MultiGradientMachine.cpp 26.6 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 16 17 18 19 20 21 22 23

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 "MultiGradientMachine.h"

#include "paddle/utils/Logging.h"

#include "paddle/utils/Stat.h"

#include "NeuralNetwork.h"
#include "ParallelNeuralNetwork.h"

24 25 26
DEFINE_bool(allow_only_one_model_on_one_gpu,
            true,
            "If true, do not allow multiple models on one GPU device");
Z
zhangjinchao01 已提交
27 28 29 30 31

namespace paddle {

// get types of the parameters which need to be merged after backward()
static void fillMergeTypes(PassType passType,
32
                           std::vector<ParameterType>* mergeTypes) {
Z
zhangjinchao01 已提交
33 34 35 36 37 38
  mergeTypes->clear();
  if (passType != PASS_TEST) {
    mergeTypes->push_back(PARAMETER_GRADIENT);
  }
}

39 40
MultiGradientMachine::MultiGradientMachine(const ModelConfig& config,
                                           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
    : useGpu_(useGpu),
      trainerBarrier_(FLAGS_trainer_count),
      allBarrier_(FLAGS_trainer_count + 1),
      inArgsCopied_(false) {
  isPassGrad_ = false;
  numThreads_ = FLAGS_trainer_count;
  if (useGpu) {
    //! TODO(yuyang18): When useGpu=false && paddle is not compiled with gpu,
    //! the hl_get_device_count will get an error result. It seems should return
    //! 0 when hppl is not compiled as gpu version.
    numDevices_ = hl_get_device_count();
  } else {
    numDevices_ = 0;
  }
  ParamInitCallback mainParamInitCb = [this](int paramId, Parameter* para) {
    // only create buf for CPU parameters
    // GPU parameters will be created in each thread
    if (para->useGpu()) return;

    if (para->isSparseRemoteUpdate()) {
61 62 63 64 65
      para->enableType(PARAMETER_VALUE,
                       FLAGS_loadsave_parameters_in_pserver
                           ? Parameter::MAT_SPARSE_ROW_PREFETCH
                           : Parameter::MAT_SPARSE_ROW_PREFETCH_FULL_SIZE);
      para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW);
Z
zhangjinchao01 已提交
66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93
    } else if (para->isGradSparseUpdate()) {
      para->enableType(PARAMETER_VALUE);
      para->enableType(PARAMETER_GRADIENT, Parameter::MAT_SPARSE_ROW_IDS);
      SparseRowIdsCpuMatrix* mat = dynamic_cast<SparseRowIdsCpuMatrix*>(
          para->getMat(PARAMETER_GRADIENT).get());
      mat->setNumOfThreads(FLAGS_trainer_count);
    } else if (para->isValueShared()) {
      para->enableType(PARAMETER_VALUE, Parameter::MAT_VALUE_SHARED);
      if (!para->isStatic()) {
        para->enableType(PARAMETER_GRADIENT);
      }
    } else {
      para->enableType(PARAMETER_VALUE);
      if (!para->isStatic()) {
        para->enableType(PARAMETER_GRADIENT);
      }
    }
  };

  NeuralNetwork* nn = NeuralNetwork::create(config);
  nn->init(config, mainParamInitCb);
  gradientMachine_.reset(nn);
  parameters_ = gradientMachine_->getParameters();

  numLogicalDevices_ = 0;
  if (useGpu_) {
    numLogicalDevices_ = 1;

94
    for (size_t pid = 0; pid < parameters_.size(); pid++) {
Z
zhangjinchao01 已提交
95 96 97 98 99
      if (parameters_[pid]->getConfig().device() + 1 > numLogicalDevices_) {
        numLogicalDevices_ = parameters_[pid]->getConfig().device() + 1;
      }
    }
    LOG(INFO) << "numLogicalDevices=" << numLogicalDevices_
100
              << " numThreads=" << numThreads_ << " numDevices=" << numDevices_;
Z
zhangjinchao01 已提交
101

102 103
    if (numLogicalDevices_ * numThreads_ > numDevices_ &&
        FLAGS_allow_only_one_model_on_one_gpu) {
Z
zhangjinchao01 已提交
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
      LOG(FATAL) << "trainer_count * num_devices_in_model "
                 << "(" << numThreads_ << "*" << numLogicalDevices_ << ")"
                 << "=" << numThreads_ * numLogicalDevices_
                 << " exceeds number of GPU devices(" << numDevices_ << ")";
    }
    numLogicalDevices_ = std::min(numLogicalDevices_, numDevices_);

