lod_tensor.cc 14.3 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2

L
Luo Tao 已提交
3 4 5
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
8

L
Luo Tao 已提交
9 10 11 12 13
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
14

F
fengjiayi 已提交
15 16 17 18 19
#include <stdint.h>
#include <string.h>
#include <algorithm>
#include <iterator>

Y
Yi Wang 已提交
20 21
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/framework.pb.h"
F
fengjiayi 已提交
22
#include "paddle/fluid/framework/lod_tensor.h"
S
sneaxiy 已提交
23
#include "paddle/fluid/framework/var_type.h"
24

Y
Yi Wang 已提交
25 26
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/memory/memory.h"
27

D
dzhwinter 已提交
28
#if defined(_WIN32)
Y
Yu Yang 已提交
29 30
#include "paddle/fluid/recordio/scanner.h"
#include "paddle/fluid/recordio/writer.h"
D
dzhwinter 已提交
31
#endif  // _WIN32
Y
Yu Yang 已提交
32

33 34 35
namespace paddle {
namespace framework {

武毅 已提交
36
std::ostream &operator<<(std::ostream &os, const LoD &lod) {
37
  os << "{";
武毅 已提交
38
  for (auto &v : lod) {
39
    os << "{";
L
Liu Yiqun 已提交
40
    bool is_first = true;
武毅 已提交
41
    for (auto &i : v) {
L
Liu Yiqun 已提交
42 43 44 45 46 47
      if (is_first) {
        os << i;
        is_first = false;
      } else {
        os << ", " << i;
      }
48 49 50 51 52 53 54 55
    }
    os << "}";
  }
  os << "}";

  return os;
}

Y
Yang Yang 已提交
56
std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
57 58
  if (!platform::is_cpu_place(t.place())) {
    LoDTensor tt;
Y
Yi Wang 已提交
59
    framework::TensorCopy(t, platform::CPUPlace(), &tt);
60 61 62 63 64 65 66 67
    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
    auto &dev_ctx = *pool.Get(t.place());
    dev_ctx.Wait();

    os << tt;
    return os;
  }

Y
Yang Yang 已提交
68 69 70 71 72 73
  os << "dim: " << t.dims() << "\n";
  os << "lod: " << t.lod() << "\n";

  // only print first ten elements
  int64_t size = t.numel() < 10 ? t.numel() : 10;
  for (int64_t i = 0; i < size; ++i) {
S
sneaxiy 已提交
74
    if (IsType<float>(t.type())) {
75
      os << t.data<float>()[i] << " ";
S
sneaxiy 已提交
76
    } else if (IsType<int64_t>(t.type())) {
77 78 79 80
      os << t.data<int64_t>()[i] << " ";
    } else {
      PADDLE_THROW("LoDTensor data type not in [float, int64_t]");
    }
Y
Yang Yang 已提交
81 82 83 84 85
  }

  return os;
}

Q
Qiao Longfei 已提交
86 87 88 89 90 91
std::string LoDToString(const LoD &lod) {
  std::ostringstream stream;
  stream << lod;
  return stream.str();
}

武毅 已提交
92
LoD SliceInLevel(const LoD &in, size_t level, size_t elem_begin,
Q
qijun 已提交
93
                 size_t elem_end) {
94
  PADDLE_ENFORCE_LT(level, in.size());
95
  PADDLE_ENFORCE_LT(elem_begin, elem_end);
96 97 98 99 100 101 102 103
  PADDLE_ENFORCE_LT(elem_end, in[level].size());

  LoD res;
  res.resize(in.size() - level);
  // copy the first level
  res[0].assign(in[level].begin() + elem_begin,
                in[level].begin() + elem_end + 1);
  for (size_t lvl = 1; lvl < res.size(); lvl++) {
武毅 已提交
104 105 106
    const auto &in_level = in[level + lvl];
    const auto &above_level = res[lvl - 1];
    auto &out_level = res[lvl];
107 108
    out_level.assign(in_level.begin() + above_level.front(),
                     in_level.begin() + above_level.back() + 1);
109
  }
110 111 112 113
  for (size_t lvl = 0; lvl < res.size(); lvl++) {
    // to make the first offset equals 0, all the elements minus the first
    // element
    size_t front = res[lvl].front();
武毅 已提交
114
    for (auto &ele : res[lvl]) {
115 116 117 118 119 120
      ele -= front;
    }
  }
  return res;
}

