downpour_worker.cc 35.1 KB
Newer Older
1
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

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 "paddle/fluid/framework/device_worker.h"
#include "paddle/fluid/platform/cpu_helper.h"
W
wanghuancoder 已提交
17

18 19 20 21
namespace pten {
class DenseTensor;
}  // namespace pten

W
wanghuancoder 已提交
22 23 24 25 26
namespace paddle {
namespace framework {
class Variable;
}  // namespace framework
}  // namespace paddle
27

28 29 30 31 32
#if defined _WIN32 || defined __APPLE__
#else
#define _LINUX
#endif

33 34 35
namespace paddle {
namespace framework {

36
void DownpourWorker::Initialize(const TrainerDesc& desc) {
37
  param_ = desc.downpour_param();
D
dongdaxiang 已提交
38
  for (int i = 0; i < param_.sparse_table_size(); ++i) {
39 40 41 42
    uint64_t table_id =
        static_cast<uint64_t>(param_.sparse_table(i).table_id());
    TableParameter table = param_.sparse_table(i);
    sparse_key_names_[table_id].resize(table.sparse_key_name_size());
D
dongdaxiang 已提交
43
    for (int j = 0; j < table.sparse_key_name_size(); ++j) {
44 45 46
      sparse_key_names_[table_id][j] = table.sparse_key_name(j);
    }
    sparse_value_names_[table_id].resize(table.sparse_value_name_size());
D
dongdaxiang 已提交
47
    for (int j = 0; j < table.sparse_value_name_size(); ++j) {
48 49 50
      sparse_value_names_[table_id][j] = table.sparse_value_name(j);
    }
    sparse_grad_names_[table_id].resize(table.sparse_grad_name_size());
D
dongdaxiang 已提交
51
    for (int j = 0; j < table.sparse_grad_name_size(); ++j) {
52 53
      sparse_grad_names_[table_id][j] = table.sparse_grad_name(j);
    }
54
    label_var_name_[table_id] = table.label_var_name();
55
    sparse_push_keys_[table_id] = std::vector<uint64_t>();
56 57
  }

D
dongdaxiang 已提交
58
  for (int i = 0; i < param_.dense_table_size(); ++i) {
59 60 61
    uint64_t table_id = static_cast<uint64_t>(param_.dense_table(i).table_id());
    auto table = param_.dense_table(i);
    dense_value_names_[table_id].resize(table.dense_value_name_size());
D
dongdaxiang 已提交
62
    for (int j = 0; j < table.dense_value_name_size(); ++j) {
63 64 65
      dense_value_names_[table_id][j] = table.dense_value_name(j);
    }
    dense_grad_names_[table_id].resize(table.dense_grad_name_size());
D
dongdaxiang 已提交
66
    for (int j = 0; j < table.dense_grad_name_size(); ++j) {
67 68 69 70
      dense_grad_names_[table_id][j] = table.dense_grad_name(j);
    }
  }

Z
zhang wenhui 已提交
71 72 73 74 75 76 77
  flag_partial_push_ = false;
  for (auto& m : param_.program_config(0).partial_pushdense_condtable_map()) {
    cond2table_map_[m.key()] = m.value();
    condvalue_set_.insert(m.value());
    flag_partial_push_ = true;
  }

78
  skip_ops_.resize(param_.skip_ops_size());
D
dongdaxiang 已提交
79
  for (int i = 0; i < param_.skip_ops_size(); ++i) {
80 81
    skip_ops_[i] = param_.skip_ops(i);
  }
82

83 84 85 86
  for (int i = 0; i < param_.stat_var_names_size(); ++i) {
    stat_var_name_map_[param_.stat_var_names(i)] = 1;
  }

87 88 89
  need_to_push_sparse_ = param_.push_sparse();
  need_to_push_dense_ = param_.push_dense();

90
  fleet_ptr_ = FleetWrapper::GetInstance();
D
dongdaxiang 已提交
91
  fetch_config_ = desc.fetch_config();
92
  use_cvm_ = desc.use_cvm();
93 94
  // for sparse value accessor, embedding only
  no_cvm_ = desc.no_cvm();
95 96
  scale_sparse_gradient_with_batch_size_ =
      desc.scale_sparse_gradient_with_batch_size();
97
  scale_datanorm_ = desc.scale_datanorm();
T
Thunderbrook 已提交
98
  dump_slot_ = desc.dump_slot();
99
  adjust_ins_weight_config_ = desc.adjust_ins_weight_config();
100 101 102
  for (int i = 0; i < desc.check_nan_var_names_size(); ++i) {
    check_nan_var_names_.push_back(desc.check_nan_var_names(i));
  }
X
xujiaqi01 已提交
103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
  copy_table_config_ = desc.copy_table_config();
  for (int i = 0; i < copy_table_config_.src_sparse_tables_size(); ++i) {
    uint64_t src_table = copy_table_config_.src_sparse_tables(i);
    uint64_t dest_table = copy_table_config_.dest_sparse_tables(i);
    VLOG(3) << "copy_sparse_tables_ push back " << src_table << "->"
            << dest_table;
    copy_sparse_tables_.push_back(std::make_pair(src_table, dest_table));
  }
  for (int i = 0; i < copy_table_config_.src_dense_tables_size(); ++i) {
    uint64_t src_table = copy_table_config_.src_dense_tables(i);
    uint64_t dest_table = copy_table_config_.dest_dense_tables(i);
    VLOG(3) << "copy_dense_tables_ push back " << src_table << "->"
            << dest_table;
    copy_dense_tables_.push_back(std::make_pair(src_table, dest_table));
  }
  for (auto& m : copy_table_config_.table_denpendency_map()) {
    if (sparse_key_names_.find(m.key()) != sparse_key_names_.end()) {
      // currently only support one dependency
      for (auto& value : m.values()) {
        table_dependency_[m.key()] = value;
      }
    }
  }
126 127
}

