tree 37b4e2980d94c0f6c2bdd5dda09f26e7bd0e3d43
parent 14691ce2e956f089d401b5bfee9fcb10e20d4755
author Jacques Pienaar <jpienaar@google.com> 1684514676 -0700
committer GitHub <noreply@github.com> 1684514676 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJkZ6d0CRBK7hj4Ov3rIwAA2kIIADQRsSc+ZSng50QghHR/eqmm
 K3ExJ4d3BTIlhPlJCDTdGB7BNVN6Mm8r2awSgNXNEciASp25Y8JBvf1I65lL7ptY
 iamvGH8OhVhSD7qHPoCrU0Wi+8E2FuVwtKWc6+NUF0uQKcThjExkxEHoRngaMO8s
 m8DJPk2wPkcBUu2rQX/y4tagX/8lSmuaEXH2c+G9NAXU5fhp9AWja0Wjb5FqqcG5
 c4ktM0SeIDdqtQ0YNfkRuG7i+CqG/GqVrhbVHIGaujX8uXVQReW4izOKLL4x2ekE
 DxRhJnIAMsCs3QPnQEFody1F52fJg/yY7aBNFkIGa7cSZsN5dNiIBFRsd2rLM1A=
 =0xkF
 -----END PGP SIGNATURE-----
 

[tosa] Add path to TOSA backed backends. (#1466)

Connect these two industry standards by way of dialect legalization from
StableHLO to TOSA. This adds basic support and testing: along with usage
of some of the equivalent canonicalization patterns from MHLO (not
included in this PR) this has been sufficient for some full models
starting from ML framework to TOSA backed. Support is not complete and
partly relies on some canonical StableHLO forms.

The legalizations are also written primarily using PDLL, but we have not
yet adopted some of the newer support there for variadics. This work
started by targeting MHLO in TensorFlow repo as StableHLO was still
young, but given StableHLO development it makes more sense to instead
start there and provide a connection for community backends.

No new repository dependency is introduced. The cmake config enables
disabling building conversion,

---------

Co-authored-by: Eugene Burmako <burmako@google.com>