Merging dev and adjusting build issues

Author: ChSonnabend

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -931,18 +931,18 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
 
             auto stop0 = std::chrono::high_resolution_clock::now();
             auto start1 = std::chrono::high_resolution_clock::now();
-            nnApplication.inferenceNetwork(clustererNN.model_class, clustererNN, iSize, clusterer.modelProbabilities);
+            nnApplication.networkInference(nnApplication.model_class, clustererNN, iSize, clustererNN.modelProbabilities, evalDtype);
             if (nnApplication.model_class.getNumOutputNodes()[0][1] == 1) {
               runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::determineClass1Labels>({GetGrid(iSize, lane, GPUReconstruction::krnlDeviceType::CPU), {iSector}}, iSector, evalDtype, 0, batchStart); // Assigning class labels
             } else {
               runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::determineClass2Labels>({GetGrid(iSize, lane, GPUReconstruction::krnlDeviceType::CPU), {iSector}}, iSector, evalDtype, 0, batchStart); // Assigning class labels
             }
 
             if (!clustererNN.nnClusterizerUseCfRegression) {
-              nnApplication.inferenceNetwork(clustererNN.model_reg_1, clustererNN, iSize, clusterer.outputDataReg1);
+              nnApplication.networkInference(nnApplication.model_reg_1, clustererNN, iSize, clustererNN.outputDataReg1, evalDtype);
               runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::publishClass1Regression>({GetGrid(iSize, lane, GPUReconstruction::krnlDeviceType::CPU), {iSector}}, iSector, evalDtype, 0, batchStart); // Running the NN for regression class 1
               if (nnApplication.model_class.getNumOutputNodes()[0][1] > 1 && nnApplication.reg_model_paths.size() > 1) {
-                nnApplication.inferenceNetwork(clustererNN.model_reg_2, clustererNN, iSize, clusterer.outputDataReg2);
+                nnApplication.networkInference(nnApplication.model_reg_2, clustererNN, iSize, clustererNN.outputDataReg2, evalDtype);
                 runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::publishClass2Regression>({GetGrid(iSize, lane, GPUReconstruction::krnlDeviceType::CPU), {iSector}}, iSector, evalDtype, 0, batchStart); // Running the NN for regression class 2
               }
             }

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerHost.cxx

@@ -54,7 +54,7 @@ GPUTPCNNClusterizerHost::GPUTPCNNClusterizerHost(const GPUSettingsProcessingNNcl
   }
 }
 
-void GPUTPCNNClusterizerHost::networkInference(o2::ml::OrtModel model, GPUTPCNNClusterizer& clusterer, size_t size, float* output)
+void GPUTPCNNClusterizerHost::networkInference(o2::ml::OrtModel model, GPUTPCNNClusterizer& clusterer, size_t size, float* output, int32_t dtype)
 {
   if (dtype == 0) {
     model.inference<OrtDataType::Float16_t, float>(clusterer.inputData16, size * clusterer.nnClusterizerElementSize, output);

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerHost.h

@@ -39,7 +39,7 @@ class GPUTPCNNClusterizerHost
   GPUTPCNNClusterizerHost() = default;
   GPUTPCNNClusterizerHost(const GPUSettingsProcessingNNclusterizer&, GPUTPCNNClusterizer&);
 
-  void networkInference(o2::ml::OrtModel model, GPUTPCNNClusterizer& clusterer, size_t size, float* output);
+  void networkInference(o2::ml::OrtModel model, GPUTPCNNClusterizer& clusterer, size_t size, float* output, int32_t dtype);
 
   std::unordered_map<std::string, std::string> OrtOptions;
   o2::ml::OrtModel model_class, model_reg_1, model_reg_2; // For splitting clusters

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerKernels.cxx

@@ -22,6 +22,12 @@ using namespace o2::gpu::tpccf;
 #include "CfUtils.h"
 #include "ClusterAccumulator.h"
 #include "ML/3rdparty/GPUORTFloat16.h"
+
+#if !defined(GPUCA_GPUCODE)
+#include "GPUHostDataTypes.h"
+#include "MCLabelAccumulator.h"
+#endif
+
 #ifdef GPUCA_GPUCODE
 #include "GPUTPCCFClusterizer.inc"
 #endif