x264 & AMD Stream SDK (GPGPU, SMP, CUDA)

[sacharja]

AMD will release its answer on NVIDIA's CUDA in 3 weeks (German translated news):
http://translate.google.com/translate?u=http%3A%2F%2Fwww.pcgameshardware.de%2Faid%2C666909%2FNews%2FAti_Stream_ab_Catalyst_812_mit_GPU-Video-Encoding_und_mehr%2F&hl=de&ie=UTF-8&sl=de&tl=en

And AMD's presentation is here:
http://www.pcgameshardware.de/&menu=browser&entity_id=154104&image_id=936621&article_id=666909&mode=article&overview=yes&page=2&show=original

AMD Stream SDK is released on 10th december with the new Catalyst 8.12 driver, free and open source to download; up to 17x acceleration for all HD 3xxx+ modells. Will some x264 developers have a look and might x264 support Stream?

[Dark Shikari]

Patches welcome, but if their performance claims are accurate, its completely useless.

They claim 5x realtime encoding speed of 320x240 video.

I mean, seriously?

x264 can 26.5x realtime on a brand new Nehalem 2.66Ghz... with one core. If their marketing is actually true, this will be slower than Badaboom by a wide margin.

[cogman]

Patches welcome, but if their performance claims are accurate, its completely useless.

They claim 5x realtime encoding speed of 320x240 video.

I mean, seriously?

x264 can 26.5x realtime on a brand new Nehalem 2.66Ghz... with one core. If their marketing is actually true, this will be slower than Badaboom by a wide margin.

Bu, bu, Its got the x in it. Anything that is 5x something must be fast :P.

On a serious note, with AMD trying to be OpenCU (or is it OpenCL) compliant, and I believe nVidia is already there. Could there be a speed boost realized in x264, and if so, where?

[Dark Shikari]

Could there be a speed boost realized in x264, and if so, where?Sure, but it would require offloading a very large portion of the code to the GPU (latency is too high to offload only a small portion), so it would take a lot of effort. Patches welcome though.

If someone wanted to try something simple, they could port the hpel filter or lowres interpolation filter to the GPU.

[sacharja]

I've corrected the links in the post above. Please have a look. They claim they're 1,5% faster than badaboom:
http://www.pcgameshardware.de/&menu=browser&entity_id=154104&image_id=936636&article_id=666909&mode=article&overview=yes&page=3&show=original. This seems realistic because AMD's shader performance is better than Nvidia's.

Here it says sth. about h.264 support, but I don't know what they mean (avivo?).
http://www.pcgameshardware.de/&menu=browser&entity_id=154104&image_id=936635&article_id=666909&mode=article&overview=yes&page=2&show=original

BTW: the AMD cards surely have a DEcoder for h.264. As you can read in the presentation Cyberlink even claimed a magnificent performance boost with Stream in the upcoming WinDVD.

[CruNcher]

For now it's still questionable if the GPU hype is all worth it currently GPU still lack in terms of Power Efficiency Imho though if you compare a Quadcore with a new Generation GPU the difference slowly is vanishing and i guess sooner or later CPU cant hold up anymore, that will be the time GPU Encoding really becomes interesting :) i made a small Power Consumption comparison (Decoding for now) Encoding (in the works) with my G92 (rather old GPU vs old CPU)

Blu-Ray (Casino Royale Construction Run)

idle = 105W
CoreAVC = 160W
GPU = 113W

This was on a X2 CNQ enabled running on it's highest multiplier @ 1.3v 2.4 GHz (load) and lowest multiplier @ 1.1v 1.1 GHz (idle)
a heavy amount of power draw comes from the GPU while in Idle state (Energy is totally inefficient blown out here, 3.2v all the time no usage adaption)

I should mention that the Decoding in this example doesn't work over CUDA but on the PV2 Logic a small ARM11 Core that's inside Nvidias GPU and runs @ 400 MHz (basicly a SOC inside the GPU, independent of it) Ati uses the same approach calling the chip logic UVD

Though there are other solutions in form of specialized low cost SOC's (first time big adoption of that gonna be seen with Microsofts Sheed Transcoding on Win7) that have the same or even better efficiency then GPUs internal used Decoding logic and would allow even more power to be saved (those solutions are mostly used in the Professional area but slowly emerge in consumer trans coding as well), especially in Netbooks those SOC's are gonna get used alot :)

[lexor]

CruNcher, I don't follow your logic. You claim GPU is power inefficient, but then provide measurements that say the opposite.

