Problem with pointwise math when output tensor is zero-strided with correct size

[fmassa]

When reusing a Tensor as the output of a pointwise math operation (like cmul, add, cdiv etc), the result is not properly computed if the output Tensor has zero stride and has the correct size.

a = torch.Tensor(1,5):expand(5,5)
b = torch.rand(5,5)
c = torch.rand(5,5)

print(a:cmul(b,c)) -- computes the wrong multiplication
print(torch.cmul(b,c)) -- correct answer

it outputs

th> print(a:cmul(b,c)) -- computes the wrong multiplication
 0.1019  0.0165  0.0757  0.7305  0.0191
 0.1019  0.0165  0.0757  0.7305  0.0191
 0.1019  0.0165  0.0757  0.7305  0.0191
 0.1019  0.0165  0.0757  0.7305  0.0191
 0.1019  0.0165  0.0757  0.7305  0.0191
[torch.DoubleTensor of size 5x5]

th> print(torch.cmul(b,c)) -- correct answer
 0.0222  0.0103  0.0174  0.2752  0.1085
 0.2826  0.2686  0.0294  0.5893  0.6549
 0.2514  0.0066  0.0279  0.0920  0.0991
 0.6117  0.4126  0.1853  0.0232  0.4656
 0.1019  0.0165  0.0757  0.7305  0.0191
[torch.DoubleTensor of size 5x5]

This happens because the THTensor_(resizeAs) function doesn't take into account the fact that even though the sizes match, there can be a zero stride somewhere.

Maybe a solution would be checking in THTensor_(isSameSizeAs) if there is one of the strides==0 and returning 0 in this case ? I don't know if there could be side effects though.

[dominikgrewe]

There are probably many functions out there that use THTensor_(isSameSizeAs) to determine if arguments have the correct size. If you now say that two arguments are different just because one has stride=0 these functions will fail (probably unnecessarily because for inputs stride=0 doesn't matter).

I agree that stride=0 is a problem for outputs though. Not sure how to fix it without checking for it in each function...

[fmassa]

The solution I proposed doesn't solve the problem. Actually, even if THTensor_(rawResize) was called, as the sizes are the same it wouldn't modify the strides. We would need for example to reset the sizes if there is a zero stride before calling THTensor_(rawResize).

The way I see it, the problem lies within THTensor_(resize) and THTensor_(resizeAs). For example, the following snippet does not work as well

a = torch.rand(1,3):expand(3,3)
b = torch.rand(3,3)
a:resizeAs(b):copy(b);

it outputs, as before

th> print(a)
 0.9823  0.1666  0.0679
 0.9823  0.1666  0.0679
 0.9823  0.1666  0.0679
[torch.DoubleTensor of size 3x3]
th> print(b)
 0.3374  0.8114  0.3300
 0.9484  0.8672  0.0940
 0.9823  0.1666  0.0679
[torch.DoubleTensor of size 3x3]

Would you expect this current behaviour to be correct ?