GPUWorkflowSpec.cxx

// 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.

/// @file   GPUWorkflowSpec.cxx
/// @author Matthias Richter, David Rohr
/// @since  2018-04-18
/// @brief  Processor spec for running TPC CA tracking

#include "GPUWorkflow/GPUWorkflowSpec.h"
#include "Headers/DataHeader.h"
#include "Framework/WorkflowSpec.h" // o2::framework::mergeInputs
#include "Framework/DataRefUtils.h"
#include "Framework/DataSpecUtils.h"
#include "Framework/DeviceSpec.h"
#include "Framework/ControlService.h"
#include "Framework/ConfigParamRegistry.h"
#include "Framework/InputRecordWalker.h"
#include "Framework/SerializationMethods.h"
#include "Framework/Logger.h"
#include "Framework/CallbackService.h"
#include "Framework/CCDBParamSpec.h"
#include "Framework/RawDeviceService.h"
#include "DataFormatsTPC/TPCSectorHeader.h"
#include "DataFormatsTPC/ClusterNative.h"
#include "DataFormatsTPC/CompressedClusters.h"
#include "DataFormatsTPC/Helpers.h"
#include "DataFormatsTPC/ZeroSuppression.h"
#include "DataFormatsTPC/RawDataTypes.h"
#include "DataFormatsTPC/WorkflowHelper.h"
#include "DataFormatsGlobalTracking/TrackTuneParams.h"
#include "TPCReconstruction/TPCTrackingDigitsPreCheck.h"
#include "TPCReconstruction/TPCFastTransformHelperO2.h"
#include "DataFormatsTPC/Digit.h"
#include "TPCFastTransform.h"
#include "DPLUtils/DPLRawParser.h"
#include "DPLUtils/DPLRawPageSequencer.h"
#include "DetectorsBase/MatLayerCylSet.h"
#include "DetectorsBase/Propagator.h"
#include "DetectorsBase/GeometryManager.h"
#include "DetectorsRaw/HBFUtils.h"
#include "DetectorsBase/GRPGeomHelper.h"
#include "CommonUtils/NameConf.h"
#include "TPCBase/RDHUtils.h"
#include "GPUO2InterfaceConfiguration.h"
#include "GPUO2InterfaceQA.h"
#include "GPUO2Interface.h"
#include "GPUO2InterfaceUtils.h"
#include "CalibdEdxContainer.h"
#include "ORTRootSerializer.h"
#include "GPUNewCalibValues.h"
#include "TPCPadGainCalib.h"
#include "TPCZSLinkMapping.h"
#include "display/GPUDisplayInterface.h"
#include "TPCBase/Sector.h"
#include "TPCBase/Utils.h"
#include "TPCBaseRecSim/CDBInterface.h"
#include "TPCCalibration/VDriftHelper.h"
#include "CorrectionMapsHelper.h"
#include "TPCCalibration/CorrectionMapsLoader.h"
#include "TPCBaseRecSim/DeadChannelMapCreator.h"
#include "SimulationDataFormat/ConstMCTruthContainer.h"
#include "SimulationDataFormat/MCCompLabel.h"
#include "Algorithm/Parser.h"
#include "DataFormatsGlobalTracking/RecoContainer.h"
#include "DataFormatsTRD/RecoInputContainer.h"
#include "TRDBase/Geometry.h"
#include "TRDBase/GeometryFlat.h"
#include "GPUTRDRecoParam.h"
#include "ITSBase/GeometryTGeo.h"
#include "CommonUtils/DebugStreamer.h"
#include "GPUReconstructionConvert.h"
#include "DetectorsRaw/RDHUtils.h"
#include "ITStracking/TrackingInterface.h"
#include "GPUWorkflowInternal.h"
#include "GPUDataTypesQA.h"
// #include "Framework/ThreadPool.h"

#include <TStopwatch.h>
#include <TObjArray.h>
#include <TH1F.h>
#include <TH2F.h>
#include <TH1D.h>
#include <TGraphAsymmErrors.h>

#include <filesystem>
#include <memory>
#include <vector>
#include <iomanip>
#include <stdexcept>
#include <regex>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <chrono>
#include <unordered_set>

using namespace o2::framework;
using namespace o2::header;
using namespace o2::gpu;
using namespace o2::base;
using namespace o2::dataformats;
using namespace o2::gpu::gpurecoworkflow_internals;

namespace o2::gpu
{

GPURecoWorkflowSpec::GPURecoWorkflowSpec(GPURecoWorkflowSpec::CompletionPolicyData* policyData, Config const& specconfig, std::vector<int32_t> const& tpcsectors, uint64_t tpcSectorMask, std::shared_ptr<o2::base::GRPGeomRequest>& ggr, std::function<bool(o2::framework::DataProcessingHeader::StartTime)>** gPolicyOrder) : o2::framework::Task(), mPolicyData(policyData), mTPCSectorMask(tpcSectorMask), mTPCSectors(tpcsectors), mSpecConfig(specconfig), mGGR(ggr)
{
  if (mSpecConfig.outputCAClusters && !mSpecConfig.caClusterer && !mSpecConfig.decompressTPC && !mSpecConfig.useFilteredOutputSpecs) {
    throw std::runtime_error("inconsistent configuration: cluster output is only possible if CA clusterer or CompCluster decompression is activated");
  }

  mConfig.reset(new GPUO2InterfaceConfiguration);
  mConfParam.reset(new GPUSettingsO2);
  mTFSettings.reset(new GPUSettingsTF);
  mTimer.reset(new TStopwatch);
  mPipeline.reset(new GPURecoWorkflowSpec_PipelineInternals);

  if (mSpecConfig.enableDoublePipeline == 1 && gPolicyOrder) {
    *gPolicyOrder = &mPolicyOrder;
  }
}

GPURecoWorkflowSpec::~GPURecoWorkflowSpec() = default;

void GPURecoWorkflowSpec::init(InitContext& ic)
{
  GRPGeomHelper::instance().setRequest(mGGR);
  GPUO2InterfaceConfiguration& config = *mConfig.get();

  // Create configuration object and fill settings
  mConfig->configGRP.solenoidBzNominalGPU = 0;
  mTFSettings->hasSimStartOrbit = 1;
  auto& hbfu = o2::raw::HBFUtils::Instance();
  mTFSettings->simStartOrbit = hbfu.getFirstIRofTF(o2::InteractionRecord(0, hbfu.orbitFirstSampled)).orbit;

  *mConfParam = mConfig->ReadConfigurableParam();

  if (mConfParam->display) {
    mDisplayFrontend.reset(GPUDisplayFrontendInterface::getFrontend(mConfig->configDisplay.displayFrontend.c_str()));
    mConfig->configProcessing.eventDisplay = mDisplayFrontend.get();
    if (mConfig->configProcessing.eventDisplay != nullptr) {
      LOG(info) << "Event display enabled";
    } else {
      throw std::runtime_error("GPU Event Display frontend could not be created!");
    }
  }
  if (mSpecConfig.enableDoublePipeline) {
    mConfig->configProcessing.doublePipeline = 1;
  }

  mAutoSolenoidBz = mConfParam->solenoidBzNominalGPU == -1e6f;
  mAutoContinuousMaxTimeBin = mConfig->configGRP.grpContinuousMaxTimeBin < 0;
  if (mAutoContinuousMaxTimeBin) {
    mConfig->configGRP.grpContinuousMaxTimeBin = GPUO2InterfaceUtils::getTpcMaxTimeBinFromNHbf(mConfParam->overrideNHbfPerTF ? mConfParam->overrideNHbfPerTF : 256);
  }
  if (mConfig->configProcessing.deviceNum == -2) {
    int32_t myId = ic.services().get<const o2::framework::DeviceSpec>().inputTimesliceId;
    int32_t idMax = ic.services().get<const o2::framework::DeviceSpec>().maxInputTimeslices;
    mConfig->configProcessing.deviceNum = myId;
    LOG(info) << "GPU device number selected from pipeline id: " << myId << " / " << idMax;
  }
  if (mConfig->configProcessing.debugLevel >= 3 && mVerbosity == 0) {
    mVerbosity = 1;
  }
  mConfig->configProcessing.runMC = mSpecConfig.processMC;
  if (mSpecConfig.outputQA) {
    if (!mSpecConfig.processMC && !mConfig->configQA.clusterRejectionHistograms) {
      throw std::runtime_error("Need MC information to create QA plots");
    }
    if (!mSpecConfig.processMC) {
      mConfig->configQA.noMC = true;
    }
    mConfig->configQA.shipToQC = true;
    if (!mConfig->configProcessing.runQA) {
      mConfig->configQA.enableLocalOutput = false;
      mQATaskMask = (mSpecConfig.processMC ? gpudatatypes::gpuqa::tasksAllMC : gpudatatypes::gpuqa::tasksNone) | (mConfig->configQA.clusterRejectionHistograms ? gpudatatypes::gpuqa::taskClusterCounts : gpudatatypes::gpuqa::tasksNone);
      mConfig->configProcessing.runQA = -mQATaskMask;
    }
  }
  mConfig->configInterface.outputToExternalBuffers = true;
  const bool runTracking = mSpecConfig.outputTracks || mSpecConfig.outputCompClustersRoot || mSpecConfig.outputCompClustersFlat;

  // Configure the "GPU workflow" i.e. which steps we run on the GPU (or CPU)
  if (runTracking) {
    mConfig->configWorkflow.steps.set(gpudatatypes::RecoStep::TPCConversion,
                                      gpudatatypes::RecoStep::TPCSectorTracking,
                                      gpudatatypes::RecoStep::TPCMerging);
    mConfig->configWorkflow.outputs.set(gpudatatypes::InOutType::TPCMergedTracks);
  }
  GPUO2Interface::ApplySyncSettings(mConfig->configProcessing, mConfig->configReconstruction, mConfig->configWorkflow.steps, mConfParam->synchronousProcessing, runTracking ? mConfParam->rundEdx : -2);

