Added type promotion logic for complex numbers (#34093)
Summary:
Issue: https://github.com/pytorch/pytorch/issues/33780
After this PR:
1. dtype promotion logic will correctly work for ops involving complex scalars
2. added alias for complex64 (cfloat) and complex128 (cdouble)
3. added an internal function get_complex_default_dtype (consciously not exposed in public API)
- sets the default complex dtype to be double if default_dtype is set to double, else float https://github.com/pytorch/pytorch/pull/34093#discussion_r392350224
>>> 1j*torch.ones(2)
tensor([(0.0000 + 1.0000j), (0.0000 + 1.0000j)], dtype=torch.complex64)
>>> torch.set_default_dtype(torch.float64)
>>> 1j*torch.ones(2)
tensor([(0.0000 + 1.0000j), (0.0000 + 1.0000j)], dtype=torch.complex128)
>>> 1j + torch.ones(2)
tensor([(1.0000 + 1.0000j), (1.0000 + 1.0000j)], dtype=torch.complex128)
>>> torch.tensor(1j) + torch.ones(2,2)
tensor([[(1.0000 + 1.0000j), (1.0000 + 1.0000j)],
[(1.0000 + 1.0000j), (1.0000 + 1.0000j)]], dtype=torch.complex128)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/34093
Differential Revision: D20537125
Pulled By: anjali411
fbshipit-source-id: 05fb1f81b8ba039d0b698cdd2c0bbf8b0ce0b767
