Jacques Pienaar
Jacques Pienaar
Benoit has good point, is this something we are concerned about and digging in?
I'd expect this to be used as form before export given spec form https://www.mlplatform.org/tosa/tosa_spec.html#_cond_if, but ops aren't isolated from above so shouldn't be needed in general AFAICS
Yes, so I think I'd do what you propose and then we could have (like we do for TFLite) a "prepare for export" pass that makes it args. But for...
Good question, I'll try disabling that one and see if only one
Only for that target so will disable downstream and lower priority here.
This should work post https://github.com/tensorflow/tensorflow/commit/763d55e7e33ca888a2aa14fa89ec413047daf1e5 (not verified yet). Note: as mentioned in that commit, it was for completeness and not necessarily performant. So more work may be needed later here.
@NatashaKnk could you look at this too?
Could you perhaps paste in the failing MHLO scatter here? (makes it easier to see before I get triage)
Local reproducer ```mlir func.func @main(%arg0: tensor, %arg1: tensor, %arg2: tensor) -> tensor { %0 = "mhlo.scatter"(%arg0, %arg1, %arg2) ({ ^bb0(%arg3: tensor, %arg4: tensor): %1 = mhlo.add %arg3, %arg4 : tensor...
Could you add link to where in the code it fails?