[sacharja]

@CruNcher
You're comparing apples with pears. On simple tasks a GPU can be 200 times faster than a CPU, i.e. a CPU with 45W would need 9000W for the same performance. Disadvantage is that the application area of a GPU is VERY limited.

@lexor
I also have a HD2600XT ;) Do you also buy the successor to RV630, the RV730 (HD4670)? Mine is coming on wednesday, I'm curious if it really consumes just 3W more while being twice as fast (and Stream of course).

[LoRd_MuldeR]

@CruNcher
You're comparing apples with pears. On simple tasks a GPU can be 200 times faster than a CPU, i.e. a CPU with 45W would need 9000W for the same performance.

I'd say you are comparing apples with pears. What is the benefit of a GPU being able to do a certain calculation 200x faster than a CPU, but there's no encoder that is actually able to make use of that theoretical speed-up? We are yet to see a GPU-based encoder that runs only 2x faster than the CPU-based one and provides equal quality at the same bitrate. So a speed-up of 200x seems to be completely unachievable in the near future. I'm pretty sure the CPU won't be unemployed that soon. It's more likely that particular GPU-friendly tasks will be moved to the GPU...

[CruNcher]

CruNcher, I don't follow your logic. You claim GPU is power inefficient, but then provide measurements that say the opposite.

No i just want to say with that, that it's not yet their in terms of state switching :) and shutting of unneeded logic to save power @ idle.
The power consumption @ work is ok @ least for the task of Video Decoding (though that is done in this case not on the GPU but on a SOC like logic inside the GPU)

@sacharja
as Lord_Mulder said show me the GPU that does 200x transcoding @ the same Quality as current CPU Encoders,
even Nvidia and Ati don't go that high and claim something like 17x, though both compare on a Core2 vs Quicktime Itunes and that is the real joke here :)

[LoRd_MuldeR]

even Nvidia and Ati don't go that high and claim something like 17x, though both compare on a Core2 vs Quicktime Itunes and that is the real joke here :)

Until they trust their own encoders enough to compare them against x264 running on Nehalem (or at least Core2 Quad), we are far away from GPU encoders taking over the scene ;)

[sacharja]

@CruNcher
As I wrote some posts before, GPUs can be the ultimate performance boost, but just for a very limited set of tasks (like folding proteins in medical applications). You can't compare them in the means of efficiency. Either they are up to 200 times more efficient, or they are totally inefficient because they can't run 99% of the programs you have on your computer.

[CruNcher]

The most efficiency of GPUs i currently see for the average consumer (and i only speak for the average consumer as this is what Nvidia and ATi market this stuff for primarily) in broad range software to not waste this idle energy (especialy of the older GPU generations) is to allow nice animated GUIs for OSes at least there you use the power consumption for something even if we can arguee about the usefullness of that (though alot of average consumers like it) :)

[lexor]

@lexor
I also have a HD2600XT ;) Do you also buy the successor to RV630, the RV730 (HD4670)? Mine is coming on wednesday, I'm curious if it really consumes just 3W more while being twice as fast (and Stream of course).

Nah, I'm still on the 7 year old AGP system here, so no more upgrades for this little puppy. Come January/February I'll finally be buying a replacement, though it will probably be Nvidia (simply because I want to get in on CUDA decoder Neruon2 got going with DGAVCDecNV). ATI dropped the ball on this, and it's the same ball AMD dropped when competing with Intel's compiler division.

[CruNcher]

Why are you so sure ATI droped the ball on giving access to UVD for average developers i wouldn't say so see the recent Stream SDK anouncement most probably there will be also a way to access UVD via it like there was for PV2 and Nvidia :)
They are late yes but i don't think that they gonna withold UVD access to average developers seeing Nvidias extreme moving forward on this matter like recently the Linux Patch for all the important OSS Video stuff http://www.phoronix.com/scan.php?page=article&item=nvidia_vdpau&num=1 http://mythtv.org/pipermail/mythtv-dev/2008-November/063677.html mplayer, libavcodec, libavutil, and ffmpeg
Ati has to react now and the API is allready implemented and waits to get used :) see http://www.phoronix.com/scan.php?page=article&item=amd_xvmc_xvba&num=1 only the tools to access it are missing and those will most probably arrive with the Stream SDK on the 10th of December, maybe ATI now under presure are even gonna bring out own OSS patches before the official SDK release :)

[9876]

I think if [AMD Stream/Nvidia CUDA] works, why not using them together, even they are slow?

