Built-in preset changes

[jbrjake]

Here's a really rough draft of what I'd like the presets to look like in the next release:
http://handbrake.fr/pastebin/pastebin.php?show=424

You'll note that the default Normal preset and all High-Profile presets go CRF, which I think are long-overdue changes.

By unanimous vote , dynaflash, sr55, and I decided to get rid of the video gaming console presets. No more PS3, 360, or PSP built-in presets. We don't have the devices, we can't support them, and with the autoleveling x264 does these days they aren't particularly necessary. They were waaaaaaaay more trouble than they were worth, from a support perspective.

The Universal preset is going to get a little faster by ditching some x264opts.

High profile presets finally lose my over-enthusiastic psy-trellis setting from the last release.

Normal gets b-adapt 2.

There might be a few more changes I'm forgetting....

Anyway, thoughts? cvk_b? I'm sure you have some, with all the testing you've been doing lately...

[refulgentis]

as cvk_b mentioned on IRC yesterday, hexagonal motion estimation provides a pretty big speed up (~15% on the sources I've tested so far, which were 1080p and 720p low motion) for no/very little (on the magnitude of .01%) increase in bitrate at a constant CRF. I think I might make pretty graphs across several 480p/720p/1080p sources to see some real numbers.

For testing, I assume keeping the same CRF and all other options and seeing how much bitrate changes as I change the motion estimation method, is the best way to see the speed versus quality tradeoff?

I was thinking about disabling AQ and psyrd and seeing how SSIM changes, but then we don't see how it would work with AQ and psyrd...

UPDATE: I ran the universal preset using the CLI, varying subme and me from default (6?) to 9 and umh to hex. To determine quality, I used the same method that d_s used in this thread: http://forum.doom9.org/showthread.php?t=127861, namely measure quality by (1 / (1 - SSIM)) / Bitrate. IMO it's in very strong support of a switch from umh to hex, or at least a switch from umh to hex:subme=9, since the switch to umh gives very little benefit in terms of speed:quality compared to the benefit of changing umh to hex:subme=9
Note that the right side of the picture is being overlaid by the forums, if you download it, you can see the full results.

[cvk_b]

refulgentis, you have tested subme8. With bframes off subme 7=6 & 9=8

jon, I'd really take a look at

Code: Select all

ref=4:bframes=3:b-adapt=2:weightb:me=umh:trellis=1:subme=7:8x8dct

for High Profile film. Just because we can turn on all those fancy options doesn't mean we should. That line, to me, represents the biggest "bang for your buck".

You might be thinking "why no mixed refs?" or "why not direct=auto?". Well, I just try to deal with big jumps in quality, speed, or size. And those other options don't represent that in my experience. I'm not sure my way of looking at options is right for HB's presets, to be honest. I looked at our film preset and asked "Can I get basically the same quality faster and smaller?"— The key word there being "basically".[/size]

[mithrandir]

refulgentis, I'd be interested to see what you get if you also tried umh:merange=8. Particularly I am interested in how that affects the high resolution clips. Reducing the merange may injure the quality of the 1080p...or it may not. However, when I was testing this stuff a month or so ago I found that umh:merange=8 provides better quality scores than hex (using the implicit default of merange=16) on DVD/low-resolution sources. Meanwhile, performance was just a little slower.

[refulgentis]

refulgentis, you have tested subme8. With bframes off subme 7=6 & 9=8

jon, I'd really take a look at

Code: Select all

ref=4:bframes=3:b-adapt=2:weightb:me=umh:trellis=1:subme=7:8x8dct

for High Profile film. Just because we can turn on all those fancy options doesn't mean we should. That line, to me, represents the biggest "bang for your buck".
[/size]

ah, thank you.
should that string have psy-rd=1.0,0.1 or 0.2 or something attached to it? i remember when i initially tried psy-rdo, trellis didn't work with it, which led to the development of psy-trellis.

