MonitorElementCollection.h

Go to the documentation of this file.

#ifndef DataFormats_Histograms_MonitorElementCollection_h
#define DataFormats_Histograms_MonitorElementCollection_h
// -*- C++ -*-
//
// Package:     DataFormats/Histograms
// Class  :     MonitorElementCollection
//
//
// Original Author:  Marcel Schneider
//         Created:  2018-05-02
//
//
#include "DataFormats/Provenance/interface/LuminosityBlockID.h"
#include "FWCore/Utilities/interface/propagate_const.h"

#include <cstdint>
#include <cassert>
#include <vector>
#include <string>

#include "TH1.h"

struct MonitorElementData {
  // This is technically a union, but the struct is safer.
  struct Scalar {
    int64_t num = 0;
    double real = 0;
    std::string str;
  };

  // Quality test result types.
  // These are inherited from DQMNet/old DQMStore, and left unchanged to avoid
  // another layer of wrapping. The APIs are used in some places in subsystem
  // code, and could be changed, but not removed.
  class QReport {
  public:
    struct QValue {
      int code;
      float qtresult;
      std::string message;
      std::string qtname;
      std::string algorithm;
    };
    struct DQMChannel {
      int binx;       //< bin # in x-axis (or bin # for 1D histogram)
      int biny;       //< bin # in y-axis (for 2D or 3D histograms)
      int binz;       //< bin # in z-axis (for 3D histograms)
      float content;  //< bin content
      float RMS;      //< RMS of bin content

      int getBin() { return getBinX(); }
      int getBinX() { return binx; }
      int getBinY() { return biny; }
      int getBinZ() { return binz; }
      float getContents() { return content; }
      float getRMS() { return RMS; }

      DQMChannel(int bx, int by, int bz, float data, float rms) {
        binx = bx;
        biny = by;
        binz = bz;
        content = data;
        RMS = rms;
      }

      DQMChannel() {
        binx = 0;
        biny = 0;
        binz = 0;
        content = 0;
        RMS = 0;
      }
    };

    QValue& getValue() { return qvalue_; };
    QValue const& getValue() const { return qvalue_; };

    int getStatus() const { return qvalue_.code; }

    float getQTresult() const { return qvalue_.qtresult; }

    const std::string& getMessage() const { return qvalue_.message; }

    const std::string& getQRName() const { return qvalue_.qtname; }

    const std::string& getAlgorithm() const { return qvalue_.algorithm; }

    const std::vector<DQMChannel>& getBadChannels() const { return badChannels_; }

    void setBadChannels(std::vector<DQMChannel> badChannels) { badChannels_ = badChannels; }

    QReport(QValue value) : qvalue_(value) {}

  private:
    QValue qvalue_;                        //< Pointer to the actual data.
    std::vector<DQMChannel> badChannels_;  //< Bad channels from QCriterion.
  };

  // These values are compatible to DQMNet, but DQMNet is not likely to exist
  // in the future.
  enum class Kind {
    INVALID = 0x0,
    INT = 0x1,
    REAL = 0x2,
    STRING = 0x3,
    TH1F = 0x10,
    TH1S = 0x11,
    TH1D = 0x12,
    TH1I = 0x13,
    TH2F = 0x20,
    TH2S = 0x21,
    TH2D = 0x22,
    TH2I = 0x23,
    TH3F = 0x30,
    TPROFILE = 0x40,
    TPROFILE2D = 0x41
  };

  // Which window of time the ME is supposed to cover.
  // There is space for a granularity level between runs and lumisections,
  // maybe blocks of 10LS or some fixed number of events or integrated
  // luminosity. We also want to be able to change the granularity centrally
  // depending on the use case. That is what the default is for, and it should
  // be used unless some specific granularity is really required.
  // We'll also need to switch the default to JOB for multi-run harvesting.
  enum Scope { JOB = 1, RUN = 2, LUMI = 3 /*, BLOCK = 4 */ };

  // The main ME data. We don't keep references/QTest results, instead we use
  // only the fields stored in DQMIO files.
  struct Value {
    Scalar scalar_;
    edm::propagate_const<std::unique_ptr<TH1>> object_;
    std::vector<QReport> qreports_;
  };

  struct Path {
  private:
    // We could use pointers to interned strings here to save some space.
    std::string dirname_;
    std::string objname_;

  public:
    enum class Type { DIR, DIR_AND_NAME };

    std::string const& getDirname() const { return dirname_; }
    std::string const& getObjectname() const { return objname_; }
    std::string getFullname() const { return dirname_ + objname_; }

    // Clean up the path and normalize it to preserve certain invariants.
    // Instead of reasoning about whatever properties of paths, we just parse
    // the thing and build a normalized instance with no slash in the beginning
    // and a slash in the end.
    // Type of string `path` could be just directory name, or
    // directory name followed by the name of the monitor element
    void set(std::string path, Path::Type type) {
      //rebuild 'path' to be in canonical form

      //remove any leading '/'
      while (not path.empty() and path.front() == '/') {
        path.erase(path.begin());
      }