武毅 已提交
121
LoD ToAbsOffset(const LoD &in) {
122 123 124
  // the lowest level stores relative offsets
  if (in.empty() || in.size() == 1) return in;
  LoD result = in;
Q
Qiao Longfei 已提交
125 126 127 128
  for (auto level = static_cast<int>(in.size() - 2); level >= 0; level--) {
    for (size_t i = 0; i < in[level].size(); ++i) {
      size_t index = in[level][i];
      result[level][i] = result[level + 1][index];
129 130 131
    }
  }
  return result;
132 133
}

武毅 已提交
134
bool operator==(const LoD &a, const LoD &b) {
135 136 137 138 139
  if (a.size() != b.size()) {
    return false;
  }

  for (size_t i = 0; i < a.size(); i++) {
武毅 已提交
140 141
    const auto &a_level = a[i];
    const auto &b_level = b[i];
142 143 144 145 146 147 148 149 150 151
    if (a_level.size() != b_level.size()) {
      return false;
    }
    for (size_t j = 0; j < a_level.size(); j++) {
      if (a_level[j] != b_level[j]) {
        return false;
      }
    }
  }
  return true;
152 153
}

Y
Yan Chunwei 已提交
154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212
bool CheckLoD(const LoD &in, int tensor_height) {
  if (in.empty()) return true;
  for (const auto &level : in) {
    // check: there should be more than 2 offsets existing in each level.
    if (level.size() < 2) return false;
    // check: the first offset(the begin offset) of each level should be 0.
    if (level.front() != 0) return false;
    // check: all the offsets in a level should be ascending(no same items
    // allows).
    if (!std::is_sorted(level.begin(), level.begin(), [](size_t a, size_t b) {
          if (a < b) return true;
          return false;
        })) {
      LOG(INFO) << "ascending error";
      return false;
    }
  }
  // check: the lowest level's last offset should equals `tensor_height` if
  //        tensor_height>0.
  if (tensor_height > 0 && (size_t)tensor_height != in.back().back())
    return false;

  // check: the higher level's last offset should equals the lower level's
  // size-1.
  // NOTE LoD store the levels from top to bottom, so the higher level goes
  // first.
  for (size_t level = 0; level < in.size() - 1; level++) {
    if (in[level].back() != in[level + 1].size() - 1) return false;
  }
  return true;
}

bool CheckAbsLoD(const LoD &in, int tensor_height) {
  if (in.empty()) return true;
  for (const auto &level : in) {
    // check: all the offsets in a level should be ascending(no same items
    // allows).
    if (!std::is_sorted(level.begin(), level.begin(), [](size_t a, size_t b) {
          if (a < b) return true;
          return false;
        })) {
      return false;
    }

    // check: there should be more than 2 offsets existing in each level.
    if (level.size() < 2) return false;

    // check: the first offset of each level should be 0, and the last should be
    // the same(the height of underlying tensor).
    if (level.front() != 0) return false;
    if (tensor_height < 0) {
      tensor_height = level.back();
    } else if ((size_t)tensor_height != level.back()) {
      return false;
    }
  }
  return true;
}

213
using LoDAndOffset = std::pair<LoD, std::pair<size_t, size_t>>;
武毅 已提交
214
LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD &lod, size_t start_idx,
215 216 217 218 219 220
                                        size_t end_idx, size_t start_level) {
  LoD sub_lod;

  for (size_t level_idx = start_level; level_idx < lod.size(); ++level_idx) {
    PADDLE_ENFORCE_LE(start_idx, end_idx);
    PADDLE_ENFORCE_LT(end_idx, lod[level_idx].size());
221 222 223 224
    std::vector<size_t> level_lens;
    for (size_t i = start_idx; i < end_idx; ++i) {
      level_lens.push_back(lod[level_idx][i + 1] - lod[level_idx][i]);
    }
225
    sub_lod.emplace_back(level_lens);
226 227 228
    start_idx = lod[level_idx][start_idx];
    end_idx = lod[level_idx][end_idx];
  }
229 230