A simple example:
I can use the C2Q to encode my video vol 1,2,3 and use the GPU to encode my video vol 4...
If they complete at the same time, it will boost performance 25% (or less little), right?

:)

[LoRd_MuldeR]

I don't think the GPU can take over the entire encoding process, making the CPU free to perform another encode concurrently.

It's more likely that the GPU will assist the CPU as a Co-Processor by moving particular parts of the encoding process from the CPU to the GPU.

Also you can easily put 100% load on your C2Q by encoding one single video with x264, no need to encode several videos concurrently...

[foxyshadis]

Moving to hardware, discussion isn't really related to x264 (until someone writes some patches).

[sacharja]

Moving to hardware, discussion isn't really related to x264 (until someone writes some patches).
Patches? Stream SDK v1.3 is released on 10th december. But it would be interesting if someone plans writing some when it's been released. AMD says implementing Stream support will be easy, I think we just need to wait and see.

[deekey777]

I think if [AMD Stream/Nvidia CUDA] works, why not using them together, even they are slow?

A simple example:
I can use the C2Q to encode my video vol 1,2,3 and use the GPU to encode my video vol 4...
If they complete at the same time, it will boost performance 25% (or less little), right?

:)


With PowerDirector 7 and HD4850 you will able to encode four videos at once.

Here is a nice interview: ATI Stream Team Interview (http://www.rage3d.com/articles/stream/index.php?p=3)
WTF is ATV? -> http://ati.amd.com/products/firepro/Siggraph_2008_video_encode_final.pdf



It's always the same: x.264 against the world, but what's about x.264 and eg. MC as a part of TMPGEnc and CUDA filters?

[LoRd_MuldeR]

With any pure software encoder I'm able to encode as many videos concurrently as I like :p

The only question is: Does it make sense ???
Instead of encoding n videos at the same time, I could simple convert them one after another (batch encode) and each video will taking only 1/n of the time...

[Jawed]

This PDF should prove to be rather sobering info for anyone who's still optimistic:

http://ati.amd.com/products/firepro/Siggraph_2008_video_encode_final.pdf

Note how the comparison is against a single core.

LOL, the thread moved on a bit while I was reading/being-distracted/posting.

Jawed

[deekey777]

With any pure software encoder I'm able to encode as many videos concurrently as I like :p

The only question is: Does it make sense ???
Instead of encoding n videos at the same time, I could simple convert them one after another (batch encode) and each video will taking only 1/n of the time...

If the converting a video for an iPhone underutilizes a HD4850, why not to convert 2-4 videos at once, if it doesn't take (much) more time (example one video -> 30 minutes, four videos -> 90 minutes)?
I've converted my Firefly DVDs for my iPhone and it took a lot of time (AutoMKV, 13 episodes*, ~ 40 minutes for every episode, [2-pass, X2 4200+], the quality is great!)


*Serenity (Pilot) was converted by Nero

[LoRd_MuldeR]

If the encode "underutilizes" the GPU, you could simply use 100% of it's processing power instead of only a part of it and finish the one video faster. Then process the next video. And so on...

So if you can finish n videos in m minutes, then you can finish 1 video in m/n minutes. Or better: You should be able to do so!

The fact that they need to encode several videos concurrently to get 100% load on the hardware shows that they have a bottleneck in their encoder, their multi-threading code scales really bad or both :rolleyes:

[CruNcher]

Yep look @ the slides carefully it will be the same situation as with Nvidia for Encoding the CPU will still be utilized for Nvidia those utilization is somewhere @ 30% most likely for ATI it will be somewhere in that region too, it's not the same situation as with Video Decoding (which is almost done entirely on the GPU, though that is done on it's own optimized logic inside of the GPU).

