FCHL19 with periodic boundary conditions and revised OpenMP parallelization

[kvkarandashev]

Adds support for periodic boundary conditions when generating FCHL19 representations. Adds the verification script for checking PBC works correctly (temporarily placed into tests).

facsf.f90 revised to avoid using REDUCTION construct in FCHL19's OpenMP parallelization while minimizing the additional CPU and memory usage compared to a serial implementation. Admittedly, I have doubts about how necessary including "!$OMP CRITICAL" was, but I cannot see a better alternative.

Unfortunately, I had trouble getting qmllib to work with OpenMP for some reason, hence correctness of fascf.f90's parallelization was verified by using the same file in my fork of original qml.

[charnley]

Unfortunately, I had trouble getting qmllib to work with OpenMP for some reason

can you expand on this? Did you have to remove the compiler flags or notice that it wasn't allocating all the resources?

[kvkarandashev]

To compile with OpenMP I edited several lines in _compile.py as follows:

COMPILER_FLAGS = ["-O3", "-fopenmp", "-m64", "-march=native", "-fPIC",
                     "-Wno-maybe-uninitialized", "-Wno-unused-function", "-Wno-cpp"]
extra_flags = ["-lgomp", "-lpthread", "-lm", "-ldl"] + COMPILER_FLAGS

flags = ["-L/usr/lib/", "-lblas", "-llapack"] + extra_flags

After compiling the code I compared its performance on a test problem to performance of the qml fork with the same fascf.f90 file. qmllib never seemed to use more than one core, as supported by the timings:
OMP_NUM_THREADS=1:
qml: 0.7533 mins
qmllib: 0.8178 mins
OMP_NUM_THREADS=20:
qml: 0.1390 mins
qmllib: 0.8145 mins

[kvkarandashev]

Found a mistake in the way gradients for PBC were calculated and verified.
Fixed the verification script and added verification via finite difference.

[charnley]

qmllib is finally updated to work with numpy2.0 and openmp flags are fixed.

• Can you pull from main?
• Use pathlib.Path instead of os and glob