  return LoDAndOffset{sub_lod, {start_idx, end_idx}};
231 232
}

武毅 已提交
233
void AppendLoD(LoD *lod, const LoD &lod_length) {
234 235
  PADDLE_ENFORCE(
      lod->empty() || lod->size() == lod_length.size(),
236
      "The lod_length should has the same size with the appended lod.");
237
  if (lod->empty()) {
Y
Yang Yu 已提交
238 239 240
    for (size_t i = 0; i < lod_length.size(); ++i) {
      lod->emplace_back(1, 0);  // size = 1, value = 0;
    }
241 242
    *lod = LoD(lod_length.size(), std::vector<size_t>({0}));
  }
243
  for (size_t i = 0; i < lod->size(); ++i) {
武毅 已提交
244
    auto &level = (*lod)[i];
245 246 247 248 249 250
    for (size_t len : lod_length[i]) {
      level.push_back(level.back() + len);
    }
  }
}

武毅 已提交
251 252
void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
                       const platform::DeviceContext &dev_ctx) {
253
  {  // the 1st field, uint32_t version for LoDTensor
武毅 已提交
254 255 256
    constexpr uint32_t version = 0;
    os.write(reinterpret_cast<const char *>(&version), sizeof(version));
  }
257 258 259 260 261 262
  {
    // the 2st field, LoD information
    // uint64_t lod_level
    // uint64_t lod_level_1 size in byte.
    // int*     lod_level_1 data
    // ...
武毅 已提交
263 264 265 266 267 268 269 270 271 272 273
    auto lod = tensor.lod();
    uint64_t size = lod.size();
    os.write(reinterpret_cast<const char *>(&size), sizeof(size));

    for (auto &each : lod) {
      size = each.size() * sizeof(framework::LoD::value_type::value_type);
      os.write(reinterpret_cast<const char *>(&size), sizeof(size));
      os.write(reinterpret_cast<const char *>(each.data()),
               static_cast<std::streamsize>(size));
    }
  }
274
  // the 3st field, Tensor
Y
Yi Wang 已提交
275
  TensorToStream(os, static_cast<Tensor>(tensor), dev_ctx);
武毅 已提交
276 277
}

Y
Yancey 已提交
278 279
void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
                           const platform::DeviceContext &dev_ctx) {
280
  {
Y
Yancey 已提交
281
    // the 1st field, unit32_t version for LoDTensor
282 283 284
    uint32_t version;
    is.read(reinterpret_cast<char *>(&version), sizeof(version));
    PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
武毅 已提交
285
  }
286 287
  {
    // the 2st field, LoD information
武毅 已提交
288 289 290 291 292 293 294 295 296 297 298 299 300
    uint64_t lod_level;
    is.read(reinterpret_cast<char *>(&lod_level), sizeof(lod_level));
    auto &lod = *tensor->mutable_lod();
    lod.resize(lod_level);
    for (uint64_t i = 0; i < lod_level; ++i) {
      uint64_t size;
      is.read(reinterpret_cast<char *>(&size), sizeof(size));
      std::vector<size_t> tmp(size / sizeof(size_t));
      is.read(reinterpret_cast<char *>(tmp.data()),
              static_cast<std::streamsize>(size));
      lod[i] = tmp;
    }
  }
301
  // the 3st filed, Tensor
Y
Yi Wang 已提交
302
  TensorFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
武毅 已提交
303 304
}

D
dzhwinter 已提交
305
#if !defined(_WIN32)
F
fengjiayi 已提交
306
void WriteToRecordIO(recordio::Writer *writer,
Y
Yu Yang 已提交
307 308 309 310 311 312 313 314
                     const std::vector<LoDTensor> &tensor,
                     const platform::DeviceContext &dev_ctx) {
  std::stringstream buffer;
  size_t sz = tensor.size();
  buffer.write(reinterpret_cast<const char *>(&sz), sizeof(uint32_t));
  for (auto &each : tensor) {
    SerializeToStream(buffer, each, dev_ctx);
  }
F
fengjiayi 已提交
315
  writer->Write(buffer.str());
Y
Yu Yang 已提交
316 317
}

