ParameterClient2.cpp 26.6 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

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 <unistd.h>

#include "ParameterClient2.h"
X
Xin Pan 已提交
18
#include "paddle/legacy/math/SparseRowMatrix.h"
Z
zhangjinchao01 已提交
19 20
#include "paddle/utils/Flags.h"
#include "paddle/utils/Stat.h"
Y
Yu Yang 已提交
21
#include "paddle/utils/StringUtil.h"
Z
zhangjinchao01 已提交
22

23 24
DEFINE_string(pservers, "127.0.0.1", "Comma separated addresses of pservers");
DEFINE_int32(parallel_thread_num, 1, "Thread number for parameter send");
Z
zhangjinchao01 已提交
25 26 27

namespace paddle {

Y
Yu Yang 已提交
28 29 30
template <typename T1, typename T2>
void copyToRepeatedField(google::protobuf::RepeatedField<T1>* dest,
                         const T2* src,
Z
zhangjinchao01 已提交
31 32 33 34 35 36 37 38 39 40 41
                         size_t size) {
  dest->Clear();
  dest->Reserve(size);
  for (size_t i = 0; i < size; ++i) {
    dest->AddAlreadyReserved(src[i]);
  }
}

ParameterClient2::ParameterClient2(bool separate, int port, int numPorts)
    : BaseClient(separate, numPorts), port_(port) {
#ifndef PADDLE_DISABLE_TIMER
42
  forwardbackwordTime_ = 0;
Z
zhangjinchao01 已提交
43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82
#endif
}

int ParameterClient2::calcParameterBlockSize(
    const std::vector<ParameterPtr>& parameters, size_t serviceNum) {
  size_t totalSize = 0;
  for (auto& para : parameters) {
    totalSize += para->getSize();
  }
  size_t perServerSize = totalSize / serviceNum;

  int sizeBits = 64 - __builtin_clzl(perServerSize);

  /// 2^10 is min block size
  /// 2^7 will be max number of blocks in one pserver
  int blockSizeBits = std::max((sizeBits - 7), 10);
  return 1 << blockSizeBits;
}

void ParameterClient2::initThreads() {
  threadNum_ = serviceNum_;
  if (FLAGS_parallel_thread_num > 1) {
    LOG(INFO) << "parallel_thread_num dosent need to set";
  }
  syncThreadPool_.reset(new SyncThreadPool(threadNum_));
  startThreads();
}

bool ParameterClient2::init(const std::vector<ParameterPtr>& parameters) {
  destroy();

  std::vector<std::string> hosts;
  str::split(FLAGS_pservers, ',', &hosts);
  serviceNum_ = hosts.size() * numPorts_;
  uint64_t denseBlockSize = calcParameterBlockSize(parameters, serviceNum_);

  /// setup prefetch matrix if exists
  for (auto& para : parameters) {
    /// set block size for each parameter
    para->getConfig().set_parameter_block_size(
83 84
        para->getConfig().sparse_remote_update() ? para->getConfig().dims(1)
                                                 : denseBlockSize);
Z
zhangjinchao01 已提交
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
  }

  for (auto& para : parameters) {
    CHECK_NE(-1UL, para->getID()) << "id in parameter is not initialized";
    parameterMap_[para->getID()] = para;
  }

  allSegments_.reserve(parameters.size());

  for (auto& para : parameters) {
    ParameterSegments segments;
    segments.name = para->getName();
    segments.id = para->getID();
    allSegments_.push_back(segments);
    if (para->getConfig().sparse_remote_update()) {
      CHECK_EQ(para->getConfig().parameter_block_size(),
101
               para->getConfig().dims(1))
Z
zhangjinchao01 已提交
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145
          << "For sparse remote update parameter,"
          << " block size is the width of each row.";
    }
  }