128
void DownpourWorker::CollectLabelInfo(size_t table_idx) {
129 130 131
  if (no_cvm_) {
    return;
  }
H
heqiaozhi 已提交
132
  uint64_t table_id = static_cast<uint64_t>(
133
      param_.program_config(0).pull_sparse_table_id(table_idx));
134

H
heqiaozhi 已提交
135 136 137 138 139 140 141
  TableParameter table;
  for (auto i : param_.sparse_table()) {
    if (i.table_id() == table_id) {
      table = i;
      break;
    }
  }
142 143 144
  auto& feature = features_[table_id];
  auto& feature_label = feature_labels_[table_id];
  feature_label.resize(feature.size());
145
  Variable* var = thread_scope_->FindVar(label_var_name_[table_id]);
146 147 148
  LoDTensor* tensor = var->GetMutable<LoDTensor>();
  int64_t* label_ptr = tensor->data<int64_t>();

D
dongdaxiang 已提交
149
  size_t global_index = 0;
150
  for (size_t i = 0; i < sparse_key_names_[table_id].size(); ++i) {
151 152
    VLOG(3) << "sparse_key_names_[" << i
            << "]: " << sparse_key_names_[table_id][i];
153
    Variable* fea_var = thread_scope_->FindVar(sparse_key_names_[table_id][i]);
154 155 156
    if (fea_var == nullptr) {
      continue;
    }
157
    LoDTensor* tensor = fea_var->GetMutable<LoDTensor>();
158 159
    CHECK(tensor != nullptr) << "tensor of var "
                             << sparse_key_names_[table_id][i] << " is null";
160 161 162 163 164 165 166 167

    // skip slots which do not have embedding
    Variable* emb_var =
        thread_scope_->FindVar(sparse_value_names_[table_id][i]);
    if (emb_var == nullptr) {
      continue;
    }

168
    int64_t* ids = tensor->data<int64_t>();
D
dongdaxiang 已提交
169
    size_t fea_idx = 0;
170
    // tensor->lod()[0].size() == batch_size + 1
171 172
    for (auto lod_idx = 1u; lod_idx < tensor->lod()[0].size(); ++lod_idx) {
      for (; fea_idx < tensor->lod()[0][lod_idx]; ++fea_idx) {
173 174 175 176
        // should be skipped feasign defined in protobuf
        if (ids[fea_idx] == 0u) {
          continue;
        }
177 178
        feature_label[global_index++] =
            static_cast<float>(label_ptr[lod_idx - 1]);
179 180 181 182 183 184 185 186
      }
    }
  }
  CHECK(global_index == feature.size())
      << "expect fea info size:" << feature.size() << " real:" << global_index;
}

void DownpourWorker::FillSparseValue(size_t table_idx) {
H
heqiaozhi 已提交
187
  uint64_t table_id = static_cast<uint64_t>(
188
      param_.program_config(0).pull_sparse_table_id(table_idx));
H
heqiaozhi 已提交
189 190 191 192 193 194 195 196

  TableParameter table;
  for (auto i : param_.sparse_table()) {
    if (i.table_id() == table_id) {
      table = i;
      break;
    }
  }
197 198 199 200

  auto& fea_value = feature_values_[table_id];
  auto fea_idx = 0u;

X
xjqbest 已提交
201
  std::vector<float> init_value(table.fea_dim());
202 203 204 205
  for (size_t i = 0; i < sparse_key_names_[table_id].size(); ++i) {
    std::string slot_name = sparse_key_names_[table_id][i];
    std::string emb_slot_name = sparse_value_names_[table_id][i];
    Variable* var = thread_scope_->FindVar(slot_name);
206 207 208
    if (var == nullptr) {
      continue;
    }
209
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
210
    CHECK(tensor != nullptr) << "tensor of var " << slot_name << " is null";
211 212 213
    int64_t* ids = tensor->data<int64_t>();
    int len = tensor->numel();
    Variable* var_emb = thread_scope_->FindVar(emb_slot_name);
214 215 216
    if (var_emb == nullptr) {
      continue;
    }
217 218 219 220 221 222 223
    LoDTensor* tensor_emb = var_emb->GetMutable<LoDTensor>();
    float* ptr = tensor_emb->mutable_data<float>({len, table.emb_dim()},
                                                 platform::CPUPlace());
    memset(ptr, 0, sizeof(float) * len * table.emb_dim());
    auto& tensor_lod = tensor->lod()[0];
    LoD data_lod{tensor_lod};
    tensor_emb->set_lod(data_lod);
224 225 226 227 228 229 230 231

    bool is_nid = (adjust_ins_weight_config_.need_adjust() &&
                   adjust_ins_weight_config_.nid_slot() == emb_slot_name);
    if (is_nid) {
      nid_show_.clear();
    }
    int nid_ins_index = 0;

