Aditya

Enable changing the residual tolerance after a batch solver is created

2

This adds an interface to change the residual tolerance for subsequent runs of a solver, so that a new factory does not have to be created just for changing the...

Currently, relative residual tolerance in batched solvers is based on ||r||/||b||. This makes sense for zero initial guess. However, the more general criterion would be based on ||r||/||Ax_0 - b||.

Cache-line aligned allocation for OpenMP

5

Currently, the OpenMP `raw_alloc` just uses `malloc`. Typically we would want to have allocated memory blocks aligned to cache-line boundaries, typically 64 bytes. Two things would be needed: - [...

Need to add an interface to a mesh partitioner for MPI runs

1

Currently, for getting multi-process performance, the cells in the mesh need to pre-ordered by partition index. An interface with Scotch is envisaged to remove this requirement and partition the mesh...

In implicit solves, reordering does not work with multi-process runs. The solver diverges with NaN/Inf in the first iteration. This is related to the fact that only "serial" or local...

When any case is run with implicit time-stepping using more than one MPI rank, using the BAIJ matrix type for the Jacobian matrices results in a runtime error by PETSc....

Needless copies related to gradient blocks in residual computation

2

Fix needless copies in * Weighted least-squares reconstruction: final assignment of gradients to output variables * ~~viscous flux computation: before passing for conversion to primitve-2~~ * ~~MUSCL Van-Albada reconstruction~~

[BUG] Building from source with PyTorch backend on AMD system fails

4

### Bug summary The build system fails to find `libamdhip64.so` despite setting `ROCM_PATH` and `ROCM_ROOT`. It searches for `/opt/rocm/lib/libamdhip64.so`, whereas `ROCM_ROOT` is `/opt/rocm-6.0.0`. I cannot create a sym-link to `/opt/rocm`....