[MfA]

Non greedy R/D optimization of MB parameters suits the GPU well, since the search space is too great for trellis based techniques (presently most of the optimization above MB level is greedy, don't think even the quantizer is done with trellis search although if you stick to just optimizing that in isolation trellis is practical).

This is not the kind of thing the commercial world or NVIDIA/ATI is likely to produce though ...

[cyberbeing]

It sounds like AMD is going to be releasing a h.264 encoder along with the Stream SDK?

If so, does anybody know if AMD's encoder will be like Badaboom and only support Baseline Profile?

[CruNcher]

thats unfortunately 1 thing they left complete unanswered and i guess they even didn't answered it to the audience most probably no one asked either ;)

[sacharja]

Because the whole Stream is open source and meant to ensure a platform independent programming and to create some standards, we might even see a x264 implementation :D


Because this is now in Hardware forum: I just got my Sapphire HD-4670 today for 77€ and it's almost twice as fast as my old HD-2600. The (passive) Accelero S2 cooler fits perfectly (even if it's not stated at the homepage). Board design is exactly the same as the Sapphire HD-2600XT, just the chip and ram is different. I have a passive system (not one fan in the whole PC system) and the card is just about 60°C at full load and under 50°C at idle, if it had a fan it would run with 3% ;) Watching movies is subjectively faster and the CPU load is with 48% less than the 90% with the old card. If Ati's postprocessing filters improve I might use the Stream decoder codecs in the near future. Haven't tested the HDMI 7.1 LPCM sound output though.

[Sharktooth]

the good news is the encoder will be included in a future version of the catalyst drivers... ;)

[sacharja]

the good news is the encoder will be included in a future version of the catalyst drivers... ;)
Do you know which encoder?

[tomos]

isnt it just going to be avivo basically but this time it will actually use the GPU? comes out on the 10th or so of Dec iirc

[Sharktooth]

all i can say is it's not the old encoder running on the GPU.

[sacharja]

Stream SDK v1.3 Beta is released: http://forums.amd.com/devforum/messageview.cfm?catid=328&threadid=104500&enterthread=y

Brook+ is AMD’s modified Brook open source compiler. CAL is AMD's forward-compatible, cross-platform interface to the GPU. Both components are required to implement applications in the Stream environment. The download comes complete with a unified installer. Installation includes programming guides and specifications for Brook+ and CAL.
Sounds good to me ;)

BTW: I tried the Catalyst 8.12 from here http://www.sapphiretech.com/us/ , seems to be the RC3 instead of the final, but I couldn't resist. Nevertheless I didn't found any h.264 encoder. CCC's Avivo is just encoding Mpeg-1, Mpeg-2, WMV, DivX and Xvid (I tested the whole drop-down box, because it's not obvious in CCC which codec is used). I also got no video while converting a WMV-HD to ipod.

[sacharja]

I was able to encode a WMV-HD trailer with ATIXcoder (http://forums.guru3d.com/showthread.php?t=226993) and H264 (not the ipod crap).

About the quality, maybe you can see a difference between x264 and AVIVO :D :
http://rapidshare.com/files/172091601/AVIVO_with_GPGPU_vs._x264.rar.html

At least it's fast, a full HD movie is converted to H264 in 5 minutes instead of 30 minutes with x264. BTW: I wasn't able to test how the GPU is used because I momentarily have no tool that can show me GPU usage, ATITool is not working with my new HD-4670.

[Dark Shikari]

At least it's fast, a full HD movie is converted to H264 in 5 minutes instead of 30 minutes with x264.Have you tried using equivalently fast settings with x264 and comparing quality?

Edit: god, what an awful comparison. If you're going for max quality, the x264 settings are awful, and if you're going for comparability to the AVIVO encode in terms of features used, they're bad as well...

x264 - core 65 r1016bm dbc5ef0 - H.264/MPEG-4 AVC codec - Copyleft 2003-2008 - h
ttp://www.videolan.org/x264.html - options: cabac=1 ref=1 deblock=1:0:0 analyse=
0x3:0x13 me=umh subme=7 psy_rd=0.0:0.0 mixed_ref=0 me_range=16 chroma_me=1 trell
is=1 8x8dct=1 cqm=0 deadzone=21,11 chroma_qp_offset=0 threads=2 thread_queue=2 n
r=0 decimate=1 mbaff=0 bframes=0 keyint=250 keyint_min=25 scenecut=40(pre) rc=ab
r bitrate=670 ratetol=1.0 qcomp=0.60 qpmin=10 qpmax=51 qpstep=4 ip_ratio=1.40 aq
=1:1.00:0:10.00

And H.264 in AVI? Single-pass encoding? 8x8dct but no B-frames? Psy off? Threads 2, despite the fact that there is no CPU in the world on which threads 2 is optimal for performance? facepalm

(For comparison, AVIVO uses baseline profile...)

