The solver needs a C++17 compiler, CMake 3.15 or newer, and the SUNDIALS library.

cd radcluster_code
cmake -S cpp_utils -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release

The binary is written to build, or to build/Release with multi-configuration generators. The bridge cpp_bridge.py looks for it in both places. The driver notebook builds the solver on first use, and builds it again when the platform changes.

Optional dependencies

Each optional dependency is detected at configure time.

Dependency	Enables	If missing
LAPACK/BLAS	the Woodbury preconditioner	it is compiled out, and the solver falls back to the Jacobi preconditioner at run time
OpenMP	the parallel right-hand side of the active-window mode	that mode runs on one thread
SuiteSparse/KLU, with SUNDIALS built with -DENABLE_KLU=ON	the sparse-direct linear solver	that option stops at run time with a diagnostic

Compiler flags

The optimization flags are -O3 -march=native, or /arch:AVX2 /GL /fp:precise with Microsoft Visual C++.

Note that fast-math is deliberately not enabled. It alters IEEE-754 semantics, and it can make the error estimator of CVODE diverge on these stiff systems.

On Windows the build copies the required run-time libraries next to the executable.

See also: Container image, Running on the Hoffman2 cluster