  /// init clients
  clients_.reserve(serviceNum_);
  recvDataMems_.resize(serviceNum_);

  for (size_t i = 0; i < hosts.size(); ++i) {
    for (int j = 0; j < numPorts_; ++j) {
      LOG(INFO) << "pserver " << i * numPorts_ + j << " " << hosts[i] << ":"
                << port_ + j;
      if (FLAGS_rdma_tcp == "rdma") {
        clients_.emplace_back(hosts[i], port_ + j, F_RDMA);
      } else {
        clients_.emplace_back(hosts[i], port_ + j, F_TCP);
      }
    }
  }

  sparseDistribution_.reset(new SparseParameterDistribution(serviceNum_));

  sleep(2);

  initThreads();

  return true;
}

ParameterClient2::~ParameterClient2() { destroy(); }

void ParameterClient2::destroy() {
  if (clients_.empty()) {
    /// this means not initialized.
    return;
  }
  finishThreads();

  parameterMap_.clear();
  allSegments_.clear();
  clients_.clear();
}

146 147
void ParameterClient2::sendParallel(int tid,
                                    size_t numThreads,
Z
zhangjinchao01 已提交
148 149 150 151 152 153 154 155 156 157
                                    ParameterType recvParameterType) {
  int numMyClients = divup(serviceNum_ - tid, numThreads);

  for (int j = 0; j < numMyClients; ++j) {
    REGISTER_TIMER("client_sendAndRecv_send");
    int i = numThreads * j + tid;
    /// Try to make different clients to send data to different pservers
    /// at the same time so that they will not flood data to the same
    /// pserver.
    i = calcClientId(i, serviceNum_);
158 159
    clients_[i].send("sendParameter",
                     sendJob_.parallelRequests[i],
Z
zhangjinchao01 已提交
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
                     sendJob_.parallelInputIovs[i]);

    /// clear large structure
    sendJob_.parallelRequests[i].Clear();
    sendJob_.parallelInputIovs[i].clear();
  }

  std::vector<void*> bufs;
  SendParameterResponse response;
  for (int j = 0; j < numMyClients; ++j) {
    REGISTER_TIMER("client_sendAndRecv_recv");
    int i = numThreads * j + tid;
    i = calcClientId(i, serviceNum_);
    auto msgReader = clients_[i].recv(&response);
    CHECK_EQ(msgReader->getNumBlocks(), (size_t)response.blocks_size());
    bufs.clear();
    bufs.reserve(response.blocks_size());
    for (auto& block : response.blocks()) {
      auto it = parameterMap_.find(block.para_id());
      CHECK(it != parameterMap_.end());
      Parameter* parameter = it->second.get();
      real* buf = nullptr;
      if (parameter->getBuf(recvParameterType)) {
        buf = parameter->getBuf(recvParameterType)->getPoint(block.begin_pos());
      } else {
        auto recvMat = dynamic_cast<SparseRowCpuMatrix*>(
            parameter->getMat(recvParameterType).get());
        CHECK(recvMat);
        size_t width = parameter->getConfig().dims(1);
T
typhoonzero 已提交
189
        // TODO(wuyi): need add lock here? may also cause resize.
Z
zhangjinchao01 已提交
190 191 192 193 194 195 196 197 198 199 200
        buf = recvMat->getLocalRow(block.begin_pos() / width);
      }
      /// sparse_id is not useful while receiving data since sparse data
      /// storage is continuous, do commit recieved data as that of dense.
      bufs.push_back(buf);
    }
    msgReader->readBlocks(bufs);
  }
}

void ParameterClient2::prepareSendData(
201 202 203 204 205 206 207 208 209
    ParameterUpdateMode updateMode,
    ParameterType parameterType,
    const std::vector<ParameterSegments>& parameterSegments,
    int64_t numSamples,
    real cost,
    bool sendBackParameter,
    ParameterType sendBackParameterType,
    BatchStatus batchStatus,
    SendJob* sendJob) {
Z
zhangjinchao01 已提交
210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226
  sendJob->parallelRequests.resize(serviceNum_);
  sendJob->parallelInputIovs.resize(serviceNum_);

