💊 CI failures summary and remediations

As of commit 398ed32 (more details on the Dr. CI page):

💚 💚 Looks good so far! There are no failures yet. 💚 💚

@mruberry The PR implements both isposinf and isneginf methods according to the numpy's APIs.
Please kindly help review it.

pytorch_xla_linux_bionic_py3_6_clang9_test job failed due to the XLA tensors do not have storage.

Jul 20 04:05:30 ERROR [0.006s]: test_isneginf_float_xla_float32 (__main__.TestTorchDeviceTypeXLA)
Jul 20 04:05:30 ----------------------------------------------------------------------
Jul 20 04:05:30 Traceback (most recent call last):
Jul 20 04:05:30   File "/opt/conda/lib/python3.6/site-packages/torch/testing/_internal/common_device_type.py", line 241, in instantiated_test
Jul 20 04:05:30     result = test(self, device_arg, dtype)
Jul 20 04:05:30   File "/var/lib/jenkins/workspace/xla/test/../../test/test_torch.py", line 6495, in test_isneginf_float
Jul 20 04:05:30     self.compare_with_numpy(torch.isneginf, np.isneginf, vals, device, dtype)
Jul 20 04:05:30   File "/opt/conda/lib/python3.6/site-packages/torch/testing/_internal/common_utils.py", line 885, in compare_with_numpy
Jul 20 04:05:30     torch_result = torch_fn(t).cpu()
Jul 20 04:05:30 RuntimeError: torch_xla/csrc/tensor_impl.cpp:144 : XLA tensors do not have storage

Please kindly let me know whether I need to update the logic for supporting XLA, or skip the XLA tests with onlyOnCPUAndCUDA decorator.

cc: @JackCaoG for the XLA failure.

Apologizes for triggering requesting reviews from code owners due to the inappropriate operations of git merge.
After investigations, I found the following Github docs that contains the suitable commands for merging upstream changes into the fork.

$ git pull https://github.com/pytorch/pytorch.git master

I will use this command in further contributions.

The CI checks all passed thanks to ailzhang's kind suggestions regarding XLA, and this PR is ready for review,

Reviewed the user of TensorIterator. You're on the right path but I suggest using a trick for bool and integer cases that will greatly simplify your code. I also have a question about how the bfloat16 dispatch path works since I don't think the cpp standard supports numeric limits for bfloat16? Maybe I'm mistaken.

The tests look good. I'm sorry I wrote something confusing about the documentation in our last discussion. I added more context. Please let me know if it makes sense now.

@mruberry I cannot thank you more for the kind help and mentoring during this PR, the kernels have been simplified significantly by adopting the suggested approach.

NumPy supports users in passing non-bool arrays to the out= parameters.

In [1]: import numpy as np
In [2]: x = np.array([float('inf'), float('inf')])
In [3]: y = np.array([10, 20])
In [4]: np.isposinf(x, out=y)
Out[4]: array([1, 1])
In [5]: y
Out[5]: array([1, 1])

Thus I kept cast_common_dtype_to_outputs and didn't add bool check for result tensors.

I also experimented with dispatching bfloat16 in my local environment.
It seems std namespace may refer to the one defined in the BFloat16-inl.h, instead of the cpp standard one.

@mruberry Thanks for the reply, I have pushed a new commit 398ed32 to restrict output tensors to be bool.
Please kindly help review.

@mruberry sure, I love to implement another PyTorch function.
Please kindly suggest which function that I should work on, I will give it a try.

@mruberry sure, I love to implement another PyTorch function.
Please kindly suggest which function that I should work on, I will give it a try.

Great!

Since you just did a pair of unary functions, how about a pair of binary functions?

numpy.maximum and numpy.minimum are another pair of functions whose implementations will have some similarities with the work you've done here, like using TensorIterator, but will require writing binary kernels, handling complex values, and dealing with different output dtypes.

Do those sound interesting?

@mruberry Definitely, they sound really interesting, and thanks for the tips regarding implementation.
I am going to start working on them.

@mruberry merged this pull request in 90074bb.

This has been merged! Congratulations, @RockingJavaBean!