I profiled some code that I've been running quite a bit, and noticed
that it was spending most of its time in IM line drawing routines,
specifically GetPixelOpacity(). So, I finally took it upon myself to
get the code running a bit faster. After two days of work, I have
30-40% speedup, which would have saved me a month of time on the
cluster (and will save me a month soon). Patch:
http://pricepages.org/magick/opti_magick.diff
Speedup will be most pronounced when you are drawing long lines.
The slowest part of the code was fetching the bounds in the beginning
of each loop in the if statements like "p->bounds.y1;". Prefetching
the data didn't seem to work, so I ended up with one main loop which
also computed the 'winding number', so it only needed to be fetched
in one pass, instead of two.
The rest of the stuff that I did was basic, like computing invariants
outside of the loop(s). Cast values only once. Avoid following
pointers. Reorder instructions to take pipelining into account, etc.
I tried to keep things readable, and not lengthen code. I wasn't sure
exactly _what_ was being computed, so I kept the logick of the code
identical. The only place the output was altered didn't make sense,
and I believe it's the optimizer doing funky things. While the output
does change, I visually don't see a difference. I put a comment
there, so you can see how it acts on your system.
Thoughts? Does it pass your test(s)?
~Seth
_______________________________________________
Magick-users mailing list
[email protected]
http://studio.imagemagick.org/mailman/listinfo/magick-users
