Method exec_micros

tensorflow/core/profiler/internal/tfprof_node.h:507–525

This is overall computation time, including both cpu and accelerator. Note, cpu and accelerator might or might not run in parallel.

  // This is overall computation time, including both cpu and accelerator.
  // Note, cpu and accelerator might or might not run in parallel.
  int64 exec_micros(int64 step) const {
    // Empty when no RunMetadata is provided.
    if (execs_.empty()) {
      return 0;
    }
    if (step >= 0) {
      auto exec = execs_.find(step);
      if (exec == execs_.end()) {
        return 0;
      }
      return exec->second.exec_micros();
    }

    int64 total_micros = 0;
    for (const auto& exec : execs_) {
      total_micros += exec.second.exec_micros();
    }
    return total_micros / execs_.size();
  }

  // This is accelerator computation time of a step, or average of
  // multiple step, when step < 0.