add and fix patches for PyTorch 1.9.0 on POWER

[Flamefire]

(created using eb --new-pr)

The patch introduced a C&P bug resulting in:

Vec256<float> C10_ALWAYS_INLINE pow(const Vec256<float>& exp) const {
    return {Sleef_powf4_u10vsx(_vec0, b._vec0), Sleef_powf4_u10vsx(_vec1, b._vec1)};
 }

I.e. b was used instead of exp.

Although in the referenced PR I left that function untouched it seems to be required as without it the test_binary_ufuncs fails but succeeds with it.
Added a commit to make the change correct.

[Micket]

lgtm

[Flamefire]

Build now works although 2 tests fail on PPC:

== building and installing PyTorch/1.9.0-foss-2020b...
== fetching files...
== ... (took 1 secs)
== creating build dir, resetting environment...
== unpacking...
== ... (took 3 secs)
== patching...
== ... (took 5 secs)
== preparing...
== ... (took 1 min 33 secs)
== configuring...
== ... (took 4 secs)
== building...
== ... (took 54 mins 31 secs)
== testing...
== ... (took 3 hours 1 min 22 secs)
== FAILED: Installation ended unsuccessfully (build directory: /dev/shm/s3248973-EasyBuild/PyTorch/1.9.0/foss-2020b): build failed (first 300 chars): 2 tests (out of 55803) failed:
* distributed/rpc/test_tensorpipe_agent
* test_binary_ufuncs (took 3 hours 57 mins 42 secs)
== Results of the build can be found in the log file(s) /tmp/easybuild-tmplog/easybuild-PyTorch-1.9.0-20220728.161426.vOeAA.log

[Micket]

lgtm (mainly because i don't know any better) (I'll try to find a machine to run build report from)

[Micket]

Oops, missed that it was marked draft!

[Flamefire]

Done now. Just noticed a mistake and wanted to wait for the custom test first. Started a job testing this on PPC now.

[Flamefire]

Test report by @Flamefire
FAILED
Build succeeded for 0 out of 1 (1 easyconfigs in total)
taurusml5 - Linux RHEL 7.6, POWER, 8335-GTX (power9le), 6 x NVIDIA Tesla V100-SXM2-32GB, 440.64.00, Python 2.7.5
See https://gist.github.com/a839027a57069711c8a3ed1f61691fdb for a full test report.

[Flamefire]

Test report by @Flamefire
FAILED
Build succeeded for 2 out of 3 (3 easyconfigs in total)
taurusml2 - Linux RHEL 7.6, POWER, 8335-GTX (power9le), 6 x NVIDIA Tesla V100-SXM2-32GB, 440.64.00, Python 2.7.5
See https://gist.github.com/11cef6c1050b7665d4124aaed082ee68 for a full test report.

[Flamefire]

Ok, this works now, the failing EC is PyTorch-1.9.0-fosscuda-2020b-imkl.eb as IMKL isn't available for PPC, so that is ok.

[Flamefire]

Test report by @Flamefire
SUCCESS
Build succeeded for 4 out of 4 (3 easyconfigs in total)
taurusi8006 - Linux CentOS Linux 7.9.2009, x86_64, AMD EPYC 7352 24-Core Processor (zen2), 8 x NVIDIA NVIDIA A100-SXM4-40GB, 470.57.02, Python 2.7.5
See https://gist.github.com/d0f4b4c8ce7f13c1410ea14a74ae7e07 for a full test report.

[boegel]

One failing test for PyTorch/1.9.0-fosscuda-2020b + PyTorch-1.9.0-fosscuda-2020b-imkl.eb on our Intel Cascade Lake V100 system (test report coming up), perhaps this one is also flaky: test_autograd

======================================================================
FAIL: test_thread_shutdown (__main__.TestAutograd)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test_autograd.py", line 4360, in test_thread_shutdown
    self.assertRegex(s, "PYTORCH_API_USAGE torch.autograd.thread_shutdown")
AssertionError: Regex didn't match: 'PYTORCH_API_USAGE torch.autograd.thread_shutdown' not found in 'PYTORCH_API_USAGE torch.python.import\nPYTORCH_API_USAGE c10d.python.import\nPYTORCH_API_USAGE tensor.create\n'

----------------------------------------------------------------------
Ran 951 tests in 23.361s

FAILED (failures=1, skipped=312, expected failures=1)
test_autograd failed!

[boegel]

Test report by @boegel
FAILED
Build succeeded for 2 out of 4 (3 easyconfigs in total)
node3307.joltik.os - Linux RHEL 8.4, x86_64, Intel(R) Xeon(R) Gold 6242 CPU @ 2.80GHz (cascadelake), 1 x NVIDIA Tesla V100-SXM2-32GB, 510.73.08, Python 3.6.8
See https://gist.github.com/3fe5074fc31e2fd614025199814ce7ab for a full test report.

[boegel]

I won't let the failing test_autograd test block this PR, since it also occurs without the changes in this PR.

@Flamefire Is this something worth following up on?

[boegel]

Going in, thanks @Flamefire!

[Flamefire]

Yes, looks worth to double check at least: #16233

[Flamefire]

Test report by @Flamefire
FAILED
Build succeeded for 2 out of 4 (3 easyconfigs in total)
taurusml27 - Linux RHEL 7.6, POWER, 8335-GTX (power9le), 6 x NVIDIA Tesla V100-SXM2-32GB, 440.64.00, Python 2.7.5
See https://gist.github.com/5b6c26a1ef289988bf456f65e14d2ce3 for a full test report.