Description

Currently, updates to weight window values are computed serially in OpenMC. This is not a problem on smaller models and/or if using a very coarse mesh, but for larger models at finer mesh resolutions (where you can easily have 100 million or more bins), the weight window update times can become pretty significant in serial.

This PR parallelizes the weight window update function across threads. As the original function made heavy usage of xtensor views and vector operations, this parallelization required a few additional lines so as to convert the vector calls to iterative for loops that could be parallelized with OpenMP.

Note that this PR should not be changing any results -- it is only intended to speed things up via parallelization.

Impact

I did some performance analysis on the JET model with this PR along with the last several optimizations on WW generation with FW-CADIS. Runtimes are given for total runtime (of the adjoint solve) on a single CPU node (for 5 inactive + 5 active batches, with extremely low ray density, and a 10cm resolution weight window mesh). Below is the summary:

As can be seen above, the new parallelization scheme gives an additional 2x overall performance boost.

Checklist

• I have performed a self-review of my own code
• I have run clang-format (version 15) on any C++ source files (if applicable)
• I have followed the style guidelines for Python source files (if applicable)
• I have made corresponding changes to the documentation (if applicable)
• I have added tests that prove my fix is effective or that my feature works (if applicable)