Feature request: conda distribution

[shimwell]

Now that the OpenMC conda install includes DAGMC I'm thinking it might be a good time to start planning a conda forge package for the wonderful double-down

In the mid term we can include it in the conda install of DAGMC and then in the longer term double-down could be included in the conda install for OpenMC.

This would help bring the speed benefits of Embree and Double-Down to a wider audience and ease the installation process for users who are unable or don't want to compile

Embree3 is on conda forge here and can be installed with

conda install -c conda-forge embree3

Perhaps in the past we found the wrong Embree package as there are quite a few but it looks like Embree3 is up to date and meets the requirements of Double-Down

To do this we would need a new Conda-forge feedstock for Double-Down, happy to pitch in on this

[pshriwise]

Definitely on board for this! As we discussed offline, we'll need to be cautious about the ISA applied in the embree3 depending on the user's machine. I'm not sure if conda has a mechanism for this or not, but I'll look into this further when I have time.

[jon-proximafusion]

Perhaps one solution is to provide different conda packages for different ISA options a bit like we do for openmc with mpi on or off

[jon-proximafusion]

I think we would need DAGMC to use cpuid to determine if the instruction set exits on the CPU architecture

[shimwell]

I understand gcc can detect the CPU architecture with attribute dispatch and compile for both then the resulting single executable can determine which one to use at run time

[matthewfeickert]

@shimwell @pshriwise following conda-forge/staged-recipes#32181, double-down now exists on conda-forge: https://github.com/conda-forge/double-down-feedstock. Can you take a look at the state of the double-down conda package as is and let me know if it will need mpi build variants in a similar manner to https://github.com/conda-forge/dagmc-feedstock and https://github.com/conda-forge/openmc-feedstock?

[shimwell]

Thanks for all your work in this area

My (potentially wrong) understanding is that double down with embree would become a new variant in the matrix of existing variants. So yes if all the matrix of combinations build there would a a mpi_ddembree and nompi_ddembree variant. But also if this is a blocker then feel free to skip as there are some people who don't use MPI but would benefit from the embree speedup.

This also might be solved but I think of of the potential issues is that one could install a ddembree variant without having AVX-512 or similar and I'm not sure if the dd+embree_dagmc gracefully and efficiently falls back to using the available instructions. I guess old x86 and arm cpus fit into this category. If dd+embree does gracefully fallback to using whatever the cpu supports then we wouldn't need a variant at all and could include dd+embree with every combination.

[shimwell]

I wanted to mention this as well in case the build script used is of any use for the conda efforts. This wheel includes DD + Embree in the DAGMC build https://github.com/shimwell/wheels