    /* Enables direct access to memory allocations on a peer device */
    for (int i = 0; i < numThreads_; i++) {
      for (int d = 0; d < numLogicalDevices_; ++d) {
        enablePeerAccess(logicalDeviceId2RealDeviceId(d, i),
                         logicalDeviceId2RealDeviceId(d, i + 1));
        enablePeerAccess(logicalDeviceId2RealDeviceId(d, i),
                         logicalDeviceId2RealDeviceId(d, i - 1));
      }
    }
  }

  for (int i = 0; i < numThreads_; ++i) {
123
    threads_.emplace_back(new TrainerThread(config, i, this));
Z
zhangjinchao01 已提交
124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
  }

  bufferSizes_.resize(numLogicalDevices_, 0);
  paraMainThread_.reserve(parameters_.size());
  int pid = 0;
  for (auto& para : parameters_) {
    if (para->isStatic() || !para->useGpu()) {
      paraMainThread_.push_back(0);
    } else {
      int end = pid++ % numThreads_;
      paraMainThread_.push_back(end);
      int paraDeviceId = para->getDeviceId();
      if (paraDeviceId == -1) paraDeviceId = 0;
      paraDeviceId = paraDeviceId % numLogicalDevices_;
      if (para->getSize() > bufferSizes_[paraDeviceId]) {
        bufferSizes_[paraDeviceId] = para->getSize();
        VLOG(1) << "bufferSize[" << paraDeviceId << "]" << para->getSize();
      }
    }
  }

  // TODO(xuwei06) Instead of using maximal buffer size, we may use a smaller
  // fixed buffer size and use pipeline to dispatch parameter value and merge
  // parameter gradient, which may be faster.

  // combination of all trainers mainPara into GradientMachine parameters
  hasNonstaticCpuParamters_ = false;
151
  for (size_t pid = 0; pid < parameters_.size(); pid++) {
Z
zhangjinchao01 已提交
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197
    if (parameters_[pid]->useGpu()) {
      parameters_[pid] = threads_[paraMainThread_[pid]]->getParameters()[pid];
    } else if (!parameters_[pid]->isStatic()) {
      hasNonstaticCpuParamters_ = true;
    }
  }

  gradBufs_.resize(numThreads_);
  for (int i = 0; i < numThreads_; ++i) {
    gradBufs_[i].resize(numLogicalDevices_);
    for (int d = 0; d < numLogicalDevices_; ++d) {
      gradBufs_[i][d].sem.post();
    }
  }

  outArgStream_ = HPPL_STREAM_1;

  for (auto& thread : threads_) {
    thread->start();
  }
}

std::vector<const std::vector<ParameterPtr>*>
MultiGradientMachine::getSlaveParameters() {
  std::vector<const std::vector<ParameterPtr>*> vec;
  vec.reserve(threads_.size());
  for (auto& thread : threads_) {
    vec.push_back(&thread->getParameters());
  }
  return vec;
}

void MultiGradientMachine::notifyGradientTransfer(int paramId) {
  gradQueue_.enqueue(paramId);
}

void MultiGradientMachine::allocGradBufs() {
  if (numLogicalDevices_ == 0) return;
  if (gradBufs_[0][0].bufs.size() >= mergeTypes_.size()) return;

  for (int i = 0; i < numThreads_; i++) {
    for (int d = 0; d < numLogicalDevices_; ++d) {
      if (bufferSizes_[d] == 0) continue;
      SetDevice device(logicalDeviceId2RealDeviceId(d, i));
      for (size_t j = 0; j < mergeTypes_.size(); j++) {
        gradBufs_[i][d].bufs.push_back(
198
            Vector::create(bufferSizes_[d], /* useGpu= */ true));
Z
zhangjinchao01 已提交
199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
      }
    }
  }
}

void MultiGradientMachine::prefetch(const std::vector<Argument>& inArgs) {
  // Each gradient machine in threads needs to do prefetch on its own
  // part of inArgs. So we need to first divide inArgs to each thread
  inArgs_ = inArgs;
  startTask(TASK_COPY_IN_ARGS);

  for (auto& para : parameters_) {
    if (para->isSparseRemoteUpdate()) {
      auto mat = dynamic_cast<SparsePrefetchRowCpuMatrix*>(
          para->getMat(PARAMETER_VALUE).get());
      mat->clearIndices();
    }
  }

  waitForCopyInArgs();

