The separate repo for testing data has been implemented and tested and is connected to the main Cholla repo as a submodule. Everything seems to work fine, I’m just waiting for Git LFS to be installed on the Pitt CRC machines before I submit a PR.
I started on implementing the MHD time step calculations but partway through discovered some inconsistencies with how the time step is calculated. On the first time step it was done entirely on the CPU and on subsequent time steps there were both global host and device arrays being allocated and deallocated each time step by the integrator, plus the GPU reduction is only done partially on the GPU.
I removed the CPU time step calculation in the first time step and replaced it with the GPU calculation. I also removed the global time step arrays and replaced them with arrays that are allocated and deallocated each time step within the time step function. My next step is to write a GPU reduction function so that the entire reduction can be done on the GPU rather than on both GPU and CPU. I’m basing my reduction code off of this presentation by Mark Harris at NVIDIA and this blog post by Sean Bone, which is basically just an easier to read version of Mark Harris’ slides.
- Implemented a basic, single rank, particle test now that the duplicate ID issue has been resolved
- Fixed an issue where the systemTestRunner class assumed that a fiducial HDF5 file was present even if the user indicated otherwise
- Added an assertion in the systemTestRunner class to make sure cholla launched properly
- Prepped for and led a git workshop