[Image-SIG] PIL 1.1.6, python 64bit on Windows XP x64

Fredrik Lundh fredrik at pythonware.com
Thu Mar 5 22:09:42 CET 2009

Thanks for the howto.  Now I just need to go find myself a 64-bit box
to play with ;-)

As for the file descriptor crashes, things like that is usually caused
by a mismatch between the C runtime libraries used for core Python and
the extension (the file handle is allocated by the Python main
process, but used by code in the extension.  If the internal structs
don't match, all sort of oddities can happen).  You can use e.g.
dumpbin /imports on the python DLL and the PIL PYD's, and verify that
they're not importing differenct MSVC-something DLL:s.


2009/3/5 Swan, Tim <Tim.Swan at disney.com>:
> Hi,
> I've been trying to get PIL 1.1.6 working on python 64bit:
> Python version : 2.5.2 (r252:60911, Feb 21 2008, 13:17:27) [MSC v.1400 64
> bit (AMD64)]
> OS : Windows XP Professional x64 SP2.
> The installer on the main download page appears to only be valid for 32bit
> python, causing a variety of import errors when used with 64bit python.
> Is there any plan to provide an installer for 64bit versions of python on 64
> bit Windows in the future?
> I've sucessfully built a version locally with DevStudio 2005, including zlib
> and jpeg-6b support. I'll detail the steps required later for posterity, it
> may be useful for someone browsing this forum at a later time - I have no
> plans to provide an installer myself!
>  I did have to work around an issue, which I thought was worth raising to
> you folks, in case there was a better way to fix this.
> After building locally (setup.py build, setup.py install) and importing
> PIL.Image, the quick test detailed below caused my python process to crash.
> Import PIL.Image
> Im = PIL.Image.open('my_image.gif')
> Im2 = im.rotate(45)
> Im2.save()
> At the point of saving, the process crashes with a windows IO exception.
> Digging into the python, and then into the C extension code, I found the
> problem appears to be due to some conflict with using the file descriptor
> that's being passed into _encode_to_file (encode.c) from _save in
> ImageFile.py (line 492 in ImageFile.py in the source I'm looking at). After
> trying various different things, I found that I could not use the file
> descriptor int (acquired via fp.fileno(), line 469 in ImageFile.py) passed
> into encode_to_file - trying to write to a file using that file descriptor
> was causing the crash (for example, line 164 in encode.c). I tried using
> _rdinfo to convert the file descriptor to a FILE pointer, and work with
> that, but that crashed in _rdinfo, and I tried a couple of different
> conversions using calls like _get_osfhandle
> , but pretty much everything I tried continued to cause crashes. My
> knowledge of Windows is pretty minimal, so I've no idea why using that file
> descriptor should be problematic (and, presumably, works just fine with the
> 32bit python version).
> After taking a step back, I realised that the code in ImageFile has two
> paths, and by putting a forced call to raise an AttributeError around line
> 469 caused the _save function to take the top code path, and this works just
> fine for me. I ended up modifying the code to look like this:
> Line 467, ImageFile.py
>     try:
>         # FIXME: force an AttributeError exception to use the python
> file-compatible object path
>         raise AttributeError
>         fh = fp.fileno()
>         fp.flush()
>     except AttributeError:
>         ...
> It's not clean, but it appears to work. Anyone got any ideas why? Can
> someone suggest a cleaner way of doing this?
> If anyone is interested, here's the steps I followed to get PIL 1.1.6
> working locally for me - DevStudio 2005 required.
> Step 1: unarchive the source to a location locally, I'll refer to this as
> the build root directory (in my case, c:\temp\Imaging-1.1.6)
> Step 2: run the visual studio 2005 command prompt for the relevant bits of
> the OS
> (in my case from Windows start menu: Programs -> (somewhere) -> Microsoft
> Visual Studio 2005 -> Visual Studio Tools -> Visual Studio 2005 x64 Win64
> Command Prompt
> http://docs.python.org/dist/module-distutils.msvccompiler.html
> Step 3: set up the following environment variables at the command prompt
> set MSSdk=1
> Step 4: edit ImageFile.py as described above
> Step 5: (optional) edit setup.py to point to the correct locations for zlib,
> jpeg-6b, etc. source and built libs
> Step 6 : build PIL using "setup.py build" in the build root directory.
> (it's worth at this point confirming that you're running setup.py with the
> correct (64bit) version of python - slap a pdb break in there and check
> sys.version if needed)
> Step 7 : embed the manifests for the two DLLS by running mt.exe in the build
> root directory:
> Mt.exe -manifest build\lib.win32-2.5_imaging.pyd.manifest
> -outputresource:build\lib.win32-2.5_imaging.pyd;2
> Mt.exe -manifest build\lib.win32-2.5_imagingmath.pyd.manifest
> -outputresource:build\lib.win32-2.5_imagingmath.pyd;2
> http://msdn.microsoft.com/en-us/library/ms235591(VS.80).aspx
> Step 8: install PIL (setup.py install)
> _______________________________________________
> Image-SIG maillist  -  Image-SIG at python.org
> http://mail.python.org/mailman/listinfo/image-sig

More information about the Image-SIG mailing list