Fixes

Author: fweig

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -1375,7 +1375,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
           if (doGPU && propagateMCLabels) {
             TransferMemoryResourceLinkToHost(RecoStep::TPCClusterFinding, clusterer.mScratchId, lane);
             SynchronizeStream(lane);
-            runKernel<GPUTPCCFClusterizer>({GetGrid(GPUTPCGeometry::NROWS, lane, GPUReconstruction::krnlDeviceType::CPU), {iSector}}, 1); // Computes MC labels
+            runKernel<GPUTPCCFHIPClusterizer>({GetGrid(GPUTPCGeometry::NROWS, lane, GPUReconstruction::krnlDeviceType::CPU), {iSector}}, 1); // Computes MC labels
           }
         }
 

GPU/GPUTracking/TPCClusterFinder/GPUTPCCFCheckPadBaseline.cxx

@@ -550,7 +550,7 @@ GPUd() void GPUTPCCFHIPClusterizer::Thread<0>(int32_t nBlocks, int32_t nThreads,
       tailStart = CAMath::Min<uint32_t>(tailStart, tail->tailStart);
       tailEnd = CAMath::Max<uint32_t>(tailEnd, tail->tailEnd);
 
-      CPU_ONLY(labelAcc.collectTail(clusterer, row, tail->pad, tail->tailStart, tail->tailEnd));
+      CPU_ONLY(labelAcc.collectTail(row, tail->pad, tail->tailStart, tail->tailEnd));
     }
 
     const float weightSum = CAMath::Max(qTot, 1.f);
@@ -569,7 +569,7 @@ GPUd() void GPUTPCCFHIPClusterizer::Thread<0>(int32_t nBlocks, int32_t nThreads,
 
     if (cn.qMax >= 1023) {
 
-      uint32_t index = 0;
+      uint32_t index;
 
       if (!onlyMC) {
         // Cut off clusters where the tail connection failed for some reason

GPU/GPUTracking/TPCClusterFinder/GPUTPCCFClusterizer.inc

@@ -39,7 +39,7 @@ GPUdii() void GPUTPCCFClusterizer::computeClustersImpl(int32_t nBlocks, int32_t
   Charge charge = chargeMap[pos].unpack();
 
   ClusterAccumulator pc;
-  CPU_ONLY(labelAcc->collect(pos, charge));
+  CPU_ONLY(labelAcc->collect(pos));
 
   buildCluster(
     calib,
@@ -112,8 +112,7 @@ GPUdii() void GPUTPCCFClusterizer::updateClusterInner(
     PackedCharge p = buf[N * lid + i];
 
     Charge q = cluster->updateInner(p, d);
-
-    CPU_ONLY(labelAcc->collect(pos.delta(d), q));
+    CPU_ONLY(labelAcc->collect(pos.delta(d)));
 
     aboveThreshold |= (uint8_t(q > calib.tpc.cfInnerThreshold) << i);
   }
@@ -139,10 +138,8 @@ GPUdii() void GPUTPCCFClusterizer::updateClusterOuter(
 
     Delta2 d = cfconsts::OuterNeighbors[i];
 
-    Charge q = cluster->updateOuter(p, d);
-    static_cast<void>(q); // Avoid unused varible warning on GPU.
-
-    CPU_ONLY(labelAcc->collect(pos.delta(d), q));
+    cluster->updateOuter(p, d);
+    CPU_ONLY(labelAcc->collect(pos.delta(d)));
   }
 }
 

GPU/GPUTracking/TPCClusterFinder/GPUTPCClusterFinder.cxx

@@ -172,7 +172,9 @@ void GPUTPCClusterFinder::PrepareMC()
   assert(mNMaxClusterPerRow > 0);
 
   clearMCMemory();
-  mPindexMap = new uint32_t[TPCMapMemoryLayout<decltype(*mPindexMap)>::items(mRec->GetProcessingSettings().overrideClusterizerFragmentLen)];
+  size_t nItems = TPCMapMemoryLayout<decltype(*mPindexMap)>::items(mRec->GetProcessingSettings().overrideClusterizerFragmentLen);
+  mPindexMap = new uint32_t[nItems];
+  std::fill_n(mPindexMap, nItems, uint32_t(-1)); // TODO check performance of this call!
   mPlabelsByRow = new GPUTPCClusterMCInterimArray[GPUTPCGeometry::NROWS];
   mPlabelsInRow = new uint32_t[GPUTPCGeometry::NROWS];
 }

GPU/GPUTracking/TPCClusterFinder/MCLabelAccumulator.cxx

@@ -16,7 +16,6 @@
 
 #include "GPUHostDataTypes.h"
 #include "GPUTPCClusterFinder.h"
-#include "PackedCharge.h"
 #include "SimulationDataFormat/MCTruthContainer.h"
 
 using namespace o2::gpu;
@@ -29,13 +28,18 @@ MCLabelAccumulator::MCLabelAccumulator(GPUTPCClusterFinder& clusterer)
 
 MCLabelAccumulator::~MCLabelAccumulator() {}
 
-void MCLabelAccumulator::collect(const CfChargePos& pos, Charge q)
+void MCLabelAccumulator::collect(const CfChargePos& pos)
 {
-  if (q == 0 || !engaged()) {
+  if (!engaged()) {
     return;
   }
 
+  // Use -1 as sentinel to indicate a missing label.
+  // Can't use zero charge, as HIP filter will zero existing digits.
   uint32_t index = mIndexMap[pos];
+  if (index == uint32_t(-1)) {
+    return;
+  }
 
   const auto& labels = mLabels->getLabels(index);
 
@@ -54,16 +58,13 @@ void MCLabelAccumulator::collect(const CfChargePos& pos, Charge q)
   }
 }
 
-void MCLabelAccumulator::collectTail(GPUTPCClusterFinder& clusterer, tpccf::Row row, tpccf::Pad pad, uint16_t tailStart, uint16_t tailEnd)
+void MCLabelAccumulator::collectTail(tpccf::Row row, tpccf::Pad pad, uint16_t tailStart, uint16_t tailEnd)
 {
-  CfArray2D<const PackedCharge> chargeMap(reinterpret_cast<PackedCharge*>(clusterer.mPchargeMap));
-
   const auto basePos = CfChargePos{row, pad, 0};
 
   for (uint16_t t = tailStart; t < tailEnd; t++) {
     const auto pos = basePos.delta({0, (int16_t)t});
-    const auto q = chargeMap[pos].unpack();
-    collect(pos, q);
+    collect(pos);
   }
 }
 

GPU/GPUTracking/TPCClusterFinder/MCLabelAccumulator.h

@@ -45,9 +45,9 @@ class MCLabelAccumulator
   MCLabelAccumulator(GPUTPCClusterFinder&);
   ~MCLabelAccumulator(); // Explicit destructor to allow forward declaring MCCompLabel with std::vector
 
-  void collect(const CfChargePos& pos, tpccf::Charge q);
+  void collect(const CfChargePos& pos);
 
-  void collectTail(GPUTPCClusterFinder& clusterer, tpccf::Row row, tpccf::Pad pad, uint16_t tailStart, uint16_t tailEnd);
+  void collectTail(tpccf::Row row, tpccf::Pad pad, uint16_t tailStart, uint16_t tailEnd);
 
   bool engaged() const { return mLabels != nullptr && mOutput != nullptr; }