[cvk_b]

should that string have psy-rd=1.0,0.1 or 0.2 or something attached to it? i remember when i initially tried psy-rdo, trellis didn't work with it, which led to the development of psy-trellis.

I just haven't found any great benefit to psy-trellis. You could add it, certainly.

[cnuro]

would you consider these new high profile presets to be scalable to 720p and 1080p encoding with no tinkering as well or were these designed strictly for dvd content?

[refulgentis]

refulgentis, I'd be interested to see what you get if you also tried umh:merange=8. Particularly I am interested in how that affects the high resolution clips. Reducing the merange may injure the quality of the 1080p...or it may not. However, when I was testing this stuff a month or so ago I found that umh:merange=8 provides better quality scores than hex (using the implicit default of merange=16) on DVD/low-resolution sources. Meanwhile, performance was just a little slower.

http://img222.imageshack.us/my.php?image=merangevp0.jpg
umh to umh:merange=8 provides the same benefit ratio as switching to hex -- impact on quality was -0.22% for both 1080p sources

[cvk_b]

Ran some interesting tests with psy-rd.

Code: Select all

===093 universal preset: level=30:cabac=0:ref=3:mixed-refs=1:analyse=all:me=umh:no-fast-pskip=1
[01:54:50] work: average encoding speed for job is 6.644102 fps
x264 [info]: slice I:5     Avg QP:18.49  size: 27326  PSNR Mean Y:44.42 U:33.03 V:33.28 Avg:35.54 Global:32.45
x264 [info]: slice P:263   Avg QP:22.59  size: 10429  PSNR Mean Y:41.23 U:33.90 V:35.01 Avg:36.05 Global:32.70
x264 [info]: mb I  I16..4: 19.6%  0.0% 80.4%
x264 [info]: mb P  I16..4:  6.5%  0.0% 16.9%  P16..4: 46.5% 19.8%  7.3%  0.4%  0.4%    skip: 2.3%
x264 [info]: ref P L0  74.8% 14.8% 10.4%
x264 [info]: SSIM Mean Y:0.9715881
x264 [info]: PSNR Mean Y:41.286 U:33.887 V:34.979 Avg:36.039 Global:32.699 kb/s:2060.75

===(proposed new universal): level=30:cabac=0:ref=2:mixed-refs=1:me=umh:no-fast-pskip=1
[02:49:41] work: average encoding speed for job is 7.969339 fps
x264 [info]: slice I:5     Avg QP:18.49  size: 27326  PSNR Mean Y:44.42 U:37.25 V:37.65 Avg:38.39 Global:34.05
x264 [info]: slice P:263   Avg QP:22.58  size: 10481  PSNR Mean Y:41.21 U:35.43 V:36.47 Avg:37.19 Global:33.40
x264 [info]: mb I  I16..4: 19.6%  0.0% 80.4%
x264 [info]: mb P  I16..4:  6.9%  0.0% 17.8%  P16..4: 45.1% 20.6%  7.4%  0.0%  0.0%    skip: 2.2%
x264 [info]: ref P L0  77.8% 22.2%
x264 [info]: SSIM Mean Y:0.9715943
x264 [info]: PSNR Mean Y:41.274 U:35.467 V:36.492 Avg:37.214 Global:33.413 kb/s:2070.59

=== cvk test1: level=30:cabac=0:ref=2:mixed-refs=1:me=umh:no-fast-pskip=1:psy-rd=0.5
[02:52:37] work: average encoding speed for job is 7.999759 fps
x264 [info]: slice I:5     Avg QP:18.49  size: 27182  PSNR Mean Y:44.43 U:37.35 V:37.69 Avg:38.42 Global:34.06
x264 [info]: slice P:263   Avg QP:22.59  size: 10199  PSNR Mean Y:41.31 U:35.52 V:36.54 Avg:37.27 Global:33.42
x264 [info]: mb I  I16..4: 23.0%  0.0% 77.0%
x264 [info]: mb P  I16..4: 10.8%  0.0% 20.2%  P16..4: 43.5% 18.7%  4.5%  0.0%  0.0%    skip: 2.4%
x264 [info]: ref P L0  78.0% 22.0%
x264 [info]: SSIM Mean Y:0.9722108
x264 [info]: PSNR Mean Y:41.365 U:35.556 V:36.563 Avg:37.288 Global:33.430 kb/s:2017.01

