@tqchen, what's your suggestion? IMO, low level IR has been there for a while, and we've had experience and understanding in low level ir. the post of unified ir to me is just a high level proposal, details needs to be discussed further, such as,
The most valuable thing to me is we can make ops white box with unified ir, that is to say, we can analyze into the ops, which is totally different with the current way, we are using ops as black box in graph framework, or separated ir. with white box ops, we don't need to care about the ops name or the formula ops are using, we can optimize them in a general way. see what's XLA doing. The abstraction of high level ir matters. you don't want to lower ops body too much from tensor expression into nested loops, since it's a big cost of analyzing ir with too much context. so I suggest we can keep moving this work on this thread meanwhile we discuss how to reuse the solution. -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/apache/incubator-tvm/issues/4468#issuecomment-563318743
