enable better depthwise conv perf on cudnn 8.2+ #58749
Closed
Conversation
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
Contributor
🔗 Helpful links
💊 CI failures summary and remediationsAs of commit f4b5067 (more details on the Dr. CI page):
4 failures not recognized by patterns:
❄️ 2 failures tentatively classified as flakybut reruns have not yet been triggered to confirm:
|
H-Huang
added
the
triaged
Contributor
Author
|
Updated result spreadsheet with V100 result. @ptrblck |
FDecaYed
force-pushed
the
deyuf/update_depthwise_cudnn
branch
from
414c0ac to
f4b5067
Compare
CI Flow Status⚛️ CI FlowRuleset - Version:
You can add a comment to the PR and tag @pytorchbot with the following commands: # ciflow rerun, "ciflow/default" will always be added automatically
@pytorchbot ciflow rerun
# ciflow rerun with additional labels "-l <ciflow/label_name>", which is equivalent to adding these labels manually and trigger the rerun
@pytorchbot ciflow rerun -l ciflow/scheduled -l ciflow/slowFor more information, please take a look at the CI Flow Wiki. |
Contributor
Author
|
@ptrblck Let's try to merge this now that cudnn 8.2 is more widely used.
|
ngimel
approved these changes
Oct 13, 2021
Collaborator
ngimel
left a comment
ngimel
left a comment
There was a problem hiding this comment.
Looks good, thanks!
But generally, since networks should be in channels-last already, that shouldn't affect much?
Contributor
|
@ngimel has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Contributor
Author
Yes this is for NCHW since channel last already call cudnn. Whoever not running channel-last should see some benefit |
fixed condition for 1D conv check
Contributor
|
@ngimel has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
facebook-github-bot
pushed a commit
that referenced
this pull request
Oct 23, 2021
Summary: There are multiple improvement of depthwise convolution speed in cudnn between 7.6 and 8.2, since #22302. This PR aim to harvest all the new improvement by enable more cudnn kernel. The workload checking logic can also be simplified now. To keep the change simple, I kept things before cudnn 8.2 unchanged. Similar to #22302, I used a script [here](https://gist.github.com/FDecaYed/e8ba98a95cd33697df2ace86fdb44897) to benchmark. Both run are using cudnn 8.2 One enhancement I did to the script is switch to event based timing. With warmup kernels to fill the launch queue ahead, this should give us accurate kernel timing even in CPU launch bound cases. Here is A100 and V100 result sorted by speedup. [Book1.xlsx](https://github.com/pytorch/pytorch/files/6530371/Book1.xlsx) Result highlights: Newly turned on 5x5 cudnn kernel show up to 6x speedup. Close to half of test sizes show >10% speedup. Fixed some corner cases that previously caused 15-20x slowdown. Only slowdown a handful of cases(~10 out of >1000) Pull Request resolved: #58749 Reviewed By: bdhirsh Differential Revision: D31613199 Pulled By: ngimel fbshipit-source-id: 883b58facad67ccd51dc9ab539368b4738d40398
langong347
pushed a commit
that referenced
this pull request
Oct 25, 2021
Summary: There are multiple improvement of depthwise convolution speed in cudnn between 7.6 and 8.2, since #22302. This PR aim to harvest all the new improvement by enable more cudnn kernel. The workload checking logic can also be simplified now. To keep the change simple, I kept things before cudnn 8.2 unchanged. Similar to #22302, I used a script [here](https://gist.github.com/FDecaYed/e8ba98a95cd33697df2ace86fdb44897) to benchmark. Both run are using cudnn 8.2 One enhancement I did to the script is switch to event based timing. With warmup kernels to fill the launch queue ahead, this should give us accurate kernel timing even in CPU launch bound cases. Here is A100 and V100 result sorted by speedup. [Book1.xlsx](https://github.com/pytorch/pytorch/files/6530371/Book1.xlsx) Result highlights: Newly turned on 5x5 cudnn kernel show up to 6x speedup. Close to half of test sizes show >10% speedup. Fixed some corner cases that previously caused 15-20x slowdown. Only slowdown a handful of cases(~10 out of >1000) Pull Request resolved: #58749 Reviewed By: bdhirsh Differential Revision: D31613199 Pulled By: ngimel fbshipit-source-id: 883b58facad67ccd51dc9ab539368b4738d40398
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 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.
There are multiple improvement of depthwise convolution speed in cudnn between 7.6 and 8.2, since #22302.
This PR aim to harvest all the new improvement by enable more cudnn kernel. The workload checking logic can also be simplified now.
To keep the change simple, I kept things before cudnn 8.2 unchanged.
Similar to #22302, I used a script here to benchmark. Both run are using cudnn 8.2
One enhancement I did to the script is switch to event based timing. With warmup kernels to fill the launch queue ahead, this should give us accurate kernel timing even in CPU launch bound cases.
Here is A100 and V100 result sorted by speedup.
Book1.xlsx
Result highlights:
Newly turned on 5x5 cudnn kernel show up to 6x speedup.
Close to half of test sizes show >10% speedup.
Fixed some corner cases that previously caused 15-20x slowdown.
Only slowdown a handful of cases(~10 out of >1000)