Is Alea relied on cuDNN 5.1?

[D-Walter]

I can not run the project DigitsCudaTest, error thrown noticed that 'Couldn't open CUDA library 'cudnn64_5''
Should i use cuDNN 5.1 or cuDNN 5.0 with CUDA 8.0?

[Sergio0694]

Hello, thank you for your interest in this project 😊
So far I've tested Alea with CUDA 9.0, and it looks like the best way to avoid issues is to just stick to the default cudnn64_5.dll file, as the 64_7 dll isn's properly supported yet (this is really up to the Alea developers, you can see I've opened a SO question here)

In order to run the library, make sure to:

• Install CUDA 9.0
• Set the ALEA_CUDA_VERSION set to 9.0 in the system environment variables (see here)
• Set the CUDA_PATH environment variable to the CUDA folder on your system. It should be something like: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0
• Download the cudnn64_5.dll file from NVIDIA and paste it in the bin folder of the CUDA\v9.0 directory, Alea will load it from there

Let me know if that works for you!

[D-Walter]

It works after I stick 64_5.dll. Thanks for your help!

[D-Walter]

I'm so sorry to disturb you again.

Though it works properly but i cannot handle the exceptions System.OutOfMemoryException noticed me that The array dimension exceeds the range of support at Alea.IL.Microsoft.Cci.Pdb.MsfDirectory..ctor(PdbReader reader, PdbFileHeader head, BitAccess bits).
It seems not fatal because it still keeps running, but i am worried whether it may cause some potential error.

Thanks for your patience.

P.S. I still use the origin project DigitsCudaTest

[Sergio0694]

No worries, happy to help!
If you get that error when starting the training, and then everything works fine (ie. if that only happens when the GPU kernels are JIT-compiled for the first time), then that's a known bug with Alea, when compiling the project using Visual Studio 17.

I've already contacted the Alea developers about that, they said it's caused by a change in the way Visual Studio handles .pdb files, which was changed from VS15 to VS17.
Anyways, that shouldn't cause any issues, and you can only see that if you directly reference the library project. That is, that error is visible for you because you're using the sample project, which has a link to the library project in the same solution, but if you create a new project in a new solution, and use the NuGet version of the NeuralNetwork.NET library, that error should not be visible at all 👍

[D-Walter]

Thx a lot, i got it