GPU clusterizer with neural networks

Common/ML/include/ML/OrtInterface.h

@@ -41,6 +41,7 @@ class OrtModel
   OrtModel(std::unordered_map<std::string, std::string> optionsMap) { reset(optionsMap); }
   void init(std::unordered_map<std::string, std::string> optionsMap) { reset(optionsMap); }
   void reset(std::unordered_map<std::string, std::string>);
+  bool isInitialized() { return mInitialized; }
 
   virtual ~OrtModel() = default;
 
@@ -55,6 +56,9 @@ class OrtModel
   template <class I, class O> // class I is the input data type, e.g. float, class O is the output data type, e.g. O2::gpu::OrtDataType::Float16_t from O2/GPU/GPUTracking/ML/convert_float16.h
   std::vector<O> inference(std::vector<std::vector<I>>&);
 
+  template <class I, class O> // class I is the input data type, e.g. float, class O is the output data type, e.g. OrtDataType::Float16_t from O2/Common/ML/include/ML/GPUORTFloat16.h
+  void inference(I*, size_t, O*);
+
   // template<class I, class T, class O> // class I is the input data type, e.g. float, class T the throughput data type and class O is the output data type
   // std::vector<O> inference(std::vector<I>&);
 
@@ -79,6 +83,7 @@ class OrtModel
   std::vector<std::vector<int64_t>> mInputShapes, mOutputShapes;
 
   // Environment settings
+  bool mInitialized = false;
   std::string modelPath, device = "cpu", dtype = "float"; // device options should be cpu, rocm, migraphx, cuda
   int intraOpNumThreads = 0, deviceId = 0, enableProfiling = 0, loggingLevel = 0, allocateDeviceMemory = 0, enableOptimizations = 0;
 
@@ -89,4 +94,4 @@ class OrtModel
 
 } // namespace o2
 
-#endif // O2_ML_ORTINTERFACE_H
+#endif // O2_ML_ORTINTERFACE_H

Common/ML/src/OrtInterface.cxx

@@ -44,17 +44,19 @@ void OrtModel::reset(std::unordered_map<std::string, std::string> optionsMap)
   if (!optionsMap.contains("model-path")) {
     LOG(fatal) << "(ORT) Model path cannot be empty!";
   }
-  modelPath = optionsMap["model-path"];
-  device = (optionsMap.contains("device") ? optionsMap["device"] : "CPU");
-  dtype = (optionsMap.contains("dtype") ? optionsMap["dtype"] : "float");
-  deviceId = (optionsMap.contains("device-id") ? std::stoi(optionsMap["device-id"]) : 0);
-  allocateDeviceMemory = (optionsMap.contains("allocate-device-memory") ? std::stoi(optionsMap["allocate-device-memory"]) : 0);
-  intraOpNumThreads = (optionsMap.contains("intra-op-num-threads") ? std::stoi(optionsMap["intra-op-num-threads"]) : 0);
-  loggingLevel = (optionsMap.contains("logging-level") ? std::stoi(optionsMap["logging-level"]) : 2);
-  enableProfiling = (optionsMap.contains("enable-profiling") ? std::stoi(optionsMap["enable-profiling"]) : 0);
-  enableOptimizations = (optionsMap.contains("enable-optimizations") ? std::stoi(optionsMap["enable-optimizations"]) : 0);
-
-  std::string dev_mem_str = "Hip";
+
+  if (!optionsMap["model-path"].empty()) {
+    modelPath = optionsMap["model-path"];
+    device = (optionsMap.contains("device") ? optionsMap["device"] : "CPU");
+    dtype = (optionsMap.contains("dtype") ? optionsMap["dtype"] : "float");
+    deviceId = (optionsMap.contains("device-id") ? std::stoi(optionsMap["device-id"]) : 0);
+    allocateDeviceMemory = (optionsMap.contains("allocate-device-memory") ? std::stoi(optionsMap["allocate-device-memory"]) : 0);
+    intraOpNumThreads = (optionsMap.contains("intra-op-num-threads") ? std::stoi(optionsMap["intra-op-num-threads"]) : 0);
+    loggingLevel = (optionsMap.contains("logging-level") ? std::stoi(optionsMap["logging-level"]) : 0);
+    enableProfiling = (optionsMap.contains("enable-profiling") ? std::stoi(optionsMap["enable-profiling"]) : 0);
+    enableOptimizations = (optionsMap.contains("enable-optimizations") ? std::stoi(optionsMap["enable-optimizations"]) : 0);
+
+    std::string dev_mem_str = "Hip";
 #if defined(ORT_ROCM_BUILD)
 #if ORT_ROCM_BUILD == 1
   if (device == "ROCM") {
@@ -93,7 +95,9 @@ void OrtModel::reset(std::unordered_map<std::string, std::string> optionsMap)
     } else if (intraOpNumThreads == 1) {
       (pImplOrt->sessionOptions).SetExecutionMode(ExecutionMode::ORT_SEQUENTIAL);
     }
-    LOG(info) << "(ORT) CPU execution provider set with " << intraOpNumThreads << " threads";
+    if (loggingLevel < 2) {
+      LOG(info) << "(ORT) CPU execution provider set with " << intraOpNumThreads << " threads";
+    }
   }
 
