GPUReconstructionTimeframe.cxx

Go to the documentation of this file.

// Copyright 2019-2020 CERN and copyright holders of ALICE O2.
// See https://alice-o2.web.cern.ch/copyright for details of the copyright holders.
// All rights not expressly granted are reserved.
//
// This software is distributed under the terms of the GNU General Public
// License v3 (GPL Version 3), copied verbatim in the file "COPYING".
//
// In applying this license CERN does not waive the privileges and immunities
// granted to it by virtue of its status as an Intergovernmental Organization
// or submit itself to any jurisdiction.
 
 
#include "GPUReconstructionTimeframe.h"
#include "GPUReconstruction.h"
#include "GPUChainTracking.h"
#include "GPUChainTrackingGetters.inc"
#include "display/GPUDisplayInterface.h"
#include "GPUQA.h"
#include "AliHLTTPCClusterMCData.h"
#include "GPUTPCMCInfo.h"
#include "GPUTPCClusterData.h"
#include "AliHLTTPCRawCluster.h"
#include "TPCFastTransform.h"
#include "CorrectionMapsHelper.h"
#include "GPUO2DataTypes.h"
#include "GPUSettings.h"
 
#include <cstdio>
#include <exception>
#include <memory>
#include <cstring>
 
using namespace o2::gpu;
 
namespace o2::gpu
{
extern GPUSettingsStandalone configStandalone;
}
static auto& config = configStandalone.TF;
 
GPUReconstructionTimeframe::GPUReconstructionTimeframe(GPUChainTracking* chain, int32_t (*read)(int32_t), int32_t nEvents) : mChain(chain), mReadEvent(read), mNEventsInDirectory(nEvents), mDisUniReal(0.f, 1.f), mRndGen1(configStandalone.seed), mRndGen2(mDisUniInt(mRndGen1))
{
  mMaxBunchesFull = TIME_ORBIT / config.bunchSpacing;
  mMaxBunches = (TIME_ORBIT - config.abortGapTime) / config.bunchSpacing;
 
  if (config.overlayRaw && chain->GetTPCTransformHelper() == nullptr) {
    GPUInfo("Overlay Raw Events requires TPC Fast Transform");
    throw std::exception();
  }
  if (config.bunchSim) {
    if (config.bunchCount * config.bunchTrainCount > mMaxBunches) {
      GPUInfo("Invalid timeframe settings: too many colliding bunches requested!");
      throw std::exception();
    }
    mTrainDist = mMaxBunches / config.bunchTrainCount;
    mCollisionProbability = (float)config.interactionRate * (float)(mMaxBunchesFull * config.bunchSpacing / 1e9f) / (float)(config.bunchCount * config.bunchTrainCount);
    GPUInfo("Timeframe settings: %d trains of %d bunches, bunch spacing: %d, train spacing: %dx%d, filled bunches %d / %d (%d), collision probability %f, mixing %d events", config.bunchTrainCount, config.bunchCount, config.bunchSpacing, mTrainDist, config.bunchSpacing,
            config.bunchCount * config.bunchTrainCount, mMaxBunches, mMaxBunchesFull, mCollisionProbability, mNEventsInDirectory);
  }
 
  mEventStride = configStandalone.seed;
  mSimBunchNoRepeatEvent = configStandalone.StartEvent;
  mEventUsed.resize(mNEventsInDirectory);
  if (config.noEventRepeat == 2) {
    memset(mEventUsed.data(), 0, mNEventsInDirectory * sizeof(mEventUsed[0]));
  }
}
 
