Add CI/CD for ARM64 using GitHub Actions

[mseth10]

• Added GitHub workflows to build, test and upload ARM64 pip wheels for Python versions 3.7, 3.8, 3.9 and 3.10
• Workflow jobs run on self-hosted runners on AWS Graviton instances
• CI workflow uploads generated pip wheels to GitHub Actions and is triggered for:
  • PR on master branch
  • PR on branch name starting with r2.
• CD workflow uploads generated pip wheels to PyPI and is triggered for:
  • New tags starting with v2.
  • Nightly on master branch at 8am UTC
• Following 18 tests currently fail and are skipped:

  tensorflow/python:nn_grad_test_cpu
  tensorflow/python:dequantize_op_test
  tensorflow/python:quantized_ops_test
  tensorflow/python/data/experimental/kernel_tests/service:cross_trainer_cache_test
  tensorflow/python/eager:forwardprop_test_cpu
  tensorflow/python/framework:node_file_writer_test_cpu
  tensorflow/python/grappler:memory_optimizer_test
  tensorflow/python/kernel_tests/linalg:linear_operator_householder_test_cpu
  tensorflow/python/kernel_tests/linalg:linear_operator_inversion_test_cpu
  tensorflow/python/kernel_tests/linalg:linear_operator_block_diag_test_cpu
  tensorflow/python/kernel_tests/linalg:linear_operator_block_lower_triangular_test_cpu
  tensorflow/python/kernel_tests/linalg:linear_operator_kronecker_test_cpu
  tensorflow/python/kernel_tests/math_ops:batch_matmul_op_test_cpu
  tensorflow/python/kernel_tests/nn_ops:conv_ops_test_cpu
  tensorflow/python/kernel_tests/nn_ops:conv2d_backprop_filter_grad_test_cpu
  tensorflow/python/kernel_tests/nn_ops:conv3d_backprop_filter_v2_grad_test_cpu
  tensorflow/python/kernel_tests/nn_ops:atrous_conv2d_test_cpu
  tensorflow/python/ops/parallel_for:math_test_cpu
  

[angerson]

Thanks a lot for your work on this, @mseth10.

The files changed here are very, very old -- at least two separate revisions of our CI process ago -- and are on the backlog of things we'd like to remove, since they have caused a lot of confusion. What led you to use these files for the tests (our CONTRIBUTING.md, perhaps)? Can we work next on upgrading to a different setup, after we get it working in its current iteration?

[mseth10]

The files changed here are very, very old -- at least two separate revisions of our CI process ago -- and are on the backlog of things we'd like to remove, since they have caused a lot of confusion. What led you to use these files for the tests (our CONTRIBUTING.md, perhaps)? Can we work next on upgrading to a different setup, after we get it working in its current iteration?

Hi @angerson , thanks for pointing that out. Are we talking about a particular file here? Or the entire tensorflow/tools/ci_build/ folder is old? I searched the repository and docs (including CONTRIBUTING.md) and found only these files to be related to pip wheel generation. Please point me to the new process, I can upgrade in the next PR. We can merge this PR in its current state to get the wheel generation process started for nightly and 2.9.0 release.

[mihaimaruseac]

In principle our release jobs are now under tensorflow/tools/ci_build/rel and we were planning to deprecare the rest, especially pip_new.sh which is inscrutable, requiring envvars to be properly set before being called.

But I would say we should go with this for now and plan a future migration.

[martin-g]

Apologies for commenting on a closed PR!
Is the work in this PR supposed to publish an aarch64 wheel for https://pypi.org/project/tensorflow-cpu/2.9.1/#files ?
If NOT then where should I look at for it ?
Thank you!

[elfringham]

@martin-g You can find it at https://pypi.org/project/tensorflow-cpu-aws/#files but once it is publicly announced you will be able to just 'pip install tensorflow' and get the appropriate wheel for your platform.

[mihaimaruseac]

Not for TF 2.9 (since we don't change branches (much) after a release), but it will in TF 2.10.

Meanwhile, they show up in nightly: https://pypi.org/project/tf-nightly/#files (actually, only a wrapper wheel that then redirects to the aarch64 wheels in a separate PyPI project to minimize need to reuse credentials, etc.)

[martin-g]

Thank you!