TMP 2022-11-24 Meeting notes

Andrew

• TensorFlow 2.11.0, 2.10.1, 2.9.3 and 2.8.4 were all released and have been built and uploaded to PyPi. 2.8.4 needed a new test exclusion as the test was broken shortly before release.

• New Docker container is having its last small teething issues being sorted out so it can be uploaded to the Linaro account on Docker Hub. This should happen very soon. Once the Jenkins CI is moved to make use of this container, we can remove the last few excludes for the Eigen build as they pass in the new container. There are just a handful of failures in the oneDNN/ACL build left.

• Will be looking for other ways to improve TensorFlow for the future now.

Nathan

• Been looking at performance issues such as why using Eigen threadpool is worse than OpenMP threads.

• Talked about looking to see if there could be a way to extend the CI for oneDNN to get some ARM64 coverage because this is not currently happening at the moment. This can result in ARM64 issues only being noticed after a release when it is imported into downstream projects. Running the benchdnn tests more regularly would be ‘a good thing’.
  
  

Crefeda

• Also looking at performance issues.

• Discussion with Theo about possible ways to reduce the workload on the ASV runs so that a reasonable coverage is maintained.

Theo

• Talked about the addition of accuracy reporting to the ASV runs but that it came at the expense of needing to stop the batch size = 1 run.

• Discussion with Crefeda suggested that reducing the number of models tested for accuracy might allow enough time to keep running batch size = 1 as well as = 16 and = 32.