GPU: Improve buffer estimates when processing CTF data with cluster rejection

davidrohr · davidrohr · commit b57b5866ead7 · 2022-11-03T23:21:20.000+01:00
diff --git a/GPU/GPUTracking/DataTypes/GPUMemorySizeScalers.h b/GPU/GPUTracking/DataTypes/GPUMemorySizeScalers.h
@@ -43,7 +43,8 @@ struct GPUMemorySizeScalers {
   double tpcTrackletsPerStartHit = 0.8;
   double tpcTrackletHitsPerHit = 5;
   double tpcSectorTracksPerHit = 0.02;
-  double tpcSectorTrackHitsPerHit = 0.8f;
+  double tpcSectorTrackHitsPerHit = 0.8;
+  double tpcSectorTrackHitsPerHitWithRejection = 1.0;
   double tpcMergedTrackPerSliceTrack = 0.9;
   double tpcMergedTrackHitPerSliceHit = 1.1;
   size_t tpcCompressedUnattachedHitsBase1024[3] = {900, 900, 500}; // No ratio, but integer fraction of 1024 for exact computation
@@ -77,7 +78,7 @@ struct GPUMemorySizeScalers {
   inline size_t NTPCTracklets(size_t tpcHits) { return getValue(tpcMaxTracklets, NTPCStartHits(tpcHits) * tpcTrackletsPerStartHit); }
   inline size_t NTPCTrackletHits(size_t tpcHits) { return getValue(tpcMaxTrackletHits, hitOffset + tpcHits * tpcTrackletHitsPerHit); }
   inline size_t NTPCSectorTracks(size_t tpcHits) { return getValue(tpcMaxSectorTracks, tpcHits * tpcSectorTracksPerHit); }
-  inline size_t NTPCSectorTrackHits(size_t tpcHits) { return getValue(tpcMaxSectorTrackHits, tpcHits * tpcSectorTrackHitsPerHit); }
+  inline size_t NTPCSectorTrackHits(size_t tpcHits, unsigned char withRejection = 0) { return getValue(tpcMaxSectorTrackHits, tpcHits * (withRejection ? tpcSectorTrackHitsPerHitWithRejection : tpcSectorTrackHitsPerHit)); }
   inline size_t NTPCMergedTracks(size_t tpcSliceTracks) { return getValue(tpcMaxMergedTracks, tpcSliceTracks * (conservative ? 1.0 : tpcMergedTrackPerSliceTrack)); }
   inline size_t NTPCMergedTrackHits(size_t tpcSliceTrackHitss) { return getValue(tpcMaxMergedTrackHits, tpcSliceTrackHitss * tpcMergedTrackHitPerSliceHit); }
   inline size_t NTPCUnattachedHitsBase1024(int type) { return (returnMaxVal || conservative) ? 1024 : std::min<size_t>(1024, tpcCompressedUnattachedHitsBase1024[type] * factor * temporaryFactor); }
diff --git a/GPU/GPUTracking/Definitions/GPUSettingsList.h b/GPU/GPUTracking/Definitions/GPUSettingsList.h
@@ -148,6 +148,7 @@ AddOption(forceMemoryPoolSize, unsigned long, 1, "memSize", 0, "Force size of al
 AddOption(forceHostMemoryPoolSize, unsigned long, 0, "hostMemSize", 0, "Force size of allocated host page locked host memory (overriding memSize)", min(0ul))
 AddOption(memoryScalingFactor, float, 1.f, "", 0, "Factor to apply to all memory scalers")
 AddOption(conservativeMemoryEstimate, bool, false, "", 0, "Use some more conservative defaults for larger buffers during TPC processing")
+AddOption(tpcInputWithClusterRejection, unsigned char, 0, "", 0, "Indicate whether the TPC input is CTF data with cluster rejection, to tune buffer estimations")
 AddOption(forceMaxMemScalers, unsigned long, 0, "", 0, "Force using the maximum values for all buffers, Set a value n > 1 to rescale all maximums to a memory size of n")
 AddOption(registerStandaloneInputMemory, bool, false, "registerInputMemory", 0, "Automatically register input memory buffers for the GPU")
 AddOption(ompThreads, int, -1, "omp", 't', "Number of OMP threads to run (-1: all)", min(-1), message("Using %s OMP threads"))
diff --git a/GPU/GPUTracking/SliceTracker/GPUTPCTracker.cxx b/GPU/GPUTracking/SliceTracker/GPUTPCTracker.cxx
@@ -163,7 +163,7 @@ void GPUTPCTracker::SetMaxData(const GPUTrackingInOutPointers& io)
   mNMaxTracklets = mRec->MemoryScalers()->NTPCTracklets(mData.NumberOfHits());
   mNMaxRowHits = mRec->MemoryScalers()->NTPCTrackletHits(mData.NumberOfHits());
   mNMaxTracks = mRec->MemoryScalers()->NTPCSectorTracks(mData.NumberOfHits());
-  mNMaxTrackHits = mRec->MemoryScalers()->NTPCSectorTrackHits(mData.NumberOfHits());
+  mNMaxTrackHits = mRec->MemoryScalers()->NTPCSectorTrackHits(mData.NumberOfHits(), mRec->GetProcessingSettings().tpcInputWithClusterRejection);
 #ifdef GPUCA_SORT_STARTHITS_GPU
   if (mRec->IsGPU()) {
     if (mNMaxStartHits > mNMaxRowStartHits * GPUCA_ROW_COUNT) {
diff --git a/prodtests/full-system-test/dpl-workflow.sh b/prodtests/full-system-test/dpl-workflow.sh
@@ -108,6 +108,7 @@ else
     ITS_CONFIG_KEY+="ITSVertexerParam.phiCut=0.5;ITSVertexerParam.clusterContributorsCut=3;ITSVertexerParam.tanLambdaCut=0.2;"
   fi
 fi
+[[ $CTFINPUT == 1 ]] && GPU_CONFIG_KEY+="GPU_proc.tpcInputWithClusterRejection=1;"
 [[ ! -z $NTRDTRKTHREADS ]] && TRD_CONFIG_KEY+="GPU_proc.ompThreads=$NTRDTRKTHREADS;"
 [[ ! -z $NGPURECOTHREADS ]] && GPU_CONFIG_KEY+="GPU_proc.ompThreads=$NGPURECOTHREADS;"
 [[ ! -z $NMFTTHREADS ]] && MFT_CONFIG+=" --nThreads $NMFTTHREADS"