  if (mSpecConfig.outputCompClustersRoot || mSpecConfig.outputCompClustersFlat) {
    mConfig->configWorkflow.steps.setBits(gpudatatypes::RecoStep::TPCCompression, true);
    mConfig->configWorkflow.outputs.setBits(gpudatatypes::InOutType::TPCCompressedClusters, true);
  }
  mConfig->configWorkflow.inputs.set(gpudatatypes::InOutType::TPCClusters);
  if (mSpecConfig.caClusterer) { // Override some settings if we have raw data as input
    mConfig->configWorkflow.inputs.set(gpudatatypes::InOutType::TPCRaw);
    mConfig->configWorkflow.steps.setBits(gpudatatypes::RecoStep::TPCClusterFinding, true);
    mConfig->configWorkflow.outputs.setBits(gpudatatypes::InOutType::TPCClusters, true);
  }
  if (mSpecConfig.decompressTPC) {
    mConfig->configWorkflow.steps.setBits(gpudatatypes::RecoStep::TPCCompression, false);
    mConfig->configWorkflow.steps.setBits(gpudatatypes::RecoStep::TPCDecompression, true);
    mConfig->configWorkflow.inputs.set(gpudatatypes::InOutType::TPCCompressedClusters);
    mConfig->configWorkflow.outputs.setBits(gpudatatypes::InOutType::TPCClusters, true);
    mConfig->configWorkflow.outputs.setBits(gpudatatypes::InOutType::TPCCompressedClusters, false);
    if (mTPCSectorMask != 0xFFFFFFFFF) {
      throw std::invalid_argument("Cannot run TPC decompression with a sector mask");
    }
  }
  if (mSpecConfig.runTRDTracking) {
    mConfig->configWorkflow.inputs.setBits(gpudatatypes::InOutType::TRDTracklets, true);
    mConfig->configWorkflow.steps.setBits(gpudatatypes::RecoStep::TRDTracking, true);
  }
  if (mSpecConfig.runITSTracking) {
    mConfig->configWorkflow.inputs.setBits(gpudatatypes::InOutType::ITSClusters, true);
    mConfig->configWorkflow.outputs.setBits(gpudatatypes::InOutType::ITSTracks, true);
    mConfig->configWorkflow.steps.setBits(gpudatatypes::RecoStep::ITSTracking, true);
  }
  if (mSpecConfig.outputSharedClusterMap) {
    mConfig->configProcessing.outputSharedClusterMap = true;
  }
  if (!mSpecConfig.outputTracks) {
    mConfig->configProcessing.createO2Output = 0; // Disable O2 TPC track format output if no track output requested
  }
  mConfig->configProcessing.param.tpcTriggerHandling = mSpecConfig.tpcTriggerHandling;

  if (mConfParam->transformationFile.size() || mConfParam->transformationSCFile.size()) {
    LOG(fatal) << "Deprecated configurable param options GPU_global.transformationFile or transformationSCFile used\n"
               << "Instead, link the corresponding file as <somedir>/TPC/Calib/CorrectionMap/snapshot.root and use it via\n"
               << "--condition-remap file://<somdir>=TPC/Calib/CorrectionMap option";
  }
  /* if (config.configProcessing.doublePipeline && ic.services().get<ThreadPool>().poolSize != 2) {
    throw std::runtime_error("double pipeline requires exactly 2 threads");
  } */
  if (config.configProcessing.doublePipeline && (mSpecConfig.readTRDtracklets || mSpecConfig.runITSTracking || !(mSpecConfig.zsOnTheFly || mSpecConfig.zsDecoder))) {
    LOG(fatal) << "GPU two-threaded pipeline works only with TPC-only processing, and with ZS input";
  }

  if (mSpecConfig.enableDoublePipeline != 2) {
    mGPUReco = std::make_unique<GPUO2Interface>();

    // initialize TPC calib objects
    initFunctionTPCCalib(ic);

    mConfig->configCalib.fastTransform = mCalibObjects.mFastTransformHelper->getCorrMap();
    mConfig->configCalib.fastTransformRef = mCalibObjects.mFastTransformHelper->getCorrMapRef();
    mConfig->configCalib.fastTransformMShape = mCalibObjects.mFastTransformHelper->getCorrMapMShape();
    mConfig->configCalib.fastTransformHelper = mCalibObjects.mFastTransformHelper.get();
    if (mConfig->configCalib.fastTransform == nullptr) {
      throw std::invalid_argument("GPU workflow: initialization of the TPC transformation failed");
    }

    if (mConfParam->matLUTFile.size()) {
      LOGP(info, "Loading matlut file {}", mConfParam->matLUTFile.c_str());
      mConfig->configCalib.matLUT = o2::base::MatLayerCylSet::loadFromFile(mConfParam->matLUTFile.c_str());
      if (mConfig->configCalib.matLUT == nullptr) {
        LOGF(fatal, "Error loading matlut file");
      }
    } else {
      mConfig->configProcessing.lateO2MatLutProvisioningSize = 50 * 1024 * 1024;
    }

    if (mSpecConfig.readTRDtracklets) {
      mTRDGeometry = std::make_unique<o2::trd::GeometryFlat>();
      mConfig->configCalib.trdGeometry = mTRDGeometry.get();

      mTRDRecoParam = std::make_unique<GPUTRDRecoParam>();
      mConfig->configCalib.trdRecoParam = mTRDRecoParam.get();
    }

    mConfig->configProcessing.willProvideO2PropagatorLate = true;
    mConfig->configProcessing.o2PropagatorUseGPUField = true;
    if (mConfig->configReconstruction.tpc.trackReferenceX == 1000.f) {
      mConfig->configReconstruction.tpc.trackReferenceX = 83.f;
    }

    if (mConfParam->printSettings && (mConfParam->printSettings > 1 || ic.services().get<const o2::framework::DeviceSpec>().inputTimesliceId == 0)) {
      mConfig->configProcessing.printSettings = true;
      if (mConfParam->printSettings > 1) {
        mConfig->PrintParam();
      }
    }

    // Configuration is prepared, initialize the tracker.
    if (mGPUReco->Initialize(config) != 0) {
      throw std::invalid_argument("GPU Reconstruction initialization failed");
    }
    if (mSpecConfig.outputQA) {
      mQA = std::make_unique<GPUO2InterfaceQA>(mConfig.get());
    }
    if (mSpecConfig.outputErrorQA) {
      mGPUReco->setErrorCodeOutput(&mErrorQA);
    }

    // initialize ITS
    if (mSpecConfig.runITSTracking) {
      initFunctionITS(ic);
    }
  }

  if (mSpecConfig.enableDoublePipeline) {
    initPipeline(ic);
    if (mConfParam->dump >= 2) {
      LOG(fatal) << "Cannot use dump-only mode with multi-threaded pipeline";
    }
  }

  auto& callbacks = ic.services().get<CallbackService>();
  callbacks.set<CallbackService::Id::RegionInfoCallback>([this](fair::mq::RegionInfo const& info) {
    if (info.size == 0) {
      return;
    }
    if (mSpecConfig.enableDoublePipeline) {
      mRegionInfos.emplace_back(info);
    }
    if (mSpecConfig.enableDoublePipeline == 2) {
      return;
    }
    if (mConfParam->registerSelectedSegmentIds != -1 && info.managed && info.id != (uint32_t)mConfParam->registerSelectedSegmentIds) {
      return;
    }
    int32_t fd = 0;
    if (mConfParam->mutexMemReg) {
      mode_t mask = S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP | S_IROTH | S_IWOTH;
      fd = open("/tmp/o2_gpu_memlock_mutex.lock", O_RDWR | O_CREAT | O_CLOEXEC, mask);
      if (fd == -1) {
        throw std::runtime_error("Error opening memlock mutex lock file");
      }
      fchmod(fd, mask);
      if (lockf(fd, F_LOCK, 0)) {
        throw std::runtime_error("Error locking memlock mutex file");
      }
    }
    std::chrono::time_point<std::chrono::high_resolution_clock> start, end;
    if (mConfParam->benchmarkMemoryRegistration) {
      start = std::chrono::high_resolution_clock::now();
    }
    if (mGPUReco->registerMemoryForGPU(info.ptr, info.size)) {
      throw std::runtime_error("Error registering memory for GPU");
    }
    if (mConfParam->benchmarkMemoryRegistration) {
      end = std::chrono::high_resolution_clock::now();
      std::chrono::duration<double> elapsed_seconds = end - start;
      LOG(info) << "Memory registration time (0x" << info.ptr << ", " << info.size << " bytes): " << elapsed_seconds.count() << " s";
    }
    if (mConfParam->mutexMemReg) {
      if (lockf(fd, F_ULOCK, 0)) {
        throw std::runtime_error("Error unlocking memlock mutex file");
      }
      close(fd);
    }
  });

  mTimer->Stop();
  mTimer->Reset();
}

void GPURecoWorkflowSpec::stop()
{
  LOGF(info, "GPU Reconstruction total timing: Cpu: %.3e Real: %.3e s in %d slots", mTimer->CpuTime(), mTimer->RealTime(), mTimer->Counter() - 1);
  handlePipelineStop();
}

void GPURecoWorkflowSpec::endOfStream(EndOfStreamContext& ec)
{
  handlePipelineEndOfStream(ec);
}

void GPURecoWorkflowSpec::finaliseCCDB(o2::framework::ConcreteDataMatcher& matcher, void* obj)
{
  if (mSpecConfig.enableDoublePipeline != 2) {
    finaliseCCDBTPC(matcher, obj);
    if (mSpecConfig.runITSTracking) {
      finaliseCCDBITS(matcher, obj);
    }
  }
  if (GRPGeomHelper::instance().finaliseCCDB(matcher, obj)) {
    mGRPGeomUpdated = true;
    return;
  }
}

template <class D, class E, class F, class G, class H, class I, class J, class K>
void GPURecoWorkflowSpec::processInputs(ProcessingContext& pc, D& tpcZSmeta, E& inputZS, F& tpcZS, G& tpcZSonTheFlySizes, bool& debugTFDump, H& compClustersDummy, I& compClustersFlatDummy, J& pCompClustersFlat, K& tmpEmptyCompClusters)
{
  if (mSpecConfig.enableDoublePipeline == 1) {
    return;
  }
  constexpr static size_t NSectors = o2::tpc::Sector::MAXSECTOR;
  constexpr static size_t NEndpoints = o2::gpu::GPUTrackingInOutZS::NENDPOINTS;

  if (mSpecConfig.zsOnTheFly || mSpecConfig.zsDecoder) {
    for (uint32_t i = 0; i < GPUTrackingInOutZS::NSECTORS; i++) {
      for (uint32_t j = 0; j < GPUTrackingInOutZS::NENDPOINTS; j++) {
        tpcZSmeta.Pointers[i][j].clear();
        tpcZSmeta.Sizes[i][j].clear();
      }
    }
  }
  if (mSpecConfig.zsOnTheFly) {
    tpcZSonTheFlySizes = {0};
    // tpcZSonTheFlySizes: #zs pages per endpoint:
    std::vector<InputSpec> filter = {{"check", ConcreteDataTypeMatcher{gDataOriginTPC, "ZSSIZES"}, Lifetime::Timeframe}};
    bool recv = false, recvsizes = false;
    for (auto const& ref : InputRecordWalker(pc.inputs(), filter)) {
      if (recvsizes) {
        throw std::runtime_error("Received multiple ZSSIZES data");
      }
      tpcZSonTheFlySizes = pc.inputs().get<std::array<uint32_t, NEndpoints * NSectors>>(ref);
      recvsizes = true;
    }
    // zs pages
    std::vector<InputSpec> filter2 = {{"check", ConcreteDataTypeMatcher{gDataOriginTPC, "TPCZS"}, Lifetime::Timeframe}};
    for (auto const& ref : InputRecordWalker(pc.inputs(), filter2)) {
      if (recv) {
        throw std::runtime_error("Received multiple TPCZS data");
      }
      inputZS = pc.inputs().get<gsl::span<o2::tpc::ZeroSuppressedContainer8kb>>(ref);
      recv = true;
    }
    if (!recv || !recvsizes) {
      throw std::runtime_error("TPC ZS on the fly data not received");
    }

