JuliaPackaging
[AOCL] Add a recipe for v5.1.0
Thanks @amd-agodbole for working in this!
cc @imciner2 @giordano @ViralBShah
I was going to give this a look over tomorrow, because I want to see if we can use the install command instead of cp for the files.
@aelobdog Are binaries available for Windows - or are these linux only?
There are Windows binaries available, but the ones on the AMD site are some sort of self-extracting installation tool that we can't just extract in our sandbox, so we can't get the binaries from them.
We either need AMD to provide a tarball redistributable package for the Windows libraries (like is done for Linux), or I guess we could repackage the three conda-forge AOCL packages to get the libraries for Windows - but I think those are only one of the versions (and I can't tell if it is LP64 or ILP64).
The Conda-forge packages are https://anaconda.org/conda-forge/aocl-blas/files, https://anaconda.org/conda-forge/aocl-lapack/files and https://anaconda.org/conda-forge/aocl-utils/files.
I looked at the feedstock for the conda forge package, and they just build the library themselves actually - but only one version of it (whatever integer size is the default for the CMake build system): https://github.com/conda-forge/aocl-blas-feedstock/blob/main/recipe/build_aoclblas.bat. So I don't think we can just repackage the conda forge one (I think it might be LP64 only actually).
Should we also be building from source - if source is available?
The source is available, and there was a start at a recipe building from source in https://github.com/JuliaPackaging/Yggdrasil/pull/11896. The thing with that source build is I was worried about the performance of the library versus compatibility with the C++ runtime library (needing too new a library), because we would need the newer GCC versions to get the architectures (GCC 12 is Zen 3, GCC 14 is Zen 5). That said, I think it could be worth doing some profiling of this build versus the source build in #11896 to see if there are any performance differences on the newer architectures (especially Zen 5) - since I think most of the critical parts would be in assembly kernels, that might not acutally be a large issue.