Re: [apache/tvm-rfcs] [RFC] Improve representation of IntSet. (PR #99)

[Tianqi Chen <no...@github.com.INVALID>]

Thank you for the proposal. There is a general tradeoff in terms between speed of solving, generality of the solution, and extra capabilities(eg ability to handle symbolic division/mod that may go beyond normal).

The current hyperbox based analysis was chosen because of its speed efficiency, of course that also comes with limitations. Additionally, there is a relaxation in analysis so to handle arithmetics that may not be exactly captured by intset.

Normal constraint solver based solution might come with some extra capabilities, the question here being whether the efficiency can be kept up. 

My recommendation would be to start with a different class first under arith for us to understand the overall space. Likely fewer passes will need such analysis, then we go from there 

-- 
Reply to this email directly or view it on GitHub:
https://github.com/apache/tvm-rfcs/pull/99#issuecomment-1427891561
You are receiving this because you are subscribed to this thread.

Message ID: <ap...@github.com>