cusp
Library of CUDA Kernels for Signal Processing
Introduction
Access to signal processing and mathematical routines implemented in Python has been implemented in several solid libraries (cuPy, cuSignal), but utilizing this functionality from c++ applications such as GNU Radio which uses its own memory management has proved challenging. The purpose of cusp (CUDA Signal Processing) is to provide a library of CUDA kernels that are wrapped in a manner that is easily callable from c++. This library is analogous to VOLK but with CUDA hardware acceleration.
Dependencies
Installation
Generally accomplished in the standard meson/ninja way:
meson setup build
cd build
ninja
ninja install
For Ubuntu 20.04 with CUDA installed via apt (CUDA v10.1), it is necessary to tell meson where to find CUDA
Also, for installing into a newsched prefix, set the --prefix and --libdir
CUDA_ROOT=/usr/ meson setup build --buildtype=debugoptimized --prefix=[PREFIX_DIR] -Denable_cuda=true --libdir=lib
Also, if CUDA is installed from tar packages, be sure to add the CUDA library to the LIBRARY_PATH, e.g. in .bashrc:
export LIBRARY_PATH=/usr/local/cuda-10.2/lib64:$LIBRARY_PATH