GPU: Remove obsolete code paths

Author: davidrohr

GPU/GPUTracking/Base/GPUReconstruction.cxx

@@ -282,21 +282,9 @@ int32_t GPUReconstruction::InitPhaseBeforeDevice()
     mProcessingSettings.nDeviceHelperThreads = 0;
   }
 
-  if (param().rec.nonConsecutiveIDs) {
-    param().rec.tpc.disableRefitAttachment = 0xFF;
-  }
-  if (!(mRecoSteps.stepsGPUMask & RecoStep::TPCMerging) || !param().rec.tpc.mergerReadFromTrackerDirectly) {
-    mProcessingSettings.fullMergerOnGPU = false;
-  }
-  if (mProcessingSettings.debugLevel > 3 || !IsGPU() || !mProcessingSettings.fullMergerOnGPU || mProcessingSettings.deterministicGPUReconstruction) {
+  if (mProcessingSettings.debugLevel > 3 || !IsGPU() || mProcessingSettings.deterministicGPUReconstruction) {
     mProcessingSettings.delayedOutput = false;
   }
-  if (!mProcessingSettings.fullMergerOnGPU && (GetRecoStepsGPU() & RecoStep::TPCMerging)) {
-    param().rec.tpc.looperInterpolationInExtraPass = 0;
-    if (param().rec.tpc.retryRefit == 1) {
-      param().rec.tpc.retryRefit = 2;
-    }
-  }
 
   UpdateAutomaticProcessingSettings();
   GPUCA_GPUReconstructionUpdateDefaults();

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -149,7 +149,6 @@ AddOptionRTC(mergerInterpolateErrors, uint8_t, 1, "", 0, "Use interpolation inst
 AddOptionRTC(mergeCE, uint8_t, 1, "", 0, "Merge tracks accross the central electrode")
 AddOptionRTC(retryRefit, int8_t, 1, "", 0, "Retry refit with seeding errors and without cluster rejection when fit fails (=2 means retry in same kernel, =1 for separate kernel")
 AddOptionRTC(looperInterpolationInExtraPass, int8_t, -1, "", 0, "Perform looper interpolation in an extra pass")
-AddOptionRTC(mergerReadFromTrackerDirectly, int8_t, 1, "", 0, "Forward data directly from tracker to merger on GPU")
 AddOptionRTC(dropSecondaryLegsInOutput, int8_t, 1, "", 0, "Do not store secondary legs of looping track in TrackTPC")
 AddOptionRTC(enablePID, int8_t, 1, "", 0, "Enable PID response")
 AddOptionRTC(PID_useNsigma, int8_t, 1, "", 0, "Use nSigma instead of absolute distance in PID response")
@@ -188,7 +187,6 @@ EndConfig()
 
 BeginSubConfig(GPUSettingsRec, rec, configStandalone, "REC", 0, "Reconstruction settings", rec)
 AddOptionRTC(maxTrackQPtB5, float, 1.f / GPUCA_MIN_TRACK_PTB5_DEFAULT, "", 0, "required max Q/Pt (==min Pt) of tracks")
-AddOptionRTC(nonConsecutiveIDs, int8_t, false, "", 0, "Non-consecutive cluster IDs as in HLT, disables features that need access to slice data in TPC merger")
 AddOptionRTC(fwdTPCDigitsAsClusters, uint8_t, 0, "", 0, "Forward TPC digits as clusters (if they pass the ZS threshold)")
 AddOptionRTC(bz0Pt10MeV, uint8_t, 60, "", 0, "Nominal Pt to set when bz = 0 (in 10 MeV)")
 AddOptionRTC(fitInProjections, int8_t, -1, "", 0, "Fit in projection, -1 to enable full fit for all but passes but the first one")
@@ -261,7 +259,6 @@ AddOption(overrideClusterizerFragmentLen, int32_t, -1, "", 0, "Force the cluster
 AddOption(trackletSelectorSlices, int8_t, -1, "", 0, "Number of slices to processes in parallel at max")
 AddOption(trackletConstructorInPipeline, int8_t, -1, "", 0, "Run tracklet constructor in the pipeline")
 AddOption(trackletSelectorInPipeline, int8_t, -1, "", 0, "Run tracklet selector in the pipeline")
-AddOption(fullMergerOnGPU, bool, true, "", 0, "Perform full TPC track merging on GPU instead of only refit")
 AddOption(delayedOutput, bool, true, "", 0, "Delay output to be parallel to track fit")
 AddOption(mergerSortTracks, int8_t, -1, "", 0, "Sort track indizes for GPU track fit")
 AddOption(alternateBorderSort, int8_t, -1, "", 0, "Alternative implementation for sorting of border tracks")