Edit again:

I am laughing hysterically. Thank you for the AVIVO encode. You have vindicated yourself. :D

Speaking of which, can I have a sane encode from AVIVO that isn't at near-CQP51? I can't reverse-engineer it with this kind of insanity ;)

[sacharja]

Just tried to use the same settings in AVIVO and x264 with the few settings I can change in ATIXcoder (in AVIVO directly you can't even change the resolution) while keeping the bitrate low to emphasize differences.

If you think you can make a better comparison go ahead ;) But AVIVO is known for its poor quality, thus the results seemed realistic to me.

I only did this comparison because I initially had the hope that AMD uses x264 instead of their cheap encoders.

Have you had a look at Stream v1.3 yet? Is there hope for a x264 with Stream support?

EDIT: Here's a detailed test of PCGH http://translate.google.com/translate?hl=de&u=http%3A%2F%2Fwww.pcgameshardware.de%2F%3Farticle_id%3D669991&sl=de&tl=en&swap=1
As the first AVIVO this seems to be just a software encoder that IS NOT using Steam.

[deekey777]

RE PCGH:
http://forum.beyond3d.com/showthread.php?p=1247436#post1247436

And here are ComputerBase' results:
http://www.computerbase.de/artikel/hardware/grafikkarten/2008/kurztest_ati_video_converter_benchmarks/3/#abschnitt_benchmarks

So, the GPUs are used.

And it's better to install ATi Tray Tools, then you get this:
http://img2.abload.de/img/unbenanntxynm.jpg


What really sucks, that AVIVO@GPU/AVT support is limited to HD4000 series. Official statement: "HD4800 delievers 5x better integer performance". Great, do I care about this? I'm an owner of the HD3850 and here are its specs:
http://ati.amd.com/products/radeonhd3800/specs.html
ATI Avivo™ HD Video and Display Platform

* Dedicated unified video decoder (UVD) for H.264/AVC and VC-1 video formats
...
* MPEG-2, MPEG-4, DivX, WMV9, VC-1, and H.264/AVC encoding and transcoding

[tomos]

just tried the avivo thing after downloading it off ATis site since it wasnt included in the 8.12 package, and right there on the page it says:

Avivo Video Converter*
*Avivo Video Converter will only work
with X1000 series products

installed it anyway and it doesnt accept any type of file i have. VOBs, M2TS, matroska and doesnt take AVS scripts either. i dont have AVIs to test it with so it's pretty much useless to me now :(

[Sharktooth]

older cards will be probably supported in future versions.

[tomos]

i have a 4850 though.

[Bitwarrior]

just tried the avivo thing after downloading it off ATis site since it wasnt included in the 8.12 package, and right there on the page it says:



installed it anyway and it doesnt accept any type of file i have. VOBs, M2TS, matroska and doesnt take AVS scripts either. i dont have AVIs to test it with so it's pretty much useless to me now :(

Unfortunately ATI did not expose the encoder as a DirectShow device. I looked at newly installed video codecs but unfortunately nothing was there. Otherwise we could use Graphedit to use it for converting from almost all other formats.

[Sharktooth]

it does.

[LoRd_MuldeR]

it does.

He probably was looking for a VfW Encoder (e.g in VirtualDub or VirtualDubMOD). DirectShow encoders won't show up there, of course.

@Bitwarrior: Try to use GraphStudio or GraphEdit.

[Bitwarrior]

He probably was looking for a VfW Encoder (e.g in VirtualDub or VirtualDubMOD). DirectShow encoders won't show up there, of course.

@Bitwarrior: Try to use GraphStudio or GraphEdit.

I did. I will try again and report here (naming is probably not that obvious). It would be nice to be able to write an alternative user interface around it and to extend the input formats it likes.