int32_t GPUReconstructionTimeframe::ReadEventShifted(int32_t iEvent, float shiftZ, float minZ, float maxZ, bool silent)
{
  mReadEvent(iEvent);
  if (config.overlayRaw) {
    float shiftTTotal = (((double)config.timeFrameLen - DRIFT_TIME) * ((double)TPCZ / (double)DRIFT_TIME) - shiftZ) / mChain->GetTPCTransformHelper()->getCorrMap()->getVDrift();
    for (uint32_t iSector = 0; iSector < NSECTORS; iSector++) {
      for (uint32_t j = 0; j < mChain->mIOPtrs.nRawClusters[iSector]; j++) {
        auto& tmp = mChain->mIOMem.rawClusters[iSector][j];
        tmp.fTime += shiftTTotal;
      }
    }
  }
  if (shiftZ != 0.f) {
    for (uint32_t iSector = 0; iSector < NSECTORS; iSector++) {
      for (uint32_t j = 0; j < mChain->mIOPtrs.nClusterData[iSector]; j++) {
        auto& tmp = mChain->mIOMem.clusterData[iSector][j];
        tmp.z += iSector < NSECTORS / 2 ? shiftZ : -shiftZ;
      }
    }
    for (uint32_t i = 0; i < mChain->mIOPtrs.nMCInfosTPC; i++) {
      auto& tmp = mChain->mIOMem.mcInfosTPC[i];
      tmp.z += i < NSECTORS / 2 ? shiftZ : -shiftZ;
    }
  }
 
  // Remove clusters outside boundaries
  uint32_t nClusters = 0;
  uint32_t removed = 0;
  if (minZ > -1e6 || maxZ > -1e6) {
    uint32_t currentClusterTotal = 0;
    for (uint32_t iSector = 0; iSector < NSECTORS; iSector++) {
      uint32_t currentClusterSector = 0;
      bool doRaw = config.overlayRaw && mChain->mIOPtrs.nClusterData[iSector] == mChain->mIOPtrs.nRawClusters[iSector];
      for (uint32_t i = 0; i < mChain->mIOPtrs.nClusterData[iSector]; i++) {
        float sign = iSector < NSECTORS / 2 ? 1 : -1;
        if (sign * mChain->mIOMem.clusterData[iSector][i].z >= minZ && sign * mChain->mIOMem.clusterData[iSector][i].z <= maxZ) {
          if (currentClusterSector != i) {
            mChain->mIOMem.clusterData[iSector][currentClusterSector] = mChain->mIOMem.clusterData[iSector][i];
            if (doRaw) {
              mChain->mIOMem.rawClusters[iSector][currentClusterSector] = mChain->mIOMem.rawClusters[iSector][i];
            }
          }
          if (mChain->mIOPtrs.nMCLabelsTPC > currentClusterTotal && nClusters != currentClusterTotal) {
            mChain->mIOMem.mcLabelsTPC[nClusters] = mChain->mIOMem.mcLabelsTPC[currentClusterTotal];
          }
          // GPUInfo("Keeping Cluster ID %d (ID in sector %d) Z=%f (sector %d) --> %d (sector %d)", currentClusterTotal, i, mChain->mIOMem.clusterData[iSector][i].fZ, iSector, nClusters, currentClusterSector);
          currentClusterSector++;
          nClusters++;
        } else {
          // GPUInfo("Removing Cluster ID %d (ID in sector %d) Z=%f (sector %d)", currentClusterTotal, i, mChain->mIOMem.clusterData[iSector][i].fZ, iSector);
          removed++;
        }
        currentClusterTotal++;
      }
      mChain->mIOPtrs.nClusterData[iSector] = currentClusterSector;
      if (doRaw) {
        mChain->mIOPtrs.nRawClusters[iSector] = currentClusterSector;
      }
    }
    if (mChain->mIOPtrs.nMCLabelsTPC) {
      mChain->mIOPtrs.nMCLabelsTPC = nClusters;
    }
  } else {
    for (uint32_t i = 0; i < NSECTORS; i++) {
      nClusters += mChain->mIOPtrs.nClusterData[i];
    }
  }
 
  if (!silent) {
    GPUInfo("Read %u Clusters with %d MC labels and %d MC tracks", nClusters, (int32_t)mChain->mIOPtrs.nMCLabelsTPC, (int32_t)mChain->mIOPtrs.nMCInfosTPC);
    if (minZ > -1e6 || maxZ > 1e6) {
      GPUInfo("\tRemoved %u / %u clusters", removed, nClusters + removed);
    }
  }
 
