Profiling managedCuda in Nsight or Visual Profiler

2
votes

I'm trying to profile an CUDA-application written in C# with managedCuda using either Nsight Visual Studio Edition or Visual Profiler. Both profilers work well with a plain C++ CUDA app. To test the profilers with managedCuda I want to profile the project "vectorAdd" in ManagedCudaSamples.

First I tried to use Nvidia Nsight Visual Studio Edition 5.0 integrated in VS 2013. I use the x64 Debug configuration. If I try to launch the app in "Application Control" in Nsight Performance Analysis I get an error message:

Analysis Session - Start Application Unable to launch 64-bit managed application '...\ManagedCudaSamples\vectorAdd\bin\x64\Debug\vectorAdd.exe'.

Additionally I tried to use Nvidia Visual Profiler 7.5 for profiling the same application. On running vectorAdd.exe nvprof console shows the following output:

==2944== NVPROF is profiling process 2944, command: ...\ManagedCudaSamples\vectorAdd\bin\x64\Debug\vectorAdd.exe ==2944== Warning: Some profiling data are not recorded. Make sure cudaProfilerStop() or cuProfilerStop() is called before application exit to flush profile data. ==2944== Generated result file: ...\nvvp_workspace\.metadata\.plugins\com.nvidia.viper\launch\7\api_2944.log

I'm new to CUDA and would be thankful for any advice how to profile managedCuda applications.

1
c#cudaprofilingmanaged-cuda

1 Answers

3
votes

You need to call CudaContext.ProfilerStop() just before you exit the application (or destroy the context) in order to flush the collected data to the profiler. The managedCuda samples don't include this call why the profiler doesn't see the collected info. This explains the second error you get.

And regarding the first error: In the release notes of Nsight 5.0 you can find a known issue:

  1. Managed applications built with the AnyCpu configuration are not supported. The target application must be built using either the Win32 or x64 configurations.

The VS-project for vectorAdd is always set to AnyCPU, regardless what the solution platform is, see the configuration manager of the managedCuda samples-solution to change that.