Skip to content

WIP: try folding arithmetic forward into polynomial.eval#3046

Open
copybara-service[bot] wants to merge 1 commit into
mainfrom
test_923460715
Open

WIP: try folding arithmetic forward into polynomial.eval#3046
copybara-service[bot] wants to merge 1 commit into
mainfrom
test_923460715

Conversation

@copybara-service

Copy link
Copy Markdown
Contributor

WIP: try folding arithmetic forward into polynomial.eval

This idea came out of the matvec + batch_norm fusing: if we can't fuse into the preceding layer, then maybe we can fuse into the next operation. If it's a polynomial.eval op from an activation approximation, then simple ops like those involved in a batch_norm become a polynomial composition, and can be folded into the polynomial.eval op's coefficients. One can also fold addition ops after a polynomial.eval into the eval.

This initial implementation was LLM-generated. I still need to do a few things to clean up and validate the approach.

This idea came out of the matvec + batch_norm fusing: if we can't fuse into the preceding layer, then maybe we can fuse into the next operation. If it's a polynomial.eval op from an activation approximation, then simple ops like those involved in a batch_norm become a polynomial composition, and can be folded into the polynomial.eval op's coefficients. One can also fold addition ops after a polynomial.eval into the eval.

This initial implementation was LLM-generated. I still need to do a few things to clean up and validate the approach.

PiperOrigin-RevId: 923460715
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant