GCC supports offloading to amdgcn/rocm targets but it needs to use LLVM for its assembler, per https://gcc.gnu.org/wiki/Offloading#For_AMD_GCN:. I have offloading with GCC + nvptx essentially working (see bug 945296) so the infra is there, just the following needs to be done: 1) define a amdgcn-amdhsa* case statement in crossdev, copying nvtpx 2) figuring out how to express BPKG=... in there for LLVM (*) 3) updating toolchain.eclass to list amdgcn-amdhsa* too 4) done? (*) It might be that we don't need to do anything for that specifically (might be able to leave it blank, possibly with some small hacks to crossdev to allow that), but we do need to wire in a way of telling it where the appropriate tools are. CCing some people interested in rocm and also people who are interested in LLVM support in crossdev. I'm hoping one of you can help handle it as I'm a bit tired from all the nvptx work. I will of course help with any issues.
(In reply to Sam James from comment #0) > https://gcc.gnu.org/wiki/Offloading#For_AMD_GCN:. > [...] > > (*) It might be that we don't need to do anything for that specifically > (might be able to leave it blank, possibly with some small hacks to crossdev > to allow that), but we do need to wire in a way of telling it where the > appropriate tools are. Per that page, it might be as simple as just making a bunch of symlinks inside the just-created /usr/amdgcn-amdhsa ROOT.
USE="-go" crossdev -t nvptx-none finished without error. This is a bit of a, well, annoying problem, since the only way this then would work regularly, is if one would put the useflag -go in a config file for the offload compiler, but then the gentoo dev teams would have to periodically check whether this is still valid. I will now test some code. Lets see....
Someone was interested in the past: https://forums.gentoo.org/viewtopic-t-1143342-start-0.html
