Argument.cpp 23.0 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15

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

    http://www.apache.org/licenses/LICENSE-2.0

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. */

#include "Argument.h"
16
#include "paddle/math/SparseMatrix.h"
Z
zhangjinchao01 已提交
17 18 19 20

#include <algorithm>

namespace paddle {
21 22 23
static void resizeAndCopy(MatrixPtr& dest,
                          const MatrixPtr& src,
                          bool useGpu,
Z
zhangjinchao01 已提交
24 25
                          hl_stream_t stream) {
  if (src) {
26 27 28 29 30 31
    if (!dest) {
      dest = src->clone(0, 0, useGpu);
    } else {
      CHECK_EQ(dest->useGpu(), useGpu);
      dest->resize(src->getHeight(), src->getWidth());
    }
Z
zhangjinchao01 已提交
32 33 34 35 36 37
    dest->copyFrom(*src, stream);
  } else {
    dest.reset();
  }
}

38 39 40
static void resizeAndCopy(IVectorPtr& dest,
                          const IVectorPtr& src,
                          bool useGpu,
Z
zhangjinchao01 已提交
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61
                          hl_stream_t stream) {
  if (src) {
    IVector::resizeOrCreate(dest, src->getSize(), useGpu);
    dest->copyFrom(*src, stream);
  } else {
    dest.reset();
  }
}

static void resizeAndCopy(ICpuGpuVectorPtr& dest,
                          const ICpuGpuVectorPtr& src,
                          bool useGpu,
                          hl_stream_t stream) {
  if (src) {
    ICpuGpuVector::resizeOrCreate(dest, src->getSize(), useGpu);
    dest->copyFrom(*src, stream);
  } else {
    dest.reset();
  }
}

62 63 64 65 66
static void resizeAndCopy(MatrixPtr& dest,
                          const MatrixPtr& src,
                          int32_t startRow,
                          int32_t copySize,
                          bool useGpu,
Z
zhangjinchao01 已提交
67 68 69 70 71
                          hl_stream_t stream = HPPL_STREAM_DEFAULT) {
  if (src) {
    CHECK_LE((size_t)startRow + copySize, src->getHeight());
    int height = copySize;
    int width = src->getWidth();
72 73 74 75 76 77
    if (!dest) {
      dest = src->clone(height, width, useGpu);
    } else {
      CHECK_EQ(dest->useGpu(), useGpu);
      dest->resize(height, width);
    }
Z
zhangjinchao01 已提交
78
    MatrixPtr submat = src->subMatrix(startRow, copySize);
79 80 81 82 83 84 85 86 87
    if (dynamic_cast<GpuSparseMatrix*>(dest.get())) {
      // copy a subMatrix of CpuSparseMatrix to GpuSparseMatrix.
      // First copy it to CPU, and then copy it to the GPU.
      MatrixPtr tmp = src->clone(height, width, false);
      tmp->copyFrom(*submat, stream);
      dest->copyFrom(*tmp, stream);
    } else {
      dest->copyFrom(*submat, stream);
    }
Z
zhangjinchao01 已提交
88 89 90 91 92
  } else {
    dest.reset();
  }
}

93 94 95 96 97
static void resizeAndCopy(IVectorPtr& dest,
                          const IVectorPtr& src,
                          int32_t startPos,
                          int32_t copySize,
                          bool useGpu,
Z
zhangjinchao01 已提交
98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
                          hl_stream_t stream = HPPL_STREAM_DEFAULT) {
  if (src) {
    CHECK_LE((size_t)startPos + copySize, src->getSize());

    int height = copySize;
    IVector::resizeOrCreate(dest, height, useGpu);
    dest->copyFrom(src->getData() + startPos, height, stream);
  } else {
    dest.reset();
  }
}

static void resizeAndCopy(ICpuGpuVectorPtr& dest,
                          const ICpuGpuVectorPtr& src,
                          int32_t startPos,
                          int32_t copySize,
                          bool useGpu,
                          hl_stream_t stream = HPPL_STREAM_DEFAULT) {
  if (src) {
    CHECK_LE((size_t)startPos + copySize, src->getSize());

