Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Daydream]

Any comments on DXVA renderless and madVR working together, as claimed by PotPlayer? And on the one comment on D9 that it's too slow for ATI cards? (may affect video card purchase decisions)

I don't know how proper this guide is for getting it to work, but by the end of it (just reading, not applying) I felt like I forgot my own name.

[Daydream]

Err... _nobody_? All the guys counting on hardware acceleration, who are you gonna leave 'em to?

[glynor]

Patience, grasshopper.

[JimH]

It's under discussion.  We'll be able to answer in a week or two?

[glynor]

It's under discussion.  We'll be able to answer in a week or two?

Awwww.... You went and ruined my sly intimation. 

[Daydream]

Very slick, Glynor, very slick! 

Thanks Jim for looking into it.

[JimH]

Awwww.... You went and ruined my sly intimation. 

Sorry, grasshopper.

[glynor]

Old man, how is it that you hear these things?

[MrC]

Old man, how is it that you hear these things?

The same way this guy could "see" ?

[JimH]

I think that's a question for madshi or nevcairiel.

[Matt]

You can select DXVA using Red October with additional filters.

If madVR accepts a DXVA connection, it should work.

[Hendrik]

madVR does not work with DXVA.

To comment on the whole "DXVA renderless" topic:

- It is surely possible (on Vista/7 at least, XP not so much), one just has to write a decoder to do it
- It is true that ATI is not necessarily fast enough, depending in which mood the driver is right now
- If that influences your buying decision - with an NVIDIA you can use LAV CUVID, which already does hardware acceleration without the DXVA renderer limitation.

