Replies: 1 comment 2 replies
-
Any reason not to take this approach? If not, I would suggest converting this to an issue soon |
Beta Was this translation helpful? Give feedback.
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Currently we use
numba.vectorize
to implement Elemwise ops. This comes with a few problems howerver:I think we could replace this by a custom numba implementation that uses the
numba.extending
hooks to directly produce llvm code. I've been experimenting a bit, and I think I got to a point where this is starting to seem doable. A basic POC that doesn't deal with broadcasting at all yet:Most of the magic happens in
codegen
, where we can use normal python code to generate the llvm ir instructions to do the actual computations.Update
With support for multiple outputs and summation of outputs:
Beta Was this translation helpful? Give feedback.
All reactions