  for (auto& request : sendJob->parallelRequests) {
#ifndef PADDLE_DISABLE_TIMER
    if (updateMode == PSERVER_UPDATE_MODE_ADD_GRADIENT) {
      request.set_forwardbackward_time(forwardbackwordTime_);
    }
#endif
    request.set_trainer_id(trainerId_);
    request.set_update_mode(updateMode);
    request.set_send_back_parameter(sendBackParameter);
    request.set_send_back_parameter_type(sendBackParameterType);
    request.set_num_samples(numSamples);
    request.set_cost(cost);
    request.set_batch_status(batchStatus);
    CHECK_EQ(request.blocks_size(), 0);
武毅 已提交
227 228 229 230 231 232 233 234
    VLOG(10) << "request: trainer_id: " << request.trainer_id()
             << " update_mode" << request.update_mode()
             << " send_back_parameter: " << request.send_back_parameter()
             << " send_back_parameter_type: "
             << request.send_back_parameter_type()
             << " num_samples: " << request.num_samples()
             << " cost: " << request.cost()
             << " batch_status: " << request.batch_status();
Z
zhangjinchao01 已提交
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
  }
  for (const auto& segments : parameterSegments) {
    const auto it = parameterMap_.find(segments.id);
    CHECK(it != parameterMap_.end());
    Parameter* parameter = it->second.get();
    CHECK(parameter != nullptr) << "parameter is nullptr";
    int64_t nameHash = std::hash<std::string>()(segments.name);
    bool sendingPara = !(updateMode == PSERVER_UPDATE_MODE_GET_PARAM ||
                         updateMode == PSERVER_UPDATE_MODE_GET_PARAM_SPARSE ||
                         updateMode == PSERVER_UPDATE_MODE_SET_PARAM_ZERO);
    bool sparseUpdate = parameter->getConfig().sparse_remote_update() &&
                        (updateMode == PSERVER_UPDATE_MODE_ADD_GRADIENT ||
                         updateMode == PSERVER_UPDATE_MODE_ASYNC_SGD ||
                         updateMode == PSERVER_UPDATE_MODE_GET_PARAM_SPARSE);

    const auto blockSize = parameter->getConfig().parameter_block_size();
    CHECK_GE(blockSize, 1LU) << "blockSize should > 0 " << blockSize;
    const auto paraSize = parameter->getSize();
    if (sparseUpdate) {
254 255
      auto prefetchMat = std::dynamic_pointer_cast<SparsePrefetchRowCpuMatrix>(
          parameter->getMat(PARAMETER_VALUE));
Z
zhangjinchao01 已提交
256 257
      CHECK(prefetchMat != nullptr) << "prefetchMat is nullptr";
      auto sendMat = dynamic_cast<SparseRowCpuMatrix*>(
258
          parameter->getMat(parameterType).get());
Z
zhangjinchao01 已提交
259 260 261
      CHECK(sendMat != nullptr) << "sendMat is nullptr";

      syncThreadPool_->exec([&](int tid, size_t numThreads) {
武毅 已提交
262
        std::lock_guard<std::mutex> guard(sparseAutoGrowthMutex_);
263
        const auto& localIndices = prefetchMat->getLocalIndices();
Z
zhangjinchao01 已提交
264 265 266 267
        /// num of sparse rows
        size_t nLocalBlocks = localIndices.size();
        uint64_t beginDim = 0;
        uint64_t endDim = 0;
武毅 已提交
268

T
typhoonzero 已提交
269 270 271
        // HACK(typhoonzero): let it resize first
        prefetchMat->getLocalRow(nLocalBlocks);
        sendMat->getLocalRow(nLocalBlocks);
武毅 已提交
272

