Can someone please double check these x264 performance results

[audyovydeo]

r800 is fast approaching so this weekend I was testing the speed improvements of the latest bunch of versions (745 >> 790).

Then I doublechecked against r655, near the time I started using x264. I got very surprising results :

clip = 1
r655 r790
------- ------------------- ------- ----------
test1 r 3 me hex 10.84 11.53
test2 r 2 me umh 8.06 9.30
test3 r 1 me esa 10.32 7.36 ???!!??


of course, x264 r655gives this :
[warning]: threads are not yet compatible with ESA


so I ran another test set :

clip = 2
threads auto (= 3)
r655 r790
------- ------------------- ------- ----------
test1 r 3 me hex 37.32 30.93
test2 r 2 me umh 31.16 36.22
test3 r 1 me esa 32.41 7.21 ???!!?


clip = 2
threads 1
r655 r790
------- ------------------- ------- ----------
test1 r 3 me hex 22.23 17.00
test2 r 2 me umh 18.04 21.44
test3 r 1 me esa 10.86 3.83 ???!!?



I find this very weird, especially considering the esa optimisations that took place :

r676 implement multithreaded me=esa
r699 previously, esa was slower with fprofile than without, since gcc thought it wasn't important. now esa benefits like anything else.
r717 1.6x faster me=esa (x86_64) or 1.3x faster (x86_32). (times consider only motion estimation. overall encode speedup may vary.)

the --ref and --me values are all I change between modes.
The rest of the cmdline :

-b 16 -m 5 -t 1 -A all -8 --direct temporal --b-pyramid -w --bime --no-fast-pskip --no-psnr --threads auto --progress -o NUL xy.avs

clip 1 is real footage from a DV/PAL camera, previously deinterlaced & saved in lagarith.
clip 2 is virtualdub's colorcube test clip (whence the huge diff in encoding times)

System is a Core2Duo (T5200) winXP/32bit.
Can someone doublecheck this ?


thanks
audyovydeo

[akupenguin]

[warning]: threads are not yet compatible with ESA
Warning messages are there for a reason, and you just ignore it and continue with the comparison?

[audyovydeo]

Warning messages are there for a reason, and you just ignore it and continue with the comparison?

Hello

yes : maybe you didn't notice I set threads to 1.

Also, it's not just the esa results that I find baffling.


cheers
a/v

[akupenguin]

Ok, the new esa has an arithmetic overflow, when it finds a perfect match and thus the next candidate must have a negative score in order to be better, while the asm does unsigned math. This case occurs all over the uniform background of colorcube.

[audyovydeo]

r798 seen & appreciated

cheers
a/v