    ICpuGpuVector::resizeOrCreate(dest, copySize, useGpu);
    dest->copyFrom(*src, startPos, copySize, useGpu, stream);
  } else {
    dest.reset();
  }
}

126 127 128
static void resizeAndCopy(SVectorPtr& dest,
                          const SVectorPtr& src,
                          bool useGpu,
Z
zhangjinchao01 已提交
129 130 131 132 133 134 135 136 137 138 139 140 141 142
                          hl_stream_t stream) {
  if (src) {
    size_t height = src->size();
    if (!dest) {
      dest = std::make_shared<std::vector<std::string>>(height);
    } else {
      dest->resize(height);
    }
    std::copy_n(src->begin(), height, dest->begin());
  } else {
    dest.reset();
  }
}

143 144 145 146 147
static void resizeAndCopy(SVectorPtr& dest,
                          const SVectorPtr& src,
                          int32_t startPos,
                          int32_t copySize,
                          bool useGpu,
Z
zhangjinchao01 已提交
148 149 150 151 152 153 154 155 156 157 158 159 160 161 162
                          hl_stream_t stream = HPPL_STREAM_DEFAULT) {
  if (src) {
    CHECK_LE((size_t)startPos + copySize, src->size());
    size_t height = copySize;
    if (!dest) {
      dest = std::make_shared<std::vector<std::string>>(height);
    } else {
      dest->resize(height);
    }
    std::copy_n(src->begin() + startPos, height, dest->begin());
  } else {
    dest.reset();
  }
}

163
void Argument::resizeAndCopyFrom(const Argument& src, bool useGpu) {
164 165
  resizeAndCopyFrom(src, useGpu, HPPL_STREAM_DEFAULT);
  hl_stream_synchronize(HPPL_STREAM_DEFAULT);
166 167
}

168 169
void Argument::resizeAndCopyFrom(const Argument& src,
                                 bool useGpu,
Z
zhangjinchao01 已提交
170 171 172 173 174 175
                                 hl_stream_t stream) {
  dataId = src.dataId;
  resizeAndCopy(value, src.value, useGpu, stream);
  resizeAndCopy(grad, src.grad, useGpu, stream);
  resizeAndCopy(in, src.in, useGpu, stream);
  resizeAndCopy(ids, src.ids, useGpu, stream);
176 177 178 179
  resizeAndCopy(sequenceStartPositions,
                src.sequenceStartPositions,
                false /* useGpu */,
                stream);
Z
zhangjinchao01 已提交
180 181
  if (src.hasSubseq()) {
    resizeAndCopy(subSequenceStartPositions,
182 183 184
                  src.subSequenceStartPositions,
                  false /* useGpu */,
                  stream);
Z
zhangjinchao01 已提交
185 186
  }
  resizeAndCopy(strs, src.strs, useGpu, stream);
L
Luo Tao 已提交
187 188
  frameWidth = src.frameWidth;
  frameHeight = src.frameHeight;
Z
zhangjinchao01 已提交
189 190
}

191 192 193 194 195 196 197 198
int32_t Argument::resizeAndCopyFrom(const Argument& src,
                                    int32_t startSeq,
                                    int32_t copySize,
                                    bool useGpu) {
  int32_t size =
      resizeAndCopyFrom(src, startSeq, copySize, useGpu, HPPL_STREAM_DEFAULT);
  hl_stream_synchronize(HPPL_STREAM_DEFAULT);
  return size;
199 200
}

201 202 203 204
int32_t Argument::resizeAndCopyFrom(const Argument& src,
                                    int32_t startSeq,
                                    int32_t copySize,
                                    bool useGpu,
Z
zhangjinchao01 已提交
205 206
                                    hl_stream_t stream) {
  dataId = src.dataId;
207 208
  frameWidth = src.frameWidth;
  frameHeight = src.frameHeight;
Z
zhangjinchao01 已提交
209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228

