Evaluation of TACO compiler

[hameerabbasi]

From all the examples I can see, the TACO compiler is good for einsum-type expressions, and any that involve tensor products, contractions and sums. I've posted about it here: tensor-compiler/taco#276

[daletovar]

This looks great to me. We'd be interfacing with cython?

[dhirschfeld]

Is TACO similar to opt_einsum?

[hameerabbasi]

@dhirschfeld It's much better than that: It allows for sparse arrays in many different formats to be represented. And it allows for adding two separate tensors, which is something einsum can't do. In the linked discussion, they're considering allowing arbitrary element-wise operations.

[dhirschfeld]

Thanks for the info! I was only vaguely aware of the project previously but it sounds like something I should keep an eye on...

[hameerabbasi]

@daletovar I think pybind11 is the better choice here.

[hameerabbasi]

After a thorough review of the options, I studied how the original TACO behaves. It emits C code, then calls the system compiler. I think we can probably do something similar: emit Python AST which can then be jitted with Numba.