Target overrides in tir/pass?

Development
1

Currently all tir passes are self-contained, i.e. the class (typically a subclass of a mutator class) and the member functions are all defined in a single .cc file and run as a unit. This doesn’t allow any of the member functions to be overridden to implement custom processing.

Is it a design choice to disallow this? Or can this be modified?

The modifications would be to create a separate header file, and let targets produce their own objects inherited from the default one, that would then be used in the pass.

2

We prefer to hide as much as possible by default. If there is a need to reuse a mutator, then we could create a header file(starting from internal header in src), for example https://github.com/apache/incubator-tvm/blob/master/src/tir/pass/storage_access.h

3

It would also be useful to think about how do we do target overrides. There are a few ways:

  • A0: Target creates its own mutator and pass
  • A1: Target registers certain callback funtions that handles a target related optimizations
  • A2: Target registers certain attributes(e.g.) can an op X be lowered to Y in this target and the pass takes that as an input.

This thread is also related Target and Attributes

One thing that I want to note is that we are going to change the way we refactor TIR passes to consoildate it with the relay via a unified IRModule and Pass (hopefully in a month). Some of the changes have already landed(e.g PrimFunc)

In the new convention:

  • The basic unit among passes will be IRModule containing a collection of tir::PrimFunc
  • Each PrimFunc could have an optional target attribute, indicating the target we intend to lower the funtion to
  • The passes should be able to read from these attributes and do some customization.
1 Like