MLIR Visualization - GSOC 2020


Can someone point to a few resources to get started on the MLIR Visualization project? And what are the rough goals of the same?


I updated the description under open projects a bit. As mentioned there are a wide variety of potential projects under there. And so the intention is to iterate with interested students on this proposal that will be submitted, which will then go to the project level (MLIR is part of both LLVM and TensorFlow’s GSOC projects) and selected.



Thanks for the reply! It looks like TF itself lacks a visualization through GraphViz which is an open issue. I feel that as the first part going through GraphViz might be good since it might be easier (this would include adding a functionality for extending it to different dialects and showing that conversion), but later on, shifting to something dynamic in html (with a layour similar to Tensorboard) would make sense to completely show the potential of MLIR.

How does this sound to you @jpienaar?

Starting statically definitely makes sense. We have a basic GraphViz export at the moment, but it is very basic and operates at only ops directly nested within a region, showing use-def chains. And as mentioned on the web page there are multiple different audiences and for some, a purely static graph may be exactly what they want (now purely static doesn’t mean just display the entire module :wink: so some thinking would be required on how to make it generally useful).