commit 1b9e84cf81836814e8fa4feb652a1e27b96a2b6f
author Nicolas Vasilache <nicolasvasilache@users.noreply.github.com> Fri Apr 08 20:30:02 2022 +0200
committer GitHub <noreply@github.com> Fri Apr 08 11:30:02 2022 -0700
tree 61779a491e9bdd701fb9e61c070364244cd230db
parent eb0d678894aa6ebb20b056d0f09bad3a16d5d26c

Adopt TransformDialectExtension and add iree_bufferize + iree_set_num_workgroups_to_one transform ops (#8821)

* Adopt TransformDialectExtension

This revision adds support for dynamic registration of ops into the transform dialect (see https://reviews.llvm.org/D123135 for more context).
This PR also add ire_set_num_workgroups_to_one and iree_bufferize as transform dialect operations.

With this layering improvement we can now run all transformations as a transform and connect
to vector lowering all the way to LLVM (inclusive or not depending on the needs).

For registration, we piggy-back on Interfaces.cpp::registerCodegenInterfaces for now. This layering should be improved in the future.

The C++-specific parts of setnumthreads and bufferization can now be removed from the interp codegen path.
The --linalg-transform-interp-disable-bufferization is now obsolete and deleted.

* Extract the implementation of SetNumWorkgroup in a free function.

This allows reusing the implementation from the transform dialect and ditch the pass application.
Passes, nesting and non-Module root ops behave very surprisingly and may easily do nothing and not report anything.

As a consequence, we would hit bug #8823 and get flaky single-thread results ..