I thought the plan was to have common code for the generic, non-critical parts like graph handling, while operators would be shipped in multiple variants, keyed and loaded at runtime by CPUID. You can even delegate that to glibc if they are packaged as shared libraries and it's a recent enough version. Maybe in 1.7?
I get the sentiment, but how many people can just casually add OpenCL to Tensorflow? For a lot of people, even if talented programmers, they may not be able to do anything but ask and beg.
My webcam doesn’t work with Linux. Should I write a driver despite not knowing the kernel API, C, or how webcams even work?
I would be better off sponsoring development somehow — this is something I’ve wondered if it’s possible to do. Someone with the skills could be paid by me, and other affected users.
0: https://github.com/ROCmSoftwarePlatform/hiptensorflow