Closed pvelesko closed 1 year ago
Maybe retarget this to the develop branch where changes should be staged before we merge in master for new releases.
@rothpc please review
Since hipBLAS is a very frequently used library, it makes much more sense to be able to build it in a single step.
This replaces the previous version which was a fork of Sabojit's branch.
MKLShim is needed by several H4I libraries (three already, expected to be more as time goes on). We do not want to build a new MKLShim with each library, so keeping them separate is the desirable approach, not building it in a single step. Better choice is to do what the chipStar-Spack project is doing - making H4I-MKLShim and the other H4I libraries as subprojects.
We do not want to build a new MKLShim with each library
I'm sure there's a large portion of users who will only need one library so it's much easier to just clone and install a single repo. I don't see any advantages of having to go through a longer process.
making H4I-MKLShim and the other H4I libraries as subprojects
which was the purpose of this PR, thanks for merging :)
making H4I-MKLShim and the other H4I libraries as subprojects
which was the purpose of this PR, thanks for merging :)
I think you misunderstood my point. I don't think MKLShim should be a subproject of H4I-HipBLAS or any of the other individual "user-callable" H4I libraries. I prefer having a separate umbrella project that has MKLShim and any existing H4I libraries that use MKLShim as subprojects. (And this is pretty much equivalent to the approach that the Spack project that builds these libraries takes.)
Since hipBLAS is a very frequently used library, it makes much more sense to be able to build it in a single step.
This replaces the previous version which was a fork of Sabojit's branch.