Skip to content

Conversation

@crazydemo
Copy link

  • add more patterns in dnnl.py.
  • add test cases, which have covered the support ops and patterns, in test_dnnl.py.
  • add implementation of new added ops and patterns in dnnl_json_runtime.cc.
  • add support to dilated conv2d with dnnl backend.

@crazydemo crazydemo force-pushed the dev-byoc_partition branch 2 times, most recently from 449b68a to ed87f71 Compare December 27, 2021 05:39
Copy link
Contributor

@jwfromm jwfromm left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for this change, it looks really excellent. I have a few small comments but am curious if youve done any benchmarking with this new pattern fusion. If so, how much do these changes speed up models you tried?

@crazydemo
Copy link
Author

Thanks for this change, it looks really excellent. I have a few small comments but am curious if youve done any benchmarking with this new pattern fusion. If so, how much do these changes speed up models you tried?

Thank you for your comments. We do have done some benchmark, and the related results are posted on RFC.

Note that, the results in RFC are collected with the registered pattern and optimization from alter_op_layout pass.
We are going to submit another two layout transform related PRs to fully support the performance gain claimed in RFC.

@masahi masahi merged commit 75cd670 into apache:main Dec 29, 2021
ylc pushed a commit to ylc/tvm that referenced this pull request Jan 7, 2022
* add unit test for byoc-dnnl

* add byoc-dnnl pattern and their test cases
ylc pushed a commit to ylc/tvm that referenced this pull request Jan 13, 2022
* add unit test for byoc-dnnl

* add byoc-dnnl pattern and their test cases
liaopeiyuan added a commit to zk-ml/tachikoma that referenced this pull request Sep 15, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants