Bruno points to an interesting read on C++ AMP.

DirectCompute is essentially Windows-only. CUDA is essentially Nvidia-only. Most importantly, all three approaches require learning not just a new API and a new way of looking at problems, but an entirely new programming language. Each of the three languages is “C-like” but is not C. Only CUDA is becoming C++-like; OpenCL and DirectCompute cannot offer C++ abstractions such as type safety and genericity. These restrictions mean that mainstream developers have generally ignored GPGPU in favor of techniques that are more generally accessible.