Y
yuyang18 已提交
318 319 320 321 322
bool ReadFromRecordIO(recordio::Scanner *scanner,
                      const platform::DeviceContext &dev_ctx,
                      std::vector<LoDTensor> *result_ptr) {
  if (!scanner->HasNext()) {
    return false;
Y
Yu Yang 已提交
323
  }
Y
yuyang18 已提交
324 325 326 327 328 329 330 331 332 333
  std::istringstream sin(scanner->Next());
  uint32_t sz;
  sin.read(reinterpret_cast<char *>(&sz), sizeof(uint32_t));
  auto &result = *result_ptr;
  result.resize(sz);
  for (uint32_t i = 0; i < sz; ++i) {
    DeserializeFromStream(sin, &result[i], dev_ctx);
  }

  return true;
Y
Yu Yang 已提交
334
}
D
dzhwinter 已提交
335 336 337 338 339
#else
class Writer {};
class Scanner {};
void WriteToRecordIO(recordio::Writer *writer,
                     const std::vector<LoDTensor> &tensor,
D
dzhwinter 已提交
340
                     const platform::DeviceContext &dev_ctx) {}
D
dzhwinter 已提交
341 342
bool ReadFromRecordIO(recordio::Scanner *scanner,
                      const platform::DeviceContext &dev_ctx,
D
dzhwinter 已提交
343
                      std::vector<LoDTensor> *result_ptr) {
D
dzhwinter 已提交
344
  PADDLE_ENFORCE("windows didn't supported recordio!.");
D
dzhwinter 已提交
345 346 347
  return true;
}
#endif  // _WIN32
Y
Yang Yang 已提交
348 349 350
std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
    const std::vector<platform::Place> places) const {
  check_memory_size();
Y
Yang Yang 已提交
351 352 353 354
  int batch_size =
      lod().empty() ? dims()[0] : static_cast<int>(lod()[0].size()) - 1;
  size_t result_size = std::min(static_cast<size_t>(batch_size), places.size());
  size_t remainder = batch_size % places.size();
Y
Yu Yang 已提交
355 356 357 358

  std::vector<LoDTensor> results;
  results.reserve(result_size);

Y
Yang Yang 已提交
359
  int step_width = static_cast<int>(batch_size / result_size);
Y
Yu Yang 已提交
360 361 362 363 364 365
  for (size_t i = 0; i < result_size; ++i) {
    int begin = static_cast<int>(i * step_width);
    int end = static_cast<int>((i + 1) * step_width);
    if (i + 1 == places.size()) {  // last
      end += remainder;
    }
Y
Yang Yang 已提交
366

367
    LoDTensor dst;
Y
Yang Yang 已提交
368 369
    if (lod().empty()) {
      auto src = Slice(begin, end);
Y
Yang Yang 已提交
370
      auto &dst_place = places[i];
Y
Yi Wang 已提交
371
      framework::TensorCopy(src, dst_place, &dst);
Y
Yang Yang 已提交
372 373 374 375 376
    } else {
      auto lod_and_offset = GetSubLoDAndAbsoluteOffset(lod(), begin, end, 0);

      auto &offset = lod_and_offset.second;
      auto src = Slice(offset.first, offset.second);
Y
Yang Yang 已提交
377
      auto &dst_place = places[i];
Y
Yi Wang 已提交
378
      framework::TensorCopy(src, dst_place, &dst);
Y
Yang Yang 已提交
379 380 381 382 383 384 385 386 387 388 389

      LoD my_lod;
      for (auto &l : lod_and_offset.first) {
        std::vector<size_t> v{0};
        for (auto &ll : l) {
          v.push_back(ll + v.back());
        }
        my_lod.emplace_back(v);
      }
      dst.set_lod(my_lod);
    }
Y
Yang Yang 已提交
390
    results.emplace_back(dst);
Y
Yang Yang 已提交
391 392
  }