   (pImplOrt->sessionOptions).DisableMemPattern();
@@ -109,6 +113,9 @@ void OrtModel::reset(std::unordered_map<std::string, std::string> optionsMap)
   } else {
     (pImplOrt->sessionOptions).DisableProfiling();
   }
+
+  mInitialized = true;
+
   (pImplOrt->sessionOptions).SetGraphOptimizationLevel(GraphOptimizationLevel(enableOptimizations));
   (pImplOrt->sessionOptions).SetLogSeverityLevel(OrtLoggingLevel(loggingLevel));
 
@@ -154,16 +161,9 @@ void OrtModel::reset(std::unordered_map<std::string, std::string> optionsMap)
   outputNamesChar.resize(mOutputNames.size(), nullptr);
   std::transform(std::begin(mOutputNames), std::end(mOutputNames), std::begin(outputNamesChar),
                  [&](const std::string& str) { return str.c_str(); });
-
-  // Print names
-  LOG(info) << "\tInput Nodes:";
-  for (size_t i = 0; i < mInputNames.size(); i++) {
-    LOG(info) << "\t\t" << mInputNames[i] << " : " << printShape(mInputShapes[i]);
   }
-
-  LOG(info) << "\tOutput Nodes:";
-  for (size_t i = 0; i < mOutputNames.size(); i++) {
-    LOG(info) << "\t\t" << mOutputNames[i] << " : " << printShape(mOutputShapes[i]);
+  if (loggingLevel < 2) {
+    LOG(info) << "(ORT) Model loaded successfully! (input: " << printShape(mInputShapes[0]) << ", output: " << printShape(mOutputShapes[0]) << ")";
   }
 }
 
@@ -187,36 +187,6 @@ std::vector<O> OrtModel::v2v(std::vector<I>& input, bool clearInput)
   }
 }
 