  // Because SparsePrefetchRowCpuMatrix can only be changed by ONE thread
  // at one time, we need to do prefetch sequentially
  for (auto& thread : threads_) {
    thread->prefetch();
  }

  for (auto& para : parameters_) {
    if (para->isSparseRemoteUpdate()) {
      auto mat = dynamic_cast<SparsePrefetchRowCpuMatrix*>(
          para->getMat(PARAMETER_VALUE).get());
      mat->setupIndices();
      auto matGrad = dynamic_cast<SparseRowCpuMatrix*>(
          para->getMat(PARAMETER_GRADIENT).get());
      matGrad->reserveStore();
    }
  }
}

238 239 240
void MultiGradientMachine::forward(const std::vector<Argument>& inArgs,
                                   std::vector<Argument>* outArgs,
                                   PassType passType) {
Z
zhangjinchao01 已提交
241 242 243
  forwardImp(inArgs, outArgs, passType, TASK_FORWARD);
}

244 245 246 247
void MultiGradientMachine::forwardImp(const std::vector<Argument>& inArgs,
                                      std::vector<Argument>* outArgs,
                                      PassType passType,
                                      TaskType taskType) {
Z
zhangjinchao01 已提交
248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268
  updateThreadParameters();
  passType_ = passType;

  if (!inArgsCopied_) {
    inArgs_ = inArgs;
    inArgsCopied_ = false;
  }

  fillMergeTypes(passType, &mergeTypes_);
  allocGradBufs();
  startTask(taskType);

  getOutArgs(outArgs, passType);
}

void MultiGradientMachine::backward(const UpdateCallback& callback) {
  backwardCallback_ = callback;
  startTask(TASK_BACKWARD);
  backwardImp(callback);
}

269 270 271 272
void MultiGradientMachine::forwardBackward(const std::vector<Argument>& inArgs,
                                           std::vector<Argument>* outArgs,
                                           PassType passType,
                                           const UpdateCallback& callback) {
Z
zhangjinchao01 已提交
273 274 275 276 277
  backwardCallback_ = callback;
  forwardImp(inArgs, outArgs, passType, TASK_FORWARD_BACKWARD);
  backwardImp(callback);
}

L
liaogang 已提交
278
Argument MultiGradientMachine::getLayerOutput(const std::string& layerName) {
L
liaogang 已提交
279 280
  std::vector<Argument> args;
  args.reserve(threads_.size());
281

L
liaogang 已提交
282 283
  for (auto& thread : threads_) {
    args.push_back(thread->getGradientMachine()->getLayerOutput(layerName));
284
  }
L
liaogang 已提交
285
  outLayerArgs_.concat(args, false /* use_gpu */, outArgStream_, passType_);
286

L
liaogang 已提交
287
  return outLayerArgs_;
288 289
}

290
void MultiGradientMachine::backwardImp(const UpdateCallback& callback) {
Z
zhangjinchao01 已提交
291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334
  for (size_t i = 0; i < parameters_.size(); i++) {
    if (!parameters_[i]->useGpu() || parameters_[i]->isStatic()) continue;
    REGISTER_TIMER("controller_dequeue");
    gradQueue_.dequeue();
  }
  if (hasNonstaticCpuParamters()) {
    waitAfterMerge();
    if (backwardCallback_) {
      for (auto& para : parameters_) {
        if (!para->useGpu() && !para->isStatic()) {
          backwardCallback_(para.get());
        }
      }
    }
  }
}

void MultiGradientMachine::updateThreadParameters() {
  for (size_t pid = 0; pid < parameters_.size(); ++pid) {
    if (!parameters_[pid]->useGpu()) continue;
    if (!parameters_[pid]->isValueUpdated()) continue;
    parameters_[pid]->clearValueUpdated();
    for (int i = 0; i < (int)threads_.size(); i++) {
      threads_[i]->incUpdateCounter();
    }
    // NotifyValueReady should happen after that all threads' incUpdateCounter()
    // are called so that the counters are correct when notifyValueReady()
    // is called.
    threads_[paraMainThread_[pid]]->notifyValueReady(pid);
  }
}

void MultiGradientMachine::onPassEnd() {
  for (auto& thread : threads_) {
    thread->onPassEnd();
  }
}

void MultiGradientMachine::finish() {
  for (auto& thread : threads_) {
    thread->stop();
  }
}