D
dongdaxiang 已提交
232
    for (int index = 0; index < len; ++index) {
233
      if (use_cvm_ || no_cvm_) {
234 235 236
        if (ids[index] == 0u) {
          memcpy(ptr + table.emb_dim() * index, init_value.data(),
                 sizeof(float) * table.emb_dim());
237 238 239 240
          if (is_nid) {
            nid_show_.push_back(-1);
            ++nid_ins_index;
          }
241 242 243 244
          continue;
        }
        memcpy(ptr + table.emb_dim() * index, fea_value[fea_idx].data(),
               sizeof(float) * table.emb_dim());
245 246
        if (is_nid &&
            static_cast<size_t>(index) == tensor->lod()[0][nid_ins_index]) {
247 248 249
          nid_show_.push_back(fea_value[fea_idx][0]);
          ++nid_ins_index;
        }
250 251 252 253 254
        fea_idx++;
      } else {
        if (ids[index] == 0u) {
          memcpy(ptr + table.emb_dim() * index, init_value.data() + 2,
                 sizeof(float) * table.emb_dim());
255 256 257 258
          if (is_nid) {
            nid_show_.push_back(-1);
            ++nid_ins_index;
          }
259 260 261
          continue;
        }
        memcpy(ptr + table.emb_dim() * index, fea_value[fea_idx].data() + 2,
262
               sizeof(float) * table.emb_dim());
263 264
        if (is_nid &&
            static_cast<size_t>(index) == tensor->lod()[0][nid_ins_index]) {
265 266 267
          nid_show_.push_back(fea_value[fea_idx][0]);
          ++nid_ins_index;
        }
268
        fea_idx++;
269 270 271 272 273
      }
    }
  }
}

274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319
void DownpourWorker::AdjustInsWeight() {
#ifdef _LINUX
  // check var and tensor not null
  if (!adjust_ins_weight_config_.need_adjust()) {
    VLOG(0) << "need_adjust=false, skip adjust ins weight";
    return;
  }
  Variable* nid_var =
      thread_scope_->FindVar(adjust_ins_weight_config_.nid_slot());
  if (nid_var == nullptr) {
    VLOG(0) << "nid slot var " << adjust_ins_weight_config_.nid_slot()
            << " is nullptr, skip adjust ins weight";
    return;
  }
  LoDTensor* nid_tensor = nid_var->GetMutable<LoDTensor>();
  if (nid_tensor == nullptr) {
    VLOG(0) << "tensor of nid slot var " << adjust_ins_weight_config_.nid_slot()
            << " is nullptr, skip adjust ins weight";
    return;
  }
  Variable* ins_weight_var =
      thread_scope_->FindVar(adjust_ins_weight_config_.ins_weight_slot());
  if (ins_weight_var == nullptr) {
    VLOG(0) << "ins weight var " << adjust_ins_weight_config_.ins_weight_slot()
            << " is nullptr, skip adjust ins weight";
    return;
  }
  LoDTensor* ins_weight_tensor = ins_weight_var->GetMutable<LoDTensor>();
  if (ins_weight_tensor == nullptr) {
    VLOG(0) << "tensor of ins weight tensor "
            << adjust_ins_weight_config_.ins_weight_slot()
            << " is nullptr, skip adjust ins weight";
    return;
  }

  float* ins_weights = ins_weight_tensor->data<float>();
  size_t len = ins_weight_tensor->numel();  // len = batch size
  // here we assume nid_show slot only has one feasign in each instance
  CHECK(len == nid_show_.size()) << "ins_weight size should be equal to "
                                 << "nid_show size, " << len << " vs "
                                 << nid_show_.size();
  float nid_adjw_threshold = adjust_ins_weight_config_.nid_adjw_threshold();
  float nid_adjw_ratio = adjust_ins_weight_config_.nid_adjw_ratio();
  int64_t nid_adjw_num = 0;
  double nid_adjw_weight = 0.0;
  size_t ins_index = 0;
320
  for (size_t i = 0; i < len; ++i) {
321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349
    float nid_show = nid_show_[i];
    VLOG(3) << "nid_show " << nid_show;
    if (nid_show < 0) {
      VLOG(3) << "nid_show < 0, continue";
      continue;
    }
    float ins_weight = 1.0;
    if (nid_show >= 0 && nid_show < nid_adjw_threshold) {
      ins_weight = log(M_E +
                       (nid_adjw_threshold - nid_show) / nid_adjw_threshold *
                           nid_adjw_ratio);
      // count nid adjw insnum and weight
      ++nid_adjw_num;
      nid_adjw_weight += ins_weight;
      // choose large ins weight
      VLOG(3) << "ins weight new " << ins_weight << ", ins weight origin "
              << ins_weights[ins_index];
      if (ins_weight > ins_weights[ins_index]) {
        VLOG(3) << "ins " << ins_index << " weight changes to " << ins_weight;
        ins_weights[ins_index] = ins_weight;
      }
      ++ins_index;
    }
  }
  VLOG(3) << "nid adjw info: total_adjw_num: " << nid_adjw_num
          << ", avg_adjw_weight: " << nid_adjw_weight;
#endif
}

