jackwish commented on issue #4277: [ARM][Topi] Improving Int8 Perf in Spatial Conv2D schedule. URL: https://github.com/apache/incubator-tvm/pull/4277#issuecomment-552157956 > @jackwish Yes, sharing code will be very helpful. For now, do you prefer changing schedule OR Legalize pass? > > Changing Schedule > > * (+) The change is contained to only one template. > * (-) Schedules are generally hard to understand. In general, I try to avoid complicating them. > > Legalize Pass > > * (+) The codebase might be cleaner (if one understands Relay passes). > * (+) Topi schedules are generally hard to understand. In general, I try to avoid complicating them. This avoids that. > * (-) Applies to all templates. For example, if I write a new template (like direct, winograd), legalize will upcast. > > I dont have any strong opinion. Hi @anijain2305 , I think a *Legalize Pass* may lead to a clean *code structure* or *architecture*, while a *Changing Schedule* contribute to a *easy to read*. I feel the same on *having no strong opinion* :(. Yet, one thing that I am considering is that, if the legalization is going to handle all, will it make the *tensorize compute pattern matching* a bit more hard to write or debug or something. By this I mean, in schedule we write *int8 cast to int32 and multiply*, legalization rewrites to *int8 cast to int16 and multiplied to int32*, so what is the compute pattern shall a tensorize to match? Not sure if this will be a problem :)
---------------------------------------------------------------- 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: [email protected] With regards, Apache Git Services