GPU/GPUTracking/Global/GPUChainTracking.cxx

@@ -185,12 +185,8 @@ bool GPUChainTracking::ValidateSteps()
     GPUError("Invalid input, TPC Clusterizer needs TPC raw input");
     return false;
   }
-  if (param().rec.tpc.mergerReadFromTrackerDirectly && (GetRecoSteps() & GPUDataTypes::RecoStep::TPCMerging) && ((GetRecoStepsInputs() & GPUDataTypes::InOutType::TPCSectorTracks) || (GetRecoStepsOutputs() & GPUDataTypes::InOutType::TPCSectorTracks) || !(GetRecoSteps() & GPUDataTypes::RecoStep::TPCConversion))) {
-    GPUError("Invalid input / output / step, mergerReadFromTrackerDirectly cannot read/store sectors tracks and needs TPC conversion");
-    return false;
-  }
-  if (!GetProcessingSettings().fullMergerOnGPU && (param().rec.tpc.mergerReadFromTrackerDirectly || GetProcessingSettings().createO2Output) && (GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCMerging)) {
-    GPUError("createO2Output and mergerReadFromTrackerDirectly works only in combination with fullMergerOnGPU if the merger is to run on GPU");
+  if ((GetRecoSteps() & GPUDataTypes::RecoStep::TPCMerging) && ((GetRecoStepsInputs() & GPUDataTypes::InOutType::TPCSectorTracks) || (GetRecoStepsOutputs() & GPUDataTypes::InOutType::TPCSectorTracks) || !(GetRecoSteps() & GPUDataTypes::RecoStep::TPCConversion))) {
+    GPUError("Invalid input / output / step, merger cannot read/store sectors tracks and needs TPC conversion");
     return false;
   }
   bool tpcClustersAvail = (GetRecoStepsInputs() & GPUDataTypes::InOutType::TPCClusters) || (GetRecoSteps() & GPUDataTypes::RecoStep::TPCClusterFinding) || (GetRecoSteps() & GPUDataTypes::RecoStep::TPCDecompression);
@@ -265,14 +261,6 @@ bool GPUChainTracking::ValidateSettings()
     GPUError("Cannot do error interpolation with NWays = 1!");
     return false;
   }
-  if ((param().rec.tpc.mergerReadFromTrackerDirectly || !param().par.earlyTpcTransform) && param().rec.nonConsecutiveIDs) {
-    GPUError("incompatible settings for non consecutive ids");
-    return false;
-  }
-  if (!param().rec.tpc.mergerReadFromTrackerDirectly && GetProcessingSettings().ompKernels) {
-    GPUError("OMP Kernels require mergerReadFromTrackerDirectly");
-    return false;
-  }
   if (param().continuousMaxTimeBin > (int32_t)GPUSettings::TPC_MAX_TF_TIME_BIN) {
     GPUError("configured max time bin exceeds 256 orbits");
     return false;
@@ -743,10 +731,6 @@ int32_t GPUChainTracking::RunChain()
     return 1;
   }
 
-  for (uint32_t i = 0; i < NSLICES; i++) {
-    // GPUInfo("slice %d clusters %d tracks %d", i, mClusterData[i].NumberOfClusters(), processors()->tpcTrackers[i].Output()->NTracks());
-    processors()->tpcMerger.SetSliceData(i, param().rec.tpc.mergerReadFromTrackerDirectly ? nullptr : processors()->tpcTrackers[i].Output());
-  }
   if (runRecoStep(RecoStep::TPCMerging, &GPUChainTracking::RunTPCTrackingMerger, false)) {
     return 1;
   }