    uint32_t offset = 0;
    for (uint32_t i = 0; i < NSectors; i++) {
      uint32_t pageSector = 0;
      for (uint32_t j = 0; j < NEndpoints; j++) {
        pageSector += tpcZSonTheFlySizes[i * NEndpoints + j];
        offset += tpcZSonTheFlySizes[i * NEndpoints + j];
      }
      if (mVerbosity >= 1) {
        LOG(info) << "GOT ZS on the fly pages FOR SECTOR " << i << " ->  pages: " << pageSector;
      }
    }
  }
  if (mSpecConfig.zsDecoder) {
    std::vector<InputSpec> filter = {{"check", ConcreteDataTypeMatcher{gDataOriginTPC, "RAWDATA"}, Lifetime::Timeframe}};
    auto isSameRdh = [](const char* left, const char* right) -> bool {
      return o2::raw::RDHUtils::getFEEID(left) == o2::raw::RDHUtils::getFEEID(right) && o2::raw::RDHUtils::getDetectorField(left) == o2::raw::RDHUtils::getDetectorField(right);
    };
    auto checkForZSData = [](const char* ptr, uint32_t subSpec) -> bool {
      const auto rdhLink = o2::raw::RDHUtils::getLinkID(ptr);
      const auto detField = o2::raw::RDHUtils::getDetectorField(ptr);
      const auto feeID = o2::raw::RDHUtils::getFEEID(ptr);
      const auto feeLinkID = o2::tpc::rdh_utils::getLink(feeID);
      // This check is not what it is supposed to be, but some MC SYNTHETIC data was generated with rdhLinkId set to feeLinkId, so we add some extra logic so we can still decode it
      return detField == o2::tpc::raw_data_types::ZS && ((feeLinkID == o2::tpc::rdh_utils::UserLogicLinkID && (rdhLink == o2::tpc::rdh_utils::UserLogicLinkID || rdhLink == 0)) ||
                                                         (feeLinkID == o2::tpc::rdh_utils::ILBZSLinkID && (rdhLink == o2::tpc::rdh_utils::UserLogicLinkID || rdhLink == o2::tpc::rdh_utils::ILBZSLinkID || rdhLink == 0)) ||
                                                         (feeLinkID == o2::tpc::rdh_utils::DLBZSLinkID && (rdhLink == o2::tpc::rdh_utils::UserLogicLinkID || rdhLink == o2::tpc::rdh_utils::DLBZSLinkID || rdhLink == 0)));
    };
    auto insertPages = [&tpcZSmeta, checkForZSData](const char* ptr, size_t count, uint32_t subSpec) -> void {
      if (checkForZSData(ptr, subSpec)) {
        int32_t rawcru = o2::tpc::rdh_utils::getCRU(ptr);
        int32_t rawendpoint = o2::tpc::rdh_utils::getEndPoint(ptr);
        tpcZSmeta.Pointers[rawcru / 10][(rawcru % 10) * 2 + rawendpoint].emplace_back(ptr);
        tpcZSmeta.Sizes[rawcru / 10][(rawcru % 10) * 2 + rawendpoint].emplace_back(count);
      }
    };
    if (DPLRawPageSequencer(pc.inputs(), filter)(isSameRdh, insertPages, checkForZSData)) {
      debugTFDump = true;
      static uint32_t nErrors = 0;
      nErrors++;
      if (nErrors == 1 || (nErrors < 100 && nErrors % 10 == 0) || nErrors % 1000 == 0 || mNTFs % 1000 == 0) {
        LOG(error) << "DPLRawPageSequencer failed to process TPC raw data - data most likely not padded correctly - Using slow page scan instead (this alarm is downscaled from now on, so far " << nErrors << " of " << mNTFs << " TFs affected)";
      }
    }

    int32_t totalCount = 0;
    for (uint32_t i = 0; i < GPUTrackingInOutZS::NSECTORS; i++) {
      for (uint32_t j = 0; j < GPUTrackingInOutZS::NENDPOINTS; j++) {
        tpcZSmeta.Pointers2[i][j] = tpcZSmeta.Pointers[i][j].data();
        tpcZSmeta.Sizes2[i][j] = tpcZSmeta.Sizes[i][j].data();
        tpcZS.sector[i].zsPtr[j] = tpcZSmeta.Pointers2[i][j];
        tpcZS.sector[i].nZSPtr[j] = tpcZSmeta.Sizes2[i][j];
        tpcZS.sector[i].count[j] = tpcZSmeta.Pointers[i][j].size();
        totalCount += tpcZSmeta.Pointers[i][j].size();
      }
    }
  } else if (mSpecConfig.decompressTPC) {
    if (mSpecConfig.decompressTPCFromROOT) {
      compClustersDummy = *pc.inputs().get<o2::tpc::CompressedClustersROOT*>("input");
      compClustersFlatDummy.setForward(&compClustersDummy);
      pCompClustersFlat = &compClustersFlatDummy;
    } else {
      pCompClustersFlat = pc.inputs().get<o2::tpc::CompressedClustersFlat*>("input").get();
    }
    if (pCompClustersFlat == nullptr) {
      tmpEmptyCompClusters.reset(new char[sizeof(o2::tpc::CompressedClustersFlat)]);
      memset(tmpEmptyCompClusters.get(), 0, sizeof(o2::tpc::CompressedClustersFlat));
      pCompClustersFlat = (o2::tpc::CompressedClustersFlat*)tmpEmptyCompClusters.get();
    }
  } else if (!mSpecConfig.zsOnTheFly) {
    if (mVerbosity) {
      LOGF(info, "running tracking for sector(s) 0x%09x", mTPCSectorMask);
    }
  }
}

int32_t GPURecoWorkflowSpec::runMain(o2::framework::ProcessingContext* pc, GPUTrackingInOutPointers* ptrs, GPUInterfaceOutputs* outputRegions, int32_t threadIndex, GPUInterfaceInputUpdate* inputUpdateCallback)
{
  int32_t retVal = 0;
  if (mConfParam->dump < 2) {
    retVal = mGPUReco->RunTracking(ptrs, outputRegions, threadIndex, inputUpdateCallback);

    if (retVal == 0 && mSpecConfig.runITSTracking) {
      retVal = runITSTracking(*pc);
    }
  }

  if (!mSpecConfig.enableDoublePipeline) { // TODO: Why is this needed for double-pipeline?
    mGPUReco->Clear(false, threadIndex);   // clean non-output memory used by GPU Reconstruction
  }
  return retVal;
}

void GPURecoWorkflowSpec::cleanOldCalibsTPCPtrs(calibObjectStruct& oldCalibObjects)
{
  if (mOldCalibObjects.size() > 0) {
    mOldCalibObjects.pop();
  }
  mOldCalibObjects.emplace(std::move(oldCalibObjects));
}

void GPURecoWorkflowSpec::run(ProcessingContext& pc)
{
  constexpr static size_t NSectors = o2::tpc::Sector::MAXSECTOR;
  constexpr static size_t NEndpoints = o2::gpu::GPUTrackingInOutZS::NENDPOINTS;

  auto cput = mTimer->CpuTime();
  auto realt = mTimer->RealTime();
  mTimer->Start(false);
  mNTFs++;

  std::vector<gsl::span<const char>> inputs;

  const o2::tpc::CompressedClustersFlat* pCompClustersFlat = nullptr;
  size_t compClustersFlatDummyMemory[(sizeof(o2::tpc::CompressedClustersFlat) + sizeof(size_t) - 1) / sizeof(size_t)];
  o2::tpc::CompressedClustersFlat& compClustersFlatDummy = reinterpret_cast<o2::tpc::CompressedClustersFlat&>(compClustersFlatDummyMemory);
  o2::tpc::CompressedClusters compClustersDummy;
  o2::gpu::GPUTrackingInOutZS tpcZS;
  GPURecoWorkflowSpec_TPCZSBuffers tpcZSmeta;
  std::array<uint32_t, NEndpoints * NSectors> tpcZSonTheFlySizes;
  gsl::span<const o2::tpc::ZeroSuppressedContainer8kb> inputZS;
  std::unique_ptr<char[]> tmpEmptyCompClusters;

  bool getWorkflowTPCInput_clusters = false, getWorkflowTPCInput_mc = false, getWorkflowTPCInput_digits = false;
  bool debugTFDump = false;

  if (mSpecConfig.processMC) {
    getWorkflowTPCInput_mc = true;
  }
  if (!mSpecConfig.decompressTPC && !mSpecConfig.caClusterer) {
    getWorkflowTPCInput_clusters = true;
  }
  if (!mSpecConfig.decompressTPC && mSpecConfig.caClusterer && ((!mSpecConfig.zsOnTheFly || mSpecConfig.processMC) && !mSpecConfig.zsDecoder)) {
    getWorkflowTPCInput_digits = true;
  }

  // ------------------------------ Handle inputs ------------------------------

  auto lockDecodeInput = std::make_unique<std::lock_guard<std::mutex>>(mPipeline->mutexDecodeInput);

  GRPGeomHelper::instance().checkUpdates(pc);
  if (mSpecConfig.enableDoublePipeline != 2) {
    if (mSpecConfig.runITSTracking && pc.inputs().getPos("itsTGeo") >= 0) {
      pc.inputs().get<o2::its::GeometryTGeo*>("itsTGeo");
    }
    if (GRPGeomHelper::instance().getGRPECS()->isDetReadOut(o2::detectors::DetID::TPC) && mConfParam->tpcTriggeredMode ^ !GRPGeomHelper::instance().getGRPECS()->isDetContinuousReadOut(o2::detectors::DetID::TPC)) {
      LOG(fatal) << "configKeyValue tpcTriggeredMode does not match GRP isDetContinuousReadOut(TPC) setting";
    }
  }