-template <class I, class O> // class I is the input data type, e.g. float, class O is the output data type, e.g. O2::gpu::OrtDataType::Float16_t from O2/GPU/GPUTracking/ML/convert_float16.h
-std::vector<O> OrtModel::inference(std::vector<I>& input)
-{
-  std::vector<int64_t> inputShape{(int64_t)(input.size() / mInputShapes[0][1]), (int64_t)mInputShapes[0][1]};
-  std::vector<Ort::Value> inputTensor;
-  inputTensor.emplace_back(Ort::Value::CreateTensor<O>(pImplOrt->memoryInfo, reinterpret_cast<O*>(input.data()), input.size(), inputShape.data(), inputShape.size()));
-  // input.clear();
-  auto outputTensors = (pImplOrt->session)->Run(pImplOrt->runOptions, inputNamesChar.data(), inputTensor.data(), inputTensor.size(), outputNamesChar.data(), outputNamesChar.size());
-  O* outputValues = reinterpret_cast<O*>(outputTensors[0].template GetTensorMutableData<O>());
-  std::vector<O> outputValuesVec{outputValues, outputValues + inputShape[0] * mOutputShapes[0][1]};
-  outputTensors.clear();
-  return outputValuesVec;
-}
-
-template <class I, class O> // class I is the input data type, e.g. float, class O is the output data type, e.g. O2::gpu::OrtDataType::Float16_t from O2/GPU/GPUTracking/ML/convert_float16.h
-std::vector<O> OrtModel::inference(std::vector<std::vector<I>>& input)
-{
-  std::vector<Ort::Value> inputTensor;
-  for (auto i : input) {
-    std::vector<int64_t> inputShape{(int64_t)(i.size() / mInputShapes[0][1]), (int64_t)mInputShapes[0][1]};
-    inputTensor.emplace_back(Ort::Value::CreateTensor<O>(pImplOrt->memoryInfo, reinterpret_cast<O*>(i.data()), i.size(), inputShape.data(), inputShape.size()));
-  }
-  // input.clear();
-  auto outputTensors = (pImplOrt->session)->Run(pImplOrt->runOptions, inputNamesChar.data(), inputTensor.data(), inputTensor.size(), outputNamesChar.data(), outputNamesChar.size());
-  O* outputValues = reinterpret_cast<O*>(outputTensors[0].template GetTensorMutableData<O>());
-  std::vector<O> outputValuesVec{outputValues, outputValues + inputTensor.size() / mInputShapes[0][1] * mOutputShapes[0][1]};
-  outputTensors.clear();
-  return outputValuesVec;
-}
-
 std::string OrtModel::printShape(const std::vector<int64_t>& v)
 {
   std::stringstream ss("");
@@ -227,78 +197,72 @@ std::string OrtModel::printShape(const std::vector<int64_t>& v)
   return ss.str();
 }
 
-template <>
-std::vector<float> OrtModel::inference<float, float>(std::vector<float>& input)
+template <class I, class O>
+std::vector<O> OrtModel::inference(std::vector<I>& input)
 {
   std::vector<int64_t> inputShape{(int64_t)(input.size() / mInputShapes[0][1]), (int64_t)mInputShapes[0][1]};
   std::vector<Ort::Value> inputTensor;
-  inputTensor.emplace_back(Ort::Value::CreateTensor<float>(pImplOrt->memoryInfo, input.data(), input.size(), inputShape.data(), inputShape.size()));
+  if constexpr (std::is_same_v<I, OrtDataType::Float16_t>) {
+    inputTensor.emplace_back(Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(input.data()), input.size(), inputShape.data(), inputShape.size()));
+  } else {
+    inputTensor.emplace_back(Ort::Value::CreateTensor<I>(pImplOrt->memoryInfo, input.data(), input.size(), inputShape.data(), inputShape.size()));
+  }
   // input.clear();
   auto outputTensors = (pImplOrt->session)->Run(pImplOrt->runOptions, inputNamesChar.data(), inputTensor.data(), inputTensor.size(), outputNamesChar.data(), outputNamesChar.size());
-  float* outputValues = outputTensors[0].template GetTensorMutableData<float>();
-  std::vector<float> outputValuesVec{outputValues, outputValues + inputShape[0] * mOutputShapes[0][1]};
+  O* outputValues = outputTensors[0].template GetTensorMutableData<O>();
+  std::vector<O> outputValuesVec{outputValues, outputValues + inputShape[0] * mOutputShapes[0][1]};
   outputTensors.clear();
   return outputValuesVec;
 }
 
-template <>
-std::vector<float> OrtModel::inference<OrtDataType::Float16_t, float>(std::vector<OrtDataType::Float16_t>& input)
-{
-  std::vector<int64_t> inputShape{(int64_t)(input.size() / mInputShapes[0][1]), (int64_t)mInputShapes[0][1]};
-  std::vector<Ort::Value> inputTensor;
-  inputTensor.emplace_back(Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(input.data()), input.size(), inputShape.data(), inputShape.size()));
-  // input.clear();
-  auto outputTensors = (pImplOrt->session)->Run(pImplOrt->runOptions, inputNamesChar.data(), inputTensor.data(), inputTensor.size(), outputNamesChar.data(), outputNamesChar.size());
-  float* outputValues = outputTensors[0].template GetTensorMutableData<float>();
-  std::vector<float> outputValuesVec{outputValues, outputValues + inputShape[0] * mOutputShapes[0][1]};
-  outputTensors.clear();
-  return outputValuesVec;
-}
+template std::vector<float> OrtModel::inference<float, float>(std::vector<float>&);
 
