I am trying to compile tensorflow with rocm support, building on bug 705712 - sci-libs/tensorflow - enable rocm backend support for AMD video cards. A lot of the patches needed seem to be around the location of rocm headers in gentoo, a file which is completely missing is rocm_version.h. I assume that it is actually difficult to clearly define the rocm version when different parts of the stack come in different versions (rocm runtimes vs numerics libraries). Still, tensorflow expects this file to be present. How is the intended use of rocm in Gentoo in that regard? Reproducible: Always Steps to Reproduce: 1. Patch tensorflow ebuild & tensorflow to enable rocm support 2. Try to emerge it
I have the same problem of missing rocm_version.h header when trying to compile caffe2/pytorch with rocm.
