pytorch/modules/observers/perf_observer.h

#pragma once

#include "caffe2/core/common.h"
#include "caffe2/core/net.h"
#include "caffe2/core/observer.h"
#include "caffe2/core/timer.h"
#include "observers/macros.h"

#include <unordered_map>

namespace caffe2 {

double getClockTimeMilliseconds();

class CAFFE2_OBSERVER_API PerfNetObserver : public NetObserver {
 public:
  explicit PerfNetObserver(NetBase* subject_);
  virtual ~PerfNetObserver();

 private:
  void Start() override;
  void Stop() override;

  caffe2::string getObserverName(const OperatorBase* op, int idx) const;

 private:
  enum LogType {
    NONE,
    OPERATOR_DELAY,
    NET_DELAY,
  };
  LogType logType_;
  unsigned int numRuns_;
  std::unordered_map<const OperatorBase*, const ObserverBase<OperatorBase>*>
      observerMap_;

  double wallMilliseconds_;
  double cpuMilliseconds_;
};

class PerfOperatorObserver : public ObserverBase<OperatorBase> {
 public:
  PerfOperatorObserver(OperatorBase* op, PerfNetObserver* netObserver);
  virtual ~PerfOperatorObserver();

  double getWallMilliseconds() const;
  double getCpuMilliseconds() const;
  std::vector<TensorShape> getTensorShapes() const;

 private:
  void Start() override;
  void Stop() override;

 private:
  // Observer of a net that owns corresponding op. We make sure net is never
  // destructed while operator observer is still alive. First operator observer
  // gets destructed, then the op, then the net and its observer.
  // We do this trick in order to get access to net's name and other fields
  // without storing inside the operator observer. Each field is memory
  // costly here and a raw pointer is a cheapest sholution
  PerfNetObserver* netObserver_;
  double wallMilliseconds_;
  double cpuMilliseconds_;
  std::vector<TensorShape> tensor_shapes_;
};
} // namespace caffe2