You are actively working on offloading some computations to GPUs using CUDA. What about OpenCL? Is it interesting for the project and the community? Does the current infrastructure allow one to switch between CUDA and OpenCL (e.g., to use OpenCL kernels if CUDA-compatible devices are not available)? Are architectural details on CUDA support available? What is the effort required to integrate kernels, excluding kernel implementation? We are thinking about contributing OpenCL kernels, so we are trying to estimate the required effort.
You are actively working on offloading some computations to GPUs using CUDA. What about OpenCL? Is it interesting for the project and the community? Does the current infrastructure allow one to switch between CUDA and OpenCL (e.g., to use OpenCL kernels if CUDA-compatible devices are not available)? Are architectural details on CUDA support available? What is the effort required to integrate kernels, excluding kernel implementation? We are thinking about contributing OpenCL kernels, so we are trying to estimate the required effort.