Recent progress in the CUDA ecosystem

Less than a week after my previous post about how "only OpenCL can expose massively parallel compute capabilities in a vendor-neutral way", NVIDIA open-sourced their CUDA compiler! Obviously they are intently watching this blog, and had to react to my nay-saying.

Jokes aside, this is a big move for CUDA as it allows anyone with the ability to create LLVM ASTs from their scripting language to use the PTX backend. Indigo already features a dynamically-compiled scripting language, Winter, and this technology could be extended to output optimised code for NVIDIA GPUs directly. Of course, this isn't necessarily an approach we'll use, however it illustrates the potential of having a powerful optimising compiler for many platforms behind your domain-specific language.

It will be exciting to see how this develops, as more back- and front-ends are added to the burgeoning LLVM framework.