Fixing one memory issue. Segfault / memory leak persists

Author: ChSonnabend

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -893,6 +893,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
           clustererNN.nnClusterizerElementSize = ((2 * nn_settings.nnClusterizerSizeInputRow + 1) * (2 * nn_settings.nnClusterizerSizeInputPad + 1) * (2 * nn_settings.nnClusterizerSizeInputTime + 1)) + (nn_settings.nnClusterizerAddIndexData ? 3 : 0);
           clustererNN.nnClusterizerBatchedMode = nn_settings.nnClusterizerBatchedMode;
           clustererNN.nnClusterizerBoundaryFillValue = nn_settings.nnClusterizerBoundaryFillValue;
+          clustererNN.nnClusterizerTotalClusters = clusterer.mPmemory->counters.nClusters;
           if (nn_settings.nnClusterizerVerbosity < 0) {
             clustererNN.nnClusterizerVerbosity = nn_settings.nnInferenceVerbosity;
           } else {
@@ -962,7 +963,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
           time_clusterizer += std::chrono::duration_cast<std::chrono::nanoseconds>(stop1 - start1).count() / 1e9;
 
           if (clustererNN.nnClusterizerVerbosity < 3) {
-            LOG(info) << "[NN CF] Apply NN (fragment " << fragment.index << ", lane: " << lane << ", slice: " << iSector << "): filling data " << time_fill << "s ; clusterizer: " << time_clusterizer << "s ; " << clusterer.mPmemory->counters.nClusters << " clusters --> " << clusterer.mPmemory->counters.nClusters / (time_fill + time_clusterizer) << " clusters/s";
+            LOG(info) << "[NN CF] Apply NN (fragment " << fragment.index << ", lane: " << lane << ", sector: " << iSector << "): filling data " << time_fill << "s ; clusterizer: " << time_clusterizer << "s ; " << clusterer.mPmemory->counters.nClusters << " clusters --> " << clusterer.mPmemory->counters.nClusters / (time_fill + time_clusterizer) << " clusters/s";
           }
 #else
           GPUFatal("Project not compiled with neural network clusterization. Aborting.");

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizer.cxx

@@ -32,7 +32,7 @@ void* GPUTPCNNClusterizer::setIOPointers(void* mem)
   computePointerWithAlignment(mem, peakPositions, nnClusterizerBatchedMode);
   computePointerWithAlignment(mem, clusterFlags, 2 * nnClusterizerBatchedMode);
   computePointerWithAlignment(mem, centralCharges, nnClusterizerBatchedMode);
-  computePointerWithAlignment(mem, outputDataClass, nnClusterizerBatchedMode);
+  computePointerWithAlignment(mem, outputDataClass, nnClusterizerTotalClusters);
   if (nnClusterizerModelClassNumOutputNodes > 0) {
     computePointerWithAlignment(mem, modelProbabilities, nnClusterizerBatchedMode * nnClusterizerModelClassNumOutputNodes);
   }

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizer.h

@@ -46,6 +46,7 @@ class GPUTPCNNClusterizer : public GPUProcessor
   bool nnSigmoidTrafoClassThreshold = 1;
   int nnClusterizerUseCfRegression = 0;
   int nnClusterizerBatchedMode = 1;
+  int nnClusterizerTotalClusters = 1;
   int nnClusterizerVerbosity = 0;
   int nnClusterizerBoundaryFillValue = -1;
   int nnClusterizerDumpDigits = 0;

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerKernels.cxx

@@ -143,6 +143,7 @@ GPUd() void GPUTPCNNClusterizerKernels::fillInputData(int32_t nBlocks, int32_t n
 
   clustererNN.peakPositions[glo_idx] = peak;
   clustererNN.centralCharges[glo_idx] = central_charge;
+  clustererNN.outputDataClass[glo_idx + batchStart] = -1;
 
   int row_offset = GPUTPCNNClusterizerKernels::rowOffset(row, clustererNN.nnClusterizerSizeInputRow);
 #ifndef GPUCA_GPUCODE