[Numpy-discussion] fast putmask implementation

[Eric Firing]

In looking at maskedarray performance, I found that the filled() 
function or method is a bottleneck.  I think it can be sped up by using 
putmask instead of indexed assignment, but I found that putmask itself 
is slower than it needs to be.  So I followed David Cournapeau's example 
of fastclip and made a similar fastputmask.  The diff relative to 
current svn (3967) is attached.

The faster version makes a factor-of-ten or larger improvement in 
putmask speed.  numpy.test() still passes.

With 10000-element integer arrays the new version reduces the times from 
136 to 15 usec for 1000 masked elements, and 445 to 18 usec for 5000 
masked elements, with a scalar value argument.  It is only slightly 
slower with an array value argument.  (Times are for Intel Core2, 2 GH, 
linux.)

I hope someone will take a look and either tell me what I need to fix or 
commit it as-is.

Thanks.

Eric
-------------- next part --------------
A non-text attachment was scrubbed...
Name: numpy_putmask.diff
Type: text/x-patch
Size: 32047 bytes
Desc: not available
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20070816/30ff93db/attachment.bin>