  if (!src.sequenceStartPositions) {
    // non-sequence input, copy samples directly
    int32_t startRow = startSeq;
    resizeAndCopy(in, src.in, startRow, copySize, useGpu, stream);
    resizeAndCopy(value, src.value, startRow, copySize, useGpu, stream);
    resizeAndCopy(grad, src.grad, startRow, copySize, useGpu, stream);
    resizeAndCopy(ids, src.ids, startRow, copySize, useGpu, stream);
    resizeAndCopy(strs, src.strs, startRow, copySize, useGpu, stream);
    return copySize;
  } else {
    // sequence input
    const int* sequence = src.sequenceStartPositions->getData(false);
    int32_t startRow = sequence[startSeq];           // sample start from here
    int32_t endRow = sequence[startSeq + copySize];  // sample end
    int32_t copyFeatureSize = endRow - startRow;     // num of samples
    resizeAndCopy(in, src.in, startRow, copyFeatureSize, useGpu, stream);
    resizeAndCopy(value, src.value, startRow, copyFeatureSize, useGpu, stream);
    resizeAndCopy(grad, src.grad, startRow, copyFeatureSize, useGpu, stream);
    resizeAndCopy(ids, src.ids, startRow, copyFeatureSize, useGpu, stream);
229 230 231 232 233 234
    resizeAndCopy(sequenceStartPositions,
                  src.sequenceStartPositions,
                  startSeq,
                  copySize + 1,
                  false,
                  stream);
Z
zhangjinchao01 已提交
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257
    // modify new sequenceStartPositions
    int* destSequences = sequenceStartPositions->getMutableData(false);
    for (int i = 0; i < copySize + 1; i++) {
      destSequences[i] -= startRow;
    }
    CHECK_EQ(destSequences[0], 0);
    CHECK_EQ(destSequences[copySize], copyFeatureSize);
    if (src.hasSubseq()) {
      // sequence has sub-sequence
      int* subSequence = src.subSequenceStartPositions->getMutableData(false);
      int32_t subStartSeq = 0;
      int32_t subEndSeq = 0;
      int numSubSequences = src.getNumSubSequences();
      for (int i = 0; i < numSubSequences + 1; i++) {
        if (subSequence[i] == startRow) {
          subStartSeq = i;
        } else if (subSequence[i] == endRow) {
          subEndSeq = i;
          break;
        }
      }
      int32_t copySubSize = subEndSeq - subStartSeq;
      resizeAndCopy(subSequenceStartPositions,
258 259 260 261 262
                    src.subSequenceStartPositions,
                    subStartSeq,
                    copySubSize + 1,
                    false,
                    stream);
Z
zhangjinchao01 已提交
263 264 265 266 267 268 269 270 271 272 273 274 275 276 277
      // modify new subSequenceStartPositions
      int* destSubSequences = subSequenceStartPositions->getMutableData(false);
      for (int i = 0; i < copySubSize + 1; i++) {
        destSubSequences[i] -= startRow;
      }
      CHECK_EQ(destSubSequences[0], 0);
      CHECK_EQ(destSubSequences[copySubSize], copyFeatureSize);
    }
    resizeAndCopy(strs, src.strs, startRow, copySize, useGpu, stream);
    return copyFeatureSize;
  }
}

void Argument::concat(const std::vector<Argument>& args,
                      const std::vector<int>& selectRows,
278
                      const std::vector<int>& seqStartPos,
279
                      const std::vector<int>& copySize,
280 281 282
                      bool useGpu,
                      hl_stream_t stream,
                      PassType passType) {
283
  CHECK(!subSequenceStartPositions)
284
      << "undefined behavior for subsequence positions";
285

286 287 288 289
  size_t batchSize = 0;
  for (size_t i = 0; i < copySize.size(); ++i)
    batchSize += copySize[i] * (seqStartPos[i + 1] - seqStartPos[i]);

290 291
  auto copyArg = [batchSize, stream](MatrixPtr& dst,
                                     MatrixPtr src,
292 293
                                     int desStartRow,
                                     int srcStartRow,
294
                                     int size,
Z
zhangjinchao01 已提交
295 296 297 298 299 300 301 302 303 304 305 306
                                     bool useGpu) {
    if (!src) {
      dst.reset();
      return;
    }
    size_t width = src->getWidth();
    if (!dst) {
      dst = src->clone(batchSize, width, useGpu);
    } else {
      dst->resize(batchSize, width);
    }

