TEST

Signed-off-by: Felix Schlepper <felix.schlepper@cern.ch>

Author: f3sch

Detectors/ITSMFT/ITS/tracking/include/ITStracking/TrackerTraits.h

@@ -54,7 +54,7 @@ class TrackerTraits
   virtual void findRoads(const int iteration);
 
   template <typename InputSeed>
-  void processNeighbours(int iteration, int iLayer, int iLevel, const bounded_vector<InputSeed>& currentCellSeed, const bounded_vector<int>& currentCellId, bounded_vector<TrackSeedN>& updatedCellSeed, bounded_vector<int>& updatedCellId);
+  void processNeighbours(int iteration, int iLayer, int iLevel, int startLevel, const bounded_vector<InputSeed>& currentCellSeed, const bounded_vector<int>& currentCellId, bounded_vector<TrackSeedN>& updatedCellSeed, bounded_vector<int>& updatedCellId, bounded_vector<TrackSeedN>* terminalOut = nullptr);
 
   void acceptTracks(int iteration, bounded_vector<TrackITSExt>& tracks, bounded_vector<bounded_vector<int>>& firstClusters, bounded_vector<bounded_vector<int>>& sharedFirstClusters);
   void markTracks(int iteration, bounded_vector<bounded_vector<int>>& sharedFirstClusters);

Detectors/ITSMFT/ITS/tracking/src/TrackerTraits.cxx

@@ -538,9 +538,15 @@ void TrackerTraits<NLayers>::findCellsNeighbours(const int iteration)
 
 template <int NLayers>
 template <typename InputSeed>
-void TrackerTraits<NLayers>::processNeighbours(int iteration, int iLayer, int iLevel, const bounded_vector<InputSeed>& currentCellSeed, const bounded_vector<int>& currentCellId, bounded_vector<TrackSeedN>& updatedCellSeeds, bounded_vector<int>& updatedCellsIds)
+void TrackerTraits<NLayers>::processNeighbours(int iteration, int iLayer, int iLevel, int startLevel, const bounded_vector<InputSeed>& currentCellSeed, const bounded_vector<int>& currentCellId, bounded_vector<TrackSeedN>& updatedCellSeeds, bounded_vector<int>& updatedCellsIds, bounded_vector<TrackSeedN>* terminalOut)
 {
   auto propagator = o2::base::Propagator::Instance();
+  const bool terminal = terminalOut != nullptr;
+  const float chi2Max = mTrkParams[iteration].MaxChi2NDF * ((startLevel + 2) * 2 - 5);
+  const auto seedFilter = [chi2Max](const TrackSeedN& seed) {
+    return seed.getQ2Pt() <= 1.e3f && seed.getChi2() <= chi2Max;
+  };
+  const size_t terminalBaseOffset = terminal ? terminalOut->size() : 0;
 
   mTaskArena->execute([&] {
     auto forCellNeighbours = [&](auto Tag, int iCell, int offset = 0) -> int {
@@ -613,14 +619,26 @@ void TrackerTraits<NLayers>::processNeighbours(int iteration, int iLayer, int iL
           seed.setSecondTrackletIndex(neighbourCell.getSecondTrackletIndex());
         }
 
+        if (terminal && !seedFilter(seed)) {
+          continue;
+        }
+
         if constexpr (decltype(Tag)::value == PassMode::OnePass::value) {
-          updatedCellSeeds.push_back(seed);
-          updatedCellsIds.push_back(neighbourCellId);
+          if (terminal) {
+            terminalOut->push_back(seed);
+          } else {
+            updatedCellSeeds.push_back(seed);
+            updatedCellsIds.push_back(neighbourCellId);
+          }
         } else if constexpr (decltype(Tag)::value == PassMode::TwoPassCount::value) {
           ++foundSeeds;
         } else if constexpr (decltype(Tag)::value == PassMode::TwoPassInsert::value) {
-          updatedCellSeeds[offset] = seed;
-          updatedCellsIds[offset++] = neighbourCellId;
+          if (terminal) {
+            (*terminalOut)[terminalBaseOffset + offset++] = seed;
+          } else {
+            updatedCellSeeds[offset] = seed;
+            updatedCellsIds[offset++] = neighbourCellId;
+          }
         } else {
           static_assert(false, "Unknown mode!");
         }
@@ -644,8 +662,12 @@ void TrackerTraits<NLayers>::processNeighbours(int iteration, int iLayer, int iL
       if (totalNeighbours == 0) {
         return;
       }
-      updatedCellSeeds.resize(totalNeighbours);
-      updatedCellsIds.resize(totalNeighbours);
+      if (terminal) {
+        terminalOut->resize(terminalBaseOffset + totalNeighbours);
+      } else {
+        updatedCellSeeds.resize(totalNeighbours);
+        updatedCellsIds.resize(totalNeighbours);
+      }
 
       tbb::parallel_for(0, nCells, [&](const int iCell) {
         int offset = perCellCount[iCell];
@@ -674,10 +696,6 @@ void TrackerTraits<NLayers>::findRoads(const int iteration)
   }
   for (int startLevel{mTrkParams[iteration].CellsPerRoad()}; startLevel >= mTrkParams[iteration].CellMinimumLevel(); --startLevel) {
 