Z
zhangjinchao01 已提交
273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293
        for (size_t row = 0; row < nLocalBlocks; ++row) {
          int64_t blockId = localIndices[row];  // local row -> sparse row
          int serverId = std::abs((blockId + nameHash) % serviceNum_);
          if (serverId % numThreads != (size_t)tid) {
            continue;
          }

          beginDim = blockId * blockSize;
          endDim = std::min<int64_t>(beginDim + blockSize, paraSize);

          auto& request = sendJob->parallelRequests[serverId];
          ParameterBlock* block = request.add_blocks();
          block->set_para_id(segments.id);
          /// global sparse row id
          block->set_block_id(blockId);
          /// local row offset
          block->set_begin_pos(row * blockSize);
          /// block len
          block->set_block_size(endDim - beginDim);
          if (sendingPara) {
            sendJob->parallelInputIovs[serverId].push_back(
294
                {sendMat->getLocalRow(row), sizeof(real) * (size_t)blockSize});
Z
zhangjinchao01 已提交
295 296
            /// detect sparse parameter distribution
            sparseDistribution_->probeDistribution(serverId,
297
                                                   sizeof(real) * blockSize);
Z
zhangjinchao01 已提交
298 299 300 301 302
          }
        }
      });

    } else {  /// parameter set for dense and sparse
303 304
      real* buf =
          sendingPara ? parameter->getBuf(parameterType)->getPoint(0) : nullptr;
Z
zhangjinchao01 已提交
305 306 307 308 309 310 311 312 313 314 315 316 317
      uint64_t endDim = 0;
      for (uint64_t beginDim = 0; beginDim < paraSize; beginDim = endDim) {
        endDim = std::min<int64_t>(beginDim + blockSize, paraSize);
        int64_t blockId = beginDim / blockSize;
        int serverId = std::abs((blockId + nameHash) % serviceNum_);

        auto& request = sendJob->parallelRequests[serverId];
        ParameterBlock* block = request.add_blocks();
        block->set_para_id(segments.id);
        block->set_block_id(blockId);
        block->set_begin_pos(beginDim);
        block->set_block_size(endDim - beginDim);
        if (buf) {
318 319
          sendJob->parallelInputIovs[serverId].push_back(
              {buf + beginDim, sizeof(real) * ((size_t)(endDim - beginDim))});
Z
zhangjinchao01 已提交
320 321 322 323 324 325 326 327 328
        }
      }
    }
  }  // parameterSegments

  sparseDistribution_->checkAndResetDistribution();
}

void ParameterClient2::sendAndReceiveParameter(
329 330 331 332 333 334 335
    ParameterUpdateMode updateMode,
    ParameterType parameterType,
    const std::vector<ParameterSegments>& parameterSegments,
    int64_t numSamples,
    real cost,
    bool sendBackParameter,
    ParameterType sendBackParameterType,
Z
zhangjinchao01 已提交
336
    ParameterType recvParameterType) {
337 338 339 340 341 342 343 344 345
  prepareSendData(updateMode,
                  parameterType,
                  parameterSegments,
                  numSamples,
                  cost,
                  sendBackParameter,
                  sendBackParameterType,
                  /*batchStatus = */ BATCH_START_AND_FINISH,
                  &sendJob_);
Z
zhangjinchao01 已提交
346 347 348 349 350 351 352

  syncThreadPool_->exec([&](int tid, size_t numThreads) {
    this->sendParallel(tid, numThreads, recvParameterType);
  });
}

void ParameterClient2::sendParameter(
353 354 355 356 357 358 359
    ParameterUpdateMode updateMode,
    ParameterType parameterType,
    const std::vector<ParameterSegments>& parameterSegments,
    int64_t numSamples,
    real cost,
    bool sendBackParameter,
    BatchStatus batchStatus) {
Z
zhangjinchao01 已提交
360
  SendJobPtr sendJob = std::make_shared<SendJob>();
361 362 363 364 365 366 367 368
  prepareSendData(updateMode,
                  parameterType,
                  parameterSegments,
                  numSamples,
                  cost,
                  sendBackParameter,
                  PARAMETER_VALUE,
                  batchStatus,
Z
zhangjinchao01 已提交
369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395
                  sendJob.get());

