Gus Smith

Some dimensions should be able to be dynamic, written as `?` in some frameworks (see Relay). This might lead to interesting optimizations: I suspect that, for example, when an access...

I think this should be doable as "tags" on dimensions.

If we changed `access-flatten` such that it only flattened the item dimensions, then no reshape would be needed in some cases. We could make a "smarter" flattening rewrite -- instead...

We have a timeline for which networks we want to support. Move that here.

I have notes on how to do this. Give it a try.

Improve the ILP problem so that it's more amenable to solving

20

The ILP problem isn't amenable to solving. This is a deep rabbit hole: there's a lot of literature about making "good" ILP problems. We only need to go deep enough...

Glenside should find a blocking strategy for Mobilenet convolutions

8

This is mostly an issue of egg search performance and the fact that Glenside's blowing up the egraph with rewrites. See comment thread for debugging chain. - [X] Add (failing)...

Switch to main TVM repo in dependencies

1

Currently pointing at mwillsey's fork, but now that his PR is merged, we can switch. Should pick a specific revision from TVM, not just point at master!

Add an access-concatenate-varargs operator

6

- [ ] Add integration test which uses new syntax Ideally I'd add a new end-to-end test that checks/proves (at least to me) that this method is faster than the...