The embrio of this development started from the algorithms in the book: Numerical Analysis: Mathematics of Scientific Computing, by David Kincaid & Ward Cheney.
Starting from sources cordialy provided by the authors, I was converging towards the implementations in T-Flows (https://github.com/DelNov/T-Flows). Initial motivation was to check wheather fill-in makes sense for ICCG solvers, but recenlty I realized that they might also be useful to dive into GPU programming.
For GPU programming, I use the OpenACC paradigm and a suite of tools which comes from Nvidia's CUDA toolkit and HPC-SDK. To be more precise, I used the following combination:
- CUDA toolkit 11.1 (https://developer.nvidia.com/cuda-toolkit)
- Nvidia HPC-SDK 20.7 (https://developer.nvidia.com/hpc-sdk)
(These two don't work in perfect harmony, but if LD_LIBRARY_PATH points to the directory from CUDA toolkit and not to the libraries which come with HPC-SDK, all seems to work fine. What will be with next release(s), no one can tell.)
A couple of hints for compilation with OpenACC.
- Compile with
-acc -Minfo=acceloption - For profiling, use:
nsys profile --trace=openacc ./SFS <option>and laterlaunch "nsight-systo load the profile file.
Set environment variable NVCOMPILER_ACC_NOTIFY to zero if you want to supress messaags while a program runs on GPUs