307 308
    MatrixPtr tmpMatrix = dst->subMatrix(desStartRow, size);
    tmpMatrix->copyFrom(*src->subMatrix(srcStartRow, size), stream);
Z
zhangjinchao01 已提交
309 310
  };

311 312
  auto copyIds = [batchSize, stream](IVectorPtr& dst,
                                     const IVectorPtr& src,
313 314
                                     int desStartRow,
                                     int srcStartRow,
315
                                     int size,
Z
zhangjinchao01 已提交
316 317 318 319 320 321
                                     bool useGpu) {
    if (!src) {
      dst.reset();
      return;
    }
    IVector::resizeOrCreate(dst, batchSize, useGpu);
322 323
    dst->subVec(desStartRow, size)
        ->copyFrom(*src->subVec(srcStartRow, size), stream);
Z
zhangjinchao01 已提交
324 325
  };

326 327
  auto copyStrs = [batchSize, stream](SVectorPtr& dst,
                                      const SVectorPtr& src,
328 329
                                      int desStartRow,
                                      int srcStartRow,
330
                                      int size,
Z
zhangjinchao01 已提交
331 332 333 334 335 336 337 338 339 340
                                      bool useGpu) {
    if (!src) {
      dst.reset();
      return;
    }
    if (!dst) {
      dst = std::make_shared<std::vector<std::string>>(batchSize);
    } else {
      dst->resize(batchSize);
    }
341 342 343
    std::copy(src->begin() + srcStartRow,
              src->begin() + srcStartRow + size,
              dst->begin() + desStartRow);
Z
zhangjinchao01 已提交
344 345 346 347
  };

  dataId = args[0].dataId;
  CHECK_NE(seqStartPos.size(), 0UL);
348 349
  int desStartRow = 0;
  for (size_t i = 0; i < copySize.size(); ++i) {
Z
zhangjinchao01 已提交
350 351 352 353 354
    int startPos = seqStartPos[i];
    int endPos = seqStartPos[i + 1];
    CHECK_GE(args.size(), static_cast<size_t>(endPos - startPos));
    for (int j = startPos; j < endPos; ++j) {
      const Argument& arg = args[j - startPos];
355 356 357 358 359
      CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have the "
                                   << "same dataId";
      const int srcStartRow = selectRows[j];
      copyArg(in, arg.in, desStartRow, srcStartRow, copySize[i], useGpu);
      copyArg(value, arg.value, desStartRow, srcStartRow, copySize[i], useGpu);
Z
zhangjinchao01 已提交
360
      if (passType != PASS_TEST) {
361
        copyArg(grad, arg.grad, desStartRow, srcStartRow, copySize[i], useGpu);
Z
zhangjinchao01 已提交
362
      }
363 364 365
      copyIds(ids, arg.ids, desStartRow, srcStartRow, copySize[i], useGpu);
      copyStrs(strs, arg.strs, desStartRow, srcStartRow, copySize[i], useGpu);
      desStartRow += copySize[i];
Z
zhangjinchao01 已提交
366 367
    }
  }
368 369 370 371
  ICpuGpuVector::resizeOrCreate(
      sequenceStartPositions, seqStartPos.size(), useGpu);
  sequenceStartPositions->copyFrom(
      seqStartPos.data(), seqStartPos.size(), useGpu);
Z
zhangjinchao01 已提交
372 373
}

374 375 376 377
void Argument::concat(const std::vector<Argument>& args,
                      bool useGpu,
                      hl_stream_t stream,
                      PassType passType) {
Z
zhangjinchao01 已提交
378 379
  int32_t batchSize = 0;
  int64_t numSequences = 0;
380
  int64_t numSubSequences = 0;
Z
zhangjinchao01 已提交
381 382 383
  for (auto& arg : args) {
    batchSize += arg.getBatchSize();
    numSequences += arg.getNumSequences();
384
    numSubSequences += arg.getNumSubSequences();
Z
zhangjinchao01 已提交
385 386
  }

387 388
  auto copyArg = [batchSize, stream](
      MatrixPtr& dst, MatrixPtr src, int startRow, bool useGpu) {
Z
zhangjinchao01 已提交
389 390 391 392 393 394 395 396 397 398 399 400 401 402 403
    if (!src) {
      dst.reset();
      return;
    }
    size_t width = src->getWidth();
    if (!dst) {
      dst = src->clone(batchSize, width, useGpu);
    } else {
      dst->resize(batchSize, width);
    }