GPU/GPUTracking/Global/GPUChainTrackingMerger.cxx

@@ -24,14 +24,14 @@ using namespace o2::gpu;
 void GPUChainTracking::RunTPCTrackingMerger_MergeBorderTracks(int8_t withinSlice, int8_t mergeMode, GPUReconstruction::krnlDeviceType deviceType)
 {
   GPUTPCGMMerger& Merger = processors()->tpcMerger;
-  bool doGPUall = GetRecoStepsGPU() & RecoStep::TPCMerging && GetProcessingSettings().fullMergerOnGPU;
-  GPUTPCGMMerger& MergerShadow = doGPUall ? processorsShadow()->tpcMerger : Merger;
+  bool doGPU = GetRecoStepsGPU() & RecoStep::TPCMerging;
+  GPUTPCGMMerger& MergerShadow = doGPU ? processorsShadow()->tpcMerger : Merger;
   if (GetProcessingSettings().deterministicGPUReconstruction) {
     uint32_t nBorderTracks = withinSlice == 1 ? NSLICES : (2 * NSLICES);
     runKernel<GPUTPCGlobalDebugSortKernels, GPUTPCGlobalDebugSortKernels::borderTracks>({{nBorderTracks, -WarpSize(), 0, deviceType}}, 0);
   }
   uint32_t n = withinSlice == -1 ? NSLICES / 2 : NSLICES;
-  if (GetProcessingSettings().alternateBorderSort && (!mRec->IsGPU() || doGPUall)) {
+  if (GetProcessingSettings().alternateBorderSort && (!mRec->IsGPU() || doGPU)) {
     TransferMemoryResourceLinkToHost(RecoStep::TPCMerging, Merger.MemoryResMemory(), 0, &mEvents->init);
     RecordMarker(&mEvents->single, 0);
     for (uint32_t i = 0; i < n; i++) {
@@ -72,7 +72,7 @@ void GPUChainTracking::RunTPCTrackingMerger_MergeBorderTracks(int8_t withinSlice
       runKernel<GPUTPCGMMergerMergeBorders, 2>(GetGridAuto(0, deviceType), i, withinSlice, mergeMode);
     }
   }
-  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPUall, Merger, &GPUTPCGMMerger::DumpMergeRanges, *mDebugFile, withinSlice, mergeMode);
+  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPU, Merger, &GPUTPCGMMerger::DumpMergeRanges, *mDebugFile, withinSlice, mergeMode);
   mRec->ReturnVolatileDeviceMemory();
 }
 
@@ -89,12 +89,11 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
 {
   mRec->PushNonPersistentMemory(qStr2Tag("TPCMERGE"));
   bool doGPU = GetRecoStepsGPU() & RecoStep::TPCMerging;
-  bool doGPUall = doGPU && GetProcessingSettings().fullMergerOnGPU;
-  GPUReconstruction::krnlDeviceType deviceType = doGPUall ? GPUReconstruction::krnlDeviceType::Auto : GPUReconstruction::krnlDeviceType::CPU;
-  uint32_t numBlocks = (!mRec->IsGPU() || doGPUall) ? BlockCount() : 1;
+  GPUReconstruction::krnlDeviceType deviceType = doGPU ? GPUReconstruction::krnlDeviceType::Auto : GPUReconstruction::krnlDeviceType::CPU;
+  uint32_t numBlocks = (!mRec->IsGPU() || doGPU) ? BlockCount() : 1;
   GPUTPCGMMerger& Merger = processors()->tpcMerger;
   GPUTPCGMMerger& MergerShadow = doGPU ? processorsShadow()->tpcMerger : Merger;
-  GPUTPCGMMerger& MergerShadowAll = doGPUall ? processorsShadow()->tpcMerger : Merger;
+  GPUTPCGMMerger& MergerShadowAll = doGPU ? processorsShadow()->tpcMerger : Merger;
   const int32_t outputStream = OutputStream();
   if (GetProcessingSettings().debugLevel >= 2) {
     GPUInfo("Running TPC Merger");
@@ -112,7 +111,7 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
 
   memset(Merger.Memory(), 0, sizeof(*Merger.Memory()));
   WriteToConstantMemory(RecoStep::TPCMerging, (char*)&processors()->tpcMerger - (char*)processors(), &MergerShadow, sizeof(MergerShadow), 0);
-  if (doGPUall) {
+  if (doGPU) {
     TransferMemoryResourcesToGPU(RecoStep::TPCMerging, &Merger, 0);
   }
 
