[TensorIterator] bug when performing inter-scalar ops on the GPU #11795
Description
When performing operations between scalar tensors of different types, type promotion happens in TensorIterator. But it seems that the results of inter-type operations are not correct when performing operations on the GPU.
The following code gives unexpected results on the GPU. CPU works fine
import torch
torch.__version__
# '1.0.0a0+91b6458', master from today
a = torch.tensor(1.5, device='cuda') # float tensor
b = torch.tensor(1, device='cuda') # int64 tensor
print(a * b)
# tensor(2.8026e-45, device='cuda:0')
a = torch.tensor(1.5, device='cuda') # float tensor
b = torch.tensor(True, device='cuda') # uint8 tensor
print(a * b)
# tensor(2.8026e-45, device='cuda:0')
print(a + b)
# tensor(1.5000, device='cuda:0')
print(a - b)
# tensor(1.5000, device='cuda:0')cc @colesbury