rfgpu

rfgpu is a CUDA-based library for GPU interferometric gridding and imaging for the realfast project. It is incorporated into a more complete processing pipeline in rfpipe.

The low-level GPU routines are written entirely in CUDA/C++, and a python interface is provided using pybind11. This is compatible with both python 2.7 and 3.x.

Installation

Build and installation is based on a simple Makefile. Two "header-only" C++ libraries are required before rfgpu can be built. These should be downloaded in the extern/ subdirectory via the provided download script:

cd extern
./download.sh

After CUB and pybind11 are dowloaded, the library can be built and installed in the src directory. This will likely require specifications of compiler and install path. For example, at NRAO a typical install into the realfast conda environment would use:

cd src
make CXX=/opt/local/compilers/gcc-5/bin/g++
make install PREFIX=/home/cbe-master/realfast/anaconda/envs/development3

At other sites, you may also need to specify the CUDA location via the CUDA_DIR variable.

Dependencies

• A C++-11 compatible C++ compiler (eg, GCC v4.8 or later)
• CUDA, including the cuFFT and cuSPARSE libraries
• CUB
• pybind11