Package level-zero on the conda-forge distribution

[ogrisel]

At the moment it is not possible to install dpctl, dpnp and/or numba-dpex from conda-forge with Intel GPU support because level-zero has not been packaged for that distribution.

Would you be interested in contributing a build recipe to conda-forge under a report such as:

• https://github.com/conda-forge/intel-level-zero-feedstock

Here is the documentation on the procedure to submit a new build recipe (a.k.a. feedstock) on the conda-forge continuous integration platform:

• https://conda-forge.org/docs/maintainer/adding_pkgs.html

[ogrisel]

Note, there is a community supported build of intel-compute-runtime to get GPU support via the OpenCL runtime at:

• https://github.com/conda-forge/intel-compute-runtime-feedstock

However the last few releases have failed builds and nobody tried to maintain them.

I will open a dedicated issue for this.

[eero-t]

Compute runtime is a bit particular about its dependencies. E.g. different LLVM versions have different level of support from it, with LLVM v13 basically not working, and v11 and v14 being best supported, and everything else in compute stack needing to be built with the same LLVM version (due to LLVM breaking its internal ABI between versions).

It may be easier to base conda support for already existing major distro packages. Both Debian/Ubuntu and Fedora provide now L0 support (see #73), in case one does not want to use Intel driver repositories.