Found the stream allocation issue. Now starting optimizations

Author: ChSonnabend

Common/ML/include/ML/OrtInterface.h

@@ -55,6 +55,7 @@ class OrtModel
   void initEnvironment();
   bool isInitialized() { return mInitialized; }
   Ort::SessionOptions& updateSessionOptions();
+  Ort::MemoryInfo& updateMemoryInfo();
   void setIO();
 
   virtual ~OrtModel() = default;

Common/ML/src/OrtInterface.cxx

@@ -40,6 +40,11 @@ Ort::SessionOptions& OrtModel::updateSessionOptions()
   return pImplOrt->sessionOptions;
 }
 
+Ort::MemoryInfo& OrtModel::updateMemoryInfo()
+{
+  return pImplOrt->memoryInfo;
+}
+
 void OrtModel::initOptions(std::unordered_map<std::string, std::string> optionsMap)
 {
   pImplOrt = new OrtVariables();

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDA.cu

@@ -697,16 +697,11 @@ void GPUReconstructionHIP::SetONNXGPUStream(Ort::SessionOptions& session_options
   const auto& api = Ort::GetApi();
   // api.GetCurrentGpuDeviceId(deviceId);
   OrtROCMProviderOptions rocm_options;
-  LOG(info) << "Creating ROCm provider options";
-  // rocm_options.has_user_compute_stream = 1; // Indicate that we are passing a user stream
-  // LOG(info) << "Setting user compute stream";
-  // rocm_options.user_compute_stream = &(mInternals->Streams[stream]);
-  // LOG(info) << "Stream is set with streamId " << stream << " and reference " << &(mInternals->Streams[stream]);
+  rocm_options.has_user_compute_stream = 1; // Indicate that we are passing a user stream
+  rocm_options.user_compute_stream = mInternals->Streams[stream];
   session_options.AppendExecutionProvider_ROCM(rocm_options);
-  LOG(info) << "Appending ROCm provider options";
   // OrtSessionOptionsAppendExecutionProvider_ROCM(session_options, *deviceId);
   // api.ReleaseROCMProviderOptions(rocm_options);
-  LOG(info) << "Releasing ROCm provider options";
 }
 
 #endif // GPUCA_HAS_ONNX

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -988,7 +988,6 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
 
             auto stop0 = std::chrono::high_resolution_clock::now();
             auto start1 = std::chrono::high_resolution_clock::now();
-            LOG(info) << "ONNX stream set. Device ID is " << deviceId << " for stream " << lane;
             nnApplication.networkInference(nnApplication.model_class, clustererNNShadow, iSize, clustererNNShadow.modelProbabilities, clustererNNShadow.nnInferenceInputDType, deviceId);
             if (nnApplication.model_class.getNumOutputNodes()[0][1] == 1) {
               runKernel<GPUTPCNNClusterizerKernels, GPUTPCNNClusterizerKernels::determineClass1Labels>({GetGrid(iSize, lane), krnlRunRangeNone}, iSector, clustererNNShadow.nnInferenceInputDType, withMC, batchStart); // Assigning class labels