wuxun-zhang commented on issue #17159: Performance regression from 1.4.1 to 1.5.1 URL: https://github.com/apache/incubator-mxnet/issues/17159#issuecomment-569084508 @jonatan1626 I just created an auto-run script to run benchmark with MXNet profiler. You can find these files in my [gist repo](https://gist.github.com/wuxun-zhang/8e1bc466fd08bc78561fd1620bfe58c3). Firstly, you need to apply the patch `symbolic_run.diff` to your original repo, and then just run `./auto_run_with_profiler.sh` to collect profiling information (will print the results into stdout). Below is what I got in my local machine. **mxnet-mkl 1.5.1** ``` operator ================= Name Total Count Time (ms) Min Time (ms) Max Time (ms) Avg Time (ms) ---- ----------- --------- ------------- ------------- ------------- _sg_mkldnn_conv 530000 93196.5547 0.0440 2.7290 0.1758 CopyCPU2CPU 10002 1750.9480 0.1330 0.2320 0.1751 _sg_mkldnn_fully_connected 10000 1408.5420 0.1180 0.4730 0.1409 BatchNorm 10000 889.8060 0.0480 0.4280 0.0890 SoftmaxOutput 10000 753.9630 0.0500 1.0160 0.0754 Pooling 20000 834.3010 0.0230 0.3370 0.0417 Flatten 10000 360.9760 0.0240 1.0100 0.0361 WaitForVar 10000 67.7780 0.0040 0.0230 0.0068 Reorder 10000 28.7770 0.0010 0.0130 0.0029 ``` **mxnet-mkl==1.4.1** ``` operator ================= Name Total Count Time (ms) Min Time (ms) Max Time (ms) Avg Time (ms) ---- ----------- --------- ------------- ------------- ------------- SoftmaxOutput 10000 310.8290 0.0240 0.0710 0.0311 DeleteVariable 20000 61.4390 0.0000 0.0120 0.0031 Reorder 30000 118.7160 0.0020 0.0120 0.0040 WaitForVar 10000 67.3600 0.0040 0.0160 0.0067 Pooling 20000 782.9750 0.0210 0.0790 0.0391 _sg_mkldnn_conv 530000 106754.1406 0.0370 4.6030 0.2014 BatchNorm 10000 517.9470 0.0430 0.0670 0.0518 FullyConnected 10000 2523.4150 0.2110 0.6170 0.2523 Flatten 10000 1566.0520 0.1320 0.4620 0.1566 CopyCPU2CPU 10002 1775.1851 0.1430 0.2000 0.1775 ```
---------------------------------------------------------------- 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: [email protected] With regards, Apache Git Services