=== cvk test2: level=30:cabac=0:ref=4:psy-rd=0.5
[02:57:44] work: average encoding speed for job is 9.944627 fps
x264 [info]: slice I:5     Avg QP:18.49  size: 27182  PSNR Mean Y:44.43 U:32.83 V:33.95 Avg:36.04 Global:32.41
x264 [info]: slice P:263   Avg QP:22.58  size: 10228  PSNR Mean Y:41.31 U:35.34 V:36.75 Avg:37.26 Global:33.47
x264 [info]: mb I  I16..4: 23.0%  0.0% 77.0%
x264 [info]: mb P  I16..4: 11.6%  0.0% 20.9%  P16..4: 47.2% 13.6%  3.0%  0.0%  0.0%    skip: 3.7%
x264 [info]: ref P L0  73.3% 13.5%  8.3%  4.9%
x264 [info]: SSIM Mean Y:0.9722490
x264 [info]: PSNR Mean Y:41.364 U:35.291 V:36.696 Avg:37.242 Global:33.447 kb/s:2022.40

[/size]

Check out that last one compared to the current Universal preset. Better quality, size, & encode time.

Maybe someone could run some of those fancy test scripts (which I should probably start paying attention to

[mithrandir]

http://img222.imageshack.us/my.php?image=merangevp0.jpg
umh to umh:merange=8 provides the same benefit ratio as switching to hex -- impact on quality was -0.22% for both 1080p sources

On the 720p-2 clip, reducing the merange from 16 to 8 (while keeping umh) increased SSIM, decreased bitrate and increased FPS...the best of all worlds. Yes, the percentages are really small but in theory this should not be possible unless (a) SSIM is an improper method of quality estimation or (b) there is a flaw somewhere in the motion estimation algorithm. So that's an interesting result.

These tests do reveal the power of umh: despite having a 50% ME range "disadvantage", umh:merange=8 beat hex:merange=16 in the quality metric by fairly substantial margins in 3 of the clips and came up short only once (by a small amount too). Of course, yes, it's also a touch slower than hex. But this does suggest those who use merange=32 or gasp, 64 are really missing the point. Even on those 1080p clips, merange=8 did not significantly degrade quality; no worse than hex.

But, yes, hex has its place because it is fast.

[jbrjake]

Things that are not happening:

- lowering the me
- lowering psy-rd

[jbrjake]

Here's a fresher draft of presets for 0.9.4.

I've streamlined things a lot.

I decided there wasn't really a pressing need for 3 different high profile presets based on content, now that they're CRF. So in addition to ditching Constant Quality Rate, the Film, TV, and Animation presets have been folded into a single High Profile preset, which I've made MP4 instead of MKV since libmp4v2 has a larger and more active development team than libmkv. The QuickTime preset has been ditched because it is similarly redundant. The Legacy preset folder is now at the root instead of nested inside Apple, and the Classic preset has migrated there. The Basic folder is now called Regular since it includes the not-at-all-basic High Profile preset.

I also replaced the ATV preset with many of the more intensive settings dynaflash has put to use for his own encodes.

Having done this, I went back and pared down the remaining presets' settings to tune for speed vs quality.