  GPUTrackingInOutPointers ptrs;
  processInputs(pc, tpcZSmeta, inputZS, tpcZS, tpcZSonTheFlySizes, debugTFDump, compClustersDummy, compClustersFlatDummy, pCompClustersFlat, tmpEmptyCompClusters);                  // Process non-digit / non-cluster inputs
  const auto& inputsClustersDigits = o2::tpc::getWorkflowTPCInput(pc, mVerbosity, getWorkflowTPCInput_mc, getWorkflowTPCInput_clusters, mTPCSectorMask, getWorkflowTPCInput_digits); // Process digit and cluster inputs

  const auto& tinfo = pc.services().get<o2::framework::TimingInfo>();
  mTFSettings->tfStartOrbit = tinfo.firstTForbit;
  mTFSettings->hasTfStartOrbit = 1;
  mTFSettings->hasNHBFPerTF = 1;
  mTFSettings->nHBFPerTF = mConfParam->overrideNHbfPerTF ? mConfParam->overrideNHbfPerTF : GRPGeomHelper::instance().getGRPECS()->getNHBFPerTF();
  mTFSettings->hasRunStartOrbit = 0;
  ptrs.settingsTF = mTFSettings.get();

  if (mSpecConfig.enableDoublePipeline != 2) {
    if (mVerbosity) {
      LOG(info) << "TF firstTForbit " << mTFSettings->tfStartOrbit << " nHBF " << mTFSettings->nHBFPerTF << " runStartOrbit " << mTFSettings->runStartOrbit << " simStartOrbit " << mTFSettings->simStartOrbit;
    }
    if (mConfParam->checkFirstTfOrbit) {
      static uint32_t lastFirstTFOrbit = -1;
      static uint32_t lastTFCounter = -1;
      if (lastFirstTFOrbit != -1 && lastTFCounter != -1) {
        int32_t diffOrbit = tinfo.firstTForbit - lastFirstTFOrbit;
        int32_t diffCounter = tinfo.tfCounter - lastTFCounter;
        if (diffOrbit != diffCounter * mTFSettings->nHBFPerTF) {
          LOG(error) << "Time frame has mismatching firstTfOrbit - Last orbit/counter: " << lastFirstTFOrbit << " " << lastTFCounter << " - Current: " << tinfo.firstTForbit << " " << tinfo.tfCounter;
        }
      }
      lastFirstTFOrbit = tinfo.firstTForbit;
      lastTFCounter = tinfo.tfCounter;
    }
  }

  o2::globaltracking::RecoContainer inputTracksTRD;
  decltype(o2::trd::getRecoInputContainer(pc, &ptrs, &inputTracksTRD)) trdInputContainer;
  if (mSpecConfig.readTRDtracklets) {
    o2::globaltracking::DataRequest dataRequestTRD;
    dataRequestTRD.requestTracks(o2::dataformats::GlobalTrackID::getSourcesMask(o2::dataformats::GlobalTrackID::NONE), false);
    inputTracksTRD.collectData(pc, dataRequestTRD);
    trdInputContainer = std::move(o2::trd::getRecoInputContainer(pc, &ptrs, &inputTracksTRD));
  }

  void* ptrEp[NSectors * NEndpoints] = {};
  bool doInputDigits = false, doInputDigitsMC = false;
  if (mSpecConfig.decompressTPC) {
    ptrs.tpcCompressedClusters = pCompClustersFlat;
  } else if (mSpecConfig.zsOnTheFly) {
    const uint64_t* buffer = reinterpret_cast<const uint64_t*>(&inputZS[0]);
    o2::gpu::GPUReconstructionConvert::RunZSEncoderCreateMeta(buffer, tpcZSonTheFlySizes.data(), *&ptrEp, &tpcZS);
    ptrs.tpcZS = &tpcZS;
    doInputDigits = doInputDigitsMC = mSpecConfig.processMC;
  } else if (mSpecConfig.zsDecoder) {
    ptrs.tpcZS = &tpcZS;
    if (mSpecConfig.processMC) {
      throw std::runtime_error("Cannot process MC information, none available");
    }
  } else if (mSpecConfig.caClusterer) {
    doInputDigits = true;
    doInputDigitsMC = mSpecConfig.processMC;
  } else {
    ptrs.clustersNative = &inputsClustersDigits->clusterIndex;
  }

  if (mTPCSectorMask != 0xFFFFFFFFF) {
    // Clean out the unused sectors, such that if they were present by chance, they are not processed, and if the values are uninitialized, we should not crash
    for (uint32_t i = 0; i < NSectors; i++) {
      if (!(mTPCSectorMask & (1ul << i))) {
        if (ptrs.tpcZS) {
          for (uint32_t j = 0; j < GPUTrackingInOutZS::NENDPOINTS; j++) {
            tpcZS.sector[i].zsPtr[j] = nullptr;
            tpcZS.sector[i].nZSPtr[j] = nullptr;
            tpcZS.sector[i].count[j] = 0;
          }
        }
      }
    }
  }

  GPUTrackingInOutDigits tpcDigitsMap;
  GPUTPCDigitsMCInput tpcDigitsMapMC;
  if (doInputDigits) {
    ptrs.tpcPackedDigits = &tpcDigitsMap;
    if (doInputDigitsMC) {
      tpcDigitsMap.tpcDigitsMC = &tpcDigitsMapMC;
    }
    for (uint32_t i = 0; i < NSectors; i++) {
      tpcDigitsMap.tpcDigits[i] = inputsClustersDigits->inputDigits[i].data();
      tpcDigitsMap.nTPCDigits[i] = inputsClustersDigits->inputDigits[i].size();
      if (doInputDigitsMC) {
        tpcDigitsMapMC.v[i] = inputsClustersDigits->inputDigitsMCPtrs[i];
      }
    }
  }

  o2::tpc::TPCSectorHeader clusterOutputSectorHeader{0};
  if (mClusterOutputIds.size() > 0) {
    clusterOutputSectorHeader.sectorBits = mTPCSectorMask;
    // subspecs [0, NSectors - 1] are used to identify sector data, we use NSectors to indicate the full TPC
    clusterOutputSectorHeader.activeSectors = mTPCSectorMask;
  }

  // ------------------------------ Prepare stage for double-pipeline before normal output preparation ------------------------------

  std::unique_ptr<GPURecoWorkflow_QueueObject> pipelineContext;
  if (mSpecConfig.enableDoublePipeline) {
    if (handlePipeline(pc, ptrs, tpcZSmeta, tpcZS, pipelineContext)) {
      return;
    }
  }

  // ------------------------------ Prepare outputs ------------------------------

  GPUInterfaceOutputs outputRegions;
  using outputDataType = char;
  using outputBufferUninitializedVector = std::decay_t<decltype(pc.outputs().make<DataAllocator::UninitializedVector<outputDataType>>(Output{"", "", 0}))>;
  using outputBufferType = std::pair<std::optional<std::reference_wrapper<outputBufferUninitializedVector>>, outputDataType*>;
  std::vector<outputBufferType> outputBuffers(GPUInterfaceOutputs::count(), {std::nullopt, nullptr});
  std::unordered_set<std::string> outputsCreated;

  auto setOutputAllocator = [this, &outputBuffers, &outputRegions, &pc, &outputsCreated](const char* name, bool condition, GPUOutputControl& region, auto&& outputSpec, size_t offset = 0) {
    if (condition) {
      auto& buffer = outputBuffers[outputRegions.getIndex(region)];
      if (mConfParam->allocateOutputOnTheFly) {
        region.allocator = [this, name, &buffer, &pc, outputSpec = std::move(outputSpec), offset, &outputsCreated](size_t size) -> void* {
          size += offset;
          if (mVerbosity) {
            LOG(info) << "ALLOCATING " << size << " bytes for " << name << ": " << std::get<DataOrigin>(outputSpec).template as<std::string>() << "/" << std::get<DataDescription>(outputSpec).template as<std::string>() << "/" << std::get<2>(outputSpec);
          }
          std::chrono::time_point<std::chrono::high_resolution_clock> start, end;
          if (mVerbosity) {
            start = std::chrono::high_resolution_clock::now();
          }
          buffer.first.emplace(pc.outputs().make<DataAllocator::UninitializedVector<outputDataType>>(std::make_from_tuple<Output>(outputSpec), size));
          outputsCreated.insert(name);
          if (mVerbosity) {
            end = std::chrono::high_resolution_clock::now();
            std::chrono::duration<double> elapsed_seconds = end - start;
            LOG(info) << "Allocation time for " << name << " (" << size << " bytes)"
                      << ": " << elapsed_seconds.count() << "s";
          }
          return (buffer.second = buffer.first->get().data()) + offset;
        };
      } else {
        buffer.first.emplace(pc.outputs().make<DataAllocator::UninitializedVector<outputDataType>>(std::make_from_tuple<Output>(outputSpec), mConfParam->outputBufferSize));
        region.ptrBase = (buffer.second = buffer.first->get().data()) + offset;
        region.size = buffer.first->get().size() - offset;
        outputsCreated.insert(name);
      }
    }
  };

  auto downSizeBuffer = [](outputBufferType& buffer, size_t size) {
    if (!buffer.first) {
      return;
    }
    if (buffer.first->get().size() < size) {
      throw std::runtime_error("Invalid buffer size requested");
    }
    buffer.first->get().resize(size);
    if (size && buffer.first->get().data() != buffer.second) {
      throw std::runtime_error("Inconsistent buffer address after downsize");
    }
  };

  /*auto downSizeBufferByName = [&outputBuffers, &outputRegions, &downSizeBuffer](GPUOutputControl& region, size_t size) {
    auto& buffer = outputBuffers[outputRegions.getIndex(region)];
    downSizeBuffer(buffer, size);
  };*/

  auto downSizeBufferToSpan = [&outputBuffers, &outputRegions, &downSizeBuffer](GPUOutputControl& region, auto span) {
    auto& buffer = outputBuffers[outputRegions.getIndex(region)];
    if (!buffer.first) {
      return;
    }
    if (span.size() && buffer.second != (char*)span.data()) {
      throw std::runtime_error("Buffer does not match span");
    }
    downSizeBuffer(buffer, span.size() * sizeof(*span.data()));
  };

  setOutputAllocator("COMPCLUSTERSFLAT", mSpecConfig.outputCompClustersFlat, outputRegions.compressedClusters, std::make_tuple(gDataOriginTPC, (DataDescription) "COMPCLUSTERSFLAT", 0));
  setOutputAllocator("CLUSTERNATIVE", mClusterOutputIds.size() > 0, outputRegions.clustersNative, std::make_tuple(gDataOriginTPC, mSpecConfig.sendClustersPerSector ? (DataDescription) "CLUSTERNATIVETMP" : (mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "CLUSTERNATIVEF" : (DataDescription) "CLUSTERNATIVE"), NSectors, clusterOutputSectorHeader), sizeof(o2::tpc::ClusterCountIndex));
  setOutputAllocator("CLSHAREDMAP", mSpecConfig.outputSharedClusterMap, outputRegions.sharedClusterMap, std::make_tuple(gDataOriginTPC, (DataDescription) "CLSHAREDMAP", 0));
  setOutputAllocator("TPCOCCUPANCYMAP", mSpecConfig.outputSharedClusterMap, outputRegions.tpcOccupancyMap, std::make_tuple(gDataOriginTPC, (DataDescription) "TPCOCCUPANCYMAP", 0));
  setOutputAllocator("TRACKS", mSpecConfig.outputTracks, outputRegions.tpcTracksO2, std::make_tuple(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "TRACKSF" : (DataDescription) "TRACKS", 0));
  setOutputAllocator("CLUSREFS", mSpecConfig.outputTracks, outputRegions.tpcTracksO2ClusRefs, std::make_tuple(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "CLUSREFSF" : (DataDescription) "CLUSREFS", 0));
  setOutputAllocator("TRACKSMCLBL", mSpecConfig.outputTracks && mSpecConfig.processMC, outputRegions.tpcTracksO2Labels, std::make_tuple(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "TRACKSMCLBLF" : (DataDescription) "TRACKSMCLBL", 0));
  setOutputAllocator("TRIGGERWORDS", mSpecConfig.caClusterer && mConfig->configProcessing.param.tpcTriggerHandling, outputRegions.tpcTriggerWords, std::make_tuple(gDataOriginTPC, (DataDescription) "TRIGGERWORDS", 0));
  o2::tpc::ClusterNativeHelper::ConstMCLabelContainerViewWithBuffer clustersMCBuffer;
  if (mSpecConfig.processMC && (mSpecConfig.caClusterer || mSpecConfig.useFilteredOutputSpecs)) {
    outputRegions.clusterLabels.allocator = [&clustersMCBuffer](size_t size) -> void* { return &clustersMCBuffer; };
  }

