tree 85effebfc54769949c42c9869e69245082d9b34e
parent 2f47c081bc9506dd583a8249e06745b83dde79b1
author Scott Todd <scotttodd@google.com> 1699560150 -0800
committer GitHub <noreply@github.com> 1699560150 -0800
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJlTTrWCRBK7hj4Ov3rIwAAhIIIAECnA5fPhIZ+sQNEjIONo64Y
 dQPXSpARX5Ok8QcqdosEDzDp9+I7V1ECBzwkk+braCF6jIqkYc2SKCKWokPYTFof
 WVts19qkr4UxNyJad7Y1SsEt94FUKRqZVgU/5oUtU+aavyTXri8dZappEqH8itle
 hWg686C+JrXsE8E9QLMD9TEwlflcrE/9Dwzn1o2evZifd8BdBOMPJO/yeWIq1OdR
 knjK2foAlnFl7cwWYCOIXe48aerl38grLUpwVyzJxpDkGw/dYGb49xGdM2C2nOYR
 3WdwKdroj6I19mohwugUwuFjDLK5WhvGL15C6htNsMrdEA5DE2wEaC2SiHq+DRM=
 =uHB+
 -----END PGP SIGNATURE-----
 

Plumb 'torch' input to auto detect via plugin interface. (#15438)

This teaches the 'auto' input conversion pipeline how to run
plugin-provided conversions automatically when plugins detect that ops
in their related dialects are present in a module. Now, programs coming
from PyTorch (torch-mlir) can omit `--iree-input-type=torch` and rely on
the default of `--iree-input-type=auto` (this fixes
https://github.com/openxla/iree/issues/15353).

The code is awkward for a few reasons, but I think it can be improved
incrementally:

* StableHLO and TOSA are "built in" to the compiler (gated on
`IREE_HAVE_STABLEHLO_INPUT` / `IREE_HAVE_TOSA_INPUT`), while Torch is
implemented via a compiler plugin
* I think we could migrate both to be plugins, which would put the input
dialects on even footing
* The new dynamic pass pipeline needs to know which dialects to register
* If all input conversions were plugins, that would feel like less of a
special case
* There is a dep cycle that I'm trying to avoid narrowly... that may be
fixed by migrating all input conversions to plugins