0.9.4 Presets Compared To Their 0.9.3 Analogues:
* iPod: Slightly Faster, Same Size, Slightly Higher Quality
* iPhone: Faster, Bigger, Higher Quality
* Normal: Slower, Much Bigger, Much Higher Quality
* Universal: Faster, Much Bigger, Higher Quality
* AppleTV: Slower, Smaller, Higher Quality
* High Profile: Slightly Faster, Much Smaller, Slightly Lower Quality

Numbers

FPS (higher is faster)

CRF093 10.741848
HiProfile094 11.134413
AppleTV094 16.391607
AppleTV093 17.307081
Universal093 17.660145
Universal094 21.312574
Normal094 24.440313
Normal093 28.941479
iPhone093 49.137417
iPhone094 52.588688
iPod093 93.280449
iPod094 94.43116

Bitrate (higher is bigger)

iPod094 702.23
iPod093 702.23
iPhone093 770.98
iPhone094 885.71
Normal093 1500.13
AppleTV094 1942.78
Universal093 1969.68
HiProfile094 2054.95
AppleTV093 2102.61
Universal094 2309.1
Normal094 2312.49
CRF093 3269.84

I-Frame QP (lower is better quality)

Normal093 18.62
iPhone093 18.03
iPhone094 17.15
Universal093 17.11
AppleTV093 16.5
Universal094 16.22
AppleTV094 16.06
HiProfile094 15.88
CRF093 15.85
Normal094 15.49
iPod093 15.35
iPod094 15.34

Preset CLI Strings

< Apple

+ Universal: -e x264 -q 20.0 -a 1,1 -E faac,ac3 -B 160,auto -R 48,Auto -6 dpl2,auto -f mp4 -X 720 -P -m -x level=30:cabac=0:ref=2:mixed-refs=1:me=umh:no-fast-pskip=1

+ iPod: -e x264 -b 700 -a 1 -E faac -B 160 -R 48 -6 dpl2 -f mp4 -I -X 320 -m -x level=30:bframes=0:cabac=0:ref=1:vbv-maxrate=768:vbv-bufsize=2000:analyse=all:me=umh:no-fast-pskip=1

+ iPhone & iPod Touch: -e x264 -q 20.0 -a 1 -E faac -B 128 -R 48 -6 dpl2 -f mp4 -X 480 -m -x level=30:cabac=0:ref=2:mixed-refs:me=umh:no-fast-pskip=1

+ AppleTV: -e x264 -q 20.0 -a 1,1 -E faac,ac3 -B 160,auto -R 48,Auto -6 dpl2,auto -f mp4 -4 -X 960 -P -m -x level=30:ref=2:mixed-refs=1:bframes=3:direct=auto:no-fast-pskip=1:me=umh:subme=7:b-adapt=2:8x8dct=1

>

< Regular

+ Normal: -e x264 -q 20.0 -a 1 -E faac -B 160 -R Auto -6 dpl2 -f mp4 -p -m -x ref=2:bframes=2:me=umh

+ High Profile: -e x264 -q 20.0 -a 1,1 -E faac,ac3 -B 160,auto -R 48,Auto -6 dpl2,auto -f mp4 --detelecine --decomb -P -m -x ref=3:mixed-refs:bframes=3:weightb:direct=auto:b-pyramid:b-adapt=2:me=umh:subme=9:analyse=all:8x8dct:no-fast-pskip=1

>

< Legacy

+ Classic: -b 1000 -a 1 -E faac -B 160 -R Auto -6 dpl2 -f mp4

+ AppleTV Legacy: -e x264 -b 2500 -a 1,1 -E faac,ac3 -B 160,auto -R 48,Auto -6 dpl2,auto -f mp4 -4 -p -m -x bframes=3:ref=1:subme=5:me=umh:no-fast-pskip=1:cabac=0

+ iPhone Legacy: -e x264 -b 960 -a 1 -E faac -B 128 -R 48 -6 dpl2 -f mp4 -I -X 480 -m -x level=30:cabac=0:ref=1:analyse=all:me=umh:no-fast-pskip=1:psy-rd=0,0

