An Early Port Of GCC To AMD's GCN Architecture

[Mangix]

Fun fact: LLVM already supports GCN.

[coder]

It's what we've been waiting for.

Is it? What do you expect it'll let you do?

HSA already supports the platform requirements for the C++11 language. So, I think this doesn't enable any new capabilities. Just a new toolchain.

[fuzz]

Is it? What do you expect it'll let you do?

HSA already supports the platform requirements for the C++11 language. So, I think this doesn't enable any new capabilities. Just a new toolchain.

The goal is support of HSA on applications that don't even target it. On Gentoo, I see that bringing a lot of benefits.

[coder]

The goal is support of HSA on applications that don't even target it. On Gentoo, I see that bringing a lot of benefits.

I'm all in favor of that. Maybe I'm mistaken, but I thought clang already supported it.

BTW, nothing wrong with having another toolchain to do the same thing. More options are better!

[S.Pam]

The goal is support of HSA on applications that don't even target it. On Gentoo, I see that bringing a lot of benefits.

So like compiling any software to run on the gpu instead, like an encoder or compressor (zlib, zstd...) etc? Or do we talk larger stuff like whole programs?

[cj.wijtmans]

I think, as far C/C++ is concerned, Clang is basically on par with GCC. And that's the problem: because Clang and GCC are equivalent there is little incentive to use Clang. It even seems that Clang's advantage in terms of faster C++ compilation is slowly going away.

As far as implementing a new backend is concerned, internals of LLVM are somewhat different from the internals of GCC.

Clang has better compiler errors. I find personally.

[coder]

So like compiling any software to run on the gpu instead, like an encoder or compressor (zlib, zstd...) etc?

Compiling generic C/C++ code for a GPU generally won't run efficiently on it. Also, you wouldn't have the standard library.

I think the best argument for providing C++11-level language support is that it makes stuff like SYCL easier. In other words, it lets the toolchain take snippets of your program and easily target them to run on the GPU, but within limits (i.e. no calling out to external code).