Pramod Kumbhar

#### Here are some notes from various experimentation / debugging attempts to get OpenMP offload working with LLVM v13.0. * On [Slack thread](https://bluebrainproject.slack.com/archives/C02CA9MDT99/p1639077956058700?thread_ts=1639077255.058400&cid=C02CA9MDT99), Olli reminded small reproducer from NERSC Hackathon...

### Summary of the IBM XL Compiler Issues reported on llvm openmp-dev mailing list: https://lists.llvm.org/pipermail/openmp-dev/2022-January/004276.html * Issue/Question # 1 : Shared library as well as Static library works * Issue/Question...

* For linux/macOS wheels, we haven't constrained numpy version for Python 3.10 https://github.com/neuronsimulator/nrn/blob/7a075811d12670d057aa6e9026f6c0698c075ce8/packaging/python/build_wheels.bash#L56 * For Windows installer we have used numpy 1.21.5 which was released in Dec 2021! https://github.com/neuronsimulator/nrn/blob/7a075811d12670d057aa6e9026f6c0698c075ce8/ci/win_install_deps.cmd#L32 So...

Edit: for windows it will be v1.21.3 because that's when windows wheel was made available.

Other goals/thoughts related to this work: * Memory efficient mode for non-interactive, batch execution? For large model simulations, memory footprint is one of the first big contraint. How the data...

We have seen that nrn_setup phase2 has race conditions! look at global hash tables which are not protected by OMP critical section.

Just for the record, following here as well https://github.com/HumanBrainProject/olfactory-bulb-3d/issues/20#issuecomment-1131569787

Is this partially improved by https://github.com/BlueBrain/CoreNeuron/pull/595?

> I lean toward the latter as it avoids an increase in file transfer size and I believe it is also possible to have better performance by modifying the NetCon...

> but one of the reasons for the c api to all this in NEURON is to allow dynamic loading of libmpi (if it exists). @nrnhines : but even if...