  mShiftedEvents.emplace_back(mChain->mIOPtrs, std::move(mChain->mIOMem), mChain->mIOPtrs.clustersNative ? *mChain->mIOPtrs.clustersNative : o2::tpc::ClusterNativeAccess());
  return nClusters;
}
 
void GPUReconstructionTimeframe::MergeShiftedEvents()
{
  mChain->ClearIOPointers();
  for (uint32_t i = 0; i < mShiftedEvents.size(); i++) {
    auto& ptr = std::get<0>(mShiftedEvents[i]);
    for (uint32_t j = 0; j < NSECTORS; j++) {
      mChain->mIOPtrs.nClusterData[j] += ptr.nClusterData[j];
      if (config.overlayRaw) {
        mChain->mIOPtrs.nRawClusters[j] += ptr.nRawClusters[j];
      }
    }
    mChain->mIOPtrs.nMCLabelsTPC += ptr.nMCLabelsTPC;
    mChain->mIOPtrs.nMCInfosTPC += ptr.nMCInfosTPC;
    mChain->mIOPtrs.nMCInfosTPCCol += ptr.nMCInfosTPCCol;
    SetDisplayInformation(i);
  }
  uint32_t nClustersTotal = 0;
  uint32_t nClustersTotalRaw = 0;
  uint32_t nClustersSectorOffset[NSECTORS] = {0};
  for (uint32_t i = 0; i < NSECTORS; i++) {
    nClustersSectorOffset[i] = nClustersTotal;
    nClustersTotal += mChain->mIOPtrs.nClusterData[i];
    nClustersTotalRaw += mChain->mIOPtrs.nRawClusters[i];
  }
  if (nClustersTotalRaw && nClustersTotalRaw != nClustersTotal) {
    GPUError("Inconsitency between raw clusters and cluster data");
    throw std::exception();
  }
  if (mChain->mIOPtrs.nMCLabelsTPC && nClustersTotal != mChain->mIOPtrs.nMCLabelsTPC) {
    GPUError("Inconsitency between TPC clusters and MC labels");
    throw std::exception();
  }
  mChain->AllocateIOMemory();
  mChain->mIOPtrs.clustersNative = nullptr;
 
  uint32_t nTrackOffset = 0;
  uint32_t nColOffset = 0;
  uint32_t nClustersEventOffset[NSECTORS] = {0};
  for (uint32_t i = 0; i < mShiftedEvents.size(); i++) {
    auto& ptr = std::get<0>(mShiftedEvents[i]);
    uint32_t inEventOffset = 0;
    for (uint32_t j = 0; j < NSECTORS; j++) {
      memcpy((void*)&mChain->mIOMem.clusterData[j][nClustersEventOffset[j]], (void*)ptr.clusterData[j], ptr.nClusterData[j] * sizeof(ptr.clusterData[j][0]));
      if (nClustersTotalRaw) {
        memcpy((void*)&mChain->mIOMem.rawClusters[j][nClustersEventOffset[j]], (void*)ptr.rawClusters[j], ptr.nRawClusters[j] * sizeof(ptr.rawClusters[j][0]));
      }
      if (mChain->mIOPtrs.nMCLabelsTPC) {
        memcpy((void*)&mChain->mIOMem.mcLabelsTPC[nClustersSectorOffset[j] + nClustersEventOffset[j]], (void*)&ptr.mcLabelsTPC[inEventOffset], ptr.nClusterData[j] * sizeof(ptr.mcLabelsTPC[0]));
      }
      for (uint32_t k = 0; k < ptr.nClusterData[j]; k++) {
        mChain->mIOMem.clusterData[j][nClustersEventOffset[j] + k].id = nClustersSectorOffset[j] + nClustersEventOffset[j] + k;
        if (mChain->mIOPtrs.nMCLabelsTPC) {
          for (int32_t l = 0; l < 3; l++) {
            auto& label = mChain->mIOMem.mcLabelsTPC[nClustersSectorOffset[j] + nClustersEventOffset[j] + k].fClusterID[l];
            if (label.fMCID >= 0) {
              label.fMCID += nTrackOffset;
            }
          }
        }
      }
 
      nClustersEventOffset[j] += ptr.nClusterData[j];
      inEventOffset += ptr.nClusterData[j];
    }
 