Y
Yu Yang 已提交
393
  return results;
Y
Yang Yang 已提交
394 395
}

Y
Yang Yang 已提交
396
void LoDTensor::MergeLoDTensor(
397 398
    const std::vector<const LoDTensor *> &lod_tensors,
    platform::Place dst_place) {
Y
Yang Yang 已提交
399
  PADDLE_ENFORCE(!lod_tensors.empty());
Y
Yang Yang 已提交
400

Y
Yang Yang 已提交
401 402
  framework::DDim new_dim = lod_tensors[0]->dims();
  std::type_index new_type = lod_tensors[0]->type();
Y
Yang Yang 已提交
403 404 405 406
  framework::DataLayout new_layout = lod_tensors[0]->layout();
  LoD new_lod = lod_tensors[0]->lod();
  for (size_t i = 1; i < lod_tensors.size(); ++i) {
    auto *t = lod_tensors[i];
S
sneaxiy 已提交
407
    PADDLE_ENFORCE_EQ(new_type, t->type());
Y
Yang Yang 已提交
408 409 410 411 412 413 414
    PADDLE_ENFORCE_EQ(new_layout, t->layout());

    PADDLE_ENFORCE_EQ(framework::product(new_dim) / new_dim[0],
                      framework::product(t->dims()) / t->dims()[0]);
    new_dim[0] += t->dims()[0];

    auto &lod = t->lod();
F
fengjiayi 已提交
415
    PADDLE_ENFORCE_EQ(new_lod.size(), lod.size());
Y
Yang Yang 已提交
416 417 418 419 420 421 422
    for (size_t j = 0; j < lod.size(); ++j) {
      auto &sub_lod = new_lod[j];
      auto &offset = sub_lod.back();
      for (size_t k = 1; k < lod[j].size(); ++k) {
        sub_lod.push_back(lod[j][k] + offset);
      }
    }
Y
Yang Yang 已提交
423 424
  }
  Resize(new_dim);
425
  set_layout(new_layout);
Y
Yang Yang 已提交
426
  set_lod(new_lod);
427
  mutable_data(dst_place, new_type);
Y
Yang Yang 已提交
428

429
  int begin = 0;
Y
Yang Yang 已提交
430
  for (auto *src : lod_tensors) {
431 432
    int end = begin + src->dims()[0];
    auto dst = Slice(begin, end);
Y
Yi Wang 已提交
433
    framework::TensorCopy(*src, dst_place, &dst);
434
    begin = end;
Y
Yang Yang 已提交
435 436 437
  }
}

438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470
LoD ConvertToLengthBasedLoD(const LoD &offset_lod) {
  LoD length_lod;
  length_lod.reserve(offset_lod.size());
  for (size_t lvl = 0; lvl < offset_lod.size(); ++lvl) {
    std::vector<size_t> level;
    if (offset_lod[lvl].size() > 0) {
      level.reserve(offset_lod[lvl].size() - 1);
    }
    for (size_t idx = 0; idx < offset_lod[lvl].size() - 1; ++idx) {
      level.push_back(offset_lod[lvl][idx + 1] - offset_lod[lvl][idx]);
    }
    length_lod.push_back(level);
  }
  return length_lod;
}

LoD ConvertToOffsetBasedLoD(const LoD &length_lod) {
  LoD offset_lod;
  offset_lod.reserve(length_lod.size());
  for (size_t lvl = 0; lvl < length_lod.size(); ++lvl) {
    std::vector<size_t> level;
    level.reserve(length_lod[lvl].size() + 1);
    size_t tmp = 0;
    level.push_back(tmp);
    for (size_t idx = 0; idx < length_lod[lvl].size(); ++idx) {
      tmp += length_lod[lvl][idx];
      level.push_back(tmp);
    }
    offset_lod.push_back(level);
  }
  return offset_lod;
}

471 472
}  // namespace framework
}  // namespace paddle