  for (int i = 0; i < threadNum_; i++) {
    sendJobQueue_[i]->enqueue(sendJob);
  }
}

void ParameterClient2::recvParameter() { recvSyncBarrier_->wait(); }

void ParameterClient2::send(int threadId) {
  int index = threadId;
  LOG(INFO) << "send thread " << threadId << " started";
  int numMyClients = divup(serviceNum_ - index, threadNum_);
  while (true) {
    SendJobPtr recvJob = sendJobQueue_[index]->dequeue();
    if (stopping_) {
      recvJobQueue_[index]->enqueue(recvJob);
      break;
    }
    for (int j = 0; j < numMyClients; ++j) {
      REGISTER_TIMER("client_send");
      int i = threadNum_ * j + index;
      /// Try to make different clients to send data to different pservers
      /// at the same time so that they will not flood data to the same
      /// pserver.
      i = calcClientId(i, serviceNum_);
      if (recvJob->parallelRequests.size()) {
396 397
        clients_[i].send("sendParameter",
                         recvJob->parallelRequests[i],
Z
zhangjinchao01 已提交
398 399
                         recvJob->parallelInputIovs[i]);
      } else {
400 401
        clients_[i].send("sendData",
                         recvJob->parallelDataRequests[i],
Z
zhangjinchao01 已提交
402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623
                         recvJob->parallelInputIovs[i]);
      }
    }
    recvJobQueue_[index]->enqueue(recvJob);
  }
}

void ParameterClient2::recv(int threadId) {
  LOG(INFO) << "recv thread " << threadId << " started";
  int index = threadId;
  int numMyClients = divup(serviceNum_ - index, threadNum_);
  while (true) {
    std::vector<void*> bufs;
    SendParameterResponse response;
    SendDataResponse dataResponse;
    SendJobPtr recvJob = recvJobQueue_[index]->dequeue();
    if (stopping_) break;
    for (int j = 0; j < numMyClients; ++j) {
      REGISTER_TIMER("client_recv");
      int i = threadNum_ * j + index;
      i = calcClientId(i, serviceNum_);
      if (recvJob->parallelRequests.size()) {
        auto msgReader = clients_[i].recv(&response);
        CHECK_EQ(msgReader->getNumBlocks(), (size_t)response.blocks_size());
        bufs.clear();
        bufs.reserve(response.blocks_size());
        for (auto& block : response.blocks()) {
          auto it = parameterMap_.find(block.para_id());
          CHECK(it != parameterMap_.end());
          Parameter* parameter = it->second.get();
          real* buf =
              parameter->getBuf(PARAMETER_VALUE)->getPoint(block.begin_pos());
          CHECK_EQ(msgReader->getBlockLength(bufs.size()),
                   sizeof(real) * (block.block_size()));
          bufs.push_back(buf);
        }
        msgReader->readBlocks(bufs);
      } else {
        auto msgReader = clients_[i].recv(&dataResponse);
        CHECK_EQ(msgReader->getNumBlocks(), (size_t)dataResponse.blocks_size());
        size_t totalLen = msgReader->getTotalLength();
        if (0 == totalLen) {
          continue;
        }
        auto& recvMem = recvDataMems_[dataResponse.server_id()];
        CHECK_EQ(dataResponse.blocks_size(), 1)
            << "Only one block currently support now!";
        auto& block = dataResponse.blocks(0);
        CHECK_EQ(totalLen % sizeof(block.data_size()), 0U);
        recvMem = std::make_shared<CpuMemoryHandle>(totalLen);
        msgReader->readNextBlock(recvMem.get()->getBuf());
      }
    }
    recvSyncBarrier_->wait();
  }
}

