Skip to content

DNN: Further optimization of Conv2D#22401

Merged
vpisarev merged 1 commit intoopencv:4.xfrom
zihaomu:conv2d_optimize
Aug 26, 2022
Merged

DNN: Further optimization of Conv2D#22401
vpisarev merged 1 commit intoopencv:4.xfrom
zihaomu:conv2d_optimize

Conversation

@zihaomu
Copy link
Copy Markdown
Member

@zihaomu zihaomu commented Aug 19, 2022 •
edited
Loading

The optimization point of this PR:

  1. Bring and adapt the latest the Ficus OpConv.fx code.
  2. Fused Conv+Add+Activation. (Currently only for Conv2D, in the future FastConv can support Conv1D and Conv3D, we will also support both.)
  3. FastConv branch still reuses the weightsMat, and remove the fastWeights.
  4. Optimize the Winograd_F63: pack tile 12 and adjust the data pipeline. (About 1 ms speedup on M1 chip, testing model: ResNet50).

TODO List:

  • Find out why MobileNetv2 is slow. It is caused by the difference between OpenMP and GCD.

Performance Test on ARM (Appel M1 Chip, 4 threads)

Model Name Wthout Patch With Patch
ReseNet 50 26.8 ms 24.6 ms
MobileNetv2 5.43 ms GCD: 5.5 ms, OpenMP: 4.7 ms

Performance Test on ARM (Raspberry Pi 4, A72, 4 threads)

Model Name Without Patch With Patch NCNN's Benchmark
ReseNet 50 440.90 ms 400.6 ms (10% faster) 330 ms
MobileNetv2 51.64 ms 51.2 ms 71 ms

Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

  • I agree to contribute to the project under Apache 2 License.
  • To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
  • The PR is proposed to the proper branch
  • There is a reference to the original bug report and related work
  • There is accuracy test, performance test and test data in opencv_extra repository, if applicable
    Patch to opencv_extra has the same branch name.
  • The feature is well documented and sample code can be built with the project CMake

@zihaomu zihaomu requested a review from vpisarev August 19, 2022 03:54
@zihaomu zihaomu mentioned this pull request Aug 19, 2022
Merged
6 tasks
@zihaomu zihaomu changed the title DNN: Further optimization of Conv2D, fused Conv_Add_Activation. DNN: Further optimization of Conv2D Aug 19, 2022
@zihaomu zihaomu marked this pull request as ready for review August 23, 2022 11:09
@vpisarev vpisarev merged commit bb64db9 into opencv:4.x Aug 26, 2022
@asenyaev
Copy link
Copy Markdown
Contributor

asenyaev commented Aug 29, 2022

Seems that merging this PR a build with coverage started to fail (log).

@zihaomu
Copy link
Copy Markdown
Member Author

zihaomu commented Aug 29, 2022

Seems that merging this PR a build with coverage started to fail (log).

Thanks, @asenyaev. I will submit a PR to fix it.

@asenyaev
Copy link
Copy Markdown
Contributor

asenyaev commented Aug 29, 2022

@zihaomu, thank you!

@zihaomu zihaomu mentioned this pull request Aug 29, 2022
Merged
6 tasks
@alalek alalek mentioned this pull request Aug 31, 2022
Merged
6 tasks
@alalek alalek mentioned this pull request Jan 8, 2023
Merged
@asmorkalov asmorkalov added this to the 4.7.0 milestone Jan 23, 2023
a-sajjad72 pushed a commit to a-sajjad72/opencv that referenced this pull request Mar 30, 2023
@zihaomu
Further optimization of Conv2D, fused Conv_Add_Activation, bring late…
40c9927 
…st code from ficus OpConv.fx. (opencv#22401)
@opencv-alalek opencv-alalek mentioned this pull request Aug 15, 2023
Merged
6 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@vpisarev vpisarev vpisarev approved these changes

Assignees

No one assigned

Labels

category: dnn optimization

Projects

None yet

Milestone

4.7.0

Development

Successfully merging this pull request may close these issues.

4 participants

@zihaomu @asenyaev @vpisarev @asmorkalov