CODEWITHSUNDEEP
rcudamicrosoft-r
user3426389
user3426389

Reputation: 23

CUDA Support with MicrosoftML in R

I want to execute the rxNeuralNet function from the MicrosoftML Package for R with the option acceleration = "gpu". However, R fails to enable the GPU accelaration as it prints the following message:

Beginning processing data.
Failed to initialize CUDA runtime. Possible reasons:
1. The machine does not have CUDA-capable card. Supported devices have compute capability 2.0 and higher.
2. Outdated graphics drivers. Please install the latest drivers from http://www.nvidia.com/Drivers.
3. CUDA runtime DLLs are missing, please see the GPU acceleration help for the installation instructions.

I do have a Nvidia GPU with the latest drivers including CUDA 8.0 installed. I also copied the required dll files into the proper as stated in the help document. The target directory for the dlls is: 

C:/Program Files/Microsoft/R Client/R_SERVER/library/MicrosoftML/mxLibs/x64

The dlls that I copied into this directory were:

cublas80_65.dll cudart80_65.dll cusparse80_65.dll cudnn80_65.dll

Did I miss something? Unfortunately CUDA is still not working for rxNeuralNet. My OS:Win7_64Bit and GPU: Nvidia 950ti

Upvotes: 0

Views: 428

Answers (2)

Gleb Krivosheev
Gleb Krivosheev

Reputation: 11

Here are basic instructions. Make sure to follow those exactly.

For GPU acceleration there are additional manual setup steps are required:

  • Download and install NVidia CUDA Toolkit 6.5 (CUDA Toolkit).
  • Download and install NVidia cuDNN v2 Library (cudnn Library).
  • Find the libs directory of the MicrosoftRML package by calling system.file("mxLibs/x64", package = "MicrosoftML").
  • Copy cublas64_65.dll, cudart64_65.dll and cusparse64_65.dll from the CUDA Toolkit 6.5 into the libs directory of the MicrosoftML package.
  • Copy cudnn64_65.dll from the cuDNN v2 Library into the libs directory of the MicrosoftML package.

Upvotes: 1

Hong Ooi
Hong Ooi

Reputation: 57696

MicrosoftML only supports CUDA 6.5 right now. This is noted in ?rxNeuralNet, if somewhat obscurely. Make sure you download the correct versions from https://developer.nvidia.com/cuda-toolkit-65 and https://developer.nvidia.com/rdp/cudnn-archive.

Upvotes: 1

Related Questions