void ParameterClient2::waitPassStart() {
  WaitPassStartRequest request;
  std::vector<WaitPassStartResponse> responses;
  multiCall(__func__, request, &responses);
}

void ParameterClient2::waitPassFinish() {
  WaitPassFinishRequest request;
  std::vector<WaitPassFinishResponse> responses;
  multiCall(__func__, request, &responses);
}

void ParameterClient2::synchronize(SyncObject syncObjectId) {
  SynchronizeRequest request;
  request.set_sync_object_id(syncObjectId);
  std::vector<SynchronizeResponse> responses;
  multiCall(__func__, request, &responses);
}

void ParameterClient2::asyncFinishPass(SyncObject syncObjectId) {
  SynchronizeRequest request;
  request.set_sync_object_id(syncObjectId);
  request.set_trainer_id(trainerId_);
  std::vector<SynchronizeResponse> responses;
  multiCall(__func__, request, &responses);
}

void ParameterClient2::setConfig(const OptimizationConfig& optConfig,
                                 const std::string& saveDir,
                                 bool isSparseServer) {
  SetConfigRequest request;
  std::vector<SetConfigResponse> responses;

  for (auto& nameAndPara : parameterMap_) {
    *request.add_param_configs() = nameAndPara.second->getConfig();
  }

  *request.mutable_opt_config() = optConfig;
  request.set_save_dir(saveDir);
  request.set_is_sparse_server(isSparseServer);

  std::vector<SetConfigRequest> requests;
  requests.resize(clients_.size());
  for (size_t i = 0; i < requests.size(); ++i) {
    requests[i].CopyFrom(request);
    requests[i].set_server_id(i);
  }

  responses.resize(clients_.size());
  size_t numClients = clients_.size();
  for (size_t i = 0; i < numClients; ++i) {
    clients_[i].send(__func__, requests[i]);
  }
  for (size_t i = 0; i < numClients; ++i) {
    clients_[i].recv(&responses[i]);
  }
}

bool ParameterClient2::inStatus(PServerStatus status) {
  GetStatusRequest request;
  std::vector<GetStatusResponse> responses;

  bool ok = true;
  multiCall("getStatus", request, &responses);
  for (auto& response : responses) {
    if (response.status() != status) {
      ok = false;
    }
  }

  return ok;
}

void ParameterClient2::setStatus(PServerStatus status) {
  SetStatusRequest request;
  request.set_status(status);
  std::vector<SetStatusResponse> responses;
  multiCall(__func__, request, &responses);
}

void ParameterClient2::waitForStatus(PServerStatus status) {
  while (!inStatus(status)) {
    sleep(1);
  }
}

template <typename Proto>
static void validateResponses(const std::vector<Proto>& responses) {
  for (auto& response : responses) {
    CHECK(response.return_message().empty())
        << "client" << &response - &responses[0]
        << " error:" << response.return_message();
  }
}

PServerVector ParameterClient2::createVector() {
  CreateVectorRequest request;
  std::vector<CreateVectorResponse> responses;
  int64_t handle = -1;

  multiCall(__func__, request, &responses);
  validateResponses(responses);

  for (auto& response : responses) {
    if (handle == -1) {
      handle = response.handle();
    } else {
      CHECK_EQ(handle, response.handle()) << "Inconsistent handle from client"
                                          << &response - &responses[0] << " "
                                          << handle << " " << response.handle();
    }
  }
  return PServerVector{handle};
}

void ParameterClient2::releaseVector(PServerVector handle) {
  ReleaseVectorRequest request;
  std::vector<ReleaseVectorResponse> responses;

  request.set_handle(handle.handle);
  multiCall(__func__, request, &responses);
  validateResponses(responses);
}

