gpuRcuda Roadmap

gpuRcuda 1.0.x

Planned Features

• nvMatrix - gpuRcuda version of gpuMatrix from gpuR referencing explicit NVIDIA

  • Will NOT inherit from gpuMatrix. This was chosen because it would require copying data to the OpenCL memory space. As such, it will be up to the user to convert to the gpuR equivalent if OpenCL is desired. A helper function should be created to facilitate this.

• cudaMatrix gpuRcuda version of vclMatrix from gpuR

• Use thrust backend.

  • This will allow users to install and use this package even if they don't have a NVIDIA GPU installed thereby facilitating greater portability.

• Use of cublas will likely be required for linear algebra routines

  • This will be implemented in gpuRcublas
  • It is not included here in order to allow the package to be installed and used on systems without CUDA
  • As such, BLAS 2 and BLAS 3 operations will likely not be included in the package until manual implementations can be created with thrust (e.g. matrix multiplication)

gpuRcuda 1.1.x

• Accept custom CUDA kernels (will require coordinating with nvcc)