  // ------------------------------ Actual processing ------------------------------

  if ((int32_t)(ptrs.tpcZS != nullptr) + (int32_t)(ptrs.tpcPackedDigits != nullptr && (ptrs.tpcZS == nullptr || ptrs.tpcPackedDigits->tpcDigitsMC == nullptr)) + (int32_t)(ptrs.clustersNative != nullptr) + (int32_t)(ptrs.tpcCompressedClusters != nullptr) != 1) {
    throw std::runtime_error("Invalid input for gpu tracking");
  }

  const auto& holdData = o2::tpc::TPCTrackingDigitsPreCheck::runPrecheck(&ptrs, mConfig.get());

  calibObjectStruct oldCalibObjects;
  doCalibUpdates(pc, oldCalibObjects);

  lockDecodeInput.reset();

  uint32_t threadIndex;
  if (mConfParam->dump) {
    if (mSpecConfig.enableDoublePipeline && pipelineContext->jobSubmitted) {
      while (pipelineContext->jobThreadIndex == -1) {
      }
      threadIndex = pipelineContext->jobThreadIndex;
    } else {
      threadIndex = 0; // TODO: Not sure if this is safe, but it is not yet known which threadIndex will pick up the enqueued job
    }

    std::string dir = "";
    if (mConfParam->dumpFolder != "") {
      dir = std::regex_replace(mConfParam->dumpFolder, std::regex("\\[P\\]"), std::to_string(getpid()));
      if (mNTFs == 1) {
        mkdir(dir.c_str(), S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH);
      }
      dir += "/";
    }
    if (mNTFs == 1) { // Must dump with first TF, since will enforce enqueued calib updates
      mGPUReco->DumpSettings(threadIndex, dir.c_str());
    }
    if (tinfo.tfCounter >= mConfParam->dumpFirst && (mConfParam->dumpLast == -1 || tinfo.tfCounter <= mConfParam->dumpLast)) {
      mGPUReco->DumpEvent(mNTFDumps, &ptrs, threadIndex, dir.c_str());
      mNTFDumps++;
    }
  }
  if (mNTFs == 1 && pc.services().get<const o2::framework::DeviceSpec>().inputTimesliceId == 0) { // TPC ConfigurableCarams are somewhat special, need to construct by hand
    o2::conf::ConfigurableParam::write(o2::base::NameConf::getConfigOutputFileName(pc.services().get<const o2::framework::DeviceSpec>().name, "rec_tpc"), "GPU_rec_tpc,GPU_rec,GPU_proc_param,GPU_proc,GPU_global,trackTuneParams");
  }

  std::unique_ptr<GPUTrackingInOutPointers> ptrsDump;
  if (mConfParam->dumpBadTFMode == 2) {
    ptrsDump.reset(new GPUTrackingInOutPointers);
    memcpy((void*)ptrsDump.get(), (const void*)&ptrs, sizeof(ptrs));
  }

  int32_t retVal = 0;
  if (mSpecConfig.enableDoublePipeline) {
    if (!pipelineContext->jobSubmitted) {
      enqueuePipelinedJob(&ptrs, &outputRegions, pipelineContext.get(), true);
    } else {
      finalizeInputPipelinedJob(&ptrs, &outputRegions, pipelineContext.get());
    }
    std::unique_lock lk(pipelineContext->jobFinishedMutex);
    pipelineContext->jobFinishedNotify.wait(lk, [context = pipelineContext.get()]() { return context->jobFinished; });
    retVal = pipelineContext->jobReturnValue;
    threadIndex = pipelineContext->jobThreadIndex;
  } else {
    // uint32_t threadIndex = pc.services().get<ThreadPool>().threadIndex;
    threadIndex = mNextThreadIndex;
    if (mConfig->configProcessing.doublePipeline) {
      mNextThreadIndex = (mNextThreadIndex + 1) % 2;
    }

    retVal = runMain(&pc, &ptrs, &outputRegions, threadIndex);
  }
  if (retVal != 0) {
    debugTFDump = true;
  }
  cleanOldCalibsTPCPtrs(oldCalibObjects);

  o2::utils::DebugStreamer::instance()->flush(); // flushing debug output to file

  if (debugTFDump && mNDebugDumps < mConfParam->dumpBadTFs) {
    mNDebugDumps++;
    if (mConfParam->dumpBadTFMode <= 1) {
      std::string filename = std::string("tpc_dump_") + std::to_string(pc.services().get<const o2::framework::DeviceSpec>().inputTimesliceId) + "_" + std::to_string(mNDebugDumps) + ".dump";
      FILE* fp = fopen(filename.c_str(), "w+b");
      std::vector<InputSpec> filter = {{"check", ConcreteDataTypeMatcher{gDataOriginTPC, "RAWDATA"}, Lifetime::Timeframe}};
      for (auto const& ref : InputRecordWalker(pc.inputs(), filter)) {
        auto data = pc.inputs().get<gsl::span<char>>(ref);
        if (mConfParam->dumpBadTFMode == 1) {
          uint64_t size = data.size();
          fwrite(&size, 1, sizeof(size), fp);
        }
        fwrite(data.data(), 1, data.size(), fp);
      }
      fclose(fp);
    } else if (mConfParam->dumpBadTFMode == 2) {
      mGPUReco->DumpEvent(mNDebugDumps - 1, ptrsDump.get(), threadIndex);
    }
  }

  if (mConfParam->dump == 2) {
    return;
  }

  // ------------------------------ Varios postprocessing steps ------------------------------

  if (mConfig->configProcessing.tpcWriteClustersAfterRejection) {
    ptrs.clustersNative = ptrs.clustersNativeReduced;
  }
  bool createEmptyOutput = false;
  if (retVal != 0) {
    if (retVal == 3 && mConfig->configProcessing.ignoreNonFatalGPUErrors) {
      if (mConfig->configProcessing.throttleAlarms) {
        LOG(warning) << "GPU Reconstruction aborted with non fatal error code, ignoring";
      } else {
        LOG(alarm) << "GPU Reconstruction aborted with non fatal error code, ignoring";
      }
      createEmptyOutput = !mConfParam->partialOutputForNonFatalErrors;
    } else {
      LOG(fatal) << "GPU Reconstruction aborted with error code " << retVal << " - errors are not ignored - terminating";
    }
  }

  std::unique_ptr<o2::tpc::ClusterNativeAccess> tmpEmptyClNative;
  if (createEmptyOutput) {
    memset(&ptrs, 0, sizeof(ptrs));
    for (uint32_t i = 0; i < outputRegions.count(); i++) {
      if (outputBuffers[i].first) {
        size_t toSize = 0;
        if (i == outputRegions.getIndex(outputRegions.compressedClusters)) {
          toSize = sizeof(*ptrs.tpcCompressedClusters);
        } else if (i == outputRegions.getIndex(outputRegions.clustersNative)) {
          toSize = sizeof(o2::tpc::ClusterCountIndex);
        }
        outputBuffers[i].first->get().resize(toSize);
        outputBuffers[i].second = outputBuffers[i].first->get().data();
        if (toSize) {
          memset(outputBuffers[i].second, 0, toSize);
        }
      }
    }
    tmpEmptyClNative = std::make_unique<o2::tpc::ClusterNativeAccess>();
    memset(tmpEmptyClNative.get(), 0, sizeof(*tmpEmptyClNative));
    ptrs.clustersNative = tmpEmptyClNative.get();
    if (mSpecConfig.processMC) {
      MCLabelContainer cont;
      cont.flatten_to(clustersMCBuffer.first);
      clustersMCBuffer.second = clustersMCBuffer.first;
      tmpEmptyClNative->clustersMCTruth = &clustersMCBuffer.second;
    }
  } else {
    gsl::span<const o2::tpc::TrackTPC> spanOutputTracks = {ptrs.outputTracksTPCO2, ptrs.nOutputTracksTPCO2};
    gsl::span<const uint32_t> spanOutputClusRefs = {ptrs.outputClusRefsTPCO2, ptrs.nOutputClusRefsTPCO2};
    gsl::span<const o2::MCCompLabel> spanOutputTracksMCTruth = {ptrs.outputTracksTPCO2MC, ptrs.outputTracksTPCO2MC ? ptrs.nOutputTracksTPCO2 : 0};
    if (!mConfParam->allocateOutputOnTheFly) {
      for (uint32_t i = 0; i < outputRegions.count(); i++) {
        if (outputRegions.asArray()[i].ptrBase) {
          if (outputRegions.asArray()[i].size == 1) {
            throw std::runtime_error("Preallocated buffer size exceeded");
          }
          outputRegions.asArray()[i].checkCurrent();
          downSizeBuffer(outputBuffers[i], (char*)outputRegions.asArray()[i].ptrCurrent - (char*)outputBuffers[i].second);
        }
      }
    }
    downSizeBufferToSpan(outputRegions.tpcTracksO2, spanOutputTracks);
    downSizeBufferToSpan(outputRegions.tpcTracksO2ClusRefs, spanOutputClusRefs);
    downSizeBufferToSpan(outputRegions.tpcTracksO2Labels, spanOutputTracksMCTruth);

