How AMD Is Taking Standard C/C++ Code to Run Directly on GPUs

The thought experiment here was "what if you treat the GPU like an embedded target, instead of using cuda style languages to program it". Works great.

I especially like this on APUs - my laptop from 2020 and a MI300A from today run it happily. Write to files, hack on data structures shared with x64, all is good. Quite a lot of libc++ works now too.

Noone seems very excited about this so far but Joseph and I are thrilled that it works. It's also to AMD's credit that they've let us run with this despite no direct commercial motivation. We did find a bunch of compiler bugs standing it up.

I wonder how the shared memory model is mapped to this approach. Looking at it now

Also check out Rust GPU, which aims to do the same for Rust: https://github.com/Rust-GPU/rust-gpu