Compiling array-based languages to hybrid CPU/GPU systems has been of
considerable interest in recent years. However, there is no reusable
toolkit to allow compiler writers to quickly build such compilers. We
present Velociraptor, a portable and reusable toolkit, specifically
designed for building just-in-time compilers for array-based languages
to hybrid CPU/GPU systems. Velociraptor takes as input language-neutral
typed AST based representation of array based programs (with full
support for flexible array indexing schemes and array operators) with
some annotations about parts of the program to be offloaded to GPU.
Velociraptor generates optimized CPU and GPU code. We will discuss
the design and internals of Velociraptor. We will also present the
integration of Velociraptor in two just-in-time compilers: one in McVM
(a VM for MATLAB initially built by other students in our research
group) and a proof-of-concept Python/NumPy compiler written by us.
Preliminary experimental results from multiple GPUs are also discussed.