    // if requested, tune TPC tracks
    if (ptrs.nOutputTracksTPCO2) {
      doTrackTuneTPC(ptrs, outputBuffers[outputRegions.getIndex(outputRegions.tpcTracksO2)].first->get().data());
    }

    if (mClusterOutputIds.size() > 0 && (void*)ptrs.clustersNative->clustersLinear != (void*)(outputBuffers[outputRegions.getIndex(outputRegions.clustersNative)].second + sizeof(o2::tpc::ClusterCountIndex))) {
      throw std::runtime_error("cluster native output ptrs out of sync"); // sanity check
    }
  }

  if (mConfig->configWorkflow.outputs.isSet(gpudatatypes::InOutType::TPCMergedTracks)) {
    LOG(info) << "found " << ptrs.nOutputTracksTPCO2 << " track(s)";
  }

  if (mSpecConfig.outputCompClustersRoot) {
    o2::tpc::CompressedClustersROOT compressedClusters = *ptrs.tpcCompressedClusters;
    pc.outputs().snapshot(Output{gDataOriginTPC, "COMPCLUSTERS", 0}, ROOTSerialized<o2::tpc::CompressedClustersROOT const>(compressedClusters));
  }

  if (mClusterOutputIds.size() > 0) {
    o2::tpc::ClusterNativeAccess const& accessIndex = *ptrs.clustersNative;
    if (mSpecConfig.sendClustersPerSector) {
      // Clusters are shipped by sector, we are copying into per-sector buffers (anyway only for ROOT output)
      for (uint32_t i = 0; i < NSectors; i++) {
        if (mTPCSectorMask & (1ul << i)) {
          DataHeader::SubSpecificationType subspec = i;
          clusterOutputSectorHeader.sectorBits = (1ul << i);
          char* buffer = pc.outputs().make<char>({gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "CLUSTERNATIVEF" : (DataDescription) "CLUSTERNATIVE", subspec, {clusterOutputSectorHeader}}, accessIndex.nClustersSector[i] * sizeof(*accessIndex.clustersLinear) + sizeof(o2::tpc::ClusterCountIndex)).data();
          o2::tpc::ClusterCountIndex* outIndex = reinterpret_cast<o2::tpc::ClusterCountIndex*>(buffer);
          memset(outIndex, 0, sizeof(*outIndex));
          for (int32_t j = 0; j < o2::tpc::constants::MAXGLOBALPADROW; j++) {
            outIndex->nClusters[i][j] = accessIndex.nClusters[i][j];
          }
          memcpy(buffer + sizeof(*outIndex), accessIndex.clusters[i][0], accessIndex.nClustersSector[i] * sizeof(*accessIndex.clustersLinear));
          if (mSpecConfig.processMC && accessIndex.clustersMCTruth) {
            MCLabelContainer cont;
            for (uint32_t j = 0; j < accessIndex.nClustersSector[i]; j++) {
              const auto& labels = accessIndex.clustersMCTruth->getLabels(accessIndex.clusterOffset[i][0] + j);
              for (const auto& label : labels) {
                cont.addElement(j, label);
              }
            }
            ConstMCLabelContainer contflat;
            cont.flatten_to(contflat);
            pc.outputs().snapshot({gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? DataDescription("CLNATIVEMCLBLF") : DataDescription("CLNATIVEMCLBL"), subspec, {clusterOutputSectorHeader}}, contflat);
          }
        }
      }
    } else {
      // Clusters are shipped as single message, fill ClusterCountIndex
      DataHeader::SubSpecificationType subspec = NSectors;
      o2::tpc::ClusterCountIndex* outIndex = reinterpret_cast<o2::tpc::ClusterCountIndex*>(outputBuffers[outputRegions.getIndex(outputRegions.clustersNative)].second);
      static_assert(sizeof(o2::tpc::ClusterCountIndex) == sizeof(accessIndex.nClusters));
      memcpy(outIndex, &accessIndex.nClusters[0][0], sizeof(o2::tpc::ClusterCountIndex));
      if (mSpecConfig.processMC && (mSpecConfig.caClusterer || mSpecConfig.useFilteredOutputSpecs) && accessIndex.clustersMCTruth) {
        pc.outputs().snapshot({gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? DataDescription("CLNATIVEMCLBLF") : DataDescription("CLNATIVEMCLBL"), subspec, {clusterOutputSectorHeader}}, clustersMCBuffer.first);
      }
    }
  }
  if (mSpecConfig.outputQA) {
    TObjArray out;
    bool sendQAOutput = !createEmptyOutput && outputRegions.qa.newQAHistsCreated;
    auto getoutput = [sendQAOutput](auto ptr) { return sendQAOutput && ptr ? *ptr : std::decay_t<decltype(*ptr)>(); };
    std::vector<TH1F> copy1 = getoutput(outputRegions.qa.hist1); // Internally, this will also be used as output, so we need a non-const copy
    std::vector<TH2F> copy2 = getoutput(outputRegions.qa.hist2);
    std::vector<TH1D> copy3 = getoutput(outputRegions.qa.hist3);
    std::vector<TGraphAsymmErrors> copy4 = getoutput(outputRegions.qa.hist4);
    if (sendQAOutput) {
      mQA->postprocessExternal(copy1, copy2, copy3, copy4, out, mQATaskMask ? mQATaskMask : -1);
    }
    pc.outputs().snapshot({gDataOriginTPC, "TRACKINGQA", 0}, out);
    if (sendQAOutput) {
      mQA->cleanup();
    }
  }
  if (mSpecConfig.outputErrorQA) {
    pc.outputs().snapshot({gDataOriginGPU, "ERRORQA", 0}, mErrorQA);
    mErrorQA.clear(); // FIXME: This is a race condition once we run multi-threaded!
  }
  if (mSpecConfig.outputSharedClusterMap && !outputsCreated.contains("TPCOCCUPANCYMAP")) {
    pc.outputs().make<DataAllocator::UninitializedVector<outputDataType>>({gDataOriginTPC, "TPCOCCUPANCYMAP", 0}, 0u);
  }
  if (mSpecConfig.tpcTriggerHandling && !outputsCreated.contains("TRIGGERWORDS")) {
    pc.outputs().make<DataAllocator::UninitializedVector<outputDataType>>(Output{gDataOriginTPC, "TRIGGERWORDS", 0}, 0u);
  }
  mTimer->Stop();
  LOG(info) << "GPU Reconstruction time for this TF " << mTimer->CpuTime() - cput << " s (cpu), " << mTimer->RealTime() - realt << " s (wall)";
}

void GPURecoWorkflowSpec::doCalibUpdates(o2::framework::ProcessingContext& pc, calibObjectStruct& oldCalibObjects)
{
  GPUCalibObjectsConst newCalibObjects;
  GPUNewCalibValues newCalibValues;
  // check for updates of TPC calibration objects
  bool needCalibUpdate = false;
  if (mGRPGeomUpdated) {
    mGRPGeomUpdated = false;
    needCalibUpdate = true;

    if (mSpecConfig.runITSTracking && !mITSGeometryCreated) {
      o2::its::GeometryTGeo* geom = o2::its::GeometryTGeo::Instance();
      geom->fillMatrixCache(o2::math_utils::bit2Mask(o2::math_utils::TransformType::T2L, o2::math_utils::TransformType::T2GRot, o2::math_utils::TransformType::T2G));
      mITSGeometryCreated = true;
    }

    if (mAutoSolenoidBz) {
      newCalibValues.newSolenoidField = true;
      newCalibValues.solenoidField = mConfig->configGRP.solenoidBzNominalGPU = GPUO2InterfaceUtils::getNominalGPUBz(*GRPGeomHelper::instance().getGRPMagField());
      // Propagator::Instance()->setBz(newCalibValues.solenoidField); // Take value from o2::Propagator::UpdateField from GRPGeomHelper
      LOG(info) << "Updating solenoid field " << newCalibValues.solenoidField;
    }
    if (mAutoContinuousMaxTimeBin) {
      newCalibValues.newContinuousMaxTimeBin = true;
      newCalibValues.continuousMaxTimeBin = mConfig->configGRP.grpContinuousMaxTimeBin = GPUO2InterfaceUtils::getTpcMaxTimeBinFromNHbf(mTFSettings->nHBFPerTF);
      LOG(info) << "Updating max time bin " << newCalibValues.continuousMaxTimeBin << " (" << mTFSettings->nHBFPerTF << " orbits)";
    }

    if (!mPropagatorInstanceCreated) {
      newCalibObjects.o2Propagator = mConfig->configCalib.o2Propagator = Propagator::Instance();
      if (mConfig->configProcessing.o2PropagatorUseGPUField) {
        mGPUReco->UseGPUPolynomialFieldInPropagator(Propagator::Instance());
      }
      mPropagatorInstanceCreated = true;
    }

    if (!mMatLUTCreated) {
      if (mConfParam->matLUTFile.size() == 0) {
        newCalibObjects.matLUT = GRPGeomHelper::instance().getMatLUT();
        LOG(info) << "Loaded material budget lookup table";
      }
      mMatLUTCreated = true;
    }
    if (mSpecConfig.readTRDtracklets) {
      if (!mTRDGeometryCreated) {
        auto gm = o2::trd::Geometry::instance();
        gm->createPadPlaneArray();
        gm->createClusterMatrixArray();
        mTRDGeometry = std::make_unique<o2::trd::GeometryFlat>(*gm);
        newCalibObjects.trdGeometry = mConfig->configCalib.trdGeometry = mTRDGeometry.get();
        LOG(info) << "Loaded TRD geometry";
        mTRDGeometryCreated = true;
      }
      if (!mTRDRecoParamCreated) {
        mTRDRecoParam = std::make_unique<GPUTRDRecoParam>();
        newCalibObjects.trdRecoParam = mConfig->configCalib.trdRecoParam = mTRDRecoParam.get();
        mTRDRecoParamCreated = true;
      }
    }
  }
  needCalibUpdate = fetchCalibsCCDBTPC(pc, newCalibObjects, oldCalibObjects) || needCalibUpdate;
  if (mSpecConfig.runITSTracking) {
    needCalibUpdate = fetchCalibsCCDBITS(pc) || needCalibUpdate;
  }
  if (mTPCCutAtTimeBin != mConfig->configGRP.tpcCutTimeBin) {
    newCalibValues.newTPCTimeBinCut = true;
    newCalibValues.tpcTimeBinCut = mConfig->configGRP.tpcCutTimeBin = mTPCCutAtTimeBin;
    needCalibUpdate = true;
  }
  if (mSpecConfig.nnLoadFromCCDB) {
    auto dumpToFile = [](const char* buffer, std::size_t validSize, const std::string& path) {
      std::ofstream out(path, std::ios::binary | std::ios::trunc);
      if (!out.is_open()) {
        throw std::runtime_error("Failed to open output file: " + path);
      }

      out.write(buffer, static_cast<std::streamsize>(validSize));
      if (!out) {
        throw std::runtime_error("Failed while writing data to: " + path);
      }
    };
    for (int i = 0; i < 3; i++) {
      newCalibObjects.nnClusterizerNetworks[i] = mConfig->configCalib.nnClusterizerNetworks[i];
      if (mSpecConfig.nnDumpToFile && newCalibObjects.nnClusterizerNetworks[i]) {
        std::string path = "tpc_nn_clusterizer_" + std::to_string(i) + ".onnx";
        dumpToFile(newCalibObjects.nnClusterizerNetworks[i]->getONNXModel(), newCalibObjects.nnClusterizerNetworks[i]->getONNXModelSize(), path);
        LOG(info) << "Dumped TPC clusterizer NN " << i << " to file " << path;
      }
    }
  }
  if (needCalibUpdate) {
    LOG(info) << "Updating GPUReconstruction calibration objects";
    mGPUReco->UpdateCalibration(newCalibObjects, newCalibValues);
  }
}

