[GitHub] [tvm] xqdan commented on pull request #7060: [WIP] Add MLIR Relay Dialect, Translate MLIR HLO to MLIR Relay to Relay

[GitBox <gi...@apache.org>]

xqdan commented on pull request #7060:
URL: https://github.com/apache/tvm/pull/7060#issuecomment-785652300


   @yongwww @tqchen @zhiics I suggest we support a kind of graph ir with boundry in relay. Let me explain a little bit. XLA is expanding big ops with fine-grained primitive ops, but losing high level information, that is the boundry of big ops, and which brings lots of trouble to fusion, since it's very difficult to do fusion for such huge graph with thoudsands of fine-grained primitive ops. But if we can keep the boundry when we expand/convert big ops into relay graph ir, we can avoid the problem above, besides we can do whatever XLA can do with this graph ir with boundry info. with this design, tvm can be a more powerful deep learning compiler, also can be a more powerful plugin optimizer for other ai frameworks.


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org