Rev 361, 2 times slower?

[Sirber]

Hi

I'm getting ~4 FPS instead of ~8 FPS like before (rev 254). When adding --subme 7 and trellis, it goes to ~1.5 FPS.

I'm on AMD64 3000+ (32bit).

Anyone having the same problem?

[Sharktooth]

trellis eats CPU time, expecially -t 2

[DryFire]

Using subme 6 encoding has actually sped up (~.75-1 fps) w/o trelis.

what exactly are the options for trelis (the options weren't listed on the build changes)?

[Sharktooth]

-t, --trellis <integer> Trellis RD quantization. Requires CABAC. [0]
- 0: disabled
- 1: enabled only on the final encode of a MB
- 2: enabled on all mode decisions

[Sirber]

Current preset Extreme:
strOptions := strOptions + '--me umh ';
strOptions := strOptions + '--subme 6 ';
strOptions := strOptions + '--8x8dct ';
strOptions := strOptions + '--trellis 1 ';
strOptions := strOptions + '--mixed-refs ';
strOptions := strOptions + '--ref 12 ';
strOptions := strOptions + '--analyse "i4x4,i8x8,p4x4,p8x8,b8x8" ';
+ other stuff

~4 FPS

Trellis mode 2 eats 50% speed :)

[Sharktooth]

yup but the gain is 0.2db ...

[ChronoCross]

only about a 20% decrease in speed here.

[DryFire]

setting adding -t 1 slowed my encode down by .25 fps on me 1:30 test clip,though it was only one run.

using -subme 7 and -t 2 i seemed to slow down ~4 fps (from 13 to 9 fps). I'll do some more tests with it tomarrow.

I'm running and A64 3000+ at 2.4ghz.

[akupenguin]

Notes: -t1 should be pretty fast, -t2 should be slow and only a little better.
The cost of either mode is directly proportional to bitrate.
Trellis currently uses 64bit math, so is faster on A64. (I intend to fix that, but it's not my 1st priority)

[ChronoCross]

Crashed in the 2nd pass with --trellis 2. I'll try to find the problem once I get off work and have a chance to look at the application error logs.

[Sirber]

No crash for me, 3 pass.

[ChronoCross]

No crash for me, 3 pass.

how many threads?

[Sirber]

unfortunatly 1 :(

[spinstate]

only about a 20% decrease in speed here.

with sirber's settings posted above, i'm also getting only a 20% decrease in speed with -t2, on my A64 2800.

[bond]

ok my own results on trellis for the usual matrix clip i test (smith interrogating morpheus, lobby shootout) with the following settings:

x264 --pass 2 --bitrate 770 --bframe 3 --b-pyramid --ref 5 --mixed-refs --filter -5:-5 --subme 5 --8x8dct --analyse all --weightb --progress -o r361_m5_trellis1.mp4 matrix_LR_720x288.avsspeed:
no trellis: 3.63 fps
trellis 1: 3.34 fps (- 8%)
trellis 2: 3.05 fps (- 16%)

average psnr:
no trellis: 44.090
trellis 1: 44.203
trellis 2: 44.216

visual quality:

i am a little bit disappointed of trellis, because it wasnt able to really show me a visual quality difference compared to when using no trellis
it was often very hard to judge which one is better because quality already varried in between single frames clearly
i had the impression that trellis in both modes clearly provided less details in highmo compared to when no trellis was used (i leave it up to everyone to decide whether this is a good thing), so for example trellis performed clearly poorly on the scene where parts of the walls are flying around. it also provided less details on neo on the "still neo in front of fast moving weapon shelf" scene
also in still closeup face scenes i wouldnt say that trellis gave more details (as i said it was very varrying)
regarding walls and fog i also didnt really see a big difference telling me that this or that is better

also i can say that trellis 2 is better than trellis 1, which doesnt mean that it beats no trellis tough

all in all i would say that people with slow pcs (like me) might better not use it as, at least on the clip i tested, the real quality difference wasnt there, so its not really worth a 8% encoding slowdown imho
people with fast pcs might want to try trellis 2 if they think trellis is worth it

[lexor]

since trellis was the second part of original topic (i.e. caused 8fps to 1.5fps drop) and the first one is about general halving of the speed, I'd like to comment on that too.

using r361, I find that I actually loose speed compared to before on r35something I had 5fps on 2nd pass with HQ-Slower, now I have 3.8fps on HQ-Slow (Sharktooth's latest profiles). So I went down in settings and dropped the speed :( (same profile will drop to 3.1)

any explanation on the speed drop? (I uninstalled previous build before installing new, always do)

[Sharktooth]

Yes, updated profiles. Now the slower profiles will be even slower coz have trellis and b-rdo enabled but quality-wise the quality will be higher.
However HQ profiles are meant to provide High-Quality at the cost of speed...

[Sirber]

General
Complete name : X:\sirber\sweet_anime;).avi
File size : 232 MiB
Format : Audio Video Interleave
Overal BitRate : 1344 KBps
PlayTime : 24mn 9s
App_Muxing : VirtualDubMod 1.5.10.1 (build 2439/release)

Video #0
Codec : XviD
Bit rate : 1203 KBps
Width : 768
Height : 432
Aspect ratio : 16/9
Frame rate : 23.976
Bits/(Pixel*Frame) : 0.151

Audio #0
Codec : MPEG1/2 L3
Bit rate : 128 KBps
Channels : 2 channels
Sampling rate : 48 KHz

Extreme Profile: strOptions := strOptions + '--me umh ';
strOptions := strOptions + '--subme 6 ';
strOptions := strOptions + '--8x8dct ';
strOptions := strOptions + '--mixed-refs ';
strOptions := strOptions + '--ref 12 ';
strOptions := strOptions + '--analyse "i4x4,i8x8,p4x4,p8x8,b8x8" ';
strOptions := strOptions + '--b-rdo ';
Glorious 2.44FPS, with LE alpha 2 (no brdo, 16 ref), ~8 FPS. odd...

What could I sue to have the same (~) quality but faster encoding?

[Sharktooth]

how many b-frame you used? (i cant see -b or --bframes option in your strOptions).
also -w (or --weightb) was not used...
are you sure you get 2.44FPS? it seems a bit low since i get 4.80FPS (2nd pass) on an athlon xp 2800+ with: --ref 12 --mixed-refs --bframes 3 --b-pyramid --filter -2,-1 --subme 6 --me umh --b-rdo --weightb --trellis 1 --analyse all --8x8dct

[Sirber]

that was the part of "profiles". I ahve 3 bframes, weighted, referenced, pyramid , 16 refs, mixedd-refs, filter +1:+1

I don't get it, only x264 is taking "juice" :confused:

[Sharktooth]

are you encoding 4:3 material or are you using heavy avisynth filters? If not, i dont really know what's wrong.

[Sagittaire]

In fact motion complexity is very important for speed too ...

With exactly the same setting, same resolution, same filter, and same source ... encoding can be at 10 fps for some part and 2 fps in other part ...

It's useless to compare your result if you don't use exactly the same source ... CQFD

[Sirber]

are you encoding 4:3 material or are you using heavy avisynth filters? If not, i dont really know what's wrong.640x480 usualy, no filtering.

[Sharktooth]

try another source and see if it gets better, i've tried different ones (movies @ 720x304) and the average FPS were between 4 and 5 for the second pass with the above settings.

[Sirber]

k, arigato gozaimasu (http://www.abcb.com/ency/a/arigato.wav)! :D

lol