X
xujiaqi01 已提交
350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397
void DownpourWorker::CopySparseTable() {
  for (size_t i = 0; i < copy_sparse_tables_.size(); ++i) {
    int64_t src_table = copy_sparse_tables_[i].first;
    int64_t dest_table = copy_sparse_tables_[i].second;
    int32_t feanum = 0;
    if (src_table == dest_table) {
      continue;
    } else if (!copy_table_config_.sparse_copy_by_feasign()) {
      if (feasign_set_.find(src_table) == feasign_set_.end()) {
        continue;
      } else if (feasign_set_[src_table].size() == 0) {
        continue;
      }
      feanum = fleet_ptr_->CopyTable(src_table, dest_table);
    } else {
      std::vector<uint64_t> fea_vec(feasign_set_[src_table].begin(),
                                    feasign_set_[src_table].end());
      feanum = fleet_ptr_->CopyTableByFeasign(src_table, dest_table, fea_vec);
      fea_vec.clear();
      std::vector<uint64_t>().swap(fea_vec);
    }
    VLOG(3) << "copy feasign from table " << src_table << " to table "
            << dest_table << ", feasign num=" << feanum;
    feasign_set_[src_table].clear();
    std::unordered_set<uint64_t>().swap(feasign_set_[src_table]);
  }
  feasign_set_.clear();
}

void DownpourWorker::CopyDenseTable() {
  if (thread_id_ != 0) {
    return;
  }
  thread_local std::vector<std::future<int32_t>> pull_dense_status;
  for (size_t i = 0; i < copy_dense_tables_.size(); ++i) {
    uint64_t src_table = copy_dense_tables_[i].first;
    uint64_t dest_table = copy_dense_tables_[i].second;
    if (src_table == dest_table) {
      continue;
    }
    int32_t dim = fleet_ptr_->CopyTable(src_table, dest_table);
    VLOG(3) << "copy param from table " << src_table << " to table "
            << dest_table << ", dim=" << dim;
    if (copy_table_config_.dense_pull_after_copy()) {
      VLOG(3) << "dense pull after copy, table=" << dest_table;
      pull_dense_status.resize(0);
      fleet_ptr_->PullDenseVarsAsync(*root_scope_, dest_table,
                                     dense_value_names_[dest_table],
T
Thunderbrook 已提交
398
                                     &pull_dense_status, true);
X
xujiaqi01 已提交
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445
      for (auto& t : pull_dense_status) {
        t.wait();
        auto status = t.get();
        if (status != 0) {
          LOG(WARNING) << "pull dense after copy table failed,"
                       << " table=" << dest_table;
        }
      }
    }
  }
}

void DownpourWorker::CopyDenseVars() {
  if (thread_id_ != 0) {
    return;
  }
  for (int i = 0; i < copy_table_config_.src_var_list_size(); ++i) {
    auto& src_var_name = copy_table_config_.src_var_list(i);
    auto& dest_var_name = copy_table_config_.dest_var_list(i);
    if (src_var_name == dest_var_name) {
      continue;
    }
    VLOG(3) << "copy dense var from " << src_var_name << " to "
            << dest_var_name;
    Variable* src_var = thread_scope_->FindVar(src_var_name);
    CHECK(src_var != nullptr) << src_var_name << " not found";  // NOLINT
    LoDTensor* src_tensor = src_var->GetMutable<LoDTensor>();
    CHECK(src_tensor != nullptr) << src_var_name
                                 << " tensor is null";  // NOLINT
    float* src_data = src_tensor->data<float>();

    Variable* dest_var = thread_scope_->FindVar(dest_var_name);
    CHECK(dest_var != nullptr) << dest_var_name << " not found";  // NOLINT
    LoDTensor* dest_tensor = dest_var->GetMutable<LoDTensor>();
    CHECK(dest_tensor != nullptr) << dest_var_name
                                  << " tensor is null";  // NOLINT
    float* dest_data = dest_tensor->data<float>();

    CHECK(src_tensor->numel() == dest_tensor->numel())
        << "tensor numel not equal," << src_tensor->numel() << " vs "
        << dest_tensor->numel();
    for (int i = 0; i < src_tensor->numel(); i++) {
      dest_data[i] = src_data[i];
    }
  }
}

446 447 448
void DownpourWorker::TrainFilesWithProfiler() {
  VLOG(3) << "Begin to train files with profiler";
  platform::SetNumThreads(1);
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
  device_reader_->Start();
  std::vector<double> op_total_time;
  std::vector<std::string> op_name;
  for (auto& op : ops_) {
    bool need_skip = false;
    for (auto t = 0u; t < skip_ops_.size(); ++t) {
      if (op->Type().find(skip_ops_[t]) != std::string::npos) {
        need_skip = true;
        break;
      }
    }
    if (!need_skip) {
      op_name.push_back(op->Type());
    }
  }

  VLOG(3) << "op name size: " << op_name.size();
  op_total_time.resize(op_name.size());
  for (size_t i = 0; i < op_total_time.size(); ++i) {
    op_total_time[i] = 0.0;
  }
  platform::Timer timeline;
  double total_time = 0.0;
  double read_time = 0.0;
  double pull_sparse_time = 0.0;
474
  double adjust_ins_weight_time = 0.0;
475 476 477 478
  double collect_label_time = 0.0;
  double fill_sparse_time = 0.0;
  double push_sparse_time = 0.0;
  double push_dense_time = 0.0;
X
xujiaqi01 已提交
479
  double copy_table_time = 0.0;
480 481
  int cur_batch;
  int batch_cnt = 0;
D
dongdaxiang 已提交
482
  uint64_t total_inst = 0;
483 484 485 486 487
  timeline.Start();
  while ((cur_batch = device_reader_->Next()) > 0) {
    timeline.Pause();
    read_time += timeline.ElapsedSec();
    total_time += timeline.ElapsedSec();
X
xujiaqi01 已提交
488 489 490 491 492 493 494 495 496 497 498 499 500 501

    timeline.Start();
    if (copy_table_config_.need_copy()) {
      VLOG(3) << "copy_sparse_tables_.size " << copy_sparse_tables_.size();
      if (batch_cnt % copy_table_config_.batch_num() == 0) {
        CopySparseTable();
        CopyDenseTable();
        CopyDenseVars();
      }
    }
    timeline.Pause();
    copy_table_time += timeline.ElapsedSec();
    total_time += timeline.ElapsedSec();

