drivanov commented on issue #16178: [WIP]improving argmax perf URL: https://github.com/apache/incubator-mxnet/pull/16178#issuecomment-534303397 I just finished comparative profiling of currently used `argmax` vs. #16178 vs. #16218 on CPU. Three versions of `argmax` were tested on - 3 axes of tensors of 24 different 3-dimentional shapes; - 2 axes of tensors of 16 different 2-dimensional shapes. (totally on 104 = 3*24 + 2*16 different combinations (shape, axis)) Here is a short summary: ``` Current: #16178 #16218 Average improvement: 1x 2.283x 4.263x Maximal improvement: n/a 6.323x 14.318x # of losses to Current Argmax: 0 22 1 # of losses to #16178 82 0 5 # of losses to #16218 103 99 0 ``` Complete results are in: [ArgMax Results 16178_vs_16218.zip](https://github.com/apache/incubator-mxnet/files/3644322/ArgMax.Results.16178_vs_16218.zip) According to these results #16218 is better. Besides, it uses the same templates for GPU version of `argmax` and ``` Average Improvement for GPU: 8.905 Maximal improvement for GPU: 119.704 # of losses to Current Argmax: 11 ```
---------------------------------------------------------------- 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
