tree 512ad891daed22f3da62bdc47207fb4f54c92ddf
parent 8c67db6065f9c8bfe87c07bcb5400e227f73d406
author MaheshRavishankar <1663364+MaheshRavishankar@users.noreply.github.com> 1634252903 -0700
committer GitHub <noreply@github.com> 1634252903 +0800
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJhaLhnCRBK7hj4Ov3rIwAApLoIAF4hwdVD6Xk1TOrN75gzEX7P
 W914+yEFVpvQIEM9yHwT170Ot5FWB9lKxebeZFj4HJDypUbuMPscfQ5jA/5c1Nsb
 mi+VpmECtqSTikbTxK9eXjEcCsDTLrVKbCR1NzV1nIwAVIMBW3ZYf3N2BcwYF1yL
 Ey66htM5LpGkKYeZ2NFHr0ZBvNzzeYIQVL64BYSGTy/Cj2JWqAqBgjU6GYWIL8Ak
 L3jT3kyomvZg+3/86uN6LsH/PSYfNJnrgtEzk9N4ViIxUzE9PcwiVDqC6Gs7ue6k
 tOYN7Jz7yjgRaxket34lVFvMQD0cJW89KwZg4l1qUUnxLG/RSjAoOhVcLCGWnd4=
 =GrKi
 -----END PGP SIGNATURE-----
 

Remove the patterns that create dispatch regions without tile+fuse+distribute. (#7328)

With TiledOpInterface most of the operations to be executed on the
device can be tiled and distributed. So the pattern that created a
non-tiled and distributed dispatch region can be removed.
Simplify the code that decides which ops are root ops.

To make all tests pass, the last operation that needs to be tiled and
distributed is tensor.extract_slice. Add an external model to
implement the tile+distribute of this op to allow deleting the fall
back path.

This is recommit of PR #7306 which was reverted in #7319 due to failure
on integrate. The changes from #7306 fix the breakages.

Resolve tensor.dim operation before legalization during dispatch
region formation.
Allow tiling of tensor.extract_slice with non-unit strides.