tqchen commented on PR #12084:
URL: https://github.com/apache/tvm/pull/12084#issuecomment-1183621971
When variable bound occurs, having an upper bound or rejecting it
ealy(assuming we don't want variable bound in compaction) are likely two
approaches. In rare case where we can do dyn shared mem allocation we might
want to keep alloca as it is without lifting(if it is only dependent on a
thread level variable).
Likely there is no one-shoe that fits all in here and perhaps we can get a
few more details. In most of the current usecases I am aware of we might want
to lean toward rejection, or fix some earlier part of the pipeline(e.g.
CompactBufferRegion) due to analysis problems.
> Variable size allocations occur because earlier passes move constant size
allocations to where the are accessed and update the allocation size to match
the access size.
This remark is interesting and it would be great to list out a TVMScript
example that it occur that motivates the change, just to see if we can reval
problems in our earlier part of the pipeline where we are supposed to fix.
--
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]