Y
Yu Yang 已提交
335
Evaluator* MultiGradientMachine::makeEvaluator() const {
Z
zhangjinchao01 已提交
336 337 338
  return threads_[0]->getGradientMachine()->makeEvaluator();
}

Y
Yu Yang 已提交
339
void MultiGradientMachine::eval(Evaluator* evaluator) const {
Z
zhangjinchao01 已提交
340 341
  for (auto& thread : threads_) {
    SetDevice device(thread->getDeviceId());
H
hedaoyuan 已提交
342 343 344
    if (thread->hasInputData()) {
      thread->getGradientMachine()->eval(evaluator);
    }
Z
zhangjinchao01 已提交
345 346 347
  }
}

348 349
void MultiGradientMachine::getOutArgs(std::vector<Argument>* outArgs,
                                      PassType passType) {
Z
zhangjinchao01 已提交
350 351 352 353
  for (auto& thread : threads_) {
    REGISTER_TIMER("waitOutArgs");
    thread->waitOutArgsReady();
  }
H
Format  
hedaoyuan 已提交
354

H
hedaoyuan 已提交
355
  outArgs_.resize(threads_[threads_.size() - 1]->getOutArgs().size());
Z
zhangjinchao01 已提交
356 357 358 359 360 361

  REGISTER_TIMER("copyOutArgs");
  for (size_t i = 0; i < outArgs_.size(); ++i) {
    std::vector<Argument> args;
    args.reserve(threads_.size());
    for (auto& thread : threads_) {
H
hedaoyuan 已提交
362 363 364 365 366
      // If the thread input is empty, then the output is empty.
      auto tmp = thread->getOutArgs();
      if (tmp.size() > 0) {
        args.push_back(tmp[i]);
      }
Z
zhangjinchao01 已提交
367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391
    }
    outArgs_[i].concat(args, useGpu_, outArgStream_, passType);
  }

  if (useGpu_) {
    hl_stream_synchronize(outArgStream_);
  }

  *outArgs = outArgs_;
}

void MultiGradientMachine::setOutputGrad(const std::vector<Argument>& args) {
  CHECK_EQ(args.size(), outArgs_.size());
  for (size_t i = 0; i < args.size(); i++) {
    outArgs_[i].grad = args[i].grad;
  }
}

void MultiGradientMachine::startTask(TaskType taskType) {
  taskType_ = taskType;
  for (auto& thread : threads_) {
    thread->notifyTaskReady();
  }
}

392 393 394
TrainerThread::TrainerThread(const ModelConfig& config,
                             int threadId,
                             MultiGradientMachine* multiMachine)
Z
zhangjinchao01 已提交
395 396 397 398 399 400 401 402 403 404 405 406 407
    : multiMachine_(multiMachine),
      config_(config),
      threadId_(threadId),
      inArgsCopied_(false) {
  int numThreads = multiMachine->getNumThreads();

  auto& mainParas = multiMachine->getParameters();

  using std::placeholders::_1;
  using std::placeholders::_2;

  partnerId_ = mod(threadId_ - 1, numThreads);

408 409 410
  deviceId_ = !multiMachine_->useGpu()
                  ? -1
                  : multiMachine_->logicalDeviceId2RealDeviceId(0, threadId_);
Z
zhangjinchao01 已提交
411 412 413 414 415 416 417 418 419
  SetDevice gpuDevice(deviceId_);

  NeuralNetwork* nn = nullptr;
  if (!multiMachine->useGpu() || !FLAGS_parallel_nn) {
    nn = NeuralNetwork::create(config);
  } else {
    nn = new ParallelNeuralNetwork();
    for (auto& paraConfig : *config_.mutable_parameters()) {
      if (paraConfig.device() != -1) {
420
        paraConfig.set_device(multiMachine_->logicalDeviceId2RealDeviceId(
Z
zhangjinchao01 已提交
421 422 423 424 425
            paraConfig.device(), threadId_));
      }
    }
    for (auto& layerConfig : *config_.mutable_layers()) {
      if (layerConfig.device() != -1) {
426
        layerConfig.set_device(multiMachine_->logicalDeviceId2RealDeviceId(
Z
zhangjinchao01 已提交
427 428 429 430 431
            layerConfig.device(), threadId_));
      }
    }
  }
  // Only GPU do not share parameter values with main paramters.
432 433
  ParamInitCallback slaveParamInitCb =
      std::bind(parameterInitNN, _1, _2, &mainParas);
Z
zhangjinchao01 已提交
434 435 436 437 438 439 440 441 442
  nn->init(config_, slaveParamInitCb);
  gradientMachine_.reset(nn);
  parameters_ = gradientMachine_->getParameters();
  if (!FLAGS_parallel_nn) {
    for (auto& para : parameters_) {
      para->setDevice(deviceId_);
    }
  }

