[CPU] SnS transform ambiguous convert #27948
Conversation
chenhu-wang
force-pushed
the
chenhu/convert_need_mark_explicite_modes_in_snippets
branch
from
d3df44f to
158d71f
Compare
| @@ -428,6 +429,7 @@ void Subgraph::data_flow_transformations(const BlockedShapeVector& blocked_input | |||
| manager.register_pass<snippets::pass::ConvertConstantsToScalars>(); | |||
|
|
|||
| manager.register_positioned_passes(backend_passes); | |||
| manager.register_pass<snippets::pass::TransformConvertToConvertTruncation>(); | |||
There was a problem hiding this comment.
We call this transformation in CommonOptimizations after subgraph tokenization to mark all original Convert ops as ConvertTruncation.
Then all other transformations which inserts Convert op (such as FQDecomposition or PropagatePrecision), should always insert utility ConvertSaturation op.
Did you find the pass after TransformConvertToConvertTruncation registration in CommonOptimizations which inserts clear Convert op (not from Snippets dialect)?
There was a problem hiding this comment.
I can't reproduce the issue locally. I checked the passes after ConvertTruncation in commonOptimization. Only found an potential possible convert in ConstantFolding, but not very sure. So I need more detailed info from the issue reporter:-)
Details:
Tickets: