commit 4a8b169993b6c9d1188cff6948c12d453edde177
author Stella Laurenzo <stellaraccident@gmail.com> Wed May 26 10:42:49 2021 -0700
committer GitHub <noreply@github.com> Wed May 26 10:42:49 2021 -0700
tree 52b1c7b0a223f518f6074973903d3ce576916547
parent ba661e87802a23de3cc02e2aefc5fda1a1c6e71c

Create new pass PromoteTensorLoads. (#6023)

* Create new pass PromoteTensorLoads.

* In the new input pipeline, the only part of
PrePartitioningConversionPass which survives is the
tensor.extract_element -> flow.tensor.load conversion.
* This conversion actually needs to be done at a couple of points during
lowering, first promoting any extracts that are introduced as part of
control flow (in the input pipeline), then allowing most of the program
to be loaded onto the device, and finally converting any remaining,
otherwise unrecognized extract elements.
* As such, I opted to make it a very specific pass that does exactly
what it says on the label. We may want to do something more
sophisticated later, and at least having one thing to see and replace
will help.
* Once the new input pipeline lands, PrePostPartitioningConversion.cpp
will be deleted.