    MatrixPtr tmpMatrix = dst->subMatrix(startRow, src->getHeight());
    tmpMatrix->copyFrom(*src, stream);
  };

404 405
  auto copyIds = [batchSize, stream](
      IVectorPtr& dst, const IVectorPtr& src, int startRow, bool useGpu) {
Z
zhangjinchao01 已提交
406 407 408 409 410 411 412 413
    if (!src) {
      dst.reset();
      return;
    }
    IVector::resizeOrCreate(dst, batchSize, useGpu);
    dst->subVec(startRow, src->getSize())->copyFrom(*src, stream);
  };

414 415
  auto copyStrs = [batchSize, stream](
      SVectorPtr& dst, const SVectorPtr& src, int startRow, bool useGpu) {
Z
zhangjinchao01 已提交
416 417 418 419 420 421 422 423 424 425 426 427
    if (!src) {
      dst.reset();
      return;
    }
    if (!dst) {
      dst = std::make_shared<std::vector<std::string>>(batchSize);
    } else {
      dst->resize(batchSize);
    }
    std::copy(src->begin(), src->end(), dst->begin() + startRow);
  };

428 429 430 431 432 433 434 435 436 437 438 439
  auto copySequencePos = [](ICpuGpuVectorPtr& dstSeq,
                            const ICpuGpuVectorPtr& srcSeq,
                            int dstNumSequences,
                            int srcNumSequences,
                            int& startSequences,
                            int startRow) {
    if (srcSeq) {
      ICpuGpuVector::resizeOrCreate(dstSeq, dstNumSequences + 1, false);
      const int* src = srcSeq->getData(false);
      int* dest = dstSeq->getMutableData(false);
      for (int i = 0; i < srcNumSequences + 1; ++i) {
        dest[i + startSequences] = src[i] + startRow;
440
      }
441 442 443 444
      startSequences += srcNumSequences;
    } else {
      dstSeq.reset();
    }
445 446
  };

Z
zhangjinchao01 已提交
447 448
  int startRow = 0;
  int startSequences = 0;
449
  int startSubSequences = 0;
Z
zhangjinchao01 已提交
450 451 452 453 454 455 456 457
  dataId = args[0].dataId;
  for (auto& arg : args) {
    CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have"
                                 << " same dataId";
    copyArg(in, arg.in, startRow, useGpu);
    copyArg(value, arg.value, startRow, useGpu);
    if (passType != PASS_TEST) copyArg(grad, arg.grad, startRow, useGpu);
    copyIds(ids, arg.ids, startRow, useGpu);
458 459 460 461 462 463 464 465 466 467 468 469
    copySequencePos(sequenceStartPositions,
                    arg.sequenceStartPositions,
                    numSequences,
                    arg.getNumSequences(),
                    startSequences,
                    startRow);
    copySequencePos(subSequenceStartPositions,
                    arg.subSequenceStartPositions,
                    numSubSequences,
                    arg.getNumSubSequences(),
                    startSubSequences,
                    startRow);
Z
zhangjinchao01 已提交
470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
    copyStrs(strs, arg.strs, startRow, useGpu);
    startRow += arg.getBatchSize();
  }
}

void Argument::splitByDataId(const std::vector<Argument>& argus,
                             std::vector<std::vector<Argument>>* arguGroups) {
  arguGroups->clear();
  int lastDataId = -1;
  for (const auto& argu : argus) {
    if (argu.dataId == -1) {
      // is -1, then create a new group
      arguGroups->emplace_back();
      lastDataId = -1;
    } else if (argu.dataId != lastDataId) {
      // not -1, also not equal to last Argument, then create a new group
      arguGroups->emplace_back();
      lastDataId = argu.dataId;
    } else {
      // not -1, and equal to last Argument, do nothing
    }
    arguGroups->back().push_back(argu);
  }
}

