You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We are using surface tuples throughout accelerate-backend kit, including decoding the EltRepr to recover the surface structure (which doesn't help us with shapes, alas).
However, there's an inconsistency here, because we count projlen in ETupProject in terms of scalar leaves. If we are sticking with surface, N-array tuples, then this should be replaced with a basic "project field N of M" operator, and this should be able to happen from when we first build SimpleAcc.
The text was updated successfully, but these errors were encountered:
I think this is related to some recent work we did. We took a different approach and are now using the
same count the leaves approach on array tuples. At least this means that it is now consistent.
Maybe Trevor has more input to offer here, but I think we got to a point where said "oh, we cannot do it the surface type way".
I don't want to close this issue with comment, but rather await your thought/review on our approach.
We are using surface tuples throughout accelerate-backend kit, including decoding the
EltRepr
to recover the surface structure (which doesn't help us with shapes, alas).However, there's an inconsistency here, because we count
projlen
inETupProject
in terms of scalar leaves. If we are sticking with surface, N-array tuples, then this should be replaced with a basic "project field N of M" operator, and this should be able to happen from when we first buildSimpleAcc
.The text was updated successfully, but these errors were encountered: