More fixes

Author: fweig

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -1059,8 +1059,9 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
           return;
         }
 
-        if (propagateMCLabels && fragment.index == 0) {
-          clusterer.PrepareMC();
+        if (propagateMCLabels) {
+          // AllocMCBuffers must be called for every fragment, to clear leftovers from previous fragment
+          clusterer.AllocMCBuffers();
           clusterer.mPinputLabels = digitsMC->v[iSector];
           if (clusterer.mPinputLabels == nullptr) {
             GPUFatal("MC label container missing, sector %d", iSector);
@@ -1455,12 +1456,12 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
       runKernel<GPUTPCCFMCLabelFlattener, GPUTPCCFMCLabelFlattener::setRowOffsets>({GetGrid(GPUTPCGeometry::NROWS, lane, GPUReconstruction::krnlDeviceType::CPU), {iSector}});
       GPUTPCCFMCLabelFlattener::setGlobalOffsetsAndAllocate(clusterer, mcLinearLabels);
       runKernel<GPUTPCCFMCLabelFlattener, GPUTPCCFMCLabelFlattener::flatten>({GetGrid(GPUTPCGeometry::NROWS, lane, GPUReconstruction::krnlDeviceType::CPU), {iSector}}, &mcLinearLabels);
-      clusterer.clearMCMemory();
+      clusterer.FreeMCBuffers();
       assert(propagateMCLabels ? mcLinearLabels.header.size() == nClsTotal : true);
     }
     if (propagateMCLabels) {
       for (int32_t lane = 0; lane < maxLane; lane++) {
-        processors()->tpcClusterer[iSectorBase + lane].clearMCMemory();
+        processors()->tpcClusterer[iSectorBase + lane].FreeMCBuffers();
       }
     }
     if (buildNativeHost && buildNativeGPU && anyLaneHasData) {

GPU/GPUTracking/TPCClusterFinder/GPUTPCCFCheckPadBaseline.cxx

@@ -518,7 +518,9 @@ GPUd() void GPUTPCCFHIPClusterizer::Thread<0>(int32_t nBlocks, int32_t nThreads,
   const auto* tails = GetHIPTails(clusterer, row);
   const auto& fragment = clusterer.mPmemory->fragment;
 
-  auto* clusterPosInRow = clusterer.mPhipClusterPosInRow + row * MaxHIPTailsPerRow;
+  auto* clusterPosInRow = clusterer.mPhipClusterPosInRow
+                            ? clusterer.mPhipClusterPosInRow + row * MaxHIPTailsPerRow
+                            : nullptr;
 
   for (uint32_t iTail = iThread + 1; iTail <= nTails; iTail += nThreads) {
 

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));
+  CPU_ONLY(labelAcc->collect(pos, charge));
 
   buildCluster(
     calib,
@@ -112,7 +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)));
+    CPU_ONLY(labelAcc->collect(pos.delta(d), q));
 
     aboveThreshold |= (uint8_t(q > calib.tpc.cfInnerThreshold) << i);
   }
@@ -138,8 +138,8 @@ GPUdii() void GPUTPCCFClusterizer::updateClusterOuter(
 
     Delta2 d = cfconsts::OuterNeighbors[i];
 
-    cluster->updateOuter(p, d);
-    CPU_ONLY(labelAcc->collect(pos.delta(d)));
+    [[maybe_unused]] Charge q = cluster->updateOuter(p, d);
+    CPU_ONLY(labelAcc->collect(pos.delta(d), q));
   }
 }
 

GPU/GPUTracking/TPCClusterFinder/GPUTPCClusterFinder.cxx

@@ -27,6 +27,8 @@
 #include "CfArray2D.h"
 #include "GPUTPCCFCheckPadBaseline.h"
 
+#include <algorithm>
+
 using namespace o2::gpu;
 using namespace o2::tpc;
 
@@ -38,7 +40,7 @@ void GPUTPCClusterFinder::InitializeProcessor()
 GPUTPCClusterFinder::~GPUTPCClusterFinder()
 {
   delete[] mMinMaxCN;
-  clearMCMemory();
+  FreeMCBuffers();
 }
 
 void* GPUTPCClusterFinder::SetPointersMemory(void* mem)
@@ -167,19 +169,19 @@ uint32_t GPUTPCClusterFinder::getNSteps(size_t items) const
   return c;
 }
 
-void GPUTPCClusterFinder::PrepareMC()
+void GPUTPCClusterFinder::AllocMCBuffers()
 {
   assert(mNMaxClusterPerRow > 0);
 
-  clearMCMemory();
+  FreeMCBuffers();
   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];
 }
 
-void GPUTPCClusterFinder::clearMCMemory()
+void GPUTPCClusterFinder::FreeMCBuffers()
 {
   delete[] mPindexMap;
   mPindexMap = nullptr;

GPU/GPUTracking/TPCClusterFinder/GPUTPCClusterFinder.h

@@ -92,8 +92,8 @@ class GPUTPCClusterFinder : public GPUProcessor
   uint32_t getNSteps(size_t items) const;
   void SetNMaxDigits(size_t nDigits, size_t nPages, size_t nDigitsFragment, size_t nDigitsEndpointMax);
 
-  void PrepareMC();
-  void clearMCMemory();
+  void AllocMCBuffers();
+  void FreeMCBuffers();
 #endif
   uint8_t* mPzs = nullptr;
   ZSOffset* mPzsOffsets = nullptr;

GPU/GPUTracking/TPCClusterFinder/MCLabelAccumulator.cxx

@@ -28,9 +28,9 @@ MCLabelAccumulator::MCLabelAccumulator(GPUTPCClusterFinder& clusterer)
 
 MCLabelAccumulator::~MCLabelAccumulator() {}
 
-void MCLabelAccumulator::collect(const CfChargePos& pos)
+void MCLabelAccumulator::collect(const CfChargePos& pos, float q)
 {
-  if (!engaged()) {
+  if (q == 0 || !engaged()) {
     return;
   }
 
@@ -60,11 +60,17 @@ void MCLabelAccumulator::collect(const CfChargePos& pos)
 
 void MCLabelAccumulator::collectTail(tpccf::Row row, tpccf::Pad pad, uint16_t tailStart, uint16_t tailEnd)
 {
+  if (!engaged()) {
+    return;
+  }
+
   const auto basePos = CfChargePos{row, pad, 0};
 
   for (uint16_t t = tailStart; t < tailEnd; t++) {
     const auto pos = basePos.delta({0, (int16_t)t});
-    collect(pos);
+    // Charge passed to collect() doesn't matter, collect() skips zero charges
+    // But we know there's an interesting value, but it was zeroed in chargeMap by tail filter
+    collect(pos, 1023.f);
   }
 }
 

GPU/GPUTracking/TPCClusterFinder/MCLabelAccumulator.h

@@ -45,7 +45,7 @@ class MCLabelAccumulator
   MCLabelAccumulator(GPUTPCClusterFinder&);
   ~MCLabelAccumulator(); // Explicit destructor to allow forward declaring MCCompLabel with std::vector
 
-  void collect(const CfChargePos& pos);
+  void collect(const CfChargePos& pos, float q);
 
   void collectTail(tpccf::Row row, tpccf::Pad pad, uint16_t tailStart, uint16_t tailEnd);