443 444
  backwardCallback_ =
      std::bind(&TrainerThread::backwardCallback, this, std::placeholders::_1);
Z
zhangjinchao01 已提交
445 446 447 448 449 450 451 452

  gradStream_ = HPPL_STREAM_2;
  valueStream_ = HPPL_STREAM_3;
  stopping_ = false;
  updateCounter_ = 0;
  parameterUpdated_ = false;
}

453
TrainerThread::~TrainerThread() { stop(); }
Z
zhangjinchao01 已提交
454 455

void TrainerThread::start() {
456
  gradientMachine_->start();
Z
zhangjinchao01 已提交
457

458
  computeThread_.reset(new std::thread([this]() { computeThread(); }));
Z
zhangjinchao01 已提交
459 460

  if (multiMachine_->useGpu()) {
461 462
    gradCollectThread_.reset(
        new std::thread([this]() { gradCollectThread(); }));
Z
zhangjinchao01 已提交
463

464 465
    valueDispatchThread_.reset(
        new std::thread([this]() { valueDispatchThread(); }));
Z
zhangjinchao01 已提交
466

467
    copyThread_.reset(new std::thread([this]() { copyGradToBufferThread(); }));
Z
zhangjinchao01 已提交
468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536
  }
}

void TrainerThread::stop() {
  if (stopping_) return;

  stopping_ = true;

  if (computeThread_) {
    taskReadySem_.post();
    computeThread_->join();
  }
  if (gradCollectThread_) {
    gradQueue_.enqueue(0);
    gradCollectThread_->join();
  }
  if (copyThread_) {
    gradBufQueue_.enqueue(0);
    copyThread_->join();
  }
  if (valueDispatchThread_) {
    valueReadyQueue_.enqueue(0);
    valueDispatchThread_->join();
  }
}

void TrainerThread::computeThread() {
  VLOG(1) << "gradComputeThread " << threadId_;

  if (deviceId_ >= 0) {
    hl_init(deviceId_);
  }

  while (true) {
    {
      REGISTER_TIMER("taskSem_wait");
      taskReadySem_.wait();
    }

    if (stopping_) break;

    switch (multiMachine_->getTaskType()) {
      case MultiGradientMachine::TASK_FORWARD_BACKWARD:
        forward();
        backward();
        break;
      case MultiGradientMachine::TASK_FORWARD:
        forward();
        break;
      case MultiGradientMachine::TASK_BACKWARD:
        backward();
        break;
      case MultiGradientMachine::TASK_COPY_IN_ARGS:
        copyInArgs();
        inArgsCopied_ = true;
        multiMachine_->waitForCopyInArgs();
        break;
    }
  }
}

void TrainerThread::prefetch() {
  SetDevice setDevice(deviceId_);
  gradientMachine_->prefetch(inArgs_);
}

void TrainerThread::forward() {
  if (!inArgsCopied_) {
    REGISTER_TIMER("copyInArgs");
H
hedaoyuan 已提交
537
    batchSize_ = copyInArgs();
Z
zhangjinchao01 已提交
538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559
  } else {
    inArgsCopied_ = false;
  }

  if (multiMachine_->getPassType() != PASS_TEST) {
    REGISTER_TIMER("clearGradient");
    // For main parameter, the user of MultiGpuSyncMachine is responsible
    // for setting the gradient to zero
    for (size_t i = 0; i < parameters_.size(); i++) {
      if (parameters_[i]->useGpu()) {
        if (multiMachine_->paraMainThread(i) != threadId_) {
          SetDevice device(parameters_[i]->getDeviceId());
          parameters_[i]->clearGradient();
        }
      } else {
        parameters_[i]->clearGradient();
      }
    }
  }

  {
    REGISTER_TIMER("wait_value");
560
    valueReadyCond_.wait([this]() { return !parameterUpdated_; });
Z
zhangjinchao01 已提交
561 562
  }

563
  { fillMergeTypes(multiMachine_->getPassType(), &mergeTypes_); }
Z
zhangjinchao01 已提交
564 565 566