Options GPURecoWorkflowSpec::options()
{
  Options opts;
  if (mSpecConfig.enableDoublePipeline) {
    bool send = mSpecConfig.enableDoublePipeline == 2;
    char* o2jobid = getenv("O2JOBID");
    char* numaid = getenv("NUMAID");
    int32_t chanid = o2jobid ? atoi(o2jobid) : (numaid ? atoi(numaid) : 0);
    std::string chan = std::string("name=gpu-prepare-channel,type=") + (send ? "push" : "pull") + ",method=" + (send ? "connect" : "bind") + ",address=ipc://@gpu-prepare-channel-" + std::to_string(chanid) + "-{timeslice0},transport=shmem,rateLogging=0";
    opts.emplace_back(o2::framework::ConfigParamSpec{"channel-config", o2::framework::VariantType::String, chan, {"Out-of-band channel config"}});
  }
  if (mSpecConfig.enableDoublePipeline == 2) {
    return opts;
  }
  if (mSpecConfig.outputTracks) {
    o2::tpc::CorrectionMapsLoader::addOptions(opts);
  }
  return opts;
}

Inputs GPURecoWorkflowSpec::inputs()
{
  Inputs inputs;
  if (mSpecConfig.zsDecoder) {
    // All ZS raw data is published with subspec 0 by the o2-raw-file-reader-workflow and DataDistribution
    // creates subspec fom CRU and endpoint id, we create one single input route subscribing to all TPC/RAWDATA
    inputs.emplace_back(InputSpec{"zsraw", ConcreteDataTypeMatcher{"TPC", "RAWDATA"}, Lifetime::Timeframe});
    if (mSpecConfig.askDISTSTF) {
      inputs.emplace_back("stdDist", "FLP", "DISTSUBTIMEFRAME", 0, Lifetime::Timeframe);
    }
  }
  if (mSpecConfig.enableDoublePipeline == 2) {
    if (!mSpecConfig.zsDecoder) {
      LOG(fatal) << "Double pipeline mode can only work with zsraw input";
    }
    return inputs;
  } else if (mSpecConfig.enableDoublePipeline == 1) {
    inputs.emplace_back("pipelineprepare", gDataOriginGPU, "PIPELINEPREPARE", 0, Lifetime::Timeframe);
  }
  if (mSpecConfig.enableDoublePipeline != 2 && (mSpecConfig.outputTracks || mSpecConfig.caClusterer)) {
    // calibration objects for TPC clusterization
    inputs.emplace_back("tpcgain", gDataOriginTPC, "PADGAINFULL", 0, Lifetime::Condition, ccdbParamSpec(o2::tpc::CDBTypeMap.at(o2::tpc::CDBType::CalPadGainFull)));
    inputs.emplace_back("tpcaltrosync", gDataOriginTPC, "ALTROSYNCSIGNAL", 0, Lifetime::Condition, ccdbParamSpec(o2::tpc::CDBTypeMap.at(o2::tpc::CDBType::AltroSyncSignal)));
  }
  if (mSpecConfig.enableDoublePipeline != 2 && mSpecConfig.outputTracks) {
    // calibration objects for TPC tracking
    const auto mapSources = mSpecConfig.tpcDeadMapSources;
    if (mapSources != 0) {
      tpc::SourcesDeadMap sources((mapSources > -1) ? static_cast<tpc::SourcesDeadMap>(mapSources) : tpc::SourcesDeadMap::All);
      if ((sources & tpc::SourcesDeadMap::IDCPadStatus) == tpc::SourcesDeadMap::IDCPadStatus) {
        inputs.emplace_back("tpcidcpadflags", gDataOriginTPC, "IDCPADFLAGS", 0, Lifetime::Condition, ccdbParamSpec(o2::tpc::CDBTypeMap.at(o2::tpc::CDBType::CalIDCPadStatusMapA), {}, 1)); // time-dependent
      }
      if ((sources & tpc::SourcesDeadMap::FEEConfig) == tpc::SourcesDeadMap::FEEConfig) {
        inputs.emplace_back("tpcruninfo", gDataOriginTPC, "TPCRUNINFO", 0, Lifetime::Condition, ccdbParamSpec(o2::tpc::CDBTypeMap.at(o2::tpc::CDBType::ConfigRunInfo)));
      }
    }

    inputs.emplace_back("tpcgainresidual", gDataOriginTPC, "PADGAINRESIDUAL", 0, Lifetime::Condition, ccdbParamSpec(o2::tpc::CDBTypeMap.at(o2::tpc::CDBType::CalPadGainResidual), {}, 1)); // time-dependent
    if (mSpecConfig.tpcUseMCTimeGain) {
      inputs.emplace_back("tpctimegain", gDataOriginTPC, "TIMEGAIN", 0, Lifetime::Condition, ccdbParamSpec(o2::tpc::CDBTypeMap.at(o2::tpc::CDBType::CalTimeGainMC), {}, 1)); // time-dependent
    } else {
      inputs.emplace_back("tpctimegain", gDataOriginTPC, "TIMEGAIN", 0, Lifetime::Condition, ccdbParamSpec(o2::tpc::CDBTypeMap.at(o2::tpc::CDBType::CalTimeGain), {}, 1)); // time-dependent
    }
    inputs.emplace_back("tpctopologygain", gDataOriginTPC, "TOPOLOGYGAIN", 0, Lifetime::Condition, ccdbParamSpec(o2::tpc::CDBTypeMap.at(o2::tpc::CDBType::CalTopologyGain)));
    inputs.emplace_back("tpcthreshold", gDataOriginTPC, "PADTHRESHOLD", 0, Lifetime::Condition, ccdbParamSpec("TPC/Config/FEEPad"));
    o2::tpc::VDriftHelper::requestCCDBInputs(inputs);
    Options optsDummy;
    o2::tpc::CorrectionMapsLoaderGloOpts gloOpts{mSpecConfig.lumiScaleType, mSpecConfig.lumiScaleMode, mSpecConfig.enableMShape, mSpecConfig.enableCTPLumi};
    mCalibObjects.mFastTransformHelper->requestCCDBInputs(inputs, optsDummy, gloOpts); // option filled here is lost
  }
  if (mSpecConfig.decompressTPC) {
    inputs.emplace_back(InputSpec{"input", ConcreteDataTypeMatcher{gDataOriginTPC, mSpecConfig.decompressTPCFromROOT ? o2::header::DataDescription("COMPCLUSTERS") : o2::header::DataDescription("COMPCLUSTERSFLAT")}, Lifetime::Timeframe});
  } else if (mSpecConfig.caClusterer) {
    // We accept digits and MC labels also if we run on ZS Raw data, since they are needed for MC label propagation
    if ((!mSpecConfig.zsOnTheFly || mSpecConfig.processMC) && !mSpecConfig.zsDecoder) {
      inputs.emplace_back(InputSpec{"input", ConcreteDataTypeMatcher{gDataOriginTPC, "DIGITS"}, Lifetime::Timeframe});
      mPolicyData->emplace_back(o2::framework::InputSpec{"digits", o2::framework::ConcreteDataTypeMatcher{"TPC", "DIGITS"}});
    }
  } else if (mSpecConfig.runTPCTracking) {
    inputs.emplace_back(InputSpec{"input", ConcreteDataTypeMatcher{gDataOriginTPC, "CLUSTERNATIVE"}, Lifetime::Timeframe});
    mPolicyData->emplace_back(o2::framework::InputSpec{"clusters", o2::framework::ConcreteDataTypeMatcher{"TPC", "CLUSTERNATIVE"}});
  }
  if (mSpecConfig.processMC) {
    if (mSpecConfig.caClusterer) {
      if (!mSpecConfig.zsDecoder) {
        inputs.emplace_back(InputSpec{"mclblin", ConcreteDataTypeMatcher{gDataOriginTPC, "DIGITSMCTR"}, Lifetime::Timeframe});
        mPolicyData->emplace_back(o2::framework::InputSpec{"digitsmc", o2::framework::ConcreteDataTypeMatcher{"TPC", "DIGITSMCTR"}});
      }
    } else if (mSpecConfig.runTPCTracking) {
      inputs.emplace_back(InputSpec{"mclblin", ConcreteDataTypeMatcher{gDataOriginTPC, "CLNATIVEMCLBL"}, Lifetime::Timeframe});
      mPolicyData->emplace_back(o2::framework::InputSpec{"clustersmc", o2::framework::ConcreteDataTypeMatcher{"TPC", "CLNATIVEMCLBL"}});
    }
  }

  if (mSpecConfig.zsOnTheFly) {
    inputs.emplace_back(InputSpec{"zsinput", ConcreteDataTypeMatcher{"TPC", "TPCZS"}, Lifetime::Timeframe});
    inputs.emplace_back(InputSpec{"zsinputsizes", ConcreteDataTypeMatcher{"TPC", "ZSSIZES"}, Lifetime::Timeframe});
  }
  if (mSpecConfig.readTRDtracklets) {
    inputs.emplace_back("trdctracklets", o2::header::gDataOriginTRD, "CTRACKLETS", 0, Lifetime::Timeframe);
    inputs.emplace_back("trdtracklets", o2::header::gDataOriginTRD, "TRACKLETS", 0, Lifetime::Timeframe);
    inputs.emplace_back("trdtriggerrec", o2::header::gDataOriginTRD, "TRKTRGRD", 0, Lifetime::Timeframe);
    inputs.emplace_back("trdtrigrecmask", o2::header::gDataOriginTRD, "TRIGRECMASK", 0, Lifetime::Timeframe);
  }

  if (mSpecConfig.runITSTracking) {
    inputs.emplace_back("compClusters", "ITS", "COMPCLUSTERS", 0, Lifetime::Timeframe);
    inputs.emplace_back("patterns", "ITS", "PATTERNS", 0, Lifetime::Timeframe);
    inputs.emplace_back("ROframes", "ITS", "CLUSTERSROF", 0, Lifetime::Timeframe);
    if (mSpecConfig.itsTriggerType == 1) {
      inputs.emplace_back("phystrig", "ITS", "PHYSTRIG", 0, Lifetime::Timeframe);
    } else if (mSpecConfig.itsTriggerType == 2) {
      inputs.emplace_back("phystrig", "TRD", "TRKTRGRD", 0, Lifetime::Timeframe);
    }
    if (mSpecConfig.enableDoublePipeline != 2) {
      if (mSpecConfig.isITS3) {
        inputs.emplace_back("cldict", "IT3", "CLUSDICT", 0, Lifetime::Condition, ccdbParamSpec("IT3/Calib/ClusterDictionary"));
        inputs.emplace_back("alppar", "ITS", "ALPIDEPARAM", 0, Lifetime::Condition, ccdbParamSpec("ITS/Config/AlpideParam"));
      } else {
        inputs.emplace_back("itscldict", "ITS", "CLUSDICT", 0, Lifetime::Condition, ccdbParamSpec("ITS/Calib/ClusterDictionary"));
        inputs.emplace_back("itsalppar", "ITS", "ALPIDEPARAM", 0, Lifetime::Condition, ccdbParamSpec("ITS/Config/AlpideParam"));
      }
      if (mSpecConfig.itsOverrBeamEst) {
        inputs.emplace_back("meanvtx", "GLO", "MEANVERTEX", 0, Lifetime::Condition, ccdbParamSpec("GLO/Calib/MeanVertex", {}, 1));
      }
    }
    if (mSpecConfig.processMC) {
      inputs.emplace_back("itsmclabels", "ITS", "CLUSTERSMCTR", 0, Lifetime::Timeframe);
      inputs.emplace_back("ITSMC2ROframes", "ITS", "CLUSTERSMC2ROF", 0, Lifetime::Timeframe);
    }
  }

