yelite commented on PR #79: URL: https://github.com/apache/tvm-rfcs/pull/79#issuecomment-1183563695
@xqdan Thanks! For your questions, > For compute fusion. With TE compute, it's easy to concate TE computes with producer-comsuer relation to get a fused compute. for example, conv + elemwise ops fusion. We should have similar function in TVM script. Which thread is related to this requirement? This will be supported naturally by the F3 (TE compute). F3 will be implemented by calling the IRBuilder APIs during parse-time evaluation. So one can write multiple TE compute side by side, like ``` x = T.compute(...) y = T.compute(...) ``` and the actual fusion is done in MetaSchedule and TIR. > For conditional lowering. We may have some attributtes in graph/relay level, which will further decide how to lower into different tir. With old ir builder/TE compute, we can do that. F4 in this RFC will ensure this,correct? Yes, this is correct. > For reducing boilerplate code. F3 is a good idea. Another one is we define a tir function (with or without host python code), and we reuse it other place. We see this in F4 which foucus on conditional lowering, however I think we should define/declare it as standalone Fearture. Mutual function call will be supported in the new TVMScript parser as well. We didn't include it in this RFC because it doesn't belong to the metaprogramming category. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