[LoRd_MuldeR]

It would be nice to be able to write an alternative user interface around it and to extend the input formats it likes.

Like this?

http://forum.doom9.org/showpost.php?p=1222130&postcount=4

[Bitwarrior]

Like this?

http://forum.doom9.org/showpost.php?p=1222130&postcount=4

:thanks: I will check it and test whether it likes DV as input. However the remarks about quality are worrying. It might not be worth the effort after all.

[qyqgpower]

although this is kinda OT, I have several questions about ATI's directshow filters.

1. Is the hardware deinterlacer accessible via graphedit? I remember someone had said he could use it in his encoding. I have found a filter named "MMACE DeInterlace" which is from ATI's dll. But only two output pins in this filter and no input pin is available.

2. what's the directshow name of xcode.dll (which is responsible for the ATI converter)?

[tre31]

although this is kinda OT, I have several questions about ATI's directshow filters.

1. Is the hardware deinterlacer accessible via graphedit? I remember someone had said he could use it in his encoding. I have found a filter named "MMACE DeInterlace" which is from ATI's dll. But only two output pins in this filter and no input pin is available.

2. what's the directshow name of xcode.dll (which is responsible for the ATI converter)?

1. Yes it is using Cyberlink PowerDVD (PDVD7) decoder, then using a directshow encoder (if you want too do it in graphedit/graphstudio) or using something that supports input from diretshow like this - http://forum.doom9.org/showthread.php?t=141441

Read more here - http://forum.doom9.org/showthread.php?t=143106
I tried too explain what I do there, however the person I explained too is still asking the same question even after I explained what too do, just no helping some people even after you give them the cleaned fish, fillet it, and even go as far as put the oil in the pan ... lol

2. Not sure I can't find a way as yet, the MPEG Encoder is available.

------
My test on ati hd2600xt using (http://www.radeon.ru/downloads/att/b/atixcoder.rar) :
input - 1920x1088 pal 16:9 m2v
output - 720x576 16:9 h264
gpu usage during encode - 18% constant
cpu usage on q6600 - 90-95% on 3 cores last was not used much

verdict on output - .avi (WTF ati? - this crashed explorer multiple times with cl264dec.ax being culprit when i even clicked on the file)
was baseline 3.0 - er .. why even bother
encoded at 100fps, but I'm sure x264 would do the same or perhaps better at such low settings.

gpu transcoding - well it was using the gpu for something, whether that was de-interlacing or resizing or encoding I really don't know, as for the output - didn't bother too check since it was crashing things, so why bother, in the end at such low settings I say atm = FAIL ... in the future if they do Main & High Profile and allow for better configuration settings - lets face it if you are going too use the gpu you want too use good encoding settings, not bare minimum, then and only then will you be able too perform a good comparison against x264.

[sacharja]

I found an article from Hexus.net (http://www.hexus.net/content/item.php?item=11725), where they were able to test GPGPU with a Radeon card and a Adobe Premiere Plugin. Performance boost versus CPU only according up to 4x:
As it happens, many media-rich applications' code is wonderfully parallelisable, meaning that, if coded for, additional plug-ins could take the load off the CPU, place it on the broad, broad shoulders of the GPU, and execute in a more-timely fashion. As an example, during the CTO conference in Amsterdam, Holland, a high-definition clip was transcoded to MPEG2 HD. Running on a Radeon HD 3870 X2, with the Adobe Premier plug-in activated, the GPU ran at 4x a quad-core CPU's speed.

AMD's presentation about AVT (Accelerated video transcoding) says the following about supported programs:
http://www.freeimagehosting.net/uploads/th.23682aeb30.png (http://www.freeimagehosting.net/image.php?23682aeb30.png)

It seems that Cyberlinks PowerDirector 7 (updated next year) will be the first application that takes full advantage of the GPGPU. I wrote a mail to Cyberlink, because the GPGPU update should've been released in Q3 08. Here is the answer I got:
Dear Customer,

Thanks for your e-mail.
This is Cyberlink Customer Support.

Due to ATI driver release issue, so PowerDirector 7 will support this feature on 2009 Q1.
Sorry for all the inconvenience you've been caused.

We sincerely apologize for any inconvenience you have been caused.
If you have any further question, please feel free to contact us.

Customer Support Dept.
Cyberlink Corp.