GPU TPC: Remove option to retry refit if cluster rejection breaks the track, we will anyway rebuild the track in the future

Author: davidrohr

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -155,7 +155,6 @@ AddOptionRTC(mergerInterpolateErrors, uint8_t, 1, "", 0, "Use interpolation inst
 AddOptionRTC(mergerInterpolateRejectAlsoOnCurrentPosition, uint8_t, 1, "", 0, "When using mergerInterpolateErrors, reject based on chi2 twice computed with interpolated and current track position starting from NDF > mergerNonInterpolateRejectMinNDF")
 AddOptionRTC(mergerNonInterpolateRejectMinNDF, uint8_t, 5, "", 0, "Minimum NDF of track for non-interpolated reject (both for chi2 and absolute distance)")
 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(enablePID, int8_t, 1, "", 0, "Enable PID response")
 AddOptionRTC(PID_useNsigma, int8_t, 1, "", 0, "Use nSigma instead of absolute distance in PID response")
 AddOptionRTC(adddEdxSubThresholdClusters, int8_t, 1, "", 0, "Add sub threshold clusters in TPC dEdx computation")

GPU/GPUTracking/Global/GPUChainTrackingMerger.cxx

@@ -227,9 +227,6 @@ int32_t GPUChainTracking::RunTPCTrackingMerger(bool synchronizeOutput)
   }
 
   runKernel<GPUTPCGMMergerTrackFit>(doGPU ? GetGrid(Merger.NMergedTracks(), 0) : GetGridAuto(0), mergerSortTracks ? 1 : 0);
-  if (param().rec.tpc.retryRefit == 1) {
-    runKernel<GPUTPCGMMergerTrackFit>(GetGridAuto(0), -1);
-  }
   runKernel<GPUTPCGMMergerFollowLoopers>(GetGridAuto(0));
 
   DoDebugAndDump(RecoStep::TPCMerging, GPUChainTrackingDebugFlags::TPCMergingRefit, Merger, &GPUTPCGMMerger::DumpRefit, *mDebugFile);

GPU/GPUTracking/Merger/GPUTPCGMMerger.cxx

@@ -402,8 +402,7 @@ void* GPUTPCGMMerger::SetPointersMerger(void* mem)
   computePointerWithAlignment(mem, mSharedCount, mNMaxClusters);
   memMax = (void*)std::max((size_t)mem, (size_t)memMax);
   mem = memBase;
-  computePointerWithAlignment(mem, mLoopData, mNMaxTracks);      // GPUTPCGMMergerTrackFit - GPUTPCGMMergerFollowLoopers
-  computePointerWithAlignment(mem, mRetryRefitIds, mNMaxTracks); // Reducing mNMaxTracks for mLoopData / mRetryRefitIds does not save memory, since the other parts are larger anyway
+  computePointerWithAlignment(mem, mLoopData, mNMaxTracks);      // GPUTPCGMMergerTrackFit - GPUTPCGMMergerFollowLoopers, Reducing mNMaxTracks for mLoopData does not save memory, other parts are larger anyway
   memMax = (void*)std::max((size_t)mem, (size_t)memMax);
   mem = memBase;
   computePointerWithAlignment(mem, mLooperCandidates, mNMaxLooperMatches); // MergeLoopers 1-3

GPU/GPUTracking/Merger/GPUTPCGMMerger.h

@@ -66,7 +66,6 @@ class GPUTPCGMMerger : public GPUProcessor
   static constexpr const int32_t NSECTORS = GPUCA_NSECTORS; //* N sectors
 