-    auto seedFilter = [&](const auto& seed) {
-      return seed.getQ2Pt() <= 1.e3 && seed.getChi2() <= mTrkParams[iteration].MaxChi2NDF * ((startLevel + 2) * 2 - 5);
-    };
-
     bounded_vector<TrackSeedN> trackSeeds(mMemoryPool.get());
     for (int startLayer{mTrkParams[iteration].NeighboursPerRoad()}; startLayer >= startLevel - 1; --startLayer) {
       if ((mTrkParams[iteration].StartLayerMask & (1 << (startLayer + 2))) == 0) {
@@ -687,23 +705,32 @@ void TrackerTraits<NLayers>::findRoads(const int iteration)
       bounded_vector<int> lastCellId(mMemoryPool.get()), updatedCellId(mMemoryPool.get());
       bounded_vector<TrackSeedN> lastCellSeed(mMemoryPool.get()), updatedCellSeed(mMemoryPool.get());
 
-      processNeighbours(iteration, startLayer, startLevel, mTimeFrame->getCells()[startLayer], lastCellId, updatedCellSeed, updatedCellId);
+      /// A call is terminal when the inner loop will not execute another iteration
+      /// after it. In that case, route the filtered seeds directly into trackSeeds
+      /// so we never materialize the full unfiltered updatedCellSeed buffer.
+      const bool firstCallTerminal = (startLayer - 1) <= 0 || startLevel <= 2;
+      processNeighbours(iteration, startLayer, startLevel, startLevel,
+                        mTimeFrame->getCells()[startLayer], lastCellId,
+                        updatedCellSeed, updatedCellId,
+                        firstCallTerminal ? &trackSeeds : nullptr);
 
       int level = startLevel;
       for (int iLayer{startLayer - 1}; iLayer > 0 && level > 2; --iLayer) {
         lastCellSeed.swap(updatedCellSeed);
         lastCellId.swap(updatedCellId);
         deepVectorClear(updatedCellSeed); /// tame the memory peaks
         deepVectorClear(updatedCellId);   /// tame the memory peaks
-        processNeighbours(iteration, iLayer, --level, lastCellSeed, lastCellId, updatedCellSeed, updatedCellId);
-      }
-      deepVectorClear(lastCellId);   /// tame the memory peaks
-      deepVectorClear(lastCellSeed); /// tame the memory peaks
-
-      if (!updatedCellSeed.empty()) {
-        trackSeeds.reserve(trackSeeds.size() + std::count_if(updatedCellSeed.begin(), updatedCellSeed.end(), seedFilter));
-        std::copy_if(updatedCellSeed.begin(), updatedCellSeed.end(), std::back_inserter(trackSeeds), seedFilter);
-      }
+        --level;
+        const bool thisCallTerminal = (iLayer - 1) <= 0 || level <= 2;
+        processNeighbours(iteration, iLayer, level, startLevel,
+                          lastCellSeed, lastCellId,
+                          updatedCellSeed, updatedCellId,
+                          thisCallTerminal ? &trackSeeds : nullptr);
+      }
+      deepVectorClear(lastCellId);      /// tame the memory peaks
+      deepVectorClear(lastCellSeed);    /// tame the memory peaks
+      deepVectorClear(updatedCellId);   /// terminal call routed to trackSeeds; drop id buffer
+      deepVectorClear(updatedCellSeed); /// terminal call routed to trackSeeds; drop seed buffer
     }
 
     if (trackSeeds.empty()) {
@@ -952,13 +979,13 @@ void TrackerTraits<NLayers>::setNThreads(int n, std::shared_ptr<tbb::task_arena>
 }
 
 template class TrackerTraits<7>;
-template void TrackerTraits<7>::processNeighbours<CellSeed>(int, int, int, const bounded_vector<CellSeed>&, const bounded_vector<int>&, bounded_vector<TrackSeed<7>>&, bounded_vector<int>&);
-template void TrackerTraits<7>::processNeighbours<TrackSeed<7>>(int, int, int, const bounded_vector<TrackSeed<7>>&, const bounded_vector<int>&, bounded_vector<TrackSeed<7>>&, bounded_vector<int>&);
+template void TrackerTraits<7>::processNeighbours<CellSeed>(int, int, int, int, const bounded_vector<CellSeed>&, const bounded_vector<int>&, bounded_vector<TrackSeed<7>>&, bounded_vector<int>&, bounded_vector<TrackSeed<7>>*);
+template void TrackerTraits<7>::processNeighbours<TrackSeed<7>>(int, int, int, int, const bounded_vector<TrackSeed<7>>&, const bounded_vector<int>&, bounded_vector<TrackSeed<7>>&, bounded_vector<int>&, bounded_vector<TrackSeed<7>>*);
 // ALICE3 upgrade
 #ifdef ENABLE_UPGRADES
 template class TrackerTraits<11>;
-template void TrackerTraits<11>::processNeighbours<CellSeed>(int, int, int, const bounded_vector<CellSeed>&, const bounded_vector<int>&, bounded_vector<TrackSeed<11>>&, bounded_vector<int>&);
-template void TrackerTraits<11>::processNeighbours<TrackSeed<11>>(int, int, int, const bounded_vector<TrackSeed<11>>&, const bounded_vector<int>&, bounded_vector<TrackSeed<11>>&, bounded_vector<int>&);
+template void TrackerTraits<11>::processNeighbours<CellSeed>(int, int, int, int, const bounded_vector<CellSeed>&, const bounded_vector<int>&, bounded_vector<TrackSeed<11>>&, bounded_vector<int>&, bounded_vector<TrackSeed<11>>*);
+template void TrackerTraits<11>::processNeighbours<TrackSeed<11>>(int, int, int, int, const bounded_vector<TrackSeed<11>>&, const bounded_vector<int>&, bounded_vector<TrackSeed<11>>&, bounded_vector<int>&, bounded_vector<TrackSeed<11>>*);
 #endif
 
 } // namespace o2::its