502
    VLOG(3) << "program config size: " << param_.program_config_size();
D
dongdaxiang 已提交
503
    for (int i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
504 505 506 507
         ++i) {
      uint64_t tid = static_cast<uint64_t>(
          param_.program_config(0).pull_sparse_table_id(i));
      TableParameter table;
508 509 510
      for (auto j : param_.sparse_table()) {
        if (j.table_id() == tid) {
          table = j;
511 512 513 514
          break;
        }
      }
      timeline.Start();
515 516 517
      fleet_ptr_->PullSparseVarsSync(
          *thread_scope_, tid, sparse_key_names_[tid], &features_[tid],
          &feature_values_[tid], table.fea_dim(), sparse_value_names_[tid]);
518 519
      timeline.Pause();
      pull_sparse_time += timeline.ElapsedSec();
D
dongdaxiang 已提交
520
      total_time += timeline.ElapsedSec();
D
dongdaxiang 已提交
521
      timeline.Start();
522 523 524
      CollectLabelInfo(i);
      timeline.Pause();
      collect_label_time += timeline.ElapsedSec();
D
dongdaxiang 已提交
525
      total_time += timeline.ElapsedSec();
526 527 528 529
      timeline.Start();
      FillSparseValue(i);
      timeline.Pause();
      fill_sparse_time += timeline.ElapsedSec();
D
dongdaxiang 已提交
530
      total_time += timeline.ElapsedSec();
531 532 533 534 535 536 537 538 539 540
      timeline.Start();
      auto nid_iter = std::find(sparse_value_names_[tid].begin(),
                                sparse_value_names_[tid].end(),
                                adjust_ins_weight_config_.nid_slot());
      if (nid_iter != sparse_value_names_[tid].end()) {
        AdjustInsWeight();
      }
      timeline.Pause();
      adjust_ins_weight_time += timeline.ElapsedSec();
      total_time += timeline.ElapsedSec();
541 542 543 544 545 546 547 548 549 550 551 552 553 554
    }
    VLOG(3) << "Fill sparse value for all sparse table done.";

    int run_op_idx = 0;
    for (auto& op : ops_) {
      bool need_skip = false;
      for (auto t = 0u; t < skip_ops_.size(); ++t) {
        if (op->Type().find(skip_ops_[t]) != std::string::npos) {
          need_skip = true;
          break;
        }
      }
      if (!need_skip) {
        timeline.Start();
555
        VLOG(3) << "Going to run op " << op_name[run_op_idx];
556
        op->Run(*thread_scope_, place_);
557
        VLOG(3) << "Op " << op_name[run_op_idx] << " Finished";
558 559 560 561 562 563
        timeline.Pause();
        op_total_time[run_op_idx++] += timeline.ElapsedSec();
        total_time += timeline.ElapsedSec();
      }
    }

564 565 566 567 568 569 570 571 572 573 574
    // check inf and nan
    for (std::string& var_name : check_nan_var_names_) {
      Variable* var = thread_scope_->FindVar(var_name);
      if (var == nullptr) {
        continue;
      }
      LoDTensor* tensor = var->GetMutable<LoDTensor>();
      if (tensor == nullptr) {
        continue;
      }
      PADDLE_ENFORCE_EQ(framework::TensorContainsInf(*tensor), false,
575 576
                        platform::errors::InvalidArgument(
                            "Tensor %s contains Inf.", var_name));
577
      PADDLE_ENFORCE_EQ(framework::TensorContainsNAN(*tensor), false,
578 579
                        platform::errors::InvalidArgument(
                            "Tensor %s contains NAN.", var_name));
580 581
    }

582
    if (need_to_push_sparse_) {
D
dongdaxiang 已提交
583 584
      for (int i = 0; i < param_.program_config(0).push_sparse_table_id_size();
           ++i) {
585 586 587 588 589 590 591 592
        uint64_t tid = static_cast<uint64_t>(
            param_.program_config(0).push_sparse_table_id(i));
        TableParameter table;
        for (auto i : param_.sparse_table()) {
          if (i.table_id() == tid) {
            table = i;
            break;
          }
593
        }
594 595 596 597
        timeline.Start();
        fleet_ptr_->PushSparseVarsWithLabelAsync(
            *thread_scope_, tid, features_[tid], feature_labels_[tid],
            sparse_key_names_[tid], sparse_grad_names_[tid], table.emb_dim(),
T
Thunderbrook 已提交
598
            &feature_grads_[tid], &push_sparse_status_, cur_batch, use_cvm_,
599 600
            dump_slot_, &sparse_push_keys_[tid], no_cvm_,
            scale_sparse_gradient_with_batch_size_);
601 602 603
        timeline.Pause();
        push_sparse_time += timeline.ElapsedSec();
        total_time += timeline.ElapsedSec();
604
      }
605 606
    }

X
xujiaqi01 已提交
607 608 609 610 611 612 613 614 615 616 617 618
#ifdef PADDLE_WITH_PSLIB
    if (copy_table_config_.need_copy()) {
      if (copy_table_config_.sparse_copy_by_feasign()) {
        for (size_t i = 0; i < copy_sparse_tables_.size(); ++i) {
          uint64_t tid = copy_sparse_tables_[i].first;
          feasign_set_[tid].insert(sparse_push_keys_[tid].begin(),
                                   sparse_push_keys_[tid].end());
        }
      }
    }
#endif

