This is a snapshot of the CUDA implementations of the matrix-free functionality of Deal.II, as of 2016-03-14.
It includes the general framework in the directory matrix_free_gpu, and an application poisson.cu which demonstrates the use of the framework. The application solves the Poisson equation with variable coefficient on a unit cube/square, using CG with a Chebyshev preconditioner. There is support for general grids with non-uniformly refined grids with hanging nodes although this is commented out in this version. bmop.cu deals with the exact same problem, but instead of solving it, it benchmarks the operator application by performing 100 successive applications. The subdirectory opbm-cpu implements the corresponding benchmark on CPU.
To build the application (require CUDA 7.5 and a CUDA-capable GPU):
mkdir build
cd build
cmake -D CMAKE_CXX_FLAGS="--std=c++11 -march=native -DDEGREE_FE=4 -DDIMENSION=3 -DMATRIX_FREE_UNIFORM_MESH" -DCMAKE_BUILD_TYPE=RELEASE -DDEAL_II_DIR=/path/to/dealii ..
make poisson
./poisson