Method TrainFilesWithProfiler

paddle/fluid/framework/hogwild_worker.cc:1235–1350

}

void HogwildWorker::TrainFilesWithProfiler() {
  platform::SetNumThreads(1);
#if defined(PADDLE_WITH_HETERPS) && \
    (defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL))
  platform::SetDeviceId(thread_id_);
#elif defined(PADDLE_WITH_HETERPS) && defined(PADDLE_WITH_XPU_BKCL)
  platform::SetXPUDeviceId(thread_id_);
#endif
  device_reader_->Start();
  std::vector<double> op_total_time;
  op_total_time.resize(ops_.size());
  for (double &op_time : op_total_time) {
    op_time = 0.0;
  }
  platform::Timer timeline;
  double total_time = 0.0;
  double read_time = 0.0;
  int cur_batch = 0;
  int batch_cnt = 0;
  if (thread_id_ == 0) {
    quit_flag_.store(false);
  }
  g_barrier.wait();

  timeline.Start();
  uint64_t total_inst = 0;

  std::unique_ptr<GarbageCollector> gc = nullptr;
  int64_t max_memory_size = GetEagerDeletionThreshold();
  if (max_memory_size >= 0) {
    gc = CreateGarbageCollector(place_, max_memory_size);
  }
  bool infer_out_of_ins = false;
  while (true) {
    cur_batch = device_reader_->Next();
    if (cur_batch <= 0 && !infer_out_of_ins) {
      break;
    }
    VLOG(3) << "read a batch in thread " << thread_id_;
    timeline.Pause();
    read_time += timeline.ElapsedSec();
    total_time += timeline.ElapsedSec();
    if (infer_out_of_ins) {
      for (size_t i = 0; i < ops_.size(); ++i) {
        timeline.Start();
        auto &op = ops_[i];
        VLOG(3) << "Going to run op " << op_names_[i];
        if (op->Type() == "c_broadcast") {
          op->Run(*thread_scope_, place_);
        }
        VLOG(3) << "Op " << op_names_[i] << " Finished";
        timeline.Pause();
        op_total_time[i] += timeline.ElapsedSec();
        total_time += timeline.ElapsedSec();
        if (gc) {
          DeleteUnusedTensors(*thread_scope_, op.get(), unused_vars_, gc.get());
        }
      }