TMP 2021-12-02 Meeting notes

Everton

• Working on Eigen presentation, and juggling the GEMMLOMP - looks like it needs a change to round to nearest - fixes the bug, then cause 3 more test cases to fail.

Crefeda

Helped to connect Andrew/Everton to Google (Thanks!) on some of the issues

Andrew

Bug spotted in the XLA compiler where it tries to read the llvm cycle counter which is forbidden on aarch64, maintainer suggests to exclude the test however this exposes the potential for a sigill

Theo

be back in Cambridge next week - focused on tf.org for now

Nathan

Working on improving ACL builds/functionality - looking to get OneDNN a flag away from being within the Tensorflow build - not quite ready to be the default build as it is missing Eigen thread pools, training. Request : have the unit test results for this build available side by side with the current. Timeframe approx January as the code makes it ways upstream (2.8 release, mainline approx December)

Tom

Fixing changes needed to tvm-bench as upstream is morphing some of the apis it uses. Finished the removal of the calls to json. Now working to not pass cpu_target to relay.build, instead using the Target constructor.

Looking at if there is something out in the wild to measure cloud inference cost of arch vs arch. In particular we’d like to see Arm servers compete / beat intel especially at cloud inference.