291
        if (out_elempack == 8) out_elemsize = 8 * 2u;
292
        if (out_elempack == 4) out_elemsize = 4 * 2u;
293
        if (out_elempack == 1) out_elemsize = 4u;
294
    }
295

296
    top_blob.create(total / out_elempack, out_elemsize, out_elempack, opt.blob_vkallocator);
297
    if (top_blob.empty())
298
        return -100;
299

300
    std::vector<VkImageMat> bindings(2);
301
    bindings[0] = bottom_blob;
302
    bindings[1] = top_blob;
303

304
    std::vector<vk_constant_type> constants(10);
305
    constants[0].i = std::min(3, bottom_blob.dims);
306
    constants[1].i = bottom_blob.w;
307
    constants[2].i = bottom_blob.h * bottom_blob.d;
308
    constants[3].i = bottom_blob.c;
309
    constants[4].i = 0; //bottom_blob.cstep;
310
    constants[5].i = std::min(3, top_blob.dims);
311
    constants[6].i = top_blob.w;
312
    constants[7].i = top_blob.h * top_blob.d;
313
    constants[8].i = top_blob.c;
314
    constants[9].i = 0; //top_blob.cstep;
315

316
    const Pipeline* pipeline = 0;
317
    if (elempack == 1 && out_elempack == 1)
318
    {
319
        pipeline = pipeline_flatten;
320
    }
321
    else if (elempack == 4 && out_elempack == 4)
322
    {
323
        pipeline = pipeline_flatten_pack4;
324
    }
325
    else if (elempack == 1 && out_elempack == 4)
326
    {
327
        pipeline = pipeline_flatten_pack1to4;
328
    }
329
    else if (elempack == 8 /*&& out_elempack == 8*/)
330
    {
331
        pipeline = pipeline_flatten_pack8;
332
    }
333
    else if (elempack == 1 && out_elempack == 8)
334
    {
335
        pipeline = pipeline_flatten_pack1to8;
336
    }
337
    else if (elempack == 4 && out_elempack == 8)
338
    {
339
        pipeline = pipeline_flatten_pack4to8;
340
    }
341

342
    cmd.record_pipeline(pipeline, bindings, constants, top_blob);
343

344
    return 0;
345
}
346

347
} // namespace ncnn
348