    memcpy((void*)&mChain->mIOMem.mcInfosTPC[nTrackOffset], (void*)ptr.mcInfosTPC, ptr.nMCInfosTPC * sizeof(ptr.mcInfosTPC[0]));
    for (uint32_t j = 0; j < ptr.nMCInfosTPCCol; j++) {
      mChain->mIOMem.mcInfosTPCCol[nColOffset + j] = ptr.mcInfosTPCCol[j];
      mChain->mIOMem.mcInfosTPCCol[nColOffset + j].first += nTrackOffset;
    }
    nTrackOffset += ptr.nMCInfosTPC;
    nColOffset += ptr.nMCInfosTPCCol;
  }
 
  GPUInfo("Merged %d events, %u clusters total", (int32_t)mShiftedEvents.size(), nClustersTotal);
 
  mShiftedEvents.clear();
}
 
int32_t GPUReconstructionTimeframe::LoadCreateTimeFrame(int32_t iEvent)
{
  if (config.nTotalEventsInTF && mNTotalCollisions >= config.nTotalEventsInTF) {
    return (2);
  }
 
  int64_t nBunch = -DRIFT_TIME / config.bunchSpacing;
  int64_t lastBunch = config.timeFrameLen / config.bunchSpacing;
  int64_t lastTFBunch = lastBunch - DRIFT_TIME / config.bunchSpacing;
  int32_t nCollisions = 0, nBorderCollisions = 0, nTrainCollissions = 0, nMultipleCollisions = 0, nTrainMultipleCollisions = 0;
  int32_t nTrain = 0;
  int32_t mcMin = -1, mcMax = -1;
  uint32_t nTotalClusters = 0;
  while (nBunch < lastBunch) {
    for (int32_t iTrain = 0; iTrain < config.bunchTrainCount && nBunch < lastBunch; iTrain++) {
      int32_t nCollisionsInTrain = 0;
      for (int32_t iBunch = 0; iBunch < config.bunchCount && nBunch < lastBunch; iBunch++) {
        const bool inTF = nBunch >= 0 && nBunch < lastTFBunch && (config.nTotalEventsInTF == 0 || nCollisions < mNTotalCollisions + config.nTotalEventsInTF);
        if (mcMin == -1 && inTF) {
          mcMin = mChain->mIOPtrs.nMCInfosTPC;
        }
        if (mcMax == -1 && nBunch >= 0 && !inTF) {
          mcMax = mChain->mIOPtrs.nMCInfosTPC;
        }
        int32_t nInBunchPileUp = 0;
        double randVal = mDisUniReal(inTF ? mRndGen2 : mRndGen1);
        double p = exp(-mCollisionProbability);
        double p2 = p;
        while (randVal > p) {
          if (config.noBorder && (nBunch < 0 || nBunch >= lastTFBunch)) {
            break;
          }
          if (nCollisionsInTrain >= mNEventsInDirectory) {
            GPUError("Error: insuffient events for mixing!");
            return (1);
          }
          if (nCollisionsInTrain == 0 && config.noEventRepeat == 0) {
            memset(mEventUsed.data(), 0, mNEventsInDirectory * sizeof(mEventUsed[0]));
          }
          if (inTF) {
            nCollisions++;
          } else {
            nBorderCollisions++;
          }
          int32_t useEvent;
          if (config.noEventRepeat == 1) {
            useEvent = mSimBunchNoRepeatEvent;
          } else {
            while (mEventUsed[useEvent = (inTF && config.eventStride ? (mEventStride += config.eventStride) : mDisUniInt(inTF ? mRndGen2 : mRndGen1)) % mNEventsInDirectory]) {
              ;
            }
          }
          if (config.noEventRepeat) {
            mSimBunchNoRepeatEvent++;
          }
          mEventUsed[useEvent] = 1;
          double shift = (double)nBunch * (double)config.bunchSpacing * (double)TPCZ / (double)DRIFT_TIME;
          int32_t nClusters = ReadEventShifted(useEvent, shift, 0, (double)config.timeFrameLen * (double)TPCZ / (double)DRIFT_TIME, true);
          if (nClusters < 0) {
            GPUError("Unexpected error");
            return (1);
          }
          nTotalClusters += nClusters;
          printf("Placing event %4d+%d (ID %4d) at z %7.3f (time %'dns) %s(collisions %4d, bunch %6ld, train %3d) (%'10d clusters, %'10d MC labels, %'10d track MC info)\n", nCollisions, nBorderCollisions, useEvent, shift, (int32_t)(nBunch * config.bunchSpacing), inTF ? " inside" : "outside",
                 nCollisions, nBunch, nTrain, nClusters, mChain->mIOPtrs.nMCLabelsTPC, mChain->mIOPtrs.nMCInfosTPC);
          nInBunchPileUp++;
          nCollisionsInTrain++;
          p2 *= mCollisionProbability / nInBunchPileUp;
          p += p2;
          if (config.noEventRepeat && mSimBunchNoRepeatEvent >= mNEventsInDirectory) {
            nBunch = lastBunch;
          }
        }
        if (nInBunchPileUp > 1) {
          nMultipleCollisions++;
        }
        nBunch++;
      }
      nBunch += mTrainDist - config.bunchCount;
      if (nCollisionsInTrain) {
        nTrainCollissions++;
      }
      if (nCollisionsInTrain > 1) {
        nTrainMultipleCollisions++;
      }
      nTrain++;
    }
    nBunch += mMaxBunchesFull - mTrainDist * config.bunchTrainCount;
  }
  mNTotalCollisions += nCollisions;
  GPUInfo("Timeframe statistics: collisions: %d+%d in %d trains (inside / outside), average rate %f (pile up: in bunch %d, in train %d)", nCollisions, nBorderCollisions, nTrainCollissions, (float)nCollisions / (float)(config.timeFrameLen - DRIFT_TIME) * 1e9, nMultipleCollisions,
          nTrainMultipleCollisions);
  MergeShiftedEvents();
  GPUInfo("\tTotal clusters: %u, MC Labels %d, MC Infos %d", nTotalClusters, (int32_t)mChain->mIOPtrs.nMCLabelsTPC, (int32_t)mChain->mIOPtrs.nMCInfosTPC);
 