-template <>
-std::vector<OrtDataType::Float16_t> OrtModel::inference<OrtDataType::Float16_t, OrtDataType::Float16_t>(std::vector<OrtDataType::Float16_t>& input)
-{
-  std::vector<int64_t> inputShape{(int64_t)(input.size() / mInputShapes[0][1]), (int64_t)mInputShapes[0][1]};
-  std::vector<Ort::Value> inputTensor;
-  inputTensor.emplace_back(Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(input.data()), input.size(), inputShape.data(), inputShape.size()));
-  // input.clear();
-  auto outputTensors = (pImplOrt->session)->Run(pImplOrt->runOptions, inputNamesChar.data(), inputTensor.data(), inputTensor.size(), outputNamesChar.data(), outputNamesChar.size());
-  OrtDataType::Float16_t* outputValues = reinterpret_cast<OrtDataType::Float16_t*>(outputTensors[0].template GetTensorMutableData<Ort::Float16_t>());
-  std::vector<OrtDataType::Float16_t> outputValuesVec{outputValues, outputValues + inputShape[0] * mOutputShapes[0][1]};
-  outputTensors.clear();
-  return outputValuesVec;
-}
+template std::vector<float> OrtModel::inference<OrtDataType::Float16_t, float>(std::vector<OrtDataType::Float16_t>&);
 
-template <>
-std::vector<OrtDataType::Float16_t> OrtModel::inference<float, OrtDataType::Float16_t>(std::vector<float>& input)
+template std::vector<OrtDataType::Float16_t> OrtModel::inference<OrtDataType::Float16_t, OrtDataType::Float16_t>(std::vector<OrtDataType::Float16_t>&);
+
+template <class I, class O>
+void OrtModel::inference(I* input, size_t input_size, O* output)
 {
-  std::vector<int64_t> inputShape{(int64_t)(input.size() / mInputShapes[0][1]), (int64_t)mInputShapes[0][1]};
-  std::vector<Ort::Value> inputTensor;
-  inputTensor.emplace_back(Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(input.data()), input.size(), inputShape.data(), inputShape.size()));
-  // input.clear();
-  auto outputTensors = (pImplOrt->session)->Run(pImplOrt->runOptions, inputNamesChar.data(), inputTensor.data(), inputTensor.size(), outputNamesChar.data(), outputNamesChar.size());
-  OrtDataType::Float16_t* outputValues = reinterpret_cast<OrtDataType::Float16_t*>(outputTensors[0].template GetTensorMutableData<Ort::Float16_t>());
-  std::vector<OrtDataType::Float16_t> outputValuesVec{outputValues, outputValues + inputShape[0] * mOutputShapes[0][1]};
-  outputTensors.clear();
-  return outputValuesVec;
+  std::vector<int64_t> inputShape{(int64_t)(input_size / mInputShapes[0][1]), (int64_t)mInputShapes[0][1]};
+  Ort::Value inputTensor = Ort::Value(nullptr);
+  if constexpr (std::is_same_v<I, OrtDataType::Float16_t>) {
+    inputTensor = Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(input), input_size, inputShape.data(), inputShape.size());
+  } else {
+    inputTensor = Ort::Value::CreateTensor<I>(pImplOrt->memoryInfo, input, input_size, inputShape.data(), inputShape.size());
+  }
+
+  std::vector<int64_t> outputShape{inputShape[0], mOutputShapes[0][1]};
+  size_t outputSize = (int64_t)(inputShape[0] * mOutputShapes[0][1]);
+  Ort::Value outputTensor = Ort::Value::CreateTensor<O>(pImplOrt->memoryInfo, output, outputSize, outputShape.data(), outputShape.size());
+
+  (pImplOrt->session)->Run(pImplOrt->runOptions, inputNamesChar.data(), &inputTensor, 1, outputNamesChar.data(), &outputTensor, 1);
 }
 
