there are several dialects in my project
I’d like to define global trait among different dialects.
MLIR defines lots of traits Traits - MLIR
is there method to create such trait which can be used in any dialect/op?
A Trait is just a C++ class, you can copy any of the traits defined in MLIR in your own project as a model.
To use them in ODS, take example on AffineScope for example (or all the others around there).
yes, trait can be reused everywhere
how about interface?
take a convolution as an example, let’s say there are FooDialect and BarDialect has convolution op
- define getWeight interface for convolution in FooDialect
then implement getWeight in FooDialect - define getWeight interface for convolution in BarDialect
then implement getWeight in BarDialect
even convolution in different dialects are almost same and has similar operands, still have to implement two similar interfaces…
if the number of dialects increases, have to do similar things again and again
that’s why I ask if there is solution which define a global interface and implement once
Interfaces don’t need to be associated with a dialect. There’s some examples here: llvm-project/mlir/include/mlir/Interfaces at main · llvm/llvm-project · GitHub
In general it is the transforms / the analysis that drives the development / design of an interface: you create an interface for the need of the pass and the dialects implement it.