Conversation
Summary: If we have a permute -> unary chain -> mean, based on the reduction dims of the mean, we can either fully remove the preceding permute or move the permute after the mean. Case 1: Dims after permute are still in same order with respect to each other, we can fully get rid of the permute and just update the reduction dims of the mean. Case 2: Not case 1. In this case, it's better to move the permute after the mean, since the permute will operate on less data. Differential Revision: D102268214
Summary: Pull Request resolved: pytorch#19106 I've seen lots of cases where we have a mean followed by a view which can either be fused into mean with keepdim True or False. These views typically become no-ops, however it's nice to remove them if we don't need them. Can potentially make work for later passes easier that are looking for mean, but view in the way. Differential Revision: D102276633
Summary: Pull Request resolved: pytorch#19120 If we have back to back slices on the same dimension, we can remove the top slice and just perform the second one. RemoveOrReplacePassInterface will handle a whole cascade if it exists. Differential Revision: D102278253 Reviewed By: abeakkas
Summary: If we have a permute -> slice, sometimes there are cases where slice would be better before the permute. If slice is not happening on dim 0, it's always better to reorder. If it is happening on dim 0, then slice is effectively a nop. But it could be better to run slice first if full_size > 2 * sliced_size (this is our cost model). Differential Revision: D102407566
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19126
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 1 New Failure, 1 Cancelled Job, 30 PendingAs of commit f708e8c with merge base 60ffe19 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
|
@DrJessop has exported this pull request. If you are a Meta employee, you can view the originating Diff in D102407566. |
This PR needs a
|
1 participant
Summary:
If we have a permute -> slice, sometimes there are cases where slice would be better before the permute.
If slice is not happening on dim 0, it's always better to reorder.
If it is happening on dim 0, then slice is effectively a nop. But it could be better to run slice first if full_size > 2 * sliced_size (this is our cost model).
Differential Revision: D102407566