notes

prior work

vs. FALCON et al

dynamic type inference used to speed up existing language
great, but what you should really do is figure out how to redesign
the language given that technique, to take full advantage of it


vs. C++ libraries

C++ is great in that it's one of few languages powerful enough
to define n-d arrays in a library and have it be fast.
our approach has 3 advantages:

1. templates are hard to read and write
2. you can't implement as big a class of rules
   must handle dimensions one at a time, recursive
3. not available at run time
   while static # of dimensions is probably sufficient for
   most applications, there is a large cliff where everything
   changes if your code is more dynamic


vs. haskell (/ repa)

has a similar recursive structure approach as templates


vs. numpy

has its own abstraction mechanisms inside its large C codebase
- compile-time code generation
- custom run-time dispatch mechanism

we believe one powerful dispatch mechanism could subsume this