To compile and run CUDA code, you need to use an external Kokkos installation that has been configured with CUDA support. This, in turn, requires a CUDA-enabled GPU, appropriate drivers, the CUDA toolkit, and the nvcc compiler. Unlike other libraries, you need special hardware and compiler to enable CUDA. Because the hardware is always evolving, older GPUs do not support all the capabilities of newer ones. In order to use CUDA with deal.II, you will need your GPU to have compute capability 3.5 or higher. Independently from the GPU itself, you also need a version of CUDA recent enough. deal.II supports CUDA 10.2 and higher. Finally to be able to configure deal.II, you will need CMake 3.9 or higher.
To configure deal.II with CUDA use the following option:
-DDEAL_II_WITH_CUDA=ONCUDA versions prior to 11.0 don't support C++17 or higher. You will have to make sure that C++17 is disabled when using earlier versions.
Several MPI implementations are able to perform MPI operations with data located in device memory directly without the need to copy to CPU memory explicitly first. This feature is commonly known as "CUDA-aware MPI". In case deal.II is compiled with MPI support and the MPI implementation supports this feature, you can tell deal.II to use it by configuring with
-DDEAL_II_WITH_CUDA=ON -DDEAL_II_WITH_MPI=ON -DDEAL_II_MPI_WITH_DEVICE_SUPPORT=ONNote, that there is no check that detects if the MPI implementation really is CUDA-aware. Activating this flag for incompatible MPI libraries will lead to segmentation faults in MPI calls.
Using CUDA in combination with architecture-specific C++ compiler flags
-march=native is known to be fragile and there might be
compatibility issues with other libraries, e.g. using CUDA 10.1 with
-DDEAL_II_CXX_FLAGS=-march=native results in compile time
/usr/lib/gcc/x86_64-linux-gnu/7/include/avx512fintrin.h(11265): error: identifier "__builtin_ia32_scalefsd_round" is undefined /usr/lib/gcc/x86_64-linux-gnu/7/include/avx512fintrin.h(11274): error: identifier "__builtin_ia32_scalefss_round" is undefinedSince vectorization in VectorizedArray is disabled when compiling with CUDA support anyway, it is recommended to drop the compile flag in that case.