Zha0q1 opened a new issue #20265: URL: https://github.com/apache/incubator-mxnet/issues/20265
I was trying to build MXNet 1.x with MKLDNN with ACL (Arm Compute Library) integration on an Arm instance. I used this [cmake config file](https://github.com/apache/incubator-mxnet/blob/v1.x/config/distribution/linux_aarch64_cpu.cmake) to integrate MKLDNN with ACL. The build was very performant and would surely benefit MXNet users hugely. I got a 3-4X boost with (16/64, 3, 512, 512) on Resnet compared to MKLDNN with no integration. However two operator unit test failed on this build: ``` test_operator.test_convolution_grouping test_operator.test_convolution_independent_gradients ``` I tried multiple mkldnn versions (v1.x now points to mkldnn release 2.0 beta 10, I also tried release 2.1 and 2.2) and ACl versions (20.08 and 21.2, 20.08 is Aug 2020 and 21.2 is the latest) and the failures persisted. This got me suspect that there is some integration issue with MXNet-MKLDNN (more possible) or MKLDNN-ACL. Would someone from the Intel team help share some insights on this? Would you help triage?@szha @leezu CC @josephevans @mseth10 @waytrue17 @sandeep-krishnamurthy -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@mxnet.apache.org For additional commands, e-mail: issues-h...@mxnet.apache.org
