AMD Working On CUDA Source Translation Support To Execute On FirePro GPUs

Early this morning I wrote a brief article about AMD working on an LLVM-based Heterogeneous Compute Compiler and since then more details have come to light.

That Heterogeneous Compute Compiler article was ran after coming across some slides from the LLVM Developers' Meeting that took place a few weeks back. SuperComputing '15 kicked off today in Austin and AMD has coincidentally revealed more details on their compiler work.

As part of HCC, AMD is also working on HIP (Heterogeneous-compute Interface for Portability) for porting CUDA source code to a common C++ programming model. This translated CUDA could could then run back through NVIDIA's NVCC CUDA compiler or be sent through AMD's new HCC compiler. It's basically translated CUDA support at the source level, but with no plans for allowing CUDA binaries to execute atop AMD GPUs.

With AMD "HIPify" tools, it should be a "simple" process for porting CUDA to HIP. The HCC compiler will support C++ 11/14 and C11 along with OpenMP 4.0 offloading. C++17 support will come. AMD is also working on Linux driver improvements as part of this latest GPU computing initative and they're working to deliver proper headless Linux support for AMD GPU compute servers.

More details can be found via AMD's press release for what they're calling the Boltzmann Initiative. Meanwhile, it seems they're leaving the Open64 compiler support left to die.