GPU: Fix copying of cluster data to GPU when not all processing steps are running on GPU

Author: davidrohr

GPU/GPUTracking/Global/GPUChainTracking.h

@@ -313,6 +313,7 @@ class GPUChainTracking : public GPUChain, GPUReconstructionHelpers::helperDelega
   void RunTPCTrackingMerger_MergeBorderTracks(int8_t withinSlice, int8_t mergeMode, GPUReconstruction::krnlDeviceType deviceType);
   void RunTPCTrackingMerger_Resolve(int8_t useOrigTrackParam, int8_t mergeAll, GPUReconstruction::krnlDeviceType deviceType);
   void RunTPCClusterFilter(o2::tpc::ClusterNativeAccess* clusters, std::function<o2::tpc::ClusterNative*(size_t)> allocator, bool applyClusterCuts);
+  bool NeedTPCClustersOnGPU();
 
   std::atomic_flag mLockAtomicOutputBuffer = ATOMIC_FLAG_INIT;
   std::mutex mMutexUpdateCalib;

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -629,7 +629,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
 
   auto* digitsMC = propagateMCLabels ? processors()->ioPtrs.tpcPackedDigits->tpcDigitsMC : nullptr;
 
-  bool buildNativeGPU = (mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCConversion) || (mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCSliceTracking) || (mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCMerging) || (mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCCompression);
+  bool buildNativeGPU = doGPU && NeedTPCClustersOnGPU();
   bool buildNativeHost = (mRec->GetRecoStepsOutputs() & GPUDataTypes::InOutType::TPCClusters) || GetProcessingSettings().deterministicGPUReconstruction; // TODO: Should do this also when clusters are needed for later steps on the host but not requested as output
 
   mInputsHost->mNClusterNative = mInputsShadow->mNClusterNative = mRec->MemoryScalers()->nTPCHits * tpcHitLowOccupancyScalingFactor;

GPU/GPUTracking/Global/GPUChainTrackingTransformation.cxx

@@ -32,6 +32,11 @@
 using namespace GPUCA_NAMESPACE::gpu;
 using namespace o2::tpc;
 
+bool GPUChainTracking::NeedTPCClustersOnGPU()
+{
+  return (mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCConversion) || (mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCSliceTracking) || (mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCMerging) || (mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCCompression);
+}
+
 int32_t GPUChainTracking::ConvertNativeToClusterData()
 {
 #ifdef GPUCA_HAVE_O2HEADERS
@@ -42,19 +47,17 @@ int32_t GPUChainTracking::ConvertNativeToClusterData()
   GPUTPCConvert& convertShadow = doGPU ? processorsShadow()->tpcConverter : convert;
 
   bool transferClusters = false;
-  if (doGPU) {
-    if (!(mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCClusterFinding)) {
-      mInputsHost->mNClusterNative = mInputsShadow->mNClusterNative = mIOPtrs.clustersNative->nClustersTotal;
-      AllocateRegisteredMemory(mInputsHost->mResourceClusterNativeBuffer);
-      processorsShadow()->ioPtrs.clustersNative = mInputsShadow->mPclusterNativeAccess;
-      WriteToConstantMemory(RecoStep::TPCConversion, (char*)&processors()->ioPtrs - (char*)processors(), &processorsShadow()->ioPtrs, sizeof(processorsShadow()->ioPtrs), 0);
-      *mInputsHost->mPclusterNativeAccess = *mIOPtrs.clustersNative;
-      mInputsHost->mPclusterNativeAccess->clustersLinear = mInputsShadow->mPclusterNativeBuffer;
-      mInputsHost->mPclusterNativeAccess->setOffsetPtrs();
-      GPUMemCpy(RecoStep::TPCConversion, mInputsShadow->mPclusterNativeBuffer, mIOPtrs.clustersNative->clustersLinear, sizeof(mIOPtrs.clustersNative->clustersLinear[0]) * mIOPtrs.clustersNative->nClustersTotal, 0, true);
-      TransferMemoryResourceLinkToGPU(RecoStep::TPCConversion, mInputsHost->mResourceClusterNativeAccess, 0);
-      transferClusters = true;
-    }
+  if (mRec->IsGPU() && !(mRec->GetRecoStepsGPU() & GPUDataTypes::RecoStep::TPCClusterFinding) && NeedTPCClustersOnGPU()) {
+    mInputsHost->mNClusterNative = mInputsShadow->mNClusterNative = mIOPtrs.clustersNative->nClustersTotal;
+    AllocateRegisteredMemory(mInputsHost->mResourceClusterNativeBuffer);
+    processorsShadow()->ioPtrs.clustersNative = mInputsShadow->mPclusterNativeAccess;
+    WriteToConstantMemory(RecoStep::TPCConversion, (char*)&processors()->ioPtrs - (char*)processors(), &processorsShadow()->ioPtrs, sizeof(processorsShadow()->ioPtrs), 0);
+    *mInputsHost->mPclusterNativeAccess = *mIOPtrs.clustersNative;
+    mInputsHost->mPclusterNativeAccess->clustersLinear = mInputsShadow->mPclusterNativeBuffer;
+    mInputsHost->mPclusterNativeAccess->setOffsetPtrs();
+    GPUMemCpy(RecoStep::TPCConversion, mInputsShadow->mPclusterNativeBuffer, mIOPtrs.clustersNative->clustersLinear, sizeof(mIOPtrs.clustersNative->clustersLinear[0]) * mIOPtrs.clustersNative->nClustersTotal, 0, true);
+    TransferMemoryResourceLinkToGPU(RecoStep::TPCConversion, mInputsHost->mResourceClusterNativeAccess, 0);
+    transferClusters = true;
   }
   if (!param().par.earlyTpcTransform) {
     if (GetProcessingSettings().debugLevel >= 3) {