-template <>
-std::vector<OrtDataType::Float16_t> OrtModel::inference<OrtDataType::Float16_t, OrtDataType::Float16_t>(std::vector<std::vector<OrtDataType::Float16_t>>& input)
+template void OrtModel::inference<OrtDataType::Float16_t, float>(OrtDataType::Float16_t*, size_t, float*);
+
+template void OrtModel::inference<float, float>(float*, size_t, float*);
+
+template <class I, class O>
+std::vector<O> OrtModel::inference(std::vector<std::vector<I>>& input)
 {
   std::vector<Ort::Value> inputTensor;
   for (auto i : input) {
     std::vector<int64_t> inputShape{(int64_t)(i.size() / mInputShapes[0][1]), (int64_t)mInputShapes[0][1]};
-    inputTensor.emplace_back(Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(i.data()), i.size(), inputShape.data(), inputShape.size()));
+    if constexpr (std::is_same_v<I, OrtDataType::Float16_t>) {
+      inputTensor.emplace_back(Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(i.data()), i.size(), inputShape.data(), inputShape.size()));
+    } else {
+      inputTensor.emplace_back(Ort::Value::CreateTensor<I>(pImplOrt->memoryInfo, i.data(), i.size(), inputShape.data(), inputShape.size()));
+    }
   }
   // input.clear();
   auto outputTensors = (pImplOrt->session)->Run(pImplOrt->runOptions, inputNamesChar.data(), inputTensor.data(), inputTensor.size(), outputNamesChar.data(), outputNamesChar.size());
-  OrtDataType::Float16_t* outputValues = reinterpret_cast<OrtDataType::Float16_t*>(outputTensors[0].template GetTensorMutableData<Ort::Float16_t>());
-  std::vector<OrtDataType::Float16_t> outputValuesVec{outputValues, outputValues + inputTensor.size() / mInputShapes[0][1] * mOutputShapes[0][1]};
+  O* outputValues = reinterpret_cast<O*>(outputTensors[0].template GetTensorMutableData<O>());
+  std::vector<O> outputValuesVec{outputValues, outputValues + inputTensor.size() / mInputShapes[0][1] * mOutputShapes[0][1]};
   outputTensors.clear();
   return outputValuesVec;
 }
 
 } // namespace ml
 
-} // namespace o2
+} // namespace o2

GPU/GPUTracking/Base/GPUConstantMem.h

@@ -34,6 +34,10 @@
 #include "GPUKernelDebugOutput.h"
 #endif
 
+#ifdef GPUCA_HAS_ONNX
+#include "GPUTPCNNClusterizer.h"
+#endif
+
 namespace o2::gpu
 {
 struct GPUConstantMem {
@@ -55,6 +59,9 @@ struct GPUConstantMem {
 #ifdef GPUCA_KERNEL_DEBUGGER_OUTPUT
   GPUKernelDebugOutput debugOutput;
 #endif
+#ifdef GPUCA_HAS_ONNX
+  GPUTPCNNClusterizer tpcNNClusterer[GPUCA_NSECTORS];
+#endif
 
   template <int32_t I>
   GPUd() auto& getTRDTracker();

GPU/GPUTracking/Base/GPUMemoryResource.h

@@ -28,6 +28,7 @@ struct GPUMemoryReuse {
   };
   enum Group : uint16_t {
     ClustererScratch,
+    NNClusterer,
     ClustererZS,
     TrackerScratch,
     TrackerDataLinks,

GPU/GPUTracking/Base/GPUReconstruction.cxx

@@ -93,6 +93,9 @@ GPUReconstruction::GPUReconstruction(const GPUSettingsDeviceBackend& cfg) : mHos
   for (uint32_t i = 0; i < NSECTORS; i++) {
     processors()->tpcTrackers[i].SetSector(i); // TODO: Move to a better place
     processors()->tpcClusterer[i].mISector = i;
+#ifdef GPUCA_HAS_ONNX
+    processors()->tpcNNClusterer[i].mISector = i;
+#endif
   }
 #ifndef GPUCA_NO_ROOT
   mROOTDump = GPUROOTDumpCore::getAndCreate();

GPU/GPUTracking/CMakeLists.txt

@@ -160,8 +160,8 @@ set(HDRS_INSTALL
 )
 
 set(SRCS_NO_CINT ${SRCS_NO_CINT} display/GPUDisplayInterface.cxx)
-set(SRCS_NO_CINT
-    ${SRCS_NO_CINT}
+
+set(SRCS_NO_CINT ${SRCS_NO_CINT}
     Global/GPUChainITS.cxx
     ITS/GPUITSFitter.cxx
     ITS/GPUITSFitterKernels.cxx
@@ -192,6 +192,10 @@ set(SRCS_NO_CINT
     Refit/GPUTrackingRefitKernel.cxx
     Merger/GPUTPCGMO2Output.cxx)
 