+ iPod Legacy: -e x264 -b 1500 -a 1 -E faac -B 160 -R 48 -6 dpl2 -f mp4 -I -X 640 -m -x level=30:bframes=0:cabac=0:ref=1:vbv-maxrate=1500:vbv-bufsize=2000:analyse=all:me=umh:no-fast-pskip=1:psy-rd=0,0

>

[rpp3po]

I like the new presets. Regarding the use of direct=auto in some of them: Back in the "Consistency of presets" thread you have cited a source about direct=auto not being well suited for 1-pass encoding. Then the decision about a B-frame's mode must be made based on the former frame's stats instead of the current (as in the 2nd pass). Has this changed for newer builds of x264?

[jbrjake]

Regarding the use of direct=auto in some of them: Back in the "Consistency of presets" thread you have cited a source about direct=auto not being well suited for 1-pass encoding.

Well, the temporal decision doesn't really work that well in 1-pass. But as far as I can tell, that just means it ends up mostly using spatial because most of the time it can't find a good temporal match. I put it in with the reasoning that a little bit of temporal is better than none, but maybe it would be better to do spatial like akupenguin suggested several years back. I'll do some tests, since I'm not entirely sure when he says "direct=auto only really works in 2pass" and D_S says "direct temporal is the one decision which is slightly better in twopass mode" they mean that it's actually harmful, or just not very useful.

[Ritsuka]

Is no-fast-pskip=1 still useful?

[jbrjake]

No-fast-pskip's not as useful since psy-RD and VAQ, no. And of course those make it difficult to do any metric tests. But for what it's worth, on the same sample I tested the above presets on (Lost season 1 episode 1 chapter 1, just like for the 0.9.3 presets), here's how default x264 options vs. -x no-fast-pskip turned out:

fast-pskip:
[18:48:52] work: average encoding speed for job is 37.559875 fps

x264 [info]: slice I:244 Avg QP:16.33 size: 25681 PSNR Mean Y:47.76 U:50.96 V:51.81 Avg:48.61 Global:47.78
x264 [info]: slice P:10594 Avg QP:20.04 size: 8683 PSNR Mean Y:44.29 U:49.18 V:49.88 Avg:45.29 Global:43.93
x264 [info]: mb I I16..4: 24.5% 0.0% 75.5%
x264 [info]: mb P I16..4: 2.4% 0.0% 8.3% P16..4: 48.0% 22.3% 11.1% 0.0% 0.0% skip: 7.8%
x264 [info]: SSIM Mean Y:0.9785912
x264 [info]: PSNR Mean Y:44.364 U:49.223 V:49.919 Avg:45.360 Global:43.992 kb/s:1738.86

no-fast-pskip:
[18:54:14] work: average encoding speed for job is 37.211720 fps

x264 [info]: slice I:244 Avg QP:16.33 size: 25681 PSNR Mean Y:47.76 U:50.96 V:51.81 Avg:48.61 Global:47.78
x264 [info]: slice P:10594 Avg QP:20.04 size: 8691 PSNR Mean Y:44.30 U:49.18 V:49.88 Avg:45.30 Global:43.94
x264 [info]: mb I I16..4: 24.5% 0.0% 75.5%
x264 [info]: mb P I16..4: 2.4% 0.0% 8.3% P16..4: 49.0% 22.4% 11.2% 0.0% 0.0% skip: 6.5%
x264 [info]: SSIM Mean Y:0.9786211
x264 [info]: PSNR Mean Y:44.378 U:49.224 V:49.920 Avg:45.373 Global:43.999 kb/s:1740.41

It doesn't really slow things down very much, and it might, if you believe such dubious and anecdotal metrics evidence, still improve P-frame quality.

...but yeah, I'm thinking about dropping it as well as direct=auto.