  // NN clusterizer
  *mConfParam = mConfig->ReadConfigurableParam();
  if (mConfig->configProcessing.nn.nnLoadFromCCDB) {

    LOG(info) << "(NN CLUS) Enabling fetching of TPC NN clusterizer from CCDB";
    mSpecConfig.nnLoadFromCCDB = true;
    mSpecConfig.nnDumpToFile = mConfig->configProcessing.nn.nnCCDBDumpToFile;
    GPUSettingsProcessingNNclusterizer& nnClusterizerSettings = mConfig->configProcessing.nn;

    std::map<std::string, std::string> metadata;
    metadata["inputDType"] = nnClusterizerSettings.nnInferenceInputDType;              // FP16 or FP32
    metadata["outputDType"] = nnClusterizerSettings.nnInferenceOutputDType;            // FP16 or FP32
    metadata["nnCCDBWithMomentum"] = nnClusterizerSettings.nnCCDBWithMomentum;         // 0, 1 -> Only for regression model
    metadata["nnCCDBLayerType"] = nnClusterizerSettings.nnCCDBClassificationLayerType; // FC, CNN
    metadata["nnCCDBInteractionRate"] = nnClusterizerSettings.nnCCDBInteractionRate;   // in kHz
    metadata["nnCCDBBeamType"] = nnClusterizerSettings.nnCCDBBeamType;                 // pp, pPb, PbPb

    auto convert_map_to_metadata = [](const std::map<std::string, std::string>& inputMap, std::vector<o2::framework::CCDBMetadata>& outputMetadata) {
      for (const auto& [key, value] : inputMap) {
        if (value != "") {
          outputMetadata.push_back({key, value});
        }
      }
    };

    mSpecConfig.nnEvalMode = o2::utils::Str::tokenize(nnClusterizerSettings.nnEvalMode, ':');
    std::vector<o2::framework::CCDBMetadata> ccdb_metadata;

    if (mConfParam->printSettings) {
      auto printSettings = [](const std::map<std::string, std::string>& settings) {
        LOG(info) << "(NN CLUS) NN Clusterizer CCDB settings:";
        for (const auto& [key, value] : settings) {
          LOG(info) << "  " << key << " : " << value;
        }
      };
      printSettings(metadata);
    }

    if (mSpecConfig.nnEvalMode[0] == "c1") {
      metadata["nnCCDBEvalType"] = "classification_c1";
      convert_map_to_metadata(metadata, ccdb_metadata);
      inputs.emplace_back("nn_classification_c1", gDataOriginTPC, "NNCLUSTERIZER_C1", 0, Lifetime::Condition, ccdbParamSpec(nnClusterizerSettings.nnCCDBPath + "/" + metadata["nnCCDBEvalType"], ccdb_metadata, 0));
    } else if (mSpecConfig.nnEvalMode[0] == "c2") {
      metadata["nnCCDBLayerType"] = nnClusterizerSettings.nnCCDBRegressionLayerType;
      metadata["nnCCDBEvalType"] = "classification_c2";
      convert_map_to_metadata(metadata, ccdb_metadata);
      inputs.emplace_back("nn_classification_c2", gDataOriginTPC, "NNCLUSTERIZER_C2", 0, Lifetime::Condition, ccdbParamSpec(nnClusterizerSettings.nnCCDBPath + "/" + metadata["nnCCDBEvalType"], ccdb_metadata, 0));
    }

    metadata["nnCCDBEvalType"] = "regression_c1";
    metadata["nnCCDBLayerType"] = nnClusterizerSettings.nnCCDBRegressionLayerType;
    convert_map_to_metadata(metadata, ccdb_metadata);
    inputs.emplace_back("nn_regression_c1", gDataOriginTPC, "NNCLUSTERIZER_R1", 0, Lifetime::Condition, ccdbParamSpec(nnClusterizerSettings.nnCCDBPath + "/" + metadata["nnCCDBEvalType"], ccdb_metadata, 0));

    if (mSpecConfig.nnEvalMode[1] == "r2") {
      metadata["nnCCDBEvalType"] = "regression_c2";
      convert_map_to_metadata(metadata, ccdb_metadata);
      inputs.emplace_back("nn_regression_c2", gDataOriginTPC, "NNCLUSTERIZER_R2", 0, Lifetime::Condition, ccdbParamSpec(nnClusterizerSettings.nnCCDBPath + "/" + metadata["nnCCDBEvalType"], ccdb_metadata, 0));
    }
  }

  return inputs;
};

Outputs GPURecoWorkflowSpec::outputs()
{
  constexpr static size_t NSectors = o2::tpc::Sector::MAXSECTOR;
  std::vector<OutputSpec> outputSpecs;
  if (mSpecConfig.enableDoublePipeline == 2) {
    outputSpecs.emplace_back(gDataOriginGPU, "PIPELINEPREPARE", 0, Lifetime::Timeframe);
    return outputSpecs;
  }
  if (mSpecConfig.outputTracks) {
    outputSpecs.emplace_back(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "TRACKSF" : (DataDescription) "TRACKS", 0, Lifetime::Timeframe);
    outputSpecs.emplace_back(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "CLUSREFSF" : (DataDescription) "CLUSREFS", 0, Lifetime::Timeframe);
  }
  if (mSpecConfig.processMC && mSpecConfig.outputTracks) {
    outputSpecs.emplace_back(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "TRACKSMCLBLF" : (DataDescription) "TRACKSMCLBL", 0, Lifetime::Timeframe);
  }
  if (mSpecConfig.outputCompClustersRoot) {
    outputSpecs.emplace_back(gDataOriginTPC, "COMPCLUSTERS", 0, Lifetime::Timeframe);
  }
  if (mSpecConfig.outputCompClustersFlat) {
    outputSpecs.emplace_back(gDataOriginTPC, "COMPCLUSTERSFLAT", 0, Lifetime::Timeframe);
  }
  if (mSpecConfig.outputCAClusters) {
    for (auto const& sector : mTPCSectors) {
      mClusterOutputIds.emplace_back(sector);
    }
    if (mSpecConfig.sendClustersPerSector) {
      outputSpecs.emplace_back(gDataOriginTPC, "CLUSTERNATIVETMP", NSectors, Lifetime::Timeframe); // Dummy buffer the TPC tracker writes the inital linear clusters to
      for (const auto sector : mTPCSectors) {
        outputSpecs.emplace_back(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "CLUSTERNATIVEF" : (DataDescription) "CLUSTERNATIVE", sector, Lifetime::Timeframe);
      }
    } else {
      outputSpecs.emplace_back(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? (DataDescription) "CLUSTERNATIVEF" : (DataDescription) "CLUSTERNATIVE", NSectors, Lifetime::Timeframe);
    }
    if (mSpecConfig.processMC) {
      if (mSpecConfig.sendClustersPerSector) {
        for (const auto sector : mTPCSectors) {
          outputSpecs.emplace_back(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? DataDescription("CLNATIVEMCLBLF") : DataDescription("CLNATIVEMCLBL"), sector, Lifetime::Timeframe);
        }
      } else {
        outputSpecs.emplace_back(gDataOriginTPC, mSpecConfig.useFilteredOutputSpecs ? DataDescription("CLNATIVEMCLBLF") : DataDescription("CLNATIVEMCLBL"), NSectors, Lifetime::Timeframe);
      }
    }
  }
  if (mSpecConfig.outputSharedClusterMap) {
    outputSpecs.emplace_back(gDataOriginTPC, "CLSHAREDMAP", 0, Lifetime::Timeframe);
    outputSpecs.emplace_back(gDataOriginTPC, "TPCOCCUPANCYMAP", 0, Lifetime::Timeframe);
  }
  if (mSpecConfig.tpcTriggerHandling) {
    outputSpecs.emplace_back(gDataOriginTPC, "TRIGGERWORDS", 0, Lifetime::Timeframe);
  }
  if (mSpecConfig.outputQA) {
    outputSpecs.emplace_back(gDataOriginTPC, "TRACKINGQA", 0, Lifetime::Timeframe);
  }
  if (mSpecConfig.outputErrorQA) {
    outputSpecs.emplace_back(gDataOriginGPU, "ERRORQA", 0, Lifetime::Timeframe);
  }

  if (mSpecConfig.runITSTracking) {
    outputSpecs.emplace_back(gDataOriginITS, "TRACKS", 0, Lifetime::Timeframe);
    outputSpecs.emplace_back(gDataOriginITS, "TRACKCLSID", 0, Lifetime::Timeframe);
    outputSpecs.emplace_back(gDataOriginITS, "ITSTrackROF", 0, Lifetime::Timeframe);
    outputSpecs.emplace_back(gDataOriginITS, "VERTICES", 0, Lifetime::Timeframe);
    outputSpecs.emplace_back(gDataOriginITS, "VERTICESROF", 0, Lifetime::Timeframe);
    outputSpecs.emplace_back(gDataOriginITS, "IRFRAMES", 0, Lifetime::Timeframe);

    if (mSpecConfig.processMC) {
      outputSpecs.emplace_back(gDataOriginITS, "VERTICESMCTR", 0, Lifetime::Timeframe);
      outputSpecs.emplace_back(gDataOriginITS, "VERTICESMCPUR", 0, Lifetime::Timeframe);
      outputSpecs.emplace_back(gDataOriginITS, "TRACKSMCTR", 0, Lifetime::Timeframe);
      outputSpecs.emplace_back(gDataOriginITS, "ITSTrackMC2ROF", 0, Lifetime::Timeframe);
    }
  }

  return outputSpecs;
};

void GPURecoWorkflowSpec::deinitialize()
{
  ExitPipeline();
  mQA.reset(nullptr);
  mDisplayFrontend.reset(nullptr);
  mGPUReco.reset(nullptr);
}

} // namespace o2::gpu