PServerMatrix ParameterClient2::createMatrix(int32_t numCols) {
  CreateMatrixRequest request;
  std::vector<CreateMatrixResponse> responses;
  int64_t handle = -1;

  request.set_num_cols(numCols);
  multiCall(__func__, request, &responses);
  validateResponses(responses);

  for (auto& response : responses) {
    if (handle == -1) {
      handle = response.handle();
    } else {
      CHECK_EQ(handle, response.handle()) << "Inconsistent handle from client"
                                          << &response - &responses[0] << " "
                                          << handle << " " << response.handle();
    }
  }
  return PServerMatrix{handle};
}

void ParameterClient2::releaseMatrix(PServerMatrix handle) {
  ReleaseMatrixRequest request;
  std::vector<ReleaseMatrixResponse> responses;

  request.set_handle(handle.handle);
  multiCall(__func__, request, &responses);
  validateResponses(responses);
}

void PreparedOperations::addOperationHelper(Operation* op, CpuVectorPtr vec) {
  ProtoVector& pvec = *op->add_vectors();
  size_t dim = vec->getSize();
  pvec.set_dim(dim);
  copyToRepeatedField(pvec.mutable_values(), vec->getData(), vec->getSize());
}

void PreparedOperations::addOperationHelper(Operation* op, CpuMatrixPtr mat) {
  ProtoMatrix& pmat = *op->add_matrices();
  pmat.set_num_cols(mat->getWidth());
  pmat.set_num_rows(mat->getHeight());
624 625
  copyToRepeatedField(
      pmat.mutable_values(), mat->getData(), pmat.num_cols() * pmat.num_rows());
Z
zhangjinchao01 已提交
626 627
}

Y
Yu Yang 已提交
628
static inline real addTwo(real a, double b) { return a + b; }
Y
Yu Yang 已提交
629

Z
zhangjinchao01 已提交
630
void ParameterClient2::doOperation(PreparedOperations& ops,
631 632
                                   bool waitForGradient,
                                   bool sendBackGradient,
Z
zhangjinchao01 已提交
633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693
                                   bool releasePass) {
  std::vector<DoOperationResponse> responses;
  ops.request_.set_wait_for_gradient(waitForGradient);
  ops.request_.set_send_back_parameter(sendBackGradient);
  ops.request_.set_release_pass(releasePass);
  multiCall(__func__, ops.request_, &responses);
  validateResponses(responses);
  size_t numPassFinishServers = 0;

  size_t numOps = ops.request_.operations_size();
  for (auto& response : responses) {
    numPassFinishServers += response.pass_finish();
    CHECK_EQ(numOps, (size_t)response.results_size());
    for (size_t opId = 0; opId < numOps; ++opId) {
      const OperationResult& result = response.results(opId);
      std::vector<real*>& resultScalars = ops.localResults_[opId].resultScalars;
      std::vector<CpuVectorPtr>& resultVectors =
          ops.localResults_[opId].resultVectors;
      std::vector<CpuMatrixPtr>& resultMatrices =
          ops.localResults_[opId].resultMatrices;

      if (&response == &responses[0]) {
        /// Initialize results to zero

        resultScalars.resize(result.scalars_size());
        for (auto p : resultScalars) {
          if (!p) continue;
          *p = 0;
        }
        size_t numVectors = result.vectors_size();
        resultVectors.resize(numVectors);
        for (size_t i = 0; i < numVectors; ++i) {
          if (!resultVectors[i]) continue;
          resultVectors[i]->resize(result.vectors(i).dim());
          resultVectors[i]->zeroMem();
        }
        size_t numMatrices = result.matrices_size();
        resultMatrices.resize(numMatrices);
        for (size_t i = 0; i < numMatrices; ++i) {
          if (!resultMatrices[i]) continue;
          resultMatrices[i]->resize(result.matrices(i).num_rows(),
                                    result.matrices(i).num_cols());
          resultMatrices[i]->zeroMem();
        }
      }

      // aggregate results from each pserver to results

      CHECK_EQ(resultScalars.size(), (size_t)result.scalars_size());
      for (ssize_t i = 0; i < result.scalars_size(); ++i) {
        real* rscalar = resultScalars[i];
        if (!rscalar) continue;
        *rscalar += result.scalars(i);
      }

