autopar and offloading

[s_j_newbury]

Like with the gentooLTO overlay I've been building my systems with LTO and Graphite optimisations for some years now, in addition I've also used GCC Auto-Parallelisation with a blacklist for packages where it causes build or run-time issues. I've now switched to the gentooLTO overlay to make my life easier.

In my gentoo-gpu overlay I've have support for GCC OpenMP/OpenACC offloading to NVidia GPUs, with the ability to easily add Intel MIC and AMD ROCm support. Unfortunately, I haven't been able to test it due to lacking suitable hardware.

If anybody with any of the below hardware is prepared to give it a try it could be quite an interesting experiment to combine LTO+autopar+offloading:

NVidia GPU - Kepler* or later with CUDA (so needs nvidia-drivers unfortunately)

Intel MIC

AMD Vega (GFX9) with pretty much any motherboard/CPU

Most AMD GFX7-8 with PCIe v3 with PCIe atomics support** (AMD Ryzen or Intel Haswell and later)

Full list of supported AMD hardware

* I'm pretty sure Kepler is the oldest supported target, when I originally worked on integrating the support my Fermi laptop was *just* unsupported!

** I have a AMD GFX8 GPU - RX470 (POLARIS10) and an AMD FX-9370 CPU so I'm really out of luck

[s_j_newbury]

So I bought a used Vega64, so I have updated my patch to support ROCm offloading for GCC along with some general clean ups.

Since there's no binutils for amdgcn-amdhsa I've made an ebuild to install a wrapper of the needed installed llvm tools with USE dependency on "LLVM_TARGETS=AMDGPU", this works fine to build the offload plugin.

I'll be updating the overlay once all fully tested.