Thanks for working on this! With sort methods I think we also have problems with float('-inf'). Is it going to be in a separate PR btw?

@ailzhang What is the problem with -inf ? It seems to work as I expect:

In [3]: x
Out[3]: tensor([3., -inf, 1.])

In [4]: x.sort()
Out[4]: (tensor([-inf, 1., 3.]), tensor([1, 2, 0]))

In [5]: x.sort(descending=True)
Out[5]: (tensor([3., 1., -inf]), tensor([0, 2, 1]))

Actually the problem was indeed caused by nan. ninf is sorted correctly. Please ignore my comment then. Thanks @umanwizard !

Both thrust and quicksort in TH require strict weak ordering comparator https://thrust.github.io/doc/group__sorting.html#ga3f47925d80f4970d5730051dba1c5603.

#define GT_OR_NAN(x, y) \
  ((x != x) || (x > y))

is not strict weak order.

@ngimel yea I agree on the solution is not strict weak order. Hmmm why do we enforce strict weak ordering btw (could you give an example) ? Thanks!

@ngimel Thank you for the hint. If I understand correctly, I can fix this by making those macros/functions return false if both the arguments are NaN.