LLVM Powers C++ AMP To OpenCL On NVIDIA

An independent developer has made improvements to the LLVM infrastructure and Clang compiler for supporting the compiling of C++ AMP code into OpenCL code with support for the NVPTX back-end so that this multi-threaded C++ code can be executed on NVIDIA GPUs.

C++ AMP, short for Accelerated Massive Parallelism, is the open specification led by Microsoft for implementing data parallelism within C++. C++ AMP is intended to be portable and to make it easier for executing C++ code on GPUs while also supporting CPU execution too.

The Accelerated Massive Parallelism support by Microsoft is implemented as a DirectX 11 library on Windows while outside support for C++ AMP is still fairly rare. Intel was working on Shevlin Park (though there hasn't been any news on it recently) for implementing C++ AMP on alternative platforms using LLVM/Clang. With not having any Shevlin Park code, an independent developer decided to implement his own C++ AMP implementation that takes the marked up code, converts it into OpenCL, and pushes it through LLVM's NVPTX back-end for execution on NVIDIA GPUs.

Interestingly, this is the first LLVM/Clang contribution by this independent developer that sought to explore LLVM/Clang internals. Right now this prototype code is just executing simple sample code by Microsoft. Using this requires changes to LLVM and Clang while an unmodified Compiler-RT is needed and there's also a Git repository for the simple working test project.

The developer behind this interesting work, Dave McFarland, explains, "It currently works only on NVIDIA GPUs, and has only been tested on my shitty old 9600GT on amd64 linux with the stable binary drivers." But right now only simple C++ AMP code is expected to work and the code is very much in a prototyping stage.

Those interested in more on this C++ AMP LLVM/Clang work can find the mailing list discussion happening this weekend.