[CI] update oneDNN to v2.6

[yangulei]

I submitted the PR to enable bfloat16 in DNNL BYOC(#11111). But the PR failed to be built due to the version of the oneDNN in CI is limited to v2.2, which is a relative old one. So I update the CI script to get and build the latest oneDNN release.
I had tested the script on my own PC and it works fine, but I don't know how to test the script in CI environment. So let me know if I did something inappropriate, thanks!

cc @Mousius @areusch @driazati

[driazati]

CI tests your changes automatically so no worries there, I would prefer that we just upgrade the oneDNN version manually (and keep the version pinned) though instead of grabbing the latest (imagine that there is a breaking update, the Docker image build would just stop working one day)

[yangulei]

CI tests your changes automatically so no worries there, I would prefer that we just upgrade the oneDNN version manually (and keep the version pinned) though instead of grabbing the latest (imagine that there is a breaking update, the Docker image build would just stop working one day)

Thanks for your comments and suggestions. I had modified the script to pin oneDNN to the latest release v2.6, and it's easy to change the pinned version or set to build the latest release.

[yangulei]

The CI failed due to ConnectTimeoutError during the installation of tensorflow-aarch64 v2.6.2:

[2022-04-28T01:43:42.770Z]` WARNING: Retrying (Retry(total=0, connect=None, read=None, redirect=None, status=None)) after connection broken by 'ConnectTimeoutError(<pip._vendor.urllib3.connection.VerifiedHTTPSConnection object at 0xffff827a3050>, 'Connection to snapshots.linaro.org timed out. (connect timeout=15)')': /ldcg/python-cache/tensorflow-aarch64/
[2022-04-28T01:43:55.126Z] ERROR: Could not find a version that satisfies the requirement tensorflow-aarch64==2.6.2 (from versions: 1.2, 2.7.0rc0, 2.7.0rc1, 2.7.0, 2.7.1, 2.8.0rc0, 2.8.0rc1, 2.8.0, 2.9.0rc0, 2.9.0rc1)
[2022-04-28T01:43:55.126Z] ERROR: No matching distribution found for tensorflow-aarch64==2.6.2
[2022-04-28T01:44:00.406Z] WARNING: You are using pip version 19.3.1; however, version 22.0.4 is available.
[2022-04-28T01:44:00.406Z] You should consider upgrading via the 'pip install --upgrade pip' command.
[2022-04-28T01:44:00.406Z] The command '/bin/sh -c bash /install/ubuntu_install_tensorflow_aarch64.sh' returned a non-zero code: 1
[2022-04-28T01:44:00.406Z] ERROR: docker build failed.
script returned exit code 1

Should I do something or just ignore this?

[areusch]

@yangulei i think this means the CI's quota to PyPI was exceeded. perhaps retry now that it's been a while?

[driazati]

Yeah that index was down for a while last week (#11156 (comment)) but it should be working now. If this happens again we might want to look into rehosting some of those assets

[yangulei]

@areusch @driazati
PyPi works fine, thanks!
Looks like we should install the package to /usr/lib instead of /usr/local in the latest main branch. I have modified my script (follow the installation of gtest) and it pass all the CI now.

[yangulei]

Hi @areusch @driazati,
PR #11111 depends on this PR and had waited for 3 weeks, so I want this PR to be merged if it's ready.
Could you please take a look at this script and approve it if no more modifications are required?
Thanks a lot.

[areusch]

thanks @yangulei and sorry for the delay here!

[masahi]

@areusch We still need to go through the actual update process for the cpu image right? I've got another request for gpu update in #11346, I can do the cpu update as well.

[yangulei]

Thank you @areusch, I will rebase #11111 once the docker image is ready.