Currently the set of libraries used for AVX2, AVX512, GPU, etc must be defined at compile time. This requires us to have separate builds for each runtime environments. Is there a way or perhaps any work on making this a runtime configurable setting?
@dawood actually no, because it involves pre compiled c++ binaries. You would make the binary sizes bigger by having different versions of it.
@saudet could chip in here but what you’re asking for is just now how code works.
No compiled code is that dynamic.
Yes, it’s possible to use cpu_features for this: https://github.com/eclipse/deeplearning4j/issues/7834#issuecomment-498170278