495
void Argument::getSeqInfo(std::vector<SeqInfo>* seqInfo) const {
Z
zhangjinchao01 已提交
496
  const int* starts = sequenceStartPositions->getData(false);
497 498
  const int* subStarts =
      hasSubseq() ? subSequenceStartPositions->getData(false) : nullptr;
499 500 501 502 503 504 505 506 507 508 509 510
  size_t numSequences = getNumSequences();
  seqInfo->reserve(numSequences);
  int subSeqEnd = 0;
  for (size_t i = 0; i < numSequences; ++i) {
    SeqInfo info;
    info.seqStart = starts[i];
    info.subLevelLength = starts[i + 1] - starts[i];
    info.seqId = i;
    if (hasSubseq()) {
      info.subSeqStart = subSeqEnd;
      while (subStarts[subSeqEnd] < starts[i + 1]) {
        ++subSeqEnd;
Z
zhangjinchao01 已提交
511
      }
512 513 514 515
      info.topLevelLength = subSeqEnd - info.subSeqStart;
    } else {
      info.topLevelLength = info.subLevelLength;
      info.subSeqStart = 0;  // not used
Z
zhangjinchao01 已提交
516
    }
517
    seqInfo->push_back(info);
Z
zhangjinchao01 已提交
518
  }
Y
Yu Yang 已提交
519 520 521 522
  std::sort(
      seqInfo->begin(), seqInfo->end(), [](const SeqInfo& a, const SeqInfo& b) {
        return a.topLevelLength > b.topLevelLength;
      });
Z
zhangjinchao01 已提交
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548
}

void Argument::checkSubset() const {
  if (getNumSequences() > getNumSubSequences()) {
    LOG(FATAL) << "numSubSequences is less than numSequences ("
               << getNumSubSequences() << " vs. " << getNumSequences() << ")";
  }
  const int* start = sequenceStartPositions->getData(false);
  const int* subStart = subSequenceStartPositions->getData(false);
  int seqId = 0;
  int subSeqId = 0;
  while (seqId < getNumSequences() && subSeqId < getNumSubSequences()) {
    if (start[seqId] > subStart[subSeqId]) {
      ++subSeqId;
    } else if (start[seqId] == subStart[subSeqId]) {
      ++subSeqId;
      ++seqId;
    } else {
      LOG(FATAL) << "seqStartPositions is not subset of subSeqStartPositions";
    }
  }
  if (seqId < getNumSequences()) {
    LOG(FATAL) << "seqStartPositions is not subset of subSeqStartPositions";
  }
}

549
void Argument::degradeSequence(const Argument& input) {
Z
zhangjinchao01 已提交
550 551 552
  CHECK_EQ(input.hasSubseq(), 1UL);
  size_t numSequences = input.getNumSequences();
  size_t numSubSequences = input.getNumSubSequences();
553 554
  ICpuGpuVector::resizeOrCreate(
      sequenceStartPositions, numSequences + 1, false);
Z
zhangjinchao01 已提交
555 556 557 558 559 560 561 562 563 564 565 566 567
  int* tgtBuf = sequenceStartPositions->getMutableData(false);
  const int* starts = input.sequenceStartPositions->getData(false);
  const int* subStarts = input.subSequenceStartPositions->getData(false);
  int seqId = 0;
  for (size_t subSeqId = 0; subSeqId < numSubSequences; ++subSeqId) {
    if (subStarts[subSeqId] == starts[seqId]) {
      tgtBuf[seqId] = subSeqId;
      seqId++;
    }
  }
  tgtBuf[numSequences] = numSubSequences;
}

568 569
void Argument::poolSequenceWithStride(const Argument& input,
                                      size_t stride,
570
                                      ICpuGpuVectorPtr* stridePostions,
L
Luo Tao 已提交
571
                                      bool reversed) {
L
Luo Tao 已提交
572 573 574 575 576
  // If input.sequenceStartPositions = [0, 9, 14, 17, 30] and stride = 5,
  // then sequenceStartPositions = [0, 2, 3, 4, 7].
  // If reversed = false, stridePostions = [0, 5, 9, 14, 17, 22, 27, 30];
  // else reversed = true, stridePostions = [0, 4, 9, 14, 17, 20, 25, 30]