  {
    REGISTER_TIMER("thread_forward");
H
hedaoyuan 已提交
567 568
    if (batchSize_ > 0) {
      gradientMachine_->forward(
H
Format  
hedaoyuan 已提交
569
          inArgs_, &outArgs_, multiMachine_->getPassType());
H
hedaoyuan 已提交
570
    } else {
H
Format  
hedaoyuan 已提交
571
      outArgs_.clear();
H
hedaoyuan 已提交
572
    }
Z
zhangjinchao01 已提交
573 574 575 576 577 578 579 580 581
  }
  outArgsReadySem_.post();
}

void TrainerThread::backward() {
  REGISTER_TIMER("thread_backward");
  if (multiMachine_->isPassGrad()) {
    copyOutputGrad();
  }
H
hedaoyuan 已提交
582 583 584 585 586 587 588
  if (batchSize_ > 0) {
    gradientMachine_->backward(backwardCallback_);
  } else {
    for (size_t i = parameters_.size(); i > 0; i--) {
      backwardCallback(parameters_[i - 1].get());
    }
  }
Z
zhangjinchao01 已提交
589 590 591 592 593 594 595 596 597 598 599 600 601
  if (multiMachine_->hasNonstaticCpuParamters()) {
    mergeCpuGradients();
  }
}

void TrainerThread::backwardCallback(Parameter* para) {
  // CPU parameters are merged in the end
  if (!para->useGpu()) return;

  int paramId = para->getID();
  if (multiMachine_->getNumThreads() == 1) {
    // no need to do merge if there is only one thread
    doCallback(paramId);
602 603
  } else if (threadId_ == mod(multiMachine_->paraMainThread(paramId) - 1,
                              multiMachine_->getNumThreads())) {
Z
zhangjinchao01 已提交
604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623
    notifyCopyGradToBuffer(paramId);
  } else {
    notifyGradientCollect(paramId);
  }
}

void TrainerThread::copyGradToBufferThread() {
  VLOG(1) << "copyGradToBufferThread " << threadId_;

  if (deviceId_ >= 0) {
    hl_init(deviceId_);
  }
  auto& partnerThread = multiMachine_->getThread(partnerId_);
  auto& gradBufs = multiMachine_->getGradBuf(partnerId_);

  while (true) {
    int pid = gradBufQueue_.dequeue();
    if (stopping_) break;

    int pdeviceId = multiMachine_->realDeviceId2LogicalDeviceId(
624
        parameters_[pid]->getDeviceId(), threadId_);
Z
zhangjinchao01 已提交
625 626 627 628 629 630 631 632 633 634 635 636 637

    auto& gradBuf = gradBufs[pdeviceId];

    {
      REGISTER_TIMER("waitBufferReady");
      gradBuf.sem.wait();
    }

    {
      REGISTER_TIMER("copyGradToBuffer");
      SetDevice setDevice(parameters_[pid]->getDeviceId());
      for (size_t i = 0; i < mergeTypes_.size(); ++i) {
        gradBuf.bufs[i]->resize(
638 639 640
            parameters_[pid]->getBuf(mergeTypes_[i])->getSize());
        gradBuf.bufs[i]->copyFrom(*parameters_[pid]->getBuf(mergeTypes_[i]),
                                  gradStream_);
Z
zhangjinchao01 已提交
641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665
      }
      hl_stream_synchronize(gradStream_);
    }
    partnerThread->notifyGradientCollect(pid);
  }
}

void TrainerThread::gradCollectThread() {
  VLOG(1) << "gradCollectThread " << threadId_;

  if (deviceId_ >= 0) {
    hl_init(deviceId_);
  }

  std::vector<size_t> gradReadyCount(parameters_.size(), 0);

  auto& gradBufs = multiMachine_->getGradBuf(threadId_);

  while (true) {
    int pid = gradQueue_.dequeue();
    if (stopping_) break;

    if (++gradReadyCount[pid] < 2) continue;
    gradReadyCount[pid] = 0;
    int pdeviceId = multiMachine_->realDeviceId2LogicalDeviceId(
666
        parameters_[pid]->getDeviceId(), threadId_);
Z
zhangjinchao01 已提交
667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739

    auto& gradBuf = gradBufs[pdeviceId];