619
    if (need_to_push_dense_) {
620
      timeline.Start();
D
dongdaxiang 已提交
621 622
      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
           ++i) {
623 624 625
        uint64_t tid = static_cast<uint64_t>(
            param_.program_config(0).push_dense_table_id(i));
        fleet_ptr_->PushDenseVarsAsync(
626 627
            *thread_scope_, tid, dense_grad_names_[tid], &push_sparse_status_,
            scale_datanorm_, cur_batch);
628
      }
629
      timeline.Pause();
630
      push_dense_time += timeline.ElapsedSec();
D
dongdaxiang 已提交
631
      total_time += timeline.ElapsedSec();
632 633 634 635 636 637 638 639 640
      VLOG(3) << "push sparse and dense gradient done.";
      int32_t tmp_push_dense_wait_times = -1;
      static uint32_t push_dense_wait_times =
          static_cast<uint32_t>(tmp_push_dense_wait_times);
      if (push_dense_status_.size() >= push_dense_wait_times) {
        for (auto& t : push_dense_status_) {
          t.wait();
        }
        push_dense_status_.resize(0);
641 642
      }

643 644
      if (tmp_push_dense_wait_times == -1) {
        push_dense_status_.resize(0);
645 646 647
      }
    }

648
    if (need_to_push_sparse_) {
649 650 651
      int32_t tmp_push_sparse_wait_times = -1;
      static uint32_t push_sparse_wait_times =
          static_cast<uint32_t>(tmp_push_sparse_wait_times);
652 653 654 655 656 657
      if (push_sparse_status_.size() >= push_sparse_wait_times) {
        for (auto& t : push_sparse_status_) {
          t.wait();
        }
        push_sparse_status_.resize(0);
      }
658

659 660 661
      if (tmp_push_sparse_wait_times == -1) {
        push_sparse_status_.resize(0);
      }
662

663 664 665
      VLOG(3) << "going to increase thread version";
      VLOG(3) << "push dense table id size: "
              << param_.program_config(0).push_dense_table_id_size();
666 667 668
    }

    if (need_to_push_dense_) {
D
dongdaxiang 已提交
669 670
      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
           ++i) {
671 672 673 674
        uint64_t tid = static_cast<uint64_t>(
            param_.program_config(0).push_dense_table_id(i));
        pull_dense_worker_->IncreaseThreadVersion(thread_id_, tid);
      }
675 676
    }

D
dongdaxiang 已提交
677
    PrintFetchVars();
678
    thread_scope_->DropKids();
D
dongdaxiang 已提交
679
    total_inst += cur_batch;
680 681 682 683 684
    ++batch_cnt;

    if (thread_id_ == 0) {
      // should be configured here
      if (batch_cnt > 0 && batch_cnt % 100 == 0) {
685 686
        double op_sum_time = 0;
        std::unordered_map<std::string, double> op_to_time;
687 688 689
        for (size_t i = 0; i < op_total_time.size(); ++i) {
          fprintf(stderr, "op_name:[%zu][%s], op_mean_time:[%fs]\n", i,
                  op_name[i].c_str(), op_total_time[i] / batch_cnt);
690 691 692 693 694 695 696 697 698
          if (op_to_time.find(op_name[i]) == op_to_time.end()) {
            op_to_time[op_name[i]] = 0.0;
          }
          op_to_time[op_name[i]] += op_total_time[i];
          op_sum_time += op_total_time[i];
        }
        for (auto& i : op_to_time) {
          fprintf(stderr, "op [%s] run total time: [%f]ms\n", i.first.c_str(),
                  i.second / batch_cnt);
699
        }
700 701 702 703 704 705 706 707 708 709 710
        fprintf(stderr, "op run total time: %fs\n", op_sum_time / batch_cnt);
        fprintf(stderr, "train total time: %fs\n", total_time / batch_cnt);
        fprintf(stderr, "pull sparse time: %fs\n",
                pull_sparse_time / batch_cnt);
        fprintf(stderr, "fill sparse time: %fs\n",
                fill_sparse_time / batch_cnt);
        fprintf(stderr, "push sparse time: %fs\n",
                push_sparse_time / batch_cnt);
        fprintf(stderr, "push dense time: %fs\n", push_dense_time / batch_cnt);
        fprintf(stderr, "collect label time: %fs\n",
                collect_label_time / batch_cnt);
711 712
        fprintf(stderr, "adjust ins weight time: %fs\n",
                adjust_ins_weight_time / batch_cnt);
X
xujiaqi01 已提交
713
        fprintf(stderr, "copy table time: %fs\n", copy_table_time / batch_cnt);
714 715
        fprintf(stderr, "mean read time: %fs\n", read_time / batch_cnt);
        fprintf(stderr, "IO percent: %f\n", read_time / total_time * 100);
716
        fprintf(stderr, "op run percent: %f\n", op_sum_time / total_time * 100);
D
dongdaxiang 已提交
717 718
        fprintf(stderr, "pull sparse time percent: %f\n",
                pull_sparse_time / total_time * 100);
719 720
        fprintf(stderr, "adjust ins weight time percent: %f\n",
                adjust_ins_weight_time / total_time * 100);
X
xujiaqi01 已提交
721 722
        fprintf(stderr, "copy table time percent: %f\n",
                copy_table_time / total_time * 100);
D
dongdaxiang 已提交
723 724 725 726 727 728 729 730
        fprintf(stderr, "collect label time percent: %f\n",
                collect_label_time / total_time * 100);
        fprintf(stderr, "fill sparse time percent: %f\n",
                fill_sparse_time / total_time * 100);
        fprintf(stderr, "push sparse time percent: %f\n",
                push_sparse_time / total_time * 100);
        fprintf(stderr, "push dense time percent: %f\n",
                push_dense_time / total_time * 100);
D
dongdaxiang 已提交
731
        fprintf(stderr, "%6.2f instances/s\n", total_inst / total_time);
732 733
      }
    }
D
dongdaxiang 已提交
734
    timeline.Start();
735
  }
X
xujiaqi01 已提交
736 737 738 739 740
  if (copy_table_config_.need_copy()) {
    CopySparseTable();
    CopyDenseTable();
    CopyDenseVars();
  }
741 742
}

