Windows 64 and numpy's dtype=int

[guyer]

As posted to the [http://thread.gmane.org/gmane.comp.python.fipy/3123 mailing list]:

The attached patch was made against the current "develop" branch. The trac issue tracker at matforge would not allow me to register in order to create a ticket for this ("The page you are looking for is temporarily unavailable"), otherwise I would have put a pull request there on a new ticket.

For some reason that I didn't have the time to dig into sufficiently, it turns out that 64-bit python on 64-bit windows still regards "int" as 32-bit, and not, as the other OSs do, 64-bit. This causes numerous problems with error messages about dtype=int64 not matching with dtype=int32. For example (the one that failed tests for me) the numpy function argmin() returns an array of dtype=int64, but if this result is assigned to an array slice that was created with dtype=int, then errors occur on win64 & py64, but not on other x64 OSs.

The attached patch defines a numerix.INT_DTYPE value that is set based on the detected platform. I then changed a handful of dtype=int references in a few other files, to refer to the new definition in numerix.

After making this change, the fipy tests pass (using the scipy solver; pysparse on py+Windows x64 appears to be afflicted by the same issue and I haven't had a chance to look at it yet) although there are some issues with double_scalars() that appear to be unrelated to the int issue:

F:\CalebHattingh\repos\fipy\examples>C:\WinPython-64bit-2.7.5.0\python-2.7.5.amd
64\python.exe test.py
................................................................................
...f:\calebhattingh\repos\fipy\fipy\variables\variable.py:1165: RuntimeWarning:
divide by zero encountered in double_scalars
  return self._BinaryOperatorVariable(lambda a,b: a/b, other)
................................................................................
.......................f:\calebhattingh\repos\fipy\fipy\solvers\scipy\linearLUSo
lver.py:71: RuntimeWarning: invalid value encountered in double_scalars
  if (numerix.sqrt(numerix.sum(errorVector**2)) / error0)  <= self.tolerance:
.....F:\CalebHattingh\repos\fipy\examples\reactiveWetting\liquidVapor1D.py:3: Ru
ntimeWarning: divide by zero encountered in log
  ##
F:\CalebHattingh\repos\fipy\examples\reactiveWetting\liquidVapor1D.py:18: Runtim
eWarning: divide by zero encountered in double_scalars
  # of their official duties.  Pursuant to title 17 Section 105 of the
.F:\CalebHattingh\repos\fipy\examples\reactiveWetting\liquidVapor2D.py:3: Runtim
eWarning: divide by zero encountered in log
  ##
F:\CalebHattingh\repos\fipy\examples\reactiveWetting\liquidVapor2D.py:22: Runtim
eWarning: divide by zero encountered in double_scalars
  # other parties, and makes no guarantees, expressed or implied, about
...
----------------------------------------------------------------------
Ran 195 tests in 264.878s

OK

Imported from trac ticket #645, created by guyer on 07-03-2013 at 09:52, last modified: 09-30-2013 at 22:27

[guyer]

guyer attached win64_numpy_int.patch on 07-03-2013 at 09:53

[fipymigrate]

In 5e6277d:

#CommitTicketReference repository="fipy" revision="5e6277d694659a05e15ee5c23f3684e6277410da"
Change `int` to `numerix.INT_DTYPE`

Deals with 32/64-bit issues on Windows.

Addresses issue #388

Trac comment by Jonathan Guyer guyer@nist.gov on 09-30-2013 at 15:04

[fipymigrate]

In 5aa276b:

#CommitTicketReference repository="fipy" revision="5aa276b24f7eb5d721967ba41bf222073e20dced"
Merge branch 'ticket645-Windows_64_and_numpys_dtype_int' into develop

Fixes issue #388

Trac comment by Jonathan Guyer guyer@nist.gov on 09-30-2013 at 22:27