@@ -136,14 +135,14 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
   if (GetProcessingSettings().deterministicGPUReconstruction) {
     runKernel<GPUTPCGlobalDebugSortKernels, GPUTPCGlobalDebugSortKernels::sectorTracks>({{GPUCA_NSLICES, -WarpSize(), 0, deviceType}}, 1);
   }
-  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPUall, Merger, &GPUTPCGMMerger::DumpSliceTracks, *mDebugFile);
+  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPU, Merger, &GPUTPCGMMerger::DumpSliceTracks, *mDebugFile);
 
   runKernel<GPUTPCGMMergerClearLinks>(GetGridAuto(0, deviceType), false);
   runKernel<GPUMemClean16>({{1, -WarpSize(), 0, deviceType, RecoStep::TPCMerging}}, MergerShadowAll.TmpCounter(), NSLICES * sizeof(*MergerShadowAll.TmpCounter()));
   runKernel<GPUTPCGMMergerMergeWithinPrepare>(GetGridAuto(0, deviceType));
   RunTPCTrackingMerger_MergeBorderTracks(1, 0, deviceType);
   RunTPCTrackingMerger_Resolve(0, 1, deviceType);
-  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPUall, Merger, &GPUTPCGMMerger::DumpMergedWithinSlices, *mDebugFile);
+  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPU, Merger, &GPUTPCGMMerger::DumpMergedWithinSlices, *mDebugFile);
 
   runKernel<GPUTPCGMMergerClearLinks>(GetGridAuto(0, deviceType), false);
   runKernel<GPUMemClean16>({{1, -WarpSize(), 0, deviceType, RecoStep::TPCMerging}}, MergerShadowAll.TmpCounter(), 2 * NSLICES * sizeof(*MergerShadowAll.TmpCounter()));
@@ -158,7 +157,7 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
   runKernel<GPUTPCGMMergerMergeSlicesPrepare>(GetGridBlk(std::max(2u, numBlocks), 0, deviceType), 0, 1, 1);
   RunTPCTrackingMerger_MergeBorderTracks(0, -1, deviceType);
   RunTPCTrackingMerger_Resolve(0, 1, deviceType);
-  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPUall, Merger, &GPUTPCGMMerger::DumpMergedBetweenSlices, *mDebugFile);
+  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPU, Merger, &GPUTPCGMMerger::DumpMergedBetweenSlices, *mDebugFile);
 
   runKernel<GPUMemClean16>({{1, -WarpSize(), 0, deviceType, RecoStep::TPCMerging}}, MergerShadowAll.TmpCounter(), 2 * NSLICES * sizeof(*MergerShadowAll.TmpCounter()));
 
@@ -168,17 +167,17 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
     runKernel<GPUTPCGlobalDebugSortKernels, GPUTPCGlobalDebugSortKernels::globalTracks1>({{1, -WarpSize(), 0, deviceType}}, 1);
     runKernel<GPUTPCGlobalDebugSortKernels, GPUTPCGlobalDebugSortKernels::globalTracks2>({{1, -WarpSize(), 0, deviceType}}, 1);
   }
-  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPUall, Merger, &GPUTPCGMMerger::DumpCollected, *mDebugFile);
+  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPU, Merger, &GPUTPCGMMerger::DumpCollected, *mDebugFile);
 
   if (param().rec.tpc.mergeCE) {
     runKernel<GPUTPCGMMergerClearLinks>(GetGridAuto(0, deviceType), true);
     RunTPCTrackingMerger_MergeBorderTracks(-1, 1, deviceType);
     RunTPCTrackingMerger_MergeBorderTracks(-1, 2, deviceType);
     runKernel<GPUTPCGMMergerMergeCE>(GetGridAuto(0, deviceType));
-    DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPUall, Merger, &GPUTPCGMMerger::DumpMergeCE, *mDebugFile);
+    DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPU, Merger, &GPUTPCGMMerger::DumpMergeCE, *mDebugFile);
   }
   int32_t waitForTransfer = 0;
-  if (doGPUall) {
+  if (doGPU) {
     TransferMemoryResourceLinkToHost(RecoStep::TPCMerging, Merger.MemoryResMemory(), 0, &mEvents->single);
     waitForTransfer = 1;
   }
@@ -189,23 +188,21 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
     runKernel<GPUTPCGMMergerSortTracks>(GetGridAuto(0, deviceType));
   }
 
