manupa-arm commented on pull request #7785: URL: https://github.com/apache/tvm/pull/7785#issuecomment-828430178
> Hi @areusch , @tqchen , @manupa-arm , > About the target vs build discussion. I think we all agree that the executor should not be part of the target, but it should be a build parameter. > > The question is if adding the build parameter now, or if doing it in another PR. The point is that the executor is really only used in `build_module.cc` , so moving it as a build parameter seems the best choice. This also avoids hacky workarounds in situations where the `target_host` is not defined (e.g., `cuda`). > > I understand the argument to leave it as a target option now and then move all the target options in an option object in a later PR. But I would prefer to reduce the number of hacks for AOT from day 1, and in a later PR try to uniform crt and link-params to AOT. In other words, if there are no drawbacks, let's try to make AOT the "proper" way and let's then address the other target options to uniform to AOT. Thoughts? My two cents for this is : for now "executor" being only consumed by relay build, I'd would prefer it be an argument over putting it in the target (between the two options of having it inside the target vs relay.build arg). I agree with the general direction we need to a way to convey compilation flags that are not strictly associated with target. However, the options described here (runtime, link-params) runs deeper than the relay.build, therefore when those could be refactored into compiler options "object", we could also make the executor part of it. What do others think ? (I think what we are discussing is where to put the "executor" until compiler options "object" is introduced) -- 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]