    {
      REGISTER_TIMER("mergeGrad");
      for (size_t i = 0; i < mergeTypes_.size(); ++i) {
        ParameterType type = mergeTypes_[i];
        const VectorPtr& localGrad = parameters_[pid]->getBuf(type);
        SetDevice setDevice(parameters_[pid]->getDeviceId());
        localGrad->add(*gradBuf.bufs[i]);
      }
    }

    gradBuf.sem.post();

    if (multiMachine_->paraMainThread(pid) == threadId_) {
      doCallback(pid);
    } else {
      notifyCopyGradToBuffer(pid);
    }
  }
}

void TrainerThread::doCallback(int pid) {
  REGISTER_TIMER("callback");
  auto& gpuThreads = multiMachine_->getAllThreads();
  if (multiMachine_->getBackwardCallback()) {
    // The callback supplied by the user of MultiGradientMachine may handle
    // the parameter update using the gradient.
    multiMachine_->getBackwardCallback()(parameters_[pid].get());
    if (parameters_[pid]->isValueUpdated()) {
      parameters_[pid]->clearValueUpdated();
      for (auto& thread : gpuThreads) {
        thread->incUpdateCounter();
      }
      notifyValueReady(pid);
    }
  }
  multiMachine_->notifyGradientTransfer(pid);
}

void TrainerThread::valueDispatchThread() {
  VLOG(1) << "valueDispatchThread " << threadId_;

  if (deviceId_ >= 0) {
    hl_init(deviceId_);
  }

  auto& thread = multiMachine_->getThread(partnerId_);

  while (true) {
    int pid;
    {
      REGISTER_TIMER("value_dequeue");
      pid = valueReadyQueue_.dequeue();
    }
    if (stopping_) break;

    if (multiMachine_->paraMainThread(pid) == partnerId_) continue;

    {
      REGISTER_TIMER("copyValue");
      SetDevice setDevice(parameters_[pid]->getDeviceId());
      thread->getValueBuf(pid)->copyFrom(*getValueBuf(pid), valueStream_);
      hl_stream_synchronize(valueStream_);
    }

    thread->notifyValueReady(pid);
  }
}

void TrainerThread::notifyValueReady(int paramId) {
  if (--updateCounter_ == 0) {
740
    valueReadyCond_.notify_all([this] { parameterUpdated_ = false; });
Z
zhangjinchao01 已提交
741 742 743 744 745
  }

  notifyValueDispatch(paramId);
}

H
hedaoyuan 已提交
746
int TrainerThread::copyInArgs() {
Z
zhangjinchao01 已提交
747
  const std::vector<Argument>& fullInArgs = multiMachine_->getInArgs();
748
  int numThreads = multiMachine_->getAllThreads().size();
Z
zhangjinchao01 已提交
749 750 751 752 753 754 755 756 757 758 759 760 761
  int32_t numSequences = fullInArgs[0].getNumSequences();
  int32_t startSeq = numSequences * threadId_ / numThreads;
  int32_t endSeq = numSequences * (threadId_ + 1) / numThreads;
  int32_t copySize = endSeq - startSeq;

  /**
   * For the first copy, need to allocate space here
   */
  if (inArgs_.size() == 0) {
    inArgs_.resize(fullInArgs.size());
  }

  if (copySize == 0) {
H
hedaoyuan 已提交
762
    return 0;
Z
zhangjinchao01 已提交
763 764
  }

765
  for (size_t i = 0; i < fullInArgs.size(); i++) {
Z
zhangjinchao01 已提交
766
    inArgs_[i].resizeAndCopyFrom(
767 768 769
        fullInArgs[i],
        startSeq,
        copySize,
Z
zhangjinchao01 已提交
770 771
        FLAGS_parallel_nn ? false : multiMachine_->useGpu());
  }
H
hedaoyuan 已提交
772
  return copySize;
Z
zhangjinchao01 已提交
773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814
}

void TrainerThread::mergeCpuGradients() {
  CHECK_EQ(mergeTypes_.size(), 1UL);
  CHECK_EQ(mergeTypes_[0], PARAMETER_GRADIENT);

  {
    REGISTER_TIMER("waitbeforeMerge");
    multiMachine_->waitBeforeMerge();
  }
  std::vector<const std::vector<ParameterPtr>*> slaveParameters =
      multiMachine_->getSlaveParameters();

