Hi!

I found that libdv library (http://libdv.sf.net/) is quite well writen and easily vectorisable, but curently don't use any potential power that is in PowerPC processors. Is there anyone interested?

I will be very happy if someone with knowledge looks at it. Kino will work much faster if this library is optimised.

Regards, Marek Szyprowski

Added to my TODO list
Though I still have a long list of things I want to check out, I'll definitely try to look at this one.

Thanks

Konstantinos

Hi markos, how are you coming on this one?

R&B

Haven't started working on this one, yet. As you probably saw, I just posted a post on MySQL. Seems that I'll have to choose another app to show off Altivec. Might as well, be this one (libdv), if you like.
Apart from the glibc stuff I mean. I just want to find an app that would benefit from Altivec and work on it, till the FTF event.

Do you think I should start on this, or should I pick sth else? Here's my TODO list:

Algorithms:
Glibc calls (mem*, str* calls)
crc32 (vectorizable, known algorithm)
mysql(hashing, maybe vectorizable)
C math library calls
umac (hashing algorithm)
gzip (parallelizable, but may or may not be vectorizable)
bzip2 (same)
gmp (definitely vectorizable, but quite hard to do)
libmcrypt (AES, RC5, DES, 3DES) (vectorizable but needs lots of testing)
SSL (maybe, I haven't checked)
STL (it's vectorizable, see macstl)
CommonC++ (some of the methods in its classes may be vectorizable)
Boost (same)
xine/mplayer (I believe they already use altivec for codecs)
Cinerella (barely works on powerpc, it's in an alpha stage yet)
X RENDER extension (X.org, that might be useful)
ogg encoding (probably vectorizable?)
libdv (probably vectorizable)

So far I have done 7-9 mem* calls and 5/20 str* calls. I could go faster doing only Altivec, but I have to do scalar optimizations to cope with smaller sizes as well.

Konstantinos

ffmpeg, xine, mplayer and so on already use altivec optimizations, but aren't present as widely as the x86 SIMD ones. There is plenty of work available =) it uses ffmpeg and other libs see before, cinelerra-cvs (is a fork/branch that is buildable w/out black magic) seems working on ppc relatively fine That could be quite useful. Each segment is 255 bytes, I guess that it could enjoy some speedup even if is just for the IO (maybe libfreevec could help)

You may have a look at evas since it is a portable canvas that is embedded ready (works already on pdas) and could enjoy altivec optimizations.

(www.get-e.org)

There is one more thing that urgently need to be (better) vectorised: mpeg2enc from mjpeg-tools. Right now it is hardly usable because of the speed. Please add it somewhere at the end of your todo list.
I know that Mpeg2 can be encoded using ffmpeg (almost 2 times faster!) but ffmpeg procudes streams that are not playable on my standalone DVD player.

added. thanks.