Bug-fixes for boundary reading

Author: ChSonnabend

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -306,7 +306,7 @@ AddOption(nnInferenceDtype, std::string, "fp32", "", 0, "(std::string) Specify t
 AddOption(nnInferenceThreadsPerNN, int, 0, "", 0, "Number of threads used to evaluate one neural network")
 AddOption(nnInferenceEnableOrtOptimization, unsigned int, 1, "", 0, "Enables graph optimizations in ONNX Runtime. Can be greater than 1!")
 AddOption(nnInferenceOrtProfiling, int, 0, "", 0, "Enables profiling of model execution in ONNX Runtime")
-AddOption(nnInferenceOrtProfilingPath, std::string, ".", "", 0, "If mmInferenceOrtProfiling is set, the path to store the profiling data")
+AddOption(nnInferenceOrtProfilingPath, std::string, ".", "", 0, "If nnInferenceOrtProfiling is set, the path to store the profiling data")
 AddOption(nnInferenceVerbosity, int, 1, "", 0, "0: No messages; 1: Warnings; 2: Warnings + major debugs; >3: All debugs")
 AddOption(nnClusterizerAddIndexData, int, 1, "", 0, "If normalized index data (sector, row, pad), should be appended to the input")
 AddOption(nnClusterizerSizeInputRow, int, 3, "", 0, "Size of the input to the NN (currently calcualted as (length-1)/2")
@@ -316,6 +316,7 @@ AddOption(nnClusterizerUseCFregression, int, 0, "", 0, "(bool, default = false)
 AddOption(nnClusterizerBatchedMode, unsigned int, 1, "", 0, "(int, default = 1) If >1, the NN is evaluated on batched input of size specified in this variable")
 AddOption(nnClusterizerVerbosity, int, -1, "", 0, "(int, default = -1) If >0, logging messages of the clusterizer will be displayed")
 AddOption(nnClusterizerBoundaryFillValue, int, -1, "", 0, "Fill value for the boundary of the input to the NN")
+AddOption(nnClusterizerApplyCfDeconvolution, int, 0, "", 0, "Applies the CFDeconvolution kernel before the digits to the network are filled")
 AddOption(nnClassificationPath, std::string, "network_class.onnx", "", 0, "The classification network path")
 AddOption(nnClassThreshold, float, 0.5, "", 0, "The cutoff at which clusters will be accepted / rejected.")
 AddOption(nnRegressionPath, std::string, "network_reg.onnx", "", 0, "The regression network path")

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -931,7 +931,9 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
               clusterer.OrtOptions["model-path"] = reg_model_paths[1];
               clusterer.model_reg_2.init(clusterer.OrtOptions);
             }
-          } else {
+          }
+          
+          if (clusterer.nnClusterizerUseCFregression || (int)(GetProcessingSettings().nnClusterizerApplyCfDeconvolution)) {
             runKernel<GPUTPCCFDeconvolution>({GetGrid(clusterer.mPmemory->counters.nPositions, lane), {iSlice}});
             DoDebugAndDump(RecoStep::TPCClusterFinding, 262144 << 4, clusterer, &GPUTPCClusterFinder::DumpChargeMap, *mDebugFile, "Split Charges");
           }
@@ -992,6 +994,8 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
             LOG(info) << "[NN CF] Apply NN (fragment " << fragment.index << ", lane: " << lane << ", slice: " << iSlice << "): 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 {
+          runKernel<GPUTPCCFDeconvolution>({GetGrid(clusterer.mPmemory->counters.nPositions, lane), {iSlice}});
+          DoDebugAndDump(RecoStep::TPCClusterFinding, 262144 << 4, clusterer, &GPUTPCClusterFinder::DumpChargeMap, *mDebugFile, "Split Charges");
           runKernel<GPUTPCCFClusterizer>({GetGrid(clusterer.mPmemory->counters.nClusters, lane, GPUReconstruction::krnlDeviceType::CPU), {iSlice}}, 0);
         }
 

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizer.cxx

@@ -108,16 +108,16 @@ bool GPUTPCNNClusterizer::isBoundary(int row, int pad, int global_shift, const G
 {
   if (pad < 0 || row < 0) { // Faster short-circuit
     return true;
-  } else if (row <= 62) {
-    if (pad < 0 || pad > geo.NPads(row)) {
+  } else if (row < 63) {
+    if (pad >= geo.NPads(row)) {
       return true;
     } else {
       return false;
     }
-  } else if (row <= 62 + global_shift) { // to account for the gap between IROC and OROC. Charge will be set to -1 in order to signal boundary to the neural network
+  } else if (row < (63 + global_shift)) { // to account for the gap between IROC and OROC. Charge will be set to -1 in order to signal boundary to the neural network
     return true;
   } else if (row <= o2::tpc::constants::MAXGLOBALPADROW - 1 + global_shift) {
-    if (pad < 0 || pad > geo.NPads(row)) {
+    if (pad >= geo.NPads(row - global_shift)) {
       return true;
     } else {
       return false;
@@ -156,7 +156,13 @@ GPUd() void GPUTPCNNClusterizer::fillInputData(int32_t nBlocks, int32_t nThreads
     for (int p = -clusterer.nnClusterizerSizeInputPad + pad_offset; p <= clusterer.nnClusterizerSizeInputPad + pad_offset; p++) {
       is_boundary = is_boundary || GPUTPCNNClusterizer::isBoundary(row + r + row_offset, pad + p, clusterer.nnClusterizerSizeInputRow, clusterer.Param().tpcGeometry);
       for (int t = -clusterer.nnClusterizerSizeInputTime; t <= clusterer.nnClusterizerSizeInputTime; t++) {
-        if (!is_boundary) {
+        if (is_boundary) {
+          if(dtype == 0){
+            clusterer.inputData16[write_idx] = (OrtDataType::Float16_t)((float)clusterer.nnClusterizerBoundaryFillValue);
+          } else {
+            clusterer.inputData32[write_idx] = (float)clusterer.nnClusterizerBoundaryFillValue;
+          }
+        } else {
           ChargePos tmp_pos(row + r, pad + p, time + t);
           if(dtype == 0){
             clusterer.inputData16[write_idx] = (OrtDataType::Float16_t)((float)chargeMap[tmp_pos].unpack() / central_charge);