743
void DownpourWorker::TrainFiles() {
D
dongdaxiang 已提交
744
  VLOG(3) << "Begin to train files";
745
  platform::SetNumThreads(1);
746
  device_reader_->Start();
747 748
  int batch_cnt = 0;
  int cur_batch;
749
  while ((cur_batch = device_reader_->Next()) > 0) {
X
xujiaqi01 已提交
750 751 752 753 754 755 756
    if (copy_table_config_.need_copy()) {
      if (batch_cnt % copy_table_config_.batch_num() == 0) {
        CopySparseTable();
        CopyDenseTable();
        CopyDenseVars();
      }
    }
757
    // pull sparse here
D
dongdaxiang 已提交
758
    for (int i = 0; i < param_.program_config(0).pull_sparse_table_id_size();
H
heqiaozhi 已提交
759 760 761 762
         ++i) {
      uint64_t tid = static_cast<uint64_t>(
          param_.program_config(0).pull_sparse_table_id(i));
      TableParameter table;
763 764 765
      for (auto j : param_.sparse_table()) {
        if (j.table_id() == tid) {
          table = j;
H
heqiaozhi 已提交
766 767 768
          break;
        }
      }
769 770 771
      fleet_ptr_->PullSparseVarsSync(
          *thread_scope_, tid, sparse_key_names_[tid], &features_[tid],
          &feature_values_[tid], table.fea_dim(), sparse_value_names_[tid]);
772 773
      CollectLabelInfo(i);
      FillSparseValue(i);
774 775 776 777 778 779
      auto nid_iter = std::find(sparse_value_names_[tid].begin(),
                                sparse_value_names_[tid].end(),
                                adjust_ins_weight_config_.nid_slot());
      if (nid_iter != sparse_value_names_[tid].end()) {
        AdjustInsWeight();
      }
780
    }
D
dongdaxiang 已提交
781
    VLOG(3) << "fill sparse value for all sparse table done.";
782 783 784

    // do computation here
    for (auto& op : ops_) {
785 786 787 788 789 790 791 792
      bool need_skip = false;
      for (auto t = 0u; t < skip_ops_.size(); ++t) {
        if (op->Type().find(skip_ops_[t]) != std::string::npos) {
          need_skip = true;
          break;
        }
      }
      if (!need_skip) {
793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834
#ifdef PADDLE_WITH_PSLIB
        try {
          op->Run(*thread_scope_, place_);
        } catch (std::exception& e) {
          fprintf(stderr, "error message: %s\n", e.what());
          auto& ins_id_vec = device_reader_->GetInsIdVec();
          size_t batch_size = device_reader_->GetCurBatchSize();
          std::string s = "";
          for (auto& ins_id : ins_id_vec) {
            if (s != "") s += ",";
            s += ins_id;
          }
          fprintf(stderr, "batch_size: %zu, ins_ids_vec: %s\n", batch_size,
                  s.c_str());
          s = "";
          for (auto& param : all_param_) {
            Variable* var = thread_scope_->FindVar(param);
            if (var == nullptr) {
              continue;
            }
            Tensor* tensor = nullptr;
            int64_t len = 0;
            if (var->IsType<framework::LoDTensor>()) {
              tensor = var->GetMutable<LoDTensor>();
              len = tensor->numel();
            } else if (var->IsType<SelectedRows>()) {
              auto selected_rows = var->GetMutable<SelectedRows>();
              tensor = selected_rows->mutable_value();
              len = tensor->numel();
            }
            if (!tensor->IsInitialized()) {
              continue;
            }
            s += param + ":" + std::to_string(len) + ":";
            s += PrintLodTensor(tensor, 0, len);
            fprintf(stderr, "%s\n", s.c_str());
            fflush(stderr);
            s = "";
          }
          throw e;
        }
#else
835
        op->Run(*thread_scope_, place_);
836
#endif
837
      }
838 839
    }

840 841 842 843 844 845 846 847 848 849 850
    // check inf and nan
    for (std::string& var_name : check_nan_var_names_) {
      Variable* var = thread_scope_->FindVar(var_name);
      if (var == nullptr) {
        continue;
      }
      LoDTensor* tensor = var->GetMutable<LoDTensor>();
      if (tensor == nullptr) {
        continue;
      }
      PADDLE_ENFORCE_EQ(framework::TensorContainsInf(*tensor), false,
851 852
                        platform::errors::InvalidArgument(
                            "Tensor %s contains Inf.", var_name));
853
      PADDLE_ENFORCE_EQ(framework::TensorContainsNAN(*tensor), false,
854 855
                        platform::errors::InvalidArgument(
                            "Tensor %s contains NAN.", var_name));
856 857
    }

