GPU TPC: Slighly augment dynamic buffer for TPC Decoding

Author: cima22

GPU/GPUTracking/DataCompression/GPUTPCDecompression.cxx

@@ -117,10 +117,12 @@ void GPUTPCDecompression::RegisterMemoryAllocation()
 
 void GPUTPCDecompression::SetMaxData(const GPUTrackingInOutPointers& io)
 {
+  GPUMemorySizeScalers scaler;
   uint32_t maxAttachedClsMargin1 = *std::max_element(mInputGPU.nSliceRowClusters, mInputGPU.nSliceRowClusters + mInputGPU.nSliceRows);
-  float clsRatio1 = (mInputGPU.nUnattachedClusters > 0 ? float(mInputGPU.nAttachedClusters) / float(mInputGPU.nUnattachedClusters) : 1.0f) * 1.5f;
+  float clsRatio1 = (mInputGPU.nUnattachedClusters > 0 ? float(mInputGPU.nAttachedClusters) / float(mInputGPU.nUnattachedClusters) : 1.0f) * scaler.tpcDecodingClusterRatioFactor1;
   maxAttachedClsMargin1 *= clsRatio1;
-  uint32_t maxAttachedClsMargin2 = mInputGPU.nSliceRows > 0 ? (mInputGPU.nAttachedClusters / mInputGPU.nSliceRows * 3.5) : 0;            // mean #attached cls per SectorRow multiplied by 3.5 (tuned)
-  mMaxNativeClustersPerBuffer = std::max({maxAttachedClsMargin1, maxAttachedClsMargin2, 1000u});                                         // take biggest margin, 1000 clusters minimum
-  mMaxNativeClustersPerBuffer = std::min(mMaxNativeClustersPerBuffer, mRec->GetProcessingSettings().tpcMaxAttachedClustersPerSectorRow); // upperbound given by configurable param
+  uint32_t maxAttachedClsMargin2 = mInputGPU.nSliceRows > 0 ? (mInputGPU.nAttachedClusters / mInputGPU.nSliceRows * scaler.tpcDecodingClusterRatioFactor2) : 0; // mean #attached cls per SectorRow multiplied by 3.5 (tuned)
+  mMaxNativeClustersPerBuffer = std::max({maxAttachedClsMargin1, maxAttachedClsMargin2, 1000u});                                                                // take biggest margin, 1000 clusters minimum
+  mMaxNativeClustersPerBuffer = std::min(mMaxNativeClustersPerBuffer, mRec->GetProcessingSettings().tpcMaxAttachedClustersPerSectorRow);                        // upperbound given by configurable param
+  mMaxNativeClustersPerBuffer += 700;
 }

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -297,7 +297,7 @@ AddOption(nnCCDBInteractionRate, std::string, "500", "", 0, "Distinguishes betwe
 AddHelp("help", 'h')
 EndConfig()
 
-// Settings steering the processing of NN Clusterization
+// Scaling factors for gpu buffer size estimation
 BeginSubConfig(GPUSettingsProcessingScaling, scaling, configStandalone.proc, "SCALING", 0, "Processing settings for neural network clusterizer", proc_scaling)
 AddOption(offset, float, 1000., "", 0, "Scaling Factor: offset")
 AddOption(hitOffset, float, 20000, "", 0, "Scaling Factor: hitOffset")
@@ -315,6 +315,8 @@ AddOption(tpcMergedTrackPerSectorTrack, float, 1.0, "", 0, "Scaling Factor: tpcM
 AddOption(tpcMergedTrackHitPerSectorHit, float, 1.1, "", 0, "Scaling Factor: tpcMergedTrackHitPerSectorHit")
 AddOptionArray(tpcCompressedUnattachedHitsBase1024, int32_t, 3, (900, 900, 500), "", 0, "Scaling Factor: tpcCompressedUnattachedHitsBase1024")
 AddOption(conservativeMemoryEstimate, bool, false, "", 0, "Use some more conservative defaults for larger buffers during TPC processing")
+AddOption(tpcDecodingClusterRatioFactor1, float, 1.5, "", 0, "Scaling Factor: for first margin of dynamic buffer allocation for attached clusters in TPC Decoding")
+AddOption(tpcDecodingClusterRatioFactor2, float, 3.5, "", 0, "Scaling Factor: for second margin of dynamic buffer allocation for attached clusters in TPC Decoding")
 AddHelp("help", 'h')
 EndConfig()