[Inductor][CPP] Fix layout for local buf in outer loop fusion #160857
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/160857
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit d62fb9f with merge base a4fc051 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
CaoE
added
topic: not user facing
CaoE
changed the title
CaoE
requested review from
jgong5 and
leslie-fang-intel
and removed request for
jgong5
| # Local Buffer is a view of global buffer | ||
| local_buffer_stride: list[int] = [] | ||
| stride = global_buffer_layout.stride[-1] | ||
| local_buffer_size = get_call_ranges(scheduler_node)[ |
There was a problem hiding this comment.
is this case scheduler_node also a view of global_buffer?
There was a problem hiding this comment.
In this case global_buffer is scheduler_node.node.
There was a problem hiding this comment.
then why can't use global_buffer_layout.size[size_offset:] directly?
There was a problem hiding this comment.
Because global_buffer_layout is the tensor layout while size_offset is the loop depth. The dimension of global_buffer_layout may not be the same as loop number, e.g, there are merged dims.
In this case, global_buffer_layout size is [5, 1, 32, 32] but the callrange is [5, 1024]. If use global_buffer_layout.size[size_offset:] to create local_buffer_layout we get the size [32] but we need [1024].
| continue | ||
| # Local Buffer is a view of global buffer | ||
| local_buffer_stride: list[int] = [] | ||
| stride = global_buffer_layout.stride[-1] |
There was a problem hiding this comment.
Will this work for a size = [] tensor?
There was a problem hiding this comment.
global_buffer is an instance of ir.ComputedBuffer. We haven't encountered a size = [] tensor yet. Is there a possible case for this? @leslie-fang-intel Could you please help on this question ?
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
…h#160857) Fixes pytorch#159154 Pull Request resolved: pytorch#160857 Approved by: https://github.com/leslie-fang-intel, https://github.com/jansel
5 participants
Fixes #159154
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben