chrishkchris opened a new pull request #708: URL: https://github.com/apache/singa/pull/708
There are several improvement in the time profiling function 1. Greatly reduced the measurement overhead: In the new version, we synchronize the cudastream only one time at the end of iteration. The previous version needs to synchronize cudastream whenever it measures each operation, so it introduced time overhead. 2. Can skip the initial iterations for time profiling. It can be set using the function SetSkipIteration, while the default is 5 iterations. 3. Modify the benchmark.py for resnet 50 benchmark time profiling. ---------------------------------------------------------------- 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
