Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

MKLDNN 3D dilate support #17915

Closed
ChaiBapchya opened this issue Mar 26, 2020 · 4 comments
Closed

MKLDNN 3D dilate support #17915

ChaiBapchya opened this issue Mar 26, 2020 · 4 comments
Labels
Feature request MKLDNN
Projects
CPU Performance and Quantization

Comments

@ChaiBapchya
Copy link
Contributor

MXNet supports 3D dilate on convolution. However, it doesn't leverage MKLDNN library.

As a result, one of our big customers faced the issue where GPU (3d conv dilate) ran in under 0.01 millisecond (as it leverages cuDNN library probably) while CPU takes ~200ms

Turns out MKLDNN since earlier versions v0.10 has been supporting conv dilate for 3d but with certain limitations. Maybe its time to port the functionality to MXNet and solve the customer problem.

Related issue
MKLDNN : oneapi-src/oneDNN#676
@sandeep-krishnamurthy
Copy link
Contributor

@PatricZhao - Can you please help us here? Thanks.

@TaoLv
Copy link
Member

TaoLv commented Mar 26, 2020

@ChaiBapchya Please try if #17884 works for you. Thanks.

@pengzhao-intel pengzhao-intel added this to To do in CPU Performance and Quantization via automation Mar 27, 2020
@pengzhao-intel
Copy link
Contributor

Thanks, @sandeep-krishnamurthy @ChaiBapchya we are working on this :)

@wuxun-zhang wuxun-zhang mentioned this issue Apr 8, 2020
Merged
3 tasks
@pengzhao-intel
Copy link
Contributor

Fixed by 17884

CPU Performance and Quantization automation moved this from To do to Done May 7, 2020
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Assignees
No one assigned
Labels
Feature request MKLDNN
Projects
No open projects
CPU Performance and Quantization
  
Done
Milestone
No milestone
Development

No branches or pull requests
4 participants
@sandeep-krishnamurthy @ChaiBapchya @TaoLv @pengzhao-intel