enable MKL-DNN/oneDNN usage for aarch64 CPUs in TensorFlow 2.5+

[Flamefire]

(created using eb --new-pr)

[boegel]

Test report by @boegel

Overview of tested easyconfigs (in order)

• FAIL (build issue) TensorFlow-2.5.0-foss-2020b.eb (partial log available at https://gist.github.com/5f40f2979bd293569ab1b0c4849533e2)
• FAIL (build issue) TensorFlow-2.6.0-foss-2021a.eb (partial log available at https://gist.github.com/de42ab10b4ea1001ed6af06cc791c45a)

Build succeeded for 0 out of 2 (2 easyconfigs in total)
select-pika-c6gd-2xlarge-0001 - Linux centos linux 8.3.2011, AArch64, ARM UNKNOWN (graviton2), Python 3.6.8
See https://gist.github.com/5a6ba346212cb80fda239c0f1a73198f for a full test report.

[Flamefire]

@boegel

gcc: fatal error: Killed signal terminated program cc1plus

Any cgroup stuff maybe?

[boegel]

@boegel

gcc: fatal error: Killed signal terminated program cc1plus

Any cgroup stuff maybe?

Or perhaps lack of memory...

This was on a c6g.2xlarge AWS instance (8 cores, 16GB of RAM), which is perhaps a bit on the light side to build TensorFlow from source (using 8 cores)?

I'll try again on a fatter node (or using less cores).

[boegel]

Test report by @boegel

Overview of tested easyconfigs (in order)

• FAIL (build issue) TensorFlow-2.5.0-foss-2020b.eb (partial log available at https://gist.github.com/e98e1014edb6d420afbd778addd1f6ac)
• FAIL (build issue) TensorFlow-2.6.0-foss-2021a.eb (partial log available at https://gist.github.com/da8ea294f219964cdf480e02b542f845)

Build succeeded for 0 out of 2 (2 easyconfigs in total)
select-pika-c6gd-4xlarge-0001 - Linux centos linux 8.3.2011, AArch64, ARM UNKNOWN (graviton2), Python 3.6.8
See https://gist.github.com/1b7f95f9ca3960d7e3d2a2de5403ce0a for a full test report.

[Flamefire]

@boegel Bug in TF with excluding based on arch and a config flag leading to ambigious, yet same select cases: https://github.com/tensorflow/tensorflow/blob/72fd2bfa42a8ad909baf8d2b7b674563d256514d/tensorflow/core/kernels/BUILD#L3028-L3034 (arm_any and no_mkldnn_contraction_kernel triggers)
Can you open a TF bug?

[boegel]

@Flamefire It's not clear to me what's going wrong here, so I'm not sure what to put in the issue. Also, I have very little bandwidth to follow up on it...

Does it basically mean that the --config=mkl_aarch64 doesn't work at all, or is it more subtle than that?

[Flamefire]

Did that: tensorflow/tensorflow#52027

More or less: yes. However for ARM we might want to use --define build_with_mkl_aarch64=true anyway, see #2577 (comment)
Or None at all, i.e. change with_mkl_dnn to False, especially as the name is likely misleading...

[Flamefire]

Closing as this isn't actually enabling mklDNN/oneDNN. See #2577 and https://groups.google.com/a/tensorflow.org/g/build/c/RZhgZst-fgQ