Why choose llvm-openmp instead of intel-openmp on Windows (e.g. MKL)?

[huweiATgithub]

Comment:

#98 switches from intel-openmp to llvm-openmp.

I am having an external library that built with intel oneAPI (thus with MKL and intel-openmp). Therefore, when used together with conda managed environment, errors occur due to incompatibility between two openmp.

So, I wonder what is the reason for switching to llvm-openmp? Is it possible to release both variants?

[jacobwilliams]

See also #108

[napetrov]

99.9% of conda-forge ecosystem is based on llvm-openmp thus there been issues with enabling MKL support with some of those.

On runtime size - llvm and intel runtimes are identical, so the only difference between them happens at build time. So with MKL itself - it's being built with intel-openmp but can work with any of those runtimes.

So the only difference is package dependency in metadata. technically it can be feasible to create openmp package similar to blas and use different backend. @isuruf - thoughts on this?

[jacobwilliams]

It seems optimistic to think we can swap out intel-openmp for llvm-openmp with no issues whatsoever. What guarantees do we have that they will be compatible or support the same OpenMP standard? When Intel updates their compiler, they will have a page like this showing what features are supported:

https://www.intel.com/content/www/us/en/developer/articles/technical/openmp-features-and-extensions-supported-in-icx.html

Does llvm-openmp always shadow that? It seems unlikely right?

I would like to see the ability to specify to use one or the other. If I build with the intel stack, then i want to use the intel stack.

[isuruf]

With https://github.com/conda-forge/intel_repack-feedstock/pull/112/files#diff-f3725a55bf339595bf865fec73bda8ac99f283b0810c205442021f29c06eea9aR120 we can switch between versions (not limited to llvm nor intel). PRs to do the same for 2024.x branch and 2025.x branch are welcome.