tree 54d4f7606768cbacb5393b4c49ba9518f0288fef
parent c232aebe272e4c9d8f25728ae7e4f6a2a4c7c5fd
author Ben Vanik <ben.vanik@gmail.com> 1697736229 -0700
committer GitHub <noreply@github.com> 1697736229 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJlMWYlCRBK7hj4Ov3rIwAAzBgIADNMBezQUCd2oXVNXXJdAWf3
 u+2l6O8ay7gcGgbh9PBYXuaBphAKj1ILt0En5yKMOeltek/tUU1L45fkbFVkKKYc
 z2AmU9ZoReM/chNyxLggJpC1wQwFdNGEnNYBZX1reA2031+Uu5vJupbeCU2GGXd0
 0wgRbzHARath/oVcO8ANePdtfVHRPJxQar4Oxu5nXRuyNQP0IzGMaLfBFdkPoTl9
 CFQadGAVpDItThFK/fVg3DRQ0dCCO/F1vKBLDDqT5paIyj7aNQPblPL+IBrbhXrm
 B4MHTr53TNBWR4eTbUOS/W5LOkMxwn8I1DqKA7KXcKRS4/Uymm7Rw03D0tkN9Bk=
 =rXZ1
 -----END PGP SIGNATURE-----
 

Improving the `*.tensor.trace` op to carry shapes/encodings. (#15228)

This avoids a bunch of IR noise when lowering into the stream dialect
and forces all traced tensors to be transfered into staging resources.
This means we no longer require mapping (or worse, emulated mapping) on
any target in order to trace as we'll have the compiler implement the
readback asynchronously.

A change I held off on here but should be done soon is making the trace
ops return their tensors so that we can asynchronously schedule them.
Then instead of emitting runtime calls on staging resources we could
batch up the transfers into a ringbuffer and trace out without
introducing dispatch-to-dispatch host synchronization.