TMP 2022-01-13 Meeting notes

Andrew

• all outstanding merge requests are merged - down to a small number of TF test exclusions - some missed the 2.8 branch but will be in for 2.9

• rc0 build and uploaded to PyPi

• Looking at benchmarks - some depend on TF contrib

• Language benchmark - get right number of packages and proper version to install run - need nightly of rust to run(!) - working with Takis

  • Nathan - this is the mlcommons one? Yes - pointed to the Arm container

Nathan

• 2.8 patches landed, should include what is needed to support build with OneDNN/ACL with flag to toggle backend

• looks like Google is switching to AWS as the official aarch64 builds

• Would like an optimized TF build while not official would showcase improvements

Everton

• Finished Merge requests for Eigen, included request for Fujitsu

• Finished study of ….

• Started to create quantized operations in Eigen - created small unit tests and benchmark -

Gustavo

Working on getting board up to date with a 5.x kernel for the smart camera project.

Theo

• Nothing for this week, brief discussion about CI - no machines to target in LAVA(Yet!)

Takis

Tom

• working with Ryan for the lab LAVA support. D05, N01, FX700 is the priority order. Need them to address that cards that have been around for quite some time.

• Focused on Yolo-v5, have it working reasonably well with TF on CPU. Porting to TF Lite to be able to take advantage of ArmNN for smart camera project. Have TVM inferencing, outputs are ‘odd’ so coming up to speed with what TVM is doing. PyTorch import into TVM is fairly awful (surprisingly)