paq8p2

[Jan Ondrus]

paq8p2 is paq8p1 with some modifications in jpeg model
- slightly simplified source
- minor changes in compression ratio for jpeg.

http://www.sendspace.com/file/t5yxdl

using -5 setting:
A10.jpg 842468
A10.jpg.paq8p1 638756
A10.jpg.paq8p2 637686
DSCN3974.jpg 1114198
DSCN3974.jpg.paq8p1 827269
DSCN3974.jpg.paq8p2 827266
football.jpg 115693
football.jpg.paq8p1 70513
football.jpg.paq8p2 70339

[encode]

[Simon Berger]

A short test from me. I compiled paq8p1 and paq8p2 by myself and without asm support because I didn't manage to do it. I thought Visual Studio 2008 can compile them but it can only handle included nasm? It's a a 3648x2736 picture of two persons with a normal camera I guess

paq8p1 - 4378752 -> 3273336 - 127.84 sec
paq8p2 - 4378752 -> 3269096 - 122.75 sec

Reading the old paq8p thread it seems paq8p1 has some bugs or am I wrong? Before the version will be continue used, these bug(s) should be fixed.

[LovePimple]

Thanks Jan!

Compiled...

[Simon Berger]

For completion I used the asm object file in Love Pimples package but compiled it with the same switches then before.
It was really much faster. 97.11 sec. I didn't retest paq8p1 if someone is interested in this result I could do. But I think the same time can be subtracted.

[Darek]

Hi!
My test scores (w/o timings, paq method -5):

f.jpg = 112'038
f.jpg.paq8p = 90'100
f.jpg.paq8p1 = 90'100
f.jpg.paq8p1_ = 90'100
f.jpg.paq8p2 = 89'750
f.jpg packed by stuffit 9.0.0.21 = 83'343
f.jpg prepacked by precomp 0.3.8 (packJPG) = 87'608
Good improvement, but still my testfile is still better compressed by precomp or stuffit - probably because it is progressive jpg.

It's possible to add progressive jpg recognition to paq jpg algorithm?


Second observation – bmp and wav compression algorithms used in paq8p2 version aren’t looked like the best as it was used in older paq8pXX variations:
Examples from my testbed:
WAV:
0.wav.paq8p = 1’342’234
0.wav.paq8p1 = 1’342’580
0.wav.paq8p1_ = 1’334’183 – best wav algorithm from paq8p variations
0.wav.paq8p2 = 1’342’234
BMP:
1.bmp.paq8p = 264’426 – best bmp algorithm from paq8p variations
1.bmp.paq8p1 = 266’187
1.bmp.paq8p1_ = 266’187
1.bmp.paq8p2 = 266’187

Of course I understand that my testbed files couldn’t be objective for broader set of test files, but if it’s a regular situation, then maybe easy way to get small improvement in next versions is to use wav algorithm from paq8p1_ version and bmp algorithm from paq8p. It’s possible?

Regards, Darek