   struct memory {
-    GPUAtomic(uint32_t) nRetryRefit;
     GPUAtomic(uint32_t) nLoopData;
     GPUAtomic(uint32_t) nUnpackedTracks;
     GPUAtomic(uint32_t) nMergedTracks;
@@ -119,7 +118,6 @@ class GPUTPCGMMerger : public GPUProcessor
   GPUhdi() GPUAtomic(uint32_t) * ClusterAttachment() const { return mClusterAttachment; }
   GPUhdi() uint32_t* TrackOrderAttach() const { return mTrackOrderAttach; }
   GPUhdi() uint32_t* TrackOrderProcess() const { return mTrackOrderProcess; }
-  GPUhdi() uint32_t* RetryRefitIds() const { return mRetryRefitIds; }
   GPUhdi() uint8_t* ClusterStateExt() const { return mClusterStateExt; }
   GPUhdi() GPUTPCGMLoopData* LoopData() const { return mLoopData; }
   GPUhdi() memory* Memory() const { return mMemory; }
@@ -287,7 +285,6 @@ class GPUTPCGMMerger : public GPUProcessor
   gputpcgmmergertypes::GPUTPCGMBorderRange* mBorderRangeMemory = nullptr; // memory for border tracks
   gputpcgmmergertypes::GPUTPCGMBorderRange* mBorderRange[NSECTORS];       // memory for border tracks
   memory* mMemory = nullptr;
-  uint32_t* mRetryRefitIds = nullptr;
   GPUTPCGMLoopData* mLoopData = nullptr;
 };
 } // namespace o2::gpu

GPU/GPUTracking/Merger/GPUTPCGMMergerGPU.cxx

@@ -21,10 +21,9 @@ using namespace o2::gpu;
 template <>
 GPUdii() void GPUTPCGMMergerTrackFit::Thread<0>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUsharedref() GPUSharedMemory& smem, processorType& GPUrestrict() merger, int32_t mode)
 {
-  const int32_t iEnd = mode == -1 ? merger.Memory()->nRetryRefit : merger.NMergedTracks();
-  GPUCA_TBB_KERNEL_LOOP(merger.GetRec(), int32_t, ii, iEnd, {
-    const int32_t i = mode == -1 ? merger.RetryRefitIds()[ii] : mode ? merger.TrackOrderProcess()[ii] : ii;
-    GPUTPCGMTrackParam::RefitTrack(merger.MergedTracks()[i], i, &merger, mode == -1);
+  GPUCA_TBB_KERNEL_LOOP(merger.GetRec(), int32_t, ii, merger.NMergedTracks(), {
+    const int32_t i = mode ? merger.TrackOrderProcess()[ii] : ii;
+    GPUTPCGMTrackParam::RefitTrack(merger.MergedTracks()[i], i, &merger);
   });
 }
 

GPU/GPUTracking/Merger/GPUTPCGMTrackParam.cxx

@@ -48,7 +48,7 @@
 using namespace o2::gpu;
 using namespace o2::tpc;
 
-GPUd() bool GPUTPCGMTrackParam::Fit(GPUTPCGMMerger* GPUrestrict() merger, int32_t iTrk, GPUTPCGMMergedTrackHit* GPUrestrict() clusters, int32_t& GPUrestrict() N, int32_t& GPUrestrict() NTolerated, float& GPUrestrict() Alpha, int32_t attempt, float maxSinPhi, GPUTPCGMMergedTrack& GPUrestrict() track)
+GPUd() bool GPUTPCGMTrackParam::Fit(GPUTPCGMMerger* GPUrestrict() merger, int32_t iTrk, GPUTPCGMMergedTrackHit* GPUrestrict() clusters, int32_t& GPUrestrict() N, int32_t& GPUrestrict() NTolerated, float& GPUrestrict() Alpha, float maxSinPhi, GPUTPCGMMergedTrack& GPUrestrict() track)
 {
   static constexpr float kDeg2Rad = M_PI / 180.f;
   CADEBUG(static constexpr float kSectAngle = 2 * M_PI / 18.f);
@@ -92,7 +92,7 @@ GPUd() bool GPUTPCGMTrackParam::Fit(GPUTPCGMMerger* GPUrestrict() merger, int32_
     const float maxSinForUpdate = CAMath::Sin(70.f * kDeg2Rad);
 
     ResetCovariance();
-    prop.SetSeedingErrors(!(refit && attempt == 0));
+    prop.SetSeedingErrors(!(refit));
     prop.SetFitInProjections(true); // param.rec.fitInProjections == -1 ? (iWay == 0) : param.rec.fitInProjections); // TODO: Reenable once fixed
     prop.SetPropagateBzOnly(param.rec.fitPropagateBzOnly == -1 ? !finalFit : param.rec.fitPropagateBzOnly);
     prop.SetMatLUT((param.rec.useMatLUT && finalFit) ? merger->GetConstantMem()->calibObjects.matLUT : nullptr);
@@ -105,7 +105,6 @@ GPUd() bool GPUTPCGMTrackParam::Fit(GPUTPCGMMerger* GPUrestrict() merger, int32_
     const bool inFlyDirection = iWay & 1;
     const int32_t wayDirection = (iWay & 1) ? -1 : 1;
 