  CHECK(slaveParameters.size());
  for (auto& para : multiMachine_->getNonStaticParameters()) {
    if (para->useGpu()) continue;
    if (para->isSparseRemoteUpdate()) {
      REGISTER_TIMER("mergeRemoteGradSparse");
      mergeGradSparseRemote(para.get(), slaveParameters);
    } else if (para->isGradSparseUpdate()) {
      REGISTER_TIMER("mergeGradSparse");
      mergeGradSparse(para.get(), slaveParameters);
    } else {
      REGISTER_TIMER("mergeGradDense");
      mergeGradDense(para.get(), slaveParameters);
    }
  }
  {
    REGISTER_TIMER("waitbeforeMerge");
    multiMachine_->waitAfterMerge();
  }
}

void TrainerThread::mergeGradSparse(
    Parameter* para,
    std::vector<const std::vector<ParameterPtr>*>& slaveParameters) {
  size_t pid = para->getID();
  SparseRowIdsCpuMatrix* mainMat = dynamic_cast<SparseRowIdsCpuMatrix*>(
      para->getMat(PARAMETER_GRADIENT).get());
  std::vector<uint32_t>& ids = mainMat->getIds(threadId_);

  for (auto slaveParams : slaveParameters) {
815 816
    SparseRowCpuMatrix* mat = dynamic_cast<SparseRowCpuMatrix*>(
        (*slaveParams)[pid]->getMat(PARAMETER_GRADIENT).get());
Z
zhangjinchao01 已提交
817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845
    mat->addTo(*mainMat, ids, threadId_, multiMachine_->getNumThreads());
    // we use a sample hash method(%) instead of range partition,
    // because range partition has balance issue sometimes,
    // when feature ids are not generated from hashcode.
  }
  uniqueIds(ids);
}

void TrainerThread::mergeGradSparseRemote(
    Parameter* para,
    std::vector<const std::vector<ParameterPtr>*>& slaveParameters) {
  size_t pid = para->getID();
  SparseRowCpuMatrix* mainMat =
      dynamic_cast<SparseRowCpuMatrix*>(para->getMat(PARAMETER_GRADIENT).get());

  mainMat->checkIndices();
  mainMat->zeroMemThread(threadId_, multiMachine_->getNumThreads());

  for (auto slaveParams : slaveParameters) {
    SparseRowCpuMatrix* mat = dynamic_cast<SparseRowCpuMatrix*>(
        (*slaveParams)[pid]->getMat(PARAMETER_GRADIENT).get());
    mat->addTo(*mainMat, threadId_, multiMachine_->getNumThreads());
  }
}

void TrainerThread::mergeGradDense(
    Parameter* para,
    std::vector<const std::vector<ParameterPtr>*>& slaveParameters) {
  size_t pid = para->getID();
846 847 848 849
  auto interval = calcSplitArrayInterval(para->getSize(),
                                         (size_t)threadId_,
                                         multiMachine_->getNumThreads(),
                                         8LU /*for avx*/);
Z
zhangjinchao01 已提交
850 851 852 853 854 855 856 857 858 859 860
  size_t startSeq = interval.first;
  size_t copySize = interval.second - interval.first;

  // setup sub bufs
  CpuVector destGrad(0, nullptr);
  destGrad.subVecFrom(*para->getBuf(PARAMETER_GRADIENT), startSeq, copySize);

  // merge
  CpuVector slaveGradSub(0, nullptr);
  for (auto slaveParams : slaveParameters) {
    slaveGradSub.subVecFrom(
861
        *(*slaveParams)[pid]->getBuf(PARAMETER_GRADIENT), startSeq, copySize);
Z
zhangjinchao01 已提交
862 863 864 865 866 867 868 869 870 871 872 873 874
    destGrad.add(slaveGradSub);
  }
}

void TrainerThread::copyOutputGrad() {
  const std::vector<Argument>& outputGradArgs = multiMachine_->outArgs_;
  int numThreads = multiMachine_->getAllThreads().size();
  int32_t numSequences = outputGradArgs[0].getNumSequences();
  int32_t startSeq = numSequences * threadId_ / numThreads;
  int32_t endSeq = numSequences * (threadId_ + 1) / numThreads;
  int32_t copySize = endSeq - startSeq;
  outArgs_.resize(outputGradArgs.size());
  for (size_t i = 0; i < outputGradArgs.size(); i++) {
875 876 877
    outArgs_[i].resizeAndCopyFrom(outputGradArgs[i],
                                  startSeq,
                                  copySize,
Z
zhangjinchao01 已提交
878 879 880 881 882 883 884 885 886
                                  multiMachine_->useGpu(),
                                  HPPL_STREAM_DEFAULT);
  }
  if (multiMachine_->useGpu()) {
    hl_stream_synchronize(HPPL_STREAM_DEFAULT);
  }
  gradientMachine_->setOutputGrad(outArgs_);
}
}  // namespace paddle