Changes for synchronization and consistency. No performance loss.

Author: ChSonnabend

Common/ML/src/OrtInterface.cxx

@@ -271,18 +271,18 @@ void OrtModel::inference(I* input, size_t input_size, O* output)
   std::vector<int64_t> inputShape{input_size, (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 * mInputShapes[0][1] * sizeof(Ort::Float16_t), inputShape.data(), inputShape.size());
+    inputTensor = Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(input), input_size * mInputShapes[0][1], inputShape.data(), inputShape.size());
   } else {
-    inputTensor = Ort::Value::CreateTensor<I>(pImplOrt->memoryInfo, input, input_size * mInputShapes[0][1] * sizeof(float), inputShape.data(), inputShape.size());
+    inputTensor = Ort::Value::CreateTensor<I>(pImplOrt->memoryInfo, input, input_size * mInputShapes[0][1], inputShape.data(), inputShape.size());
   }
   (pImplOrt->ioBinding)->BindInput(mInputNames[0].c_str(), inputTensor);
 
   std::vector<int64_t> outputShape{input_size, mOutputShapes[0][1]};
   Ort::Value outputTensor = Ort::Value(nullptr);
   if constexpr (std::is_same_v<O, OrtDataType::Float16_t>) {
-    outputTensor = Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(output), input_size * mOutputShapes[0][1] * sizeof(Ort::Float16_t), outputShape.data(), outputShape.size());
+    outputTensor = Ort::Value::CreateTensor<Ort::Float16_t>(pImplOrt->memoryInfo, reinterpret_cast<Ort::Float16_t*>(output), input_size * mOutputShapes[0][1], outputShape.data(), outputShape.size());
   } else {
-    outputTensor = Ort::Value::CreateTensor<O>(pImplOrt->memoryInfo, output, input_size * mOutputShapes[0][1] * sizeof(float), outputShape.data(), outputShape.size());
+    outputTensor = Ort::Value::CreateTensor<O>(pImplOrt->memoryInfo, output, input_size * mOutputShapes[0][1], outputShape.data(), outputShape.size());
   }
   (pImplOrt->ioBinding)->BindOutput(mOutputNames[0].c_str(), outputTensor);
 

GPU/GPUTracking/Global/GPUChainTrackingClusterizer.cxx

@@ -630,23 +630,23 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
     mRec->runParallelOuterLoop(doGPU, numLanes, [&](uint32_t lane) {
       nnApplications[lane].init(nn_settings);
       if (nnApplications[lane].modelsUsed[0]) {
-        SetONNXGPUStream((nnApplications[lane].model_class).getSessionOptions(), lane + numLanes, &deviceId);
+        SetONNXGPUStream((nnApplications[lane].model_class).getSessionOptions(), lane, &deviceId);
         (nnApplications[lane].model_class).setDeviceId(deviceId);
         if (nnApplications[lane].model_class.getIntraOpNumThreads() > maxThreads) {
           nnApplications[lane].model_class.setIntraOpNumThreads(maxThreads);
         }
         (nnApplications[lane].model_class).initEnvironment();
       }
       if (nnApplications[lane].modelsUsed[1]) {
-        SetONNXGPUStream((nnApplications[lane].model_reg_1).getSessionOptions(), lane + 2*numLanes, &deviceId);
+        SetONNXGPUStream((nnApplications[lane].model_reg_1).getSessionOptions(), lane, &deviceId);
         (nnApplications[lane].model_reg_1).setDeviceId(deviceId);
         if (nnApplications[lane].model_reg_1.getIntraOpNumThreads() > maxThreads) {
           nnApplications[lane].model_reg_1.setIntraOpNumThreads(maxThreads);
         }
         (nnApplications[lane].model_reg_1).initEnvironment();
       }
       if (nnApplications[lane].modelsUsed[2]) {
-        SetONNXGPUStream((nnApplications[lane].model_reg_2).getSessionOptions(), lane + 3*numLanes, &deviceId);
+        SetONNXGPUStream((nnApplications[lane].model_reg_2).getSessionOptions(), lane, &deviceId);
         (nnApplications[lane].model_reg_2).setDeviceId(deviceId);
         if (nnApplications[lane].model_reg_2.getIntraOpNumThreads() > maxThreads) {
           nnApplications[lane].model_reg_2.setIntraOpNumThreads(maxThreads);
@@ -1039,7 +1039,6 @@ int32_t GPUChainTracking::RunTPCClusterizer(bool synchronizeOutput)
             }
             LOG(info) << "[NN CF] Apply NN (fragment " << fragment.index << ", lane: " << lane << ", sector: " << iSector << "): filling data " << time_fill << "s ; networks: " << time_networks << "s ; clusterizer: " << time_clusterizer << "s ; " << clusterer.mPmemory->counters.nClusters << " clusters, " << acceptedClusters << " accepted. --> " << (int32_t)clusterer.mPmemory->counters.nClusters / (time_fill + time_clusterizer) << " clusters/s";
           }
-          TransferMemoryResourcesToHost(RecoStep::TPCClusterFinding, &clustererNN, lane);
 #else
           GPUFatal("Project not compiled with neural network clusterization. Aborting.");
 #endif