[Image-SIG] Speeding up image.resize()

[Jeff Epler]

After seeing how slow image.resize() is, especially ANTIALIAS, I looked
at the underlying code in libImaging.

I examined the object code produced for Antialias.c.  On my system
(linux on intel) not only is there a function call for XXX_filter, but
antialias_filter has more function calls inside as well (including a
call to libm's sin())

To improve the speed, I made all the XXX_filter functions (including
sinc_filter) 'static inline', and moved most of the body of
ImagingStretch into an include file, which is included once for each
filter.  ImagingStretch now dispatches to one of the 4 _XXX_filter
functions, instead of merely setting some function pointers.

I also set some extra_compile_args appropriate for my system and
compiler:
    exts = [(Extension(
        "_imaging", files, libraries=libs, define_macros=defs,
        extra_compile_args=['-O3', '-march=pentium4', '-mfpmath=sse',
                            '-msse2', '-ffast-math', '-mno-ieee-fp']
        ))]

In their current form, I doubt these changes are suitable for inclusion
in Imaging (because of the use of the "inline" extension and specialized gcc
compiler flags).  However, I thought this might be helpful to someone.
If there's interest, I can produce a patch.

How much difference?  I used 'timeit' on Imaging-1.1.5b1 and my modified
version of it.  I tested on a laptop with a 1.5GHz Pentium-M CPU, with
SpeedStep disabled.

The script:
    import Image
    i=Image.open('test.jpg')
    i.thumbnail((1024,768), Image.ANTIALIAS)
    i.load()
test.jpg is a 3072x2048 digital camera image, which (I assume) is
loaded as 1532x1024 and resized to 1024x682.

1.1.5b1 gives:
    10 loops, best of 3: 1.24e+06 usec per loop  (1240 msec)

My version gives:
    10 loops, best of 3: 6.99e+05 usec per loop  (699 msec)
That's a 43% speedup.

Here's data for all filters:
            (times in msec)
Filter      Vanilla     Patched     Speedup
ANTIALIAS   1240        699         43%
BICUBIC     783         626         20%
BILINEAR    576         439         23%
NEAREST     411         517         -26%  (!)

Jeff
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://mail.python.org/pipermail/image-sig/attachments/20041018/5c5d5b68/attachment.pgp