Lower LogSigmoid, LogSigmoidBackward, Elu, and EluBackward

[wonjoo-wj]

Lower LogSigmoid, LogSigmoidBackward, Elu, and EluBackward

As part of #3527, lowering existing ops that rely on IR level computation and removing the use of ScopePusher.

Testing unit tests succeeded locally.

[wonjoo-wj]

Unit tests failing:

[  FAILED  ] AtenXlaTensorTest.TestSelu
[  FAILED  ] AtenXlaTensorTest.TestSeluInPlace
[  FAILED  ] AtenXlaTensorTest.TestCelu
[  FAILED  ] AtenXlaTensorTest.TestCeluInPlace

Seems related to this PR, looking into it.

[wonjoo-wj]

This is a bit odd, all these unit tests are succeeding locally.

[JackCaoG]

can you do a full cpp test run using https://github.com/pytorch/xla/blob/master/test/cpp/run_tests.sh (./test/cpp/run_tests.sh) should work

[wonjoo-wj]

can you do a full cpp test run using https://github.com/pytorch/xla/blob/master/test/cpp/run_tests.sh (./test/cpp/run_tests.sh) should work

Thanks! Doing that enabled me to reproduce the failures locally.

I'm a bit confused here, according to https://github.com/pytorch/xla/blob/master/xla_native_functions.yaml, we don't lower selu or celu yet we have cpp tests for these ops. Not sure how this PR affected these tests. I'll try to build locally on master and check the difference.

[JackCaoG]

seems like torch::selu will call elu under the hood.

[wonjoo-wj]

Seemingly unrelated tests failing:

======================================================================
FAIL: test_addcmul_promotion_xla (__main__.TestTypePromotionXLA)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_internal/common_device_type.py", line 376, in instantiated_test
    result = test(self, **param_kwargs)
  File "/tmp/pytorch/xla/test/../../test/test_type_promotion.py", line 849, in test_addcmul_promotion
    self._ternary_promotion_common(device, op1, op2)
  File "/tmp/pytorch/xla/test/../../test/test_type_promotion.py", line 830, in _ternary_promotion_common
    self.assertEqual(res1, res2)
  File "/tmp/pytorch/xla/test/pytorch_test_base.py", line 628, in assertEqual
    return DeviceTypeTestBase.assertEqual(self, x, y, *args, **kwargs)
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_internal/common_utils.py", line 2240, in assertEqual
    msg=msg,
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_comparison.py", line 1074, in assert_equal
    raise error_metas[0].to_error()
AssertionError: The values for attribute 'dtype' do not match: torch.float64 != torch.complex128.

----------------------------------------------------------------------

Looking into it.

[JackCaoG]

You can rerun the test, upstream pr should be reverted.

[wonjoo-wj]

@JackCaoG, this should be ready for a quick review now. Thanks!

[JackCaoG]

Thanks!