-  uint32_t maxId = param().rec.nonConsecutiveIDs ? Merger.Memory()->nOutputTrackClusters : Merger.NMaxClusters();
+  uint32_t maxId = Merger.NMaxClusters();
   if (maxId > Merger.NMaxClusters()) {
     throw std::runtime_error("mNMaxClusters too small");
   }
-  if (!param().rec.nonConsecutiveIDs) {
-    runKernel<GPUMemClean16>({{numBlocks, -ThreadCount(), 0, deviceType, RecoStep::TPCMerging}}, MergerShadowAll.SharedCount(), maxId * sizeof(*MergerShadowAll.SharedCount()));
-    runKernel<GPUMemClean16>({{numBlocks, -ThreadCount(), 0, deviceType, RecoStep::TPCMerging}}, MergerShadowAll.ClusterAttachment(), maxId * sizeof(*MergerShadowAll.ClusterAttachment()));
-    runKernel<GPUTPCGMMergerPrepareClusters, 0>(GetGridAuto(0, deviceType));
-    CondWaitEvent(waitForTransfer, &mEvents->single);
-    runKernel<GPUTPCGMMergerSortTracksQPt>(GetGridAuto(0, deviceType));
-    runKernel<GPUTPCGMMergerPrepareClusters, 1>(GetGridAuto(0, deviceType));
-    runKernel<GPUTPCGMMergerPrepareClusters, 2>(GetGridAuto(0, deviceType));
-  }
+  runKernel<GPUMemClean16>({{numBlocks, -ThreadCount(), 0, deviceType, RecoStep::TPCMerging}}, MergerShadowAll.SharedCount(), maxId * sizeof(*MergerShadowAll.SharedCount()));
+  runKernel<GPUMemClean16>({{numBlocks, -ThreadCount(), 0, deviceType, RecoStep::TPCMerging}}, MergerShadowAll.ClusterAttachment(), maxId * sizeof(*MergerShadowAll.ClusterAttachment()));
+  runKernel<GPUTPCGMMergerPrepareClusters, 0>(GetGridAuto(0, deviceType));
+  CondWaitEvent(waitForTransfer, &mEvents->single);
+  runKernel<GPUTPCGMMergerSortTracksQPt>(GetGridAuto(0, deviceType));
+  runKernel<GPUTPCGMMergerPrepareClusters, 1>(GetGridAuto(0, deviceType));
+  runKernel<GPUTPCGMMergerPrepareClusters, 2>(GetGridAuto(0, deviceType));
 
-  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPUall, Merger, &GPUTPCGMMerger::DumpFitPrepare, *mDebugFile);
+  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPU, Merger, &GPUTPCGMMerger::DumpFitPrepare, *mDebugFile);
 
-  if (doGPUall) {
+  if (doGPU) {
     CondWaitEvent(waitForTransfer, &mEvents->single);
     if (waitForTransfer) {
       ReleaseEvent(mEvents->single);
@@ -228,29 +225,23 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
   if (param().rec.tpc.looperInterpolationInExtraPass) {
     runKernel<GPUTPCGMMergerFollowLoopers>(GetGridAuto(0));
   }
-  if (doGPU && !doGPUall) {
-    TransferMemoryResourcesToHost(RecoStep::TPCMerging, &Merger, 0);
-    SynchronizeStream(0);
-  }
 
   DoDebugAndDump(RecoStep::TPCMerging, 2048, Merger, &GPUTPCGMMerger::DumpRefit, *mDebugFile);
   runKernel<GPUTPCGMMergerFinalize, 0>(GetGridAuto(0, deviceType));
-  if (!param().rec.nonConsecutiveIDs) {
-    runKernel<GPUTPCGMMergerFinalize, 1>(GetGridAuto(0, deviceType));
-    runKernel<GPUTPCGMMergerFinalize, 2>(GetGridAuto(0, deviceType));
-  }
+  runKernel<GPUTPCGMMergerFinalize, 1>(GetGridAuto(0, deviceType));
+  runKernel<GPUTPCGMMergerFinalize, 2>(GetGridAuto(0, deviceType));
   if (param().rec.tpc.mergeLoopersAfterburner) {
-    runKernel<GPUTPCGMMergerMergeLoopers, 0>(doGPUall ? GetGrid(Merger.NOutputTracks(), 0, deviceType) : GetGridAuto(0, deviceType));
+    runKernel<GPUTPCGMMergerMergeLoopers, 0>(doGPU ? GetGrid(Merger.NOutputTracks(), 0, deviceType) : GetGridAuto(0, deviceType));
     if (doGPU) {
       TransferMemoryResourceLinkToHost(RecoStep::TPCMerging, Merger.MemoryResMemory(), 0);
       SynchronizeStream(0); // TODO: could probably synchronize on an event after runKernel<GPUTPCGMMergerMergeLoopers, 1>
     }
     runKernel<GPUTPCGMMergerMergeLoopers, 1>(GetGridAuto(0, deviceType));
-    runKernel<GPUTPCGMMergerMergeLoopers, 2>(doGPUall ? GetGrid(Merger.Memory()->nLooperMatchCandidates, 0, deviceType) : GetGridAuto(0, deviceType));
+    runKernel<GPUTPCGMMergerMergeLoopers, 2>(doGPU ? GetGrid(Merger.Memory()->nLooperMatchCandidates, 0, deviceType) : GetGridAuto(0, deviceType));
   }
-  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPUall, Merger, &GPUTPCGMMerger::DumpFinal, *mDebugFile);
+  DoDebugAndDump(RecoStep::TPCMerging, 2048, doGPU, Merger, &GPUTPCGMMerger::DumpFinal, *mDebugFile);
 
