Uh oh!
There was an error while loading. Please reload this page.
[Paddle] Eliminate amax update bubbles by using custom_ops - #436
Merged
timmoon10 merged 2 commits intoSep 22, 2023
Conversation
zlsh80826force-pushed
the
rewang/eliminate-amax-update-bubbles
branch
from
September 18, 2023 15:48
cdd2a51 to
4d3b5a8Comparezlsh80826
commented
Sep 18, 2023
CollaboratorAuthor
/te-ci |
zlsh80826force-pushed
the
rewang/eliminate-amax-update-bubbles
branch
from
September 19, 2023 06:50
4d3b5a8 to
81fcae4CompareSigned-off-by: rewang <rewang@nvidia.com>
zlsh80826force-pushed
the
rewang/eliminate-amax-update-bubbles
branch
from
September 19, 2023 06:52
81fcae4 to
7b55091Comparezlsh80826
marked this pull request as ready for review
September 19, 2023 07:01
zlsh80826
commented
Sep 19, 2023
CollaboratorAuthor
/te-ci |
Uh oh!
There was an error while loading. Please reload this page.
Signed-off-by: rewang <rewang@nvidia.com>
zlsh80826
commented
Sep 20, 2023
CollaboratorAuthor
/te-ci |
ptrendx pushed a commit
that referenced
this pull request
Oct 2, 2023
* Eliminate amax_and_scale_update bubbles Signed-off-by: rewang <rewang@nvidia.com> * Add CUDA check Signed-off-by: rewang <rewang@nvidia.com> --------- Signed-off-by: rewang <rewang@nvidia.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Paddle dynamic graph inserted a few bubbles when using assign operators/slice operators.
This PR moves
amax_and_scale_updateandcopy_amax_from_bufferto custom_ops in c level to eliminate those tiny kernels and bubbles.copy_amax_from_buffer: There was a slice kernel for accessingamax_history[0]and multiple kernels for=operator. By using the customizedcopy_amax_from_bufferop, all of above kernels can be reduced as a single memcpy. The spent time reduced from 18us to 1.8us.amax_and_scale_update: There were multiple tiny kernels for updating amax/scale/scale_inv/history. This PR fused a few assign/elementwise operators to a single kernel to reduce the total 13 kernels to 4 kernels. The spent time reduced from ~39us to ~15us.