Fixes

Author: ChSonnabend

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDA.cu

@@ -16,7 +16,6 @@
 #include "GPUReconstructionCUDAIncludesHost.h"
 
 #include <cuda_profiler_api.h>
-#include "ML/OrtInterface.h"
 
 #include "GPUReconstructionCUDA.h"
 #include "GPUReconstructionCUDAInternals.h"
@@ -692,7 +691,7 @@ void* GPUReconstructionHIP::getGPUPointer(void* ptr)
 }
 
 #ifdef GPUCA_HAS_ONNX
-int32_t GPUReconstructionCUDA::SetONNXGPUStream(OrtSessionOptions* session_options, int32_t stream)
+int32_t GPUReconstructionHIP::SetONNXGPUStream(OrtSessionOptions* session_options, int32_t stream)
 {
   // Create ROCm provider options
   const auto& api = Ort::GetApi();

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -640,7 +640,7 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
         clustererNN.nnClusterizerVerbosity = nn_settings.nnClusterizerVerbosity;
       }
       clustererNN.nnClusterizerDtype = nn_settings.nnInferenceDtype.find("32") != std::string::npos;
-      GPUTPCNNClusterizerHost nnApplication(nn_settings, clustererNN);
+      GPUTPCNNClusterizerHost nnApplication(nn_settings, clustererNN, iSector);
       AllocateRegisteredMemory(clustererNN.mMemoryId);
     }
   }

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerHost.cxx

@@ -19,12 +19,12 @@
 
 using namespace o2::gpu;
 
-GPUTPCNNClusterizerHost::GPUTPCNNClusterizerHost(const GPUSettingsProcessingNNclusterizer& settings, GPUTPCNNClusterizer& clusterer)
+GPUTPCNNClusterizerHost::GPUTPCNNClusterizerHost(const GPUSettingsProcessingNNclusterizer& settings, GPUTPCNNClusterizer& clusterer, int32_t streamId)
 {
   OrtOptions = {
     {"model-path", settings.nnClassificationPath},
     {"device", settings.nnInferenceDevice},
-    {"device-id", std::to_string(settings.nnInferenceDeviceId)},
+    {"stream-id", std::to_string(streamId)},
     {"allocate-device-memory", std::to_string(settings.nnInferenceAllocateDevMem)},
     {"dtype", settings.nnInferenceDtype},
     {"intra-op-num-threads", std::to_string(settings.nnInferenceIntraOpNumThreads)},

GPU/GPUTracking/TPCClusterFinder/GPUTPCNNClusterizerHost.h

@@ -37,7 +37,7 @@ class GPUTPCNNClusterizerHost
 {
  public:
   GPUTPCNNClusterizerHost() = default;
-  GPUTPCNNClusterizerHost(const GPUSettingsProcessingNNclusterizer&, GPUTPCNNClusterizer&);
+  GPUTPCNNClusterizerHost(const GPUSettingsProcessingNNclusterizer&, GPUTPCNNClusterizer&, int32_t = 0);
 
   void networkInference(o2::ml::OrtModel model, GPUTPCNNClusterizer& clusterer, size_t size, float* output, int32_t dtype);