A CUDA/C++ implementation of the Discontinuous Galerkin method as presented in the book:
Nodal Discontinuous Galerkin Methods - Algorithms, Analysis, and Applications, Jan S. Hesthaven and Tim Warburton, Springer, 2008
The code, built by g++ 7.5.0 and CUDA 10.2, was tested on NVIDIA Volta GPU (CUDA Capability 7.2). It can be configured to run on GPU or CPU-only by the flag USE_CPU_ONLY defined in /src/config.h.
The code adopts the "one-thread-per-element" strategy for parallelization - each thread performs computations for a single element. Layouts of degree-of-freedoms (DOFs) in the solution vectors promote coalesced memory access. CUDA constant memory is used to store small-sized, but frequently accessed, data to improve performance.
The code contains only header files in the /src folder and there is no third-party dependencies except CUDA. A few example problems are provided in the /examples folder. To build and run these examples, go to the individual subfolder and run make. If your CUDA installation directory is different from /usr/local/cuda-10.2, change the path of CUDA_PATH accordingly in the makefile.
In the subfolder of each example, there is a PDF file describing the problem, the boundary and initial conditions, the analytical solutions (if any), the meshes, and the performance results.
This code focuses on 2D problems on unstructured triangular meshes. For 3D problems, see my other repository, rdg-and-fr, at https://github.com/hsongxa/rdg-and-fr.