-    int32_t goodRows = 0;
     for (int32_t ihit = ihitStart; ihit >= 0 && ihit < maxN; ihit += wayDirection) {
       const bool crossCE = lastSector != 255 && ((lastSector < 18) ^ (clusters[ihit].sector < 18));
       if (crossCE) {
@@ -179,11 +178,6 @@ GPUd() bool GPUTPCGMTrackParam::Fit(GPUTPCGMMerger* GPUrestrict() merger, int32_
           retValProp = prop.PropagateToXAlpha(xx, clAlpha, inFlyDirection);
         }
       }
-      if (lastRow == 255 || CAMath::Abs((int32_t)lastRow - (int32_t)cluster.row) > 5 || lastSector != cluster.sector || (param.rec.tpc.trackFitRejectMode < 0 && -nMissed <= param.rec.tpc.trackFitRejectMode)) {
-        goodRows = 0;
-      } else {
-        goodRows++;
-      }
       if (retValProp == 0) {
         lastRow = cluster.row;
         lastSector = cluster.sector;
@@ -228,17 +222,13 @@ GPUd() bool GPUTPCGMTrackParam::Fit(GPUTPCGMMerger* GPUrestrict() merger, int32_
         retValUpd = GPUTPCGMPropagator::updateErrorClusterRejectedDistance;
       } else {
         int8_t rejectChi2 = 0;
-        if (attempt == 0) {
-          if (param.rec.tpc.mergerInterpolateErrors && CAMath::Abs(ihit - ihitMergeFirst) <= 1) {
-            if (iWay == nWays - 3) {
-              rejectChi2 = GPUTPCGMPropagator::rejectInterFill;
-            } else if (iWay == nWays - 2) {
-              rejectChi2 = GPUTPCGMPropagator::rejectInterReject;
-            } else if (iWay == nWays - 1) {
-              rejectChi2 = (param.rec.tpc.mergerInterpolateRejectAlsoOnCurrentPosition && GetNDF() > (int32_t)param.rec.tpc.mergerNonInterpolateRejectMinNDF) ? GPUTPCGMPropagator::rejectDirect : 0;
-            }
-          } else {
-            rejectChi2 = allowChangeClusters && goodRows > 5;
+        if (param.rec.tpc.mergerInterpolateErrors && CAMath::Abs(ihit - ihitMergeFirst) <= 1) {
+          if (iWay == nWays - 3) {
+            rejectChi2 = GPUTPCGMPropagator::rejectInterFill;
+          } else if (iWay == nWays - 2) {
+            rejectChi2 = GPUTPCGMPropagator::rejectInterReject;
+          } else if (iWay == nWays - 1) {
+            rejectChi2 = (param.rec.tpc.mergerInterpolateRejectAlsoOnCurrentPosition && GetNDF() > (int32_t)param.rec.tpc.mergerNonInterpolateRejectMinNDF) ? GPUTPCGMPropagator::rejectDirect : 0;
           }
         }
 
@@ -905,7 +895,7 @@ GPUd() bool GPUTPCGMTrackParam::CheckNumericalQuality(float overrideCovYY) const
   return ok;
 }
 
-GPUdii() void GPUTPCGMTrackParam::RefitTrack(GPUTPCGMMergedTrack& GPUrestrict() track, int32_t iTrk, GPUTPCGMMerger* GPUrestrict() merger, int32_t attempt) // VS: GPUd changed to GPUdii. No change in output and no performance penalty.
+GPUdii() void GPUTPCGMTrackParam::RefitTrack(GPUTPCGMMergedTrack& GPUrestrict() track, int32_t iTrk, GPUTPCGMMerger* GPUrestrict() merger) // VS: GPUd changed to GPUdii. No change in output and no performance penalty.
 {
   if (!track.OK()) {
     return;
@@ -920,27 +910,10 @@ GPUdii() void GPUTPCGMTrackParam::RefitTrack(GPUTPCGMMergedTrack& GPUrestrict()
   GPUTPCGMTrackParam t = track.Param();
   float Alpha = track.Alpha();
   CADEBUG(int32_t nTrackHitsOld = nTrackHits; float ptOld = t.QPt());
-  bool ok = t.Fit(merger, iTrk, merger->Clusters() + track.FirstClusterRef(), nTrackHits, NTolerated, Alpha, attempt, GPUCA_MAX_SIN_PHI, track);
+  bool ok = t.Fit(merger, iTrk, merger->Clusters() + track.FirstClusterRef(), nTrackHits, NTolerated, Alpha, GPUCA_MAX_SIN_PHI, track);
   CADEBUG(printf("Finished Fit Track %d\n", iTrk));
   CADEBUG(printf("OUTPUT hits %d -> %d+%d = %d, QPt %f -> %f, SP %f, OK %d chi2 %f chi2ndf %f\n", nTrackHitsOld, nTrackHits, NTolerated, nTrackHits + NTolerated, ptOld, t.QPt(), t.SinPhi(), (int32_t)ok, t.Chi2(), t.Chi2() / CAMath::Max(1, nTrackHits)));
 
