CUDA SIPG matrix-free

This code solves the Poisson problem with Dirichlet border condition with the Symmetric Interior Penalty Galerkin method.

The code has been written in CUDA C++.

Requisites

• a CUDA device with compute capability 2.0 or higher.
• CUDA Toolkit 5.0 or higher.
• cmake 2.8 or higher.
• lapacke library is needed to run some preliminary validation tests on CPU, it is not necessary for the solution of the problem. The command sudo apt-get install liblapacke-dev installs it on Debian or Ubuntu.
• doxygen for build the documentation
• GCL library is needed in order to run the code on multiple GPUs in a MPI environment. GCL requires Boost library.

Build

    $ cd source-dir
    $ mkdir build
    $ cd build 
    $ cmake -D COMPUTE_CAPABILITY=sm_xx ..

Where sm_xx is the compute capability of your device, e.g. sm_20, sm_35... In order to employ GCL library, you need to set GCL_USAGE=ON and GCL_SOURCE_DIR.

Now, you can compile and run the binary with

    $ make binary-name
    $ ./binary-name

The command make help shows the full list of available binaries.

A Makefile in the source root directory builds the documentation with doxygen

    $ cd source-dir
    $ mkdir doc
    $ make

Now, you can find the generated documentation in doc directory.

Binary List

The most important binaries are

• sipg_2d_h_adaptivity_multigpu is the two-dimensional Poisson problem validation tests. It uses multiple GPUs.
• sipg_2d_h_adaptivity and sipg_2d_p_adaptivity are the two-dimensional Poisson problem validation tests. They use the GPU.
• sipg_sem_1d_class_validation_test is a one-dimensional Poisson problem preliminary test. It uses the CPU and lapacke.
• sem_1d_nitsche_bc_validation_test and sem_2d_nitsche_bc_validation_test are one and two-dimensional SEM basis validation tests. They use the CPU and lapacke.
• constant_dof_change_degree_and_noe, flux_kernels_performance_test, one_iteration_of_sem_sipg_solved_on_gpu_test, volume_kernel_performance_test are run with nvprof to test the performance.