  if (!config.noBorder && mChain->GetQA()) {
    mChain->GetQA()->SetMCTrackRange(mcMin, mcMax);
  }
  return (0);
}
 
int32_t GPUReconstructionTimeframe::LoadMergedEvents(int32_t iEvent)
{
  for (int32_t iEventInTimeframe = 0; iEventInTimeframe < config.nMerge; iEventInTimeframe++) {
    float shift;
    if (config.shiftFirstEvent || iEventInTimeframe) {
      if (config.randomizeDistance) {
        shift = mDisUniReal(mRndGen2);
        if (config.shiftFirstEvent) {
          shift = (iEventInTimeframe + shift) * config.averageDistance;
        } else {
          if (iEventInTimeframe == 0) {
            shift = 0;
          } else {
            shift = (iEventInTimeframe - 0.5f + shift) * config.averageDistance;
          }
        }
      } else {
        if (config.shiftFirstEvent) {
          shift = config.averageDistance * (iEventInTimeframe + 0.5f);
        } else {
          shift = config.averageDistance * (iEventInTimeframe);
        }
      }
    } else {
      shift = 0.f;
    }
 
    if (ReadEventShifted(iEvent * config.nMerge + iEventInTimeframe, shift) < 0) {
      return (1);
    }
  }
  MergeShiftedEvents();
  return (0);
}
 
void GPUReconstructionTimeframe::SetDisplayInformation(int32_t iCol)
{
  if (mChain->GetEventDisplay()) {
    for (uint32_t sl = 0; sl < NSECTORS; sl++) {
      mChain->GetEventDisplay()->SetCollisionFirstCluster(iCol, sl, mChain->mIOPtrs.nClusterData[sl]);
    }
    mChain->GetEventDisplay()->SetCollisionFirstCluster(iCol, NSECTORS, mChain->mIOPtrs.nMCInfosTPC);
  }
}