+if(NOT ALIGPU_BUILD_TYPE STREQUAL "Standalone")
+  list(APPEND SRCS_NO_CINT TPCClusterFinder/GPUTPCNNClusterizerKernels.cxx TPCClusterFinder/GPUTPCNNClusterizer.cxx TPCClusterFinder/GPUTPCNNClusterizerHost.cxx)
+endif()
+
 set(SRCS_DATATYPES
     ${SRCS_DATATYPES}
     DataTypes/TPCPadGainCalib.cxx
@@ -271,9 +275,11 @@ if(ALIGPU_BUILD_TYPE STREQUAL "O2")
                                        O2::GPUCommon
                                        O2::ReconstructionDataFormats
                                        O2::TPCFastTransformation
+                                       O2::ML
                  PRIVATE_LINK_LIBRARIES O2::DataFormatsTPC
                  SOURCES ${SRCS_DATATYPES})
-  target_compile_definitions(${targetName} PRIVATE GPUCA_O2_LIB GPUCA_TPC_GEOMETRY_O2)
+  target_compile_definitions(${targetName} PRIVATE GPUCA_O2_LIB GPUCA_TPC_GEOMETRY_O2 GPUCA_HAS_ONNX=1)
+
   o2_target_root_dictionary(GPUDataTypes
                             HEADERS ${HDRS_CINT_DATATYPES} ${HDRS_CINT_O2_ADDITIONAL}
                             LINKDEF GPUTrackingLinkDef_O2_DataTypes.h)

GPU/GPUTracking/Definitions/GPUDefGPUParameters.h

@@ -489,6 +489,9 @@
   #ifndef GPUCA_LB_GPUTPCCFClusterizer
     #define GPUCA_LB_GPUTPCCFClusterizer 512
   #endif
+  #ifndef GPUCA_LB_GPUTPCNNClusterizerKernels
+    #define GPUCA_LB_GPUTPCNNClusterizerKernels 512
+  #endif
   #ifndef GPUCA_LB_GPUTrackingRefitKernel_mode0asGPU
     #define GPUCA_LB_GPUTrackingRefitKernel_mode0asGPU 256
   #endif
@@ -507,6 +510,16 @@
 
 #define GPUCA_LB_GPUTPCCFNoiseSuppression_noiseSuppression GPUCA_LB_GPUTPCCFNoiseSuppression
 #define GPUCA_LB_GPUTPCCFNoiseSuppression_updatePeaks GPUCA_LB_GPUTPCCFNoiseSuppression
+
+#ifdef GPUCA_HAS_ONNX
+#define GPUCA_LB_GPUTPCNNClusterizerKernels_runCfClusterizer GPUCA_LB_GPUTPCNNClusterizerKernels
+#define GPUCA_LB_GPUTPCNNClusterizerKernels_fillInputNN GPUCA_LB_GPUTPCNNClusterizerKernels
+#define GPUCA_LB_GPUTPCNNClusterizerKernels_determineClass1Labels GPUCA_LB_GPUTPCNNClusterizerKernels
+#define GPUCA_LB_GPUTPCNNClusterizerKernels_determineClass2Labels GPUCA_LB_GPUTPCNNClusterizerKernels
+#define GPUCA_LB_GPUTPCNNClusterizerKernels_publishClass1Regression GPUCA_LB_GPUTPCNNClusterizerKernels
+#define GPUCA_LB_GPUTPCNNClusterizerKernels_publishClass2Regression GPUCA_LB_GPUTPCNNClusterizerKernels
+#endif
+
 #define GPUCA_LB_GPUTPCCFStreamCompaction_scanStart GPUCA_THREAD_COUNT_SCAN
 #define GPUCA_LB_GPUTPCCFStreamCompaction_scanUp GPUCA_THREAD_COUNT_SCAN
 #define GPUCA_LB_GPUTPCCFStreamCompaction_scanTop GPUCA_THREAD_COUNT_SCAN