hogwild_worker.cc 6.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.

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. */

15
#include "paddle/fluid/framework/data_type.h"
16
#include "paddle/fluid/framework/device_worker.h"
17
#include "paddle/fluid/framework/device_worker_factory.h"
18
#include "paddle/fluid/platform/cpu_helper.h"
D
dongdaxiang 已提交
19
#include "paddle/fluid/platform/lodtensor_printer.h"
20 21 22 23

namespace paddle {
namespace framework {

24
void HogwildWorker::Initialize(const TrainerDesc &desc) {
D
dongdaxiang 已提交
25
  fetch_config_ = desc.fetch_config();
26 27
  param_ = desc.hogwild_param();
  skip_ops_.resize(param_.skip_ops_size());
28
  for (int i = 0; i < param_.skip_ops_size(); ++i) {
29 30
    skip_ops_[i] = param_.skip_ops(i);
  }
31
  use_cvm_ = desc.use_cvm();
D
dongdaxiang 已提交
32 33
}

34 35
void HogwildWorker::CreateThreadOperators(const ProgramDesc &program) {
  auto &block = program.Block(0);
36
  op_names_.clear();
37
  for (auto &op_desc : block.AllOps()) {
38 39
    std::unique_ptr<OperatorBase> local_op = OpRegistry::CreateOp(*op_desc);
    op_names_.push_back(op_desc->Type());
40
    OperatorBase *local_op_ptr = local_op.release();
41 42 43 44 45
    ops_.push_back(local_op_ptr);
    continue;
  }
}

46 47
void HogwildWorker::CreateThreadScope(const ProgramDesc &program) {
  auto &block = program.Block(0);
48 49 50 51 52

  PADDLE_ENFORCE_NOT_NULL(
      root_scope_, "root_scope should be set before creating thread scope");

  thread_scope_ = &root_scope_->NewScope();
53 54

  for (auto &var : block.AllVars()) {
55
    if (var->Persistable()) {
56
      auto *ptr = root_scope_->Var(var->Name());
57
      InitializeVariable(ptr, var->GetType());
58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74
      if (stat_var_name_map_.find(var->Name()) != stat_var_name_map_.end() &&
          thread_id_ != 0) {
        int tensor_dim =
            root_scope_->FindVar(var->Name())->GetMutable<LoDTensor>()->numel();
        auto *ptr1 = thread_scope_->Var(var->Name());
        InitializeVariable(ptr1, var->GetType());
        LoDTensor *thread_tensor = ptr1->GetMutable<LoDTensor>();
        LoDTensor *root_tensor =
            root_scope_->FindVar(var->Name())->GetMutable<LoDTensor>();
#define MemsetCallback(cpp_type, proto_type)                     \
  do {                                                           \
    if (root_tensor->type() == proto_type) {                     \
      SetZero<cpp_type>(thread_tensor, root_tensor, tensor_dim); \
    }                                                            \
  } while (0)
        _ForEachDataType_(MemsetCallback);
      }
75
    } else {
76
      auto *ptr = thread_scope_->Var(var->Name());
77 78 79 80 81
      InitializeVariable(ptr, var->GetType());
    }
  }
}

82 83 84 85 86 87 88
template <typename T>
void HogwildWorker::SetZero(LoDTensor *tensor, LoDTensor *root_tensor,
                            int tensor_dim) {
  T *ptr = tensor->mutable_data<T>(root_tensor->dims(), platform::CPUPlace());
  memset(ptr, 0, sizeof(T) * tensor_dim);
}

89
void HogwildWorker::BindingDataFeedMemory() {
90
  const std::vector<std::string> &input_feed =
91
      device_reader_->GetUseSlotAlias();
92
  for (auto name : input_feed) {
93
    device_reader_->AddFeedVar(thread_scope_->FindVar(name), name);
94 95 96
  }
}