      //handle '..' and '//'
      // the 'dir' tokens are separate by a single '/'
      std::string::size_type tokenStartPos = 0;
      while (tokenStartPos < path.size()) {
        auto tokenEndPos = path.find('/', tokenStartPos);
        if (tokenEndPos == std::string::npos) {
          tokenEndPos = path.size();
        }
        if (0 == tokenEndPos - tokenStartPos) {
          //we are sitting on a '/'
          path.erase(path.begin() + tokenStartPos);
          continue;
        } else if (2 == tokenEndPos - tokenStartPos) {
          if (path[tokenStartPos] == '.' and path[tokenStartPos + 1] == '.') {
            //need to go backwards and remove previous directory
            auto endOfLastToken = tokenStartPos;
            if (tokenStartPos > 1) {
              endOfLastToken -= 2;
            }
            auto startOfLastToken = path.rfind('/', endOfLastToken);
            if (startOfLastToken == std::string::npos) {
              //we are at the very beginning of 'path' since no '/' found
              path.erase(path.begin(), path.begin() + tokenEndPos);
              tokenStartPos = 0;
            } else {
              path.erase(path.begin() + startOfLastToken + 1, path.begin() + tokenEndPos);
              tokenStartPos = startOfLastToken + 1;
            }
            continue;
          }
        }
        tokenStartPos = tokenEndPos + 1;
      }

      //separate into objname_ and dirname_;
      objname_.clear();
      if (type == Path::Type::DIR) {
        if (not path.empty() and path.back() != '/') {
          path.append(1, '/');
        }
        dirname_ = std::move(path);
      } else {
        auto lastSlash = path.rfind('/');
        if (lastSlash == std::string::npos) {
          objname_ = std::move(path);
          dirname_.clear();
        } else {
          objname_ = path.substr(lastSlash + 1);
          path.erase(path.begin() + lastSlash + 1, path.end());
          dirname_ = std::move(path);
        }
      }
    }

    bool operator==(Path const& other) const {
      return this->dirname_ == other.dirname_ && this->objname_ == other.objname_;
    }
  };

  // Metadata about the ME. The range is included here in case we have e.g.
  // multiple per-lumi histograms in one collection. For a logical comparison,
  // one should look only at the name.
  struct Key {
    Path path_;

    // Run number (and optionally lumi number) that the ME belongs to.
    edm::LuminosityBlockID id_;
    Scope scope_;
    Kind kind_;

    bool operator<(Key const& other) const {
      auto makeKeyTuple = [](Key const& k) {
        return std::make_tuple(
            k.path_.getDirname(), k.path_.getObjectname(), k.scope_, k.id_.run(), k.id_.luminosityBlock());
      };

      return makeKeyTuple(*this) < makeKeyTuple(other);
    }
  };

  bool operator<(MonitorElementData const& other) const { return this->key_ < other.key_; }

  // The only non class/struct members
  Key key_;
  Value value_;
};

// For now, no additional (meta-)data is needed apart from the MEs themselves.
// The framework will take care of tracking the plugin and LS/run that the MEs
// belong to.
// Unused for now.
class MonitorElementCollection {
  std::vector<std::unique_ptr<const MonitorElementData>> data_;

public:
  void push_back(std::unique_ptr<const MonitorElementData> value) {
    // enforce ordering
    assert(data_.empty() || data_[data_.size() - 1] <= value);
    data_.push_back(std::move(value));
  }

  void swap(MonitorElementCollection& other) { data_.swap(other.data_); }

  auto begin() const { return data_.begin(); }

  auto end() const { return data_.end(); }

  bool mergeProduct(MonitorElementCollection const& product) {
    // discussion: https://twiki.cern.ch/twiki/bin/view/CMSPublic/SWGuidePerRunAndPerLumiBlockData#Merging_Run_and_Luminosity_Block
    assert(!"Not implemented yet.");
    return false;
    // Things to decide:
    // - Should we allow merging collections of different sets of MEs? (probably not.) [0]
    // - Should we assume the MEs to be ordered? (probably yes.)
    // - How to handle incompatible MEs (different binning)? (fail hard.) [1]
    // - Can multiple MEs with same (dirname, objname) exist? (probably yes.) [2]
    // - Shall we modify the (immutable?) ROOT objects? (probably yes.)
    //
    // [0] Merging should increase the statistics, but not change the number of
    // MEs, at least with the current workflows. It might be convenient to
    // allow it, but for the beginning, it would only mask errors.
    // [1] The DQM framework should guarantee same booking parameters as long
    // as we stay within the Scope of the MEs.
    // [2] To implement e.g. MEs covering blocks of 10LS, we'd store them in a
    // run product, but have as many MEs with same name but different range as
    // needed to perserve the wanted granularity. Merging then can merge or
    // concatenate as possible/needed.
    // Problem: We need to keep copies in memory until the end of run, even
    // though we could save them to the output file as soon as it is clear that
    // the nexe LS will not fall into the same block. Instead, we could drop
    // them into the next lumi block we see; the semantics would be weird (the
    // MEs in the lumi block don't actually correspond to the lumi block they
    // are in) but the DQMIO output should be able to handle that.
  }
};

#endif