I read up on where we need primitive vs characteristic reconstruction with VL+CT. We can use either but unlike CTU+CT we don’t need to characteristic tracing step to update the hydro variables a half time step before running them through the Riemann solver. So the plan is to implement all the Cholla reconstruction methods for MHD but for now I’m just doing Piecewise Constant Method (PCM) and once everything else works I’ll add linear and parabolic methods.
I added MHD support to PCM and refactored the current PCM kernel to achieve roughly an 60% performance improvement by reducing the number of reads from device memory. Now the 3D Van Leer integrator has MHD support up through the first Riemann solve. Now that I’ve actually been working in the integrator I decided to format the interface state and fluxes differently than I had originally intended when writing the HLLD solver. This required a minor refactor and simplification of the HLLD solver which went pretty easily. The process of updating the tests for the HLLD solver was a lot trickier. With the new data format several of the things I was doing in the testing didn’t really make sense to do anymore but it took a lot of care, and fighting with
std::rotate to remove them cleanly and correctly. However, the testing code is much shorter and clearer now and I found a few logical bugs and a minor memory leak while I was at it.
Git Hash & Compile Time Macros
I made it so that when Cholla knows what commit hash it was compiled at and which preprocessor macros were passed at compile time with the
-D argument. I’ve got it working and writing that information out to the terminal, log files, ASCII, and HDF5 output. PR #138
- Started on Beck 2015
- Finished the git introduction I started two weeks ago