-  if (doGPUall) {
+  if (doGPU) {
     RecordMarker(&mEvents->single, 0);
     auto* waitEvent = &mEvents->single;
     if (GetProcessingSettings().keepDisplayMemory || GetProcessingSettings().createO2Output <= 1 || mFractionalQAEnabled) {
@@ -302,7 +293,7 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
     TransferMemoryResourceLinkToHost(RecoStep::TPCMerging, Merger.MemoryResMemory(), 0, &mEvents->single);
     runKernel<GPUTPCGMO2Output, GPUTPCGMO2Output::sort>(GetGridAuto(0, deviceType));
     mRec->ReturnVolatileDeviceMemory();
-    SynchronizeEventAndRelease(mEvents->single, doGPUall);
+    SynchronizeEventAndRelease(mEvents->single, doGPU);
 
     if (GetProcessingSettings().clearO2OutputFromGPU) {
       mRec->AllocateVolatileDeviceMemory(0); // make future device memory allocation volatile
@@ -316,7 +307,7 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
       AllocateRegisteredMemory(Merger.MemoryResOutputO2MC(), mSubOutputControls[GPUTrackingOutputs::getIndex(&GPUTrackingOutputs::tpcTracksO2Labels)]);
       TransferMemoryResourcesToHost(RecoStep::TPCMerging, &Merger, -1, true);
       runKernel<GPUTPCGMO2Output, GPUTPCGMO2Output::mc>(GetGridAuto(0, GPUReconstruction::krnlDeviceType::CPU));
-    } else if (doGPUall) {
+    } else if (doGPU) {
       RecordMarker(&mEvents->single, 0);
       TransferMemoryResourceLinkToHost(RecoStep::TPCMerging, Merger.MemoryResOutputO2(), outputStream, nullptr, &mEvents->single);
       TransferMemoryResourceLinkToHost(RecoStep::TPCMerging, Merger.MemoryResOutputO2Clus(), outputStream);

GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx

@@ -30,11 +30,8 @@ int32_t GPUChainTracking::GlobalTracking(uint32_t iSlice, int32_t threadId, bool
     GPUInfo("GPU Tracker running Global Tracking for slice %u on thread %d\n", iSlice, threadId);
   }
 
-  GPUReconstruction::krnlDeviceType deviceType = GetProcessingSettings().fullMergerOnGPU ? GPUReconstruction::krnlDeviceType::Auto : GPUReconstruction::krnlDeviceType::CPU;
-  runKernel<GPUTPCGlobalTracking>({GetGridBlk(256, iSlice % mRec->NStreams(), deviceType), {iSlice}});
-  if (GetProcessingSettings().fullMergerOnGPU) {
-    TransferMemoryResourceLinkToHost(RecoStep::TPCSliceTracking, processors()->tpcTrackers[iSlice].MemoryResCommon(), iSlice % mRec->NStreams());
-  }
+  runKernel<GPUTPCGlobalTracking>({GetGridBlk(256, iSlice % mRec->NStreams()), {iSlice}});
+  TransferMemoryResourceLinkToHost(RecoStep::TPCSliceTracking, processors()->tpcTrackers[iSlice].MemoryResCommon(), iSlice % mRec->NStreams());
   if (synchronizeOutput) {
     SynchronizeStream(iSlice % mRec->NStreams());
   }
@@ -450,7 +447,7 @@ int32_t GPUChainTracking::RunTPCTrackingSlices_internal()
             blocking[tmpSlice * mRec->NStreams() + sliceRight % mRec->NStreams()] = true;
           }
         }
-        GlobalTracking(tmpSlice, 0, !GetProcessingSettings().fullMergerOnGPU);
+        GlobalTracking(tmpSlice, 0, false);
       }
     }
     for (uint32_t iSlice = 0; iSlice < NSLICES; iSlice++) {