-  if (!ok && attempt == 0 && merger->Param().rec.tpc.retryRefit) {
-    for (uint32_t i = 0; i < track.NClusters(); i++) {
-      merger->Clusters()[track.FirstClusterRef() + i].state &= GPUTPCGMMergedTrackHit::clustererAndSharedFlags;
-    }
-    CADEBUG(printf("Track rejected, marking for retry\n"));
-    if (merger->Param().rec.tpc.retryRefit == 2) {
-      nTrackHits = track.NClusters();
-      NTolerated = 0; // Clusters not fit but tollerated for track length cut
-      t = track.Param();
-      Alpha = track.Alpha();
-      ok = t.Fit(merger, iTrk, merger->Clusters() + track.FirstClusterRef(), nTrackHits, NTolerated, Alpha, 1, GPUCA_MAX_SIN_PHI, track);
-    } else {
-      uint32_t nRefit = CAMath::AtomicAdd(&merger->Memory()->nRetryRefit, 1u);
-      merger->RetryRefitIds()[nRefit] = iTrk;
-      return;
-    }
-  }
   if (CAMath::Abs(t.QPt()) < 1.e-4f) {
     t.QPt() = 1.e-4f;
   }

GPU/GPUTracking/Merger/GPUTPCGMTrackParam.h

@@ -141,7 +141,7 @@ class GPUTPCGMTrackParam
   GPUd() bool CheckNumericalQuality(float overrideCovYY = -1.f) const;
   GPUd() bool CheckCov() const;
 
-  GPUd() bool Fit(GPUTPCGMMerger* merger, int32_t iTrk, GPUTPCGMMergedTrackHit* clusters, int32_t& N, int32_t& NTolerated, float& Alpha, int32_t attempt, float maxSinPhi, GPUTPCGMMergedTrack& track);
+  GPUd() bool Fit(GPUTPCGMMerger* merger, int32_t iTrk, GPUTPCGMMergedTrackHit* clusters, int32_t& N, int32_t& NTolerated, float& Alpha, float maxSinPhi, GPUTPCGMMergedTrack& track);
   GPUd() void MoveToReference(GPUTPCGMPropagator& prop, const GPUParam& param, float& alpha);
   GPUd() void MirrorTo(GPUTPCGMPropagator& prop, float toY, float toZ, bool inFlyDirection, const GPUParam& param, uint8_t row, uint8_t clusterState, bool mirrorParameters, int8_t sector);
   GPUd() int32_t MergeDoubleRowClusters(int32_t& ihit, int32_t wayDirection, GPUTPCGMMergedTrackHit* clusters, const GPUTPCGMMerger* merger, GPUTPCGMPropagator& prop, float& xx, float& yy, float& zz, int32_t maxN, float clAlpha, uint8_t& clusterState, bool rejectChi2);
@@ -200,7 +200,7 @@ class GPUTPCGMTrackParam
     }
   }
 
-  GPUd() static void RefitTrack(GPUTPCGMMergedTrack& track, int32_t iTrk, GPUTPCGMMerger* merger, int32_t attempt);
+  GPUd() static void RefitTrack(GPUTPCGMMergedTrack& track, int32_t iTrk, GPUTPCGMMerger* merger);
 
   GPUdi() void ConstrainSinPhi(float limit = GPUCA_MAX_SIN_PHI)
   {