858 859
    if (need_to_push_sparse_) {
      // push gradients here
D
dongdaxiang 已提交
860 861
      for (int i = 0; i < param_.program_config(0).push_sparse_table_id_size();
           ++i) {
862 863 864 865 866 867 868 869
        uint64_t tid = static_cast<uint64_t>(
            param_.program_config(0).push_sparse_table_id(i));
        TableParameter table;
        for (auto i : param_.sparse_table()) {
          if (i.table_id() == tid) {
            table = i;
            break;
          }
H
heqiaozhi 已提交
870
        }
871 872 873
        fleet_ptr_->PushSparseVarsWithLabelAsync(
            *thread_scope_, tid, features_[tid], feature_labels_[tid],
            sparse_key_names_[tid], sparse_grad_names_[tid], table.emb_dim(),
T
Thunderbrook 已提交
874
            &feature_grads_[tid], &push_sparse_status_, cur_batch, use_cvm_,
875 876
            dump_slot_, &sparse_push_keys_[tid], no_cvm_,
            scale_sparse_gradient_with_batch_size_);
H
heqiaozhi 已提交
877
      }
878 879
    }

X
xujiaqi01 已提交
880 881 882 883 884 885 886 887 888 889 890 891
#ifdef PADDLE_WITH_PSLIB
    if (copy_table_config_.need_copy()) {
      if (copy_table_config_.sparse_copy_by_feasign()) {
        for (size_t i = 0; i < copy_sparse_tables_.size(); ++i) {
          uint64_t tid = copy_sparse_tables_[i].first;
          feasign_set_[tid].insert(sparse_push_keys_[tid].begin(),
                                   sparse_push_keys_[tid].end());
        }
      }
    }
#endif

892
    if (need_to_push_dense_) {
Z
zhang wenhui 已提交
893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926
      if (flag_partial_push_) {
        Variable* var = (*thread_scope_).FindVar("cond_tag");
        LoDTensor* tensor = var->GetMutable<LoDTensor>();
        // check type in python code
        int64_t* cond_value_batch = tensor->data<int64_t>();

        for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
             ++i) {
          uint64_t tid = static_cast<uint64_t>(
              param_.program_config(0).push_dense_table_id(i));
          if (condvalue_set_.find(tid) != condvalue_set_.end()) {
            // common dense table must push dense
            if (cond2table_map_[cond_value_batch[0]] != tid) {
              // can't push dense
              continue;
            }
          }

          VLOG(3) << "push multitask dense gradient " << tid;
          fleet_ptr_->PushDenseVarsAsync(
              *thread_scope_, tid, dense_grad_names_[tid], &push_sparse_status_,
              scale_datanorm_, cur_batch);
        }

      } else {
        for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
             ++i) {
          uint64_t tid = static_cast<uint64_t>(
              param_.program_config(0).push_dense_table_id(i));

          fleet_ptr_->PushDenseVarsAsync(
              *thread_scope_, tid, dense_grad_names_[tid], &push_sparse_status_,
              scale_datanorm_, cur_batch);
        }
927
      }
Z
zhang wenhui 已提交
928

929
      VLOG(3) << "push dense gradient done.";
930

931 932 933 934 935
      // the following code should be more precise and clean
      // TODO(guru4elephant)
      int32_t tmp_push_dense_wait_times = -1;
      static uint32_t push_dense_wait_times =
          static_cast<uint32_t>(tmp_push_dense_wait_times);
936

937 938 939 940 941
      if (push_dense_status_.size() >= push_dense_wait_times) {
        for (auto& t : push_dense_status_) {
          t.wait();
        }
        push_dense_status_.resize(0);
942 943
      }

944 945 946
      if (tmp_push_dense_wait_times == -1) {
        push_dense_status_.resize(0);
      }
947 948
    }

949 950 951 952 953 954 955 956 957 958
    if (need_to_push_sparse_) {
      VLOG(3) << "push sparse gradient done.";
      int32_t tmp_push_sparse_wait_times = -1;
      static uint32_t push_sparse_wait_times =
          static_cast<uint32_t>(tmp_push_sparse_wait_times);
      if (push_sparse_status_.size() >= push_sparse_wait_times) {
        for (auto& t : push_sparse_status_) {
          t.wait();
        }
        push_sparse_status_.resize(0);
959 960
      }

961 962 963
      if (tmp_push_sparse_wait_times == -1) {
        push_sparse_status_.resize(0);
      }
964 965
    }

966
    if (need_to_push_dense_) {
D
dongdaxiang 已提交
967 968
      for (int i = 0; i < param_.program_config(0).push_dense_table_id_size();
           ++i) {
969 970 971 972
        uint64_t tid = static_cast<uint64_t>(
            param_.program_config(0).push_dense_table_id(i));
        pull_dense_worker_->IncreaseThreadVersion(thread_id_, tid);
      }
973
    }
974
    if (need_dump_field_) {
H
hutuxian 已提交
975 976 977 978
      DumpField(*thread_scope_, dump_mode_, dump_interval_);
    }
    if (need_dump_param_ && thread_id_ == 0) {
      DumpParam(*thread_scope_, batch_cnt);
979
    }
980

D
dongdaxiang 已提交
981
    PrintFetchVars();
982 983 984
    thread_scope_->DropKids();
    ++batch_cnt;
  }
H
hutuxian 已提交
985
  if (need_dump_field_ || need_dump_param_) {
986 987
    writer_.Flush();
  }
X
xujiaqi01 已提交
988 989 990 991 992
  if (copy_table_config_.need_copy()) {
    CopySparseTable();
    CopyDenseTable();
    CopyDenseVars();
  }
993 994 995 996
}

}  // end namespace framework
}  // end namespace paddle