      CHECK_EQ(resultVectors.size(), (size_t)result.vectors_size());
      for (auto& vec : result.vectors()) {
        int i = &vec - &result.vectors(0);
        CpuVectorPtr rvec = resultVectors[i];
        if (!rvec) continue;
        CHECK_EQ(rvec->getSize(), (size_t)vec.dim());
Y
Yu Yang 已提交
694 695 696 697
        std::transform(rvec->getData(),
                       rvec->getData() + rvec->getSize(),
                       vec.values().data(),
                       rvec->getData(),
Y
Yu Yang 已提交
698
                       addTwo);
Z
zhangjinchao01 已提交
699 700 701 702 703 704 705 706 707
      }

      CHECK_EQ(resultMatrices.size(), (size_t)result.matrices_size());
      for (auto& mat : result.matrices()) {
        int i = &mat - &result.matrices(0);
        CpuMatrixPtr rmat = resultMatrices[i];
        if (!rmat) continue;
        CHECK_EQ(rmat->getHeight(), (size_t)mat.num_rows());
        CHECK_EQ(rmat->getWidth(), (size_t)mat.num_cols());
Y
Yu Yang 已提交
708 709 710 711 712

        std::transform(rmat->getData(),
                       rmat->getData() + rmat->getElementCnt(),
                       mat.values().data(),
                       rmat->getData(),
Y
Yu Yang 已提交
713
                       addTwo);
Z
zhangjinchao01 已提交
714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745
      }
    }
  }
  passFinish_ = numPassFinishServers == clients_.size();
}

real ParameterClient2::vectorDotProduct(PServerVector u, PServerVector v) {
  real result = 0.0;
  PreparedOperations ops;
  ops.addOperation(PSERVER_OP_utv, u, v)(&result);
  doOperation(ops, false, false);
  return result;
}

void ParameterClient2::vectorScale(PServerVector u, real a) {
  PreparedOperations ops;
  ops.addOperation(PSERVER_OP_au, u, a);
  doOperation(ops, false, false);
}

void ParameterClient2::vectorCopy(PServerVector src, PServerVector dst) {
  PreparedOperations ops;
  ops.addOperation(PSERVER_OP_COPY, src, dst);
  doOperation(ops, false, false);
}

void ParameterClient2::vectorAddMult(PServerVector u, PServerVector v, real a) {
  PreparedOperations ops;
  ops.addOperation(PSERVER_OP_au_bv, v, u, a, (real)1);
  doOperation(ops, false, false);
}

746 747 748 749
void ParameterClient2::vectorAddMultInto(PServerVector u,
                                         PServerVector v,
                                         PServerVector w,
                                         real a) {
Z
zhangjinchao01 已提交
750 751 752 753 754
  PreparedOperations ops;
  ops.addOperation(PSERVER_OP_au_bv_cw, v, w, u, (real)1, a, (real)0);
  doOperation(ops, false, false);
}

755 756
void ParameterClient2::vectorScaleInto(PServerVector u,
                                       PServerVector v,
Z
zhangjinchao01 已提交
757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781
                                       real a) {
  PreparedOperations ops;
  ops.addOperation(PSERVER_OP_au_bv, v, u, a, (real)0);
  doOperation(ops, false, false);
}

void ParameterClient2::loadValueVector(const std::string& dirName) {
  LoadValueRequest request;
  request.set_dir_name(dirName);
  std::vector<LoadValueResponse> responses;

  multiCall(__func__, request, &responses);
  validateResponses(responses);
}

void ParameterClient2::saveValueVector(const std::string& dirName) {
  SaveValueRequest request;
  request.set_dir_name(dirName);
  std::vector<SaveValueResponse> responses;

  multiCall(__func__, request, &responses);
  validateResponses(responses);
}

}  // namespace paddle