Oh yeah, here's a test of direct=auto on the same sample, but this time with more options applied:

jonathon-rubins-macbook-pro:build jonr$ ./HandBrakeCLI -i ~/Movies/Samples/lost.vob -o ~/Desktop/autotest.m4v -e x264 -q 20.0 -a 1,1 -E faac,ac3 -B 160,auto -R 48,Auto -6 dpl2,auto -f mp4 -4 -X 960 -P -m -x level=30:ref=2:mixed-refs=1:bframes=3:direct=auto:no-fast-pskip=1:me=umh:subme=7:b-adapt=2:8x8dct=1 -v2
x264 [info]: slice I:251 Avg QP:16.25 size: 28880 PSNR Mean Y:48.29 U:51.20 V:52.14 Avg:49.09 Global:48.30
x264 [info]: slice P:4517 Avg QP:19.16 size: 15064 PSNR Mean Y:45.65 U:49.89 V:50.69 Avg:46.58 Global:44.75
x264 [info]: slice B:6071 Avg QP:21.36 size: 5551 PSNR Mean Y:43.91 U:49.35 V:50.07 Avg:44.98 Global:43.96
x264 [info]: consecutive B-frames: 8.3% 23.6% 66.6% 1.5%
x264 [info]: mb I I16..4: 5.2% 86.6% 8.2%
x264 [info]: mb P I16..4: 1.5% 23.6% 2.9% P16..4: 38.3% 19.2% 10.0% 0.0% 0.0% skip: 4.5%
x264 [info]: mb B I16..4: 0.2% 2.6% 0.4% B16..8: 57.3% 3.1% 4.0% direct: 7.5% skip:24.7% L0:35.8% L1:51.6% BI:12.6%
x264 [info]: 8x8 transform intra:84.0% inter:68.6%
x264 [info]: direct mvs spatial:98.4% temporal:1.6%
x264 [info]: ref P L0 83.7% 16.3%
x264 [info]: SSIM Mean Y:0.9793430
x264 [info]: PSNR Mean Y:44.737 U:49.619 V:50.376 Avg:45.741 Global:44.342 kb/s:1928.73

jonathon-rubins-macbook-pro:build jonr$ ./HandBrakeCLI -i ~/Movies/Samples/lost.vob -o ~/Desktop/noautotest.m4v -e x264 -q 20.0 -a 1,1 -E faac,ac3 -B 160,auto -R 48,Auto -6 dpl2,auto -f mp4 -4 -X 960 -P -m -x level=30:ref=2:mixed-refs=1:bframes=3:no-fast-pskip=1:me=umh:subme=7:b-adapt=2:8x8dct=1 -v2
x264 [info]: slice I:251 Avg QP:16.25 size: 28880 PSNR Mean Y:48.29 U:51.20 V:52.14 Avg:49.09 Global:48.30
x264 [info]: slice P:4517 Avg QP:19.16 size: 15064 PSNR Mean Y:45.65 U:49.89 V:50.69 Avg:46.58 Global:44.75
x264 [info]: slice B:6071 Avg QP:21.37 size: 5552 PSNR Mean Y:43.91 U:49.35 V:50.07 Avg:44.98 Global:43.96
x264 [info]: consecutive B-frames: 8.3% 23.6% 66.6% 1.5%
x264 [info]: mb I I16..4: 5.2% 86.6% 8.2%
x264 [info]: mb P I16..4: 1.5% 23.6% 2.9% P16..4: 38.3% 19.2% 10.0% 0.0% 0.0% skip: 4.5%
x264 [info]: mb B I16..4: 0.2% 2.6% 0.4% B16..8: 57.4% 3.1% 4.0% direct: 7.5% skip:24.7% L0:35.8% L1:51.6% BI:12.6%
x264 [info]: 8x8 transform intra:84.0% inter:68.6%
x264 [info]: ref P L0 83.7% 16.3%
x264 [info]: SSIM Mean Y:0.9793431
x264 [info]: PSNR Mean Y:44.738 U:49.619 V:50.376 Avg:45.742 Global:44.342 kb/s:1928.83