577 578
  CHECK(input.sequenceStartPositions);
  CHECK_EQ(input.hasSubseq(), 0UL);
Y
Yu Yang 已提交
579
  CHECK_GT(stride, 0UL) << "stride must larger than 0";
580 581 582 583 584 585 586
  size_t numSequences = input.getNumSequences();
  ICpuGpuVector::resizeOrCreate(
      sequenceStartPositions, numSequences + 1, false);
  const int* starts = input.sequenceStartPositions->getData(false);
  int* tgtBuf = sequenceStartPositions->getMutableData(false);
  // first index of target sequence and stride positions are both 0
  tgtBuf[0] = 0;
587
  std::vector<int> stridePos;
588 589
  for (size_t seqId = 0; seqId < numSequences; ++seqId) {
    size_t seqLength = starts[seqId + 1] - starts[seqId];
590
    stridePos.emplace_back(starts[seqId]);
591 592 593 594
    if (seqLength == 0) {
      // empty sequence
      tgtBuf[seqId + 1] = tgtBuf[seqId];
    } else {
L
Luo Tao 已提交
595 596
      int size = ceil((float)seqLength / stride);
      tgtBuf[seqId + 1] = tgtBuf[seqId] + size;
L
Luo Tao 已提交
597
      for (int i = 0; i < size - 1; ++i) {
L
Luo Tao 已提交
598 599 600
        int cur = reversed ? starts[seqId + 1] - (size - 1 - i) * stride
                           : stridePos.back() + stride;
        stridePos.emplace_back(cur);
601 602 603
      }
    }
  }
604 605 606
  stridePos.emplace_back(starts[numSequences]);
  int size = stridePos.size();
  CHECK_EQ(size - 1, tgtBuf[numSequences]);
607 608
  ICpuGpuVector::resizeOrCreate(*stridePostions, size, false);
  (*stridePostions)->getMutableVector(false)->copyFrom(stridePos.data(), size);
609 610
}

611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636
void Argument::getValueString(
    std::unordered_map<std::string, std::string>* out) const {
  if (value) {
    std::ostringstream os;
    value->print(os);
    out->insert({"value", os.str()});
  }
  if (ids) {
    std::ostringstream os;
    ids->print(os, ids->getSize());
    out->insert({"ids", os.str()});
  }
  if (sequenceStartPositions) {
    std::ostringstream os;
    sequenceStartPositions->getVector(false)->print(
        os, sequenceStartPositions->getSize());
    out->insert({"sequence pos", os.str()});
  }
  if (subSequenceStartPositions) {
    std::ostringstream os;
    subSequenceStartPositions->getVector(false)->print(
        os, subSequenceStartPositions->getSize());
    out->insert({"sub-sequence pos", os.str()});
  }
}

637 638 639 640
void Argument::printValueString(std::ostream& stream,
                                const std::string& prefix) const {
  std::unordered_map<std::string, std::string> out;
  getValueString(&out);
641
  for (auto field : {"value", "ids", "sequence pos", "sub-sequence pos"}) {
642 643 644 645 646 647 648
    auto it = out.find(field);
    if (it != out.end()) {
      stream << prefix << field << ":\n" << it->second;
    }
  }
}

649 650 651 652 653 654 655 656 657
void Argument::subArgFrom(const Argument& input,
                          size_t offset,
                          size_t height,
                          size_t width,
                          bool useGpu,
                          bool trans,
                          bool seqFlag,
                          size_t seqStart,
                          size_t seqSize) {
658
  if (input.value) {
659 660
    value = Matrix::create(
        input.value->getData() + offset * width, height, width, trans, useGpu);
661 662 663 664
  }
  if (input.ids) {
    ids = IVector::create(input.ids->getData() + offset, height, useGpu);
  }
Z
zhangjinchao01 已提交
665
  if (input.grad) {
666 667
    grad = Matrix::create(
        input.grad->getData() + offset * width, height, width, trans, useGpu);
Z
zhangjinchao01 已提交
668 669 670
  }
  if (seqFlag) {
    sequenceStartPositions = std::make_shared<ICpuGpuVector>(
671
        *(input.sequenceStartPositions), seqStart, seqSize);
Z
zhangjinchao01 已提交
672 673 674 675
  }
}

}  // namespace paddle