elvin-n commented on PR #11357:
URL: https://github.com/apache/tvm/pull/11357#issuecomment-1165504740

   @mbs-octoml  - several more answers on initial comment
   >    New annotator pass, which perhaps should also be made target-specific 
since the rules for deriving scope from shape seem pretty target specific. 
   
   The annotation pass consist of two parts - generic one and target specific. 
The generic one goes by graph and detects the targets and then construct name 
of the function including all found targets. It is not well robust for several 
targets, but works quite deterministic for one any target. We are introducing 
Adreno specific transformation, but other can be easily added. If we need to 
move adreno specific part into target specific directory, I will appreciate if 
you can suggest the proper place for target specific relay transformations.
   
   > I'm not sure how to reconcile the fairly generic multi-dimensional buffer 
support we now have with the hard 1d vs 2d distinction you're adding in graph 
executor. Is there a discussion about that someplace I've missed?
   
   There is no fairly generic multi-dimensional buffer in the memory planner - 
memory manager operates for now by continuous 1d memory blocks or other words 
flatten 1d buffers. Later on in tir this flatten memory can be managed as 
multidimensional array using tvm arithmetic. But on the memory management stage 
it is represented always as 1d memory. Until this PR.
   
   > I can't quite tell if your StorageInfo pass supports tuples & tuple 
projection. That whole pass needs commenting. But in any case worth mentioning 
you've eschewed the collect-and-solve constraint approach used by plan devices 
in favor of an eager back propagation of scopes from argument to constants. 
Fair enough, but please document the subset of relay this supports.
   
   It is still open question how to deal with tuple, I need some examples of 
ops producing tuples and not be merged into prim function. Could you please 
share which ops generate tuple of tensors as its output?


-- 
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]

Reply via email to