Deformable convolution and other custom ops

Seems unfortunately neither Tatiana nor Stephan can be tagged here, they were interested in reaching out when I asked last. I’ll send a reminder (I think it may work easiest at one of your SIG meetings).

Re: DSL. There are some investigation that way but not plan of record yet. I think it is quite valuable and personally interested there (Dex is an interesting data point there IMHO). The biggest open question is around support surface if one goes that way. The current plan still circles around expressing such new ops via decomposition using golden set of ops (@Eugene_Burmako may be able to say more there shortly).

I understand this but It could be hard in ecosystem python repos to interface with the AVG contributor for a decomposition check that I suppose It means to be able to achieve the contribution just with the compositional golden ops path.

Often I don’t think that this could be done easly, “on the fly”, when they are going to propose an unvectorized PR or if the compositional path is not memory or computational performant enought.
At least we need some kind of interface/github issue labeling support from the compilers team.
This mean that we need a clear support ownesrhip on request also if probably lower than to maintain a full DSL.

That is very valid yes. Else it is just you guessing.

If you want another real use cases in an ecosystem python repo please take a look at:

We are hitting again the same old topic but this time in Keras-cv

/cc @Magnus as I think he could be interested in this old long thread.

For who is still interested, as this is going to cross the TF bridge(s) bounderies, I’ve opened a parallel topic at: