TODO.txt

General future goals:
* Fake heterogeneous compute: Prefer on-GPU compute, but perform CPU compute if that isn't present. Declare input and output buffer sizes ahead of time, load input up with byte[], then call a method on the compute procedure to make the buffer available either on-GPU (which can be faked from-CPU by uploading it as a Buffer) or on-CPU (which can be faked from-GPU by downloading the buffer to the client). If both are called, one is effectively ignored.

* Face streaming

* Indexed model support