"We already have code generators for turning kernels into Com- mon Lisp code and for turning a subset of kernels into C++ or CUDA code. Right now, our strategy is to use C++ and GCC when possible, and Common Lisp code otherwise. In the future, we plan to make the C++ generator obsolete by using SIMD optimized Common Lisp instead. Doing so would build on our previous work on s b - s i m d that we presented at the 15th European Lisp Symposium in Porto"

My take from that is, that for CUDA external tools are used today. Still, interesting work.