97
void HogwildWorker::CreateDeviceResource(const ProgramDesc &main_prog) {
98 99 100 101 102 103
  CreateThreadScope(main_prog);
  CreateThreadOperators(main_prog);
}

void HogwildWorker::TrainFilesWithProfiler() {
  platform::SetNumThreads(1);
104
  device_reader_->Start();
105 106
  std::vector<double> op_total_time;
  std::vector<std::string> op_name;
107
  for (auto &op : ops_) {
108 109 110 111 112 113 114 115 116 117 118 119
    op_name.push_back(op->Type());
  }
  op_total_time.resize(ops_.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;
  int cur_batch;
  int batch_cnt = 0;
  timeline.Start();
D
dongdaxiang 已提交
120
  uint64_t total_inst = 0;
121
  while ((cur_batch = device_reader_->Next()) > 0) {
122
    VLOG(3) << "read a batch in thread " << thread_id_;
123 124 125 126
    timeline.Pause();
    read_time += timeline.ElapsedSec();
    total_time += timeline.ElapsedSec();
    for (size_t i = 0; i < ops_.size(); ++i) {
127 128 129 130 131 132 133
      bool need_skip = false;
      for (auto t = 0u; t < skip_ops_.size(); ++t) {
        if (ops_[i]->Type().find(skip_ops_[t]) != std::string::npos) {
          need_skip = true;
          break;
        }
      }
134
      timeline.Start();
135
      VLOG(3) << "Going to run op " << op_name[i];
136 137 138
      if (!need_skip) {
        ops_[i]->Run(*thread_scope_, place_);
      }
139
      VLOG(3) << "Op " << op_name[i] << " Finished";
140 141 142 143
      timeline.Pause();
      op_total_time[i] += timeline.ElapsedSec();
      total_time += timeline.ElapsedSec();
    }
D
dongdaxiang 已提交
144
    total_inst += cur_batch;
145
    ++batch_cnt;
D
dongdaxiang 已提交
146
    PrintFetchVars();
147 148 149 150 151 152 153
    if (thread_id_ == 0) {
      if (batch_cnt > 0 && batch_cnt % 100 == 0) {
        for (size_t i = 0; i < ops_.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);
        }
        fprintf(stderr, "mean read time: %fs\n", read_time / batch_cnt);
D
dongdaxiang 已提交
154
        fprintf(stderr, "IO percent: %f\n", read_time / total_time * 100);
D
dongdaxiang 已提交
155
        fprintf(stderr, "%6.2f instances/s\n", total_inst / total_time);
156 157
      }
    }
D
dongdaxiang 已提交
158
    thread_scope_->DropKids();
159 160 161 162 163 164 165 166
    timeline.Start();
  }
}

void HogwildWorker::TrainFiles() {
  platform::SetNumThreads(1);

  // how to accumulate fetched values here
167
  device_reader_->Start();
168
  int cur_batch;
169
  while ((cur_batch = device_reader_->Next()) > 0) {
170
    for (auto &op : ops_) {
171 172 173 174 175 176 177 178 179 180
      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->Run(*thread_scope_, place_);
      }
181 182
    }

D
dongdaxiang 已提交
183
    PrintFetchVars();
D
dongdaxiang 已提交
184
    thread_scope_->DropKids();
185 186 187
  }
}

D
dongdaxiang 已提交
188 189 190 191
void HogwildWorker::PrintFetchVars() {
  // call count
  batch_num_++;
  int batch_per_print = fetch_config_.print_period();
D
dongdaxiang 已提交
192
  if (thread_id_ == 0) {
D
dongdaxiang 已提交
193 194
    if (batch_num_ % batch_per_print == 0) {
      int fetch_var_num = fetch_config_.fetch_var_names_size();
D
dongdaxiang 已提交
195
      for (int i = 0; i < fetch_var_num; ++i) {
D
dongdaxiang 已提交
196
        platform::PrintVar(thread_scope_, fetch_config_.fetch_var_names(i),
D
dongdaxiang 已提交
197
                           fetch_config_.fetch_var_str_format(i));
D
dongdaxiang 已提交
198 199 200 201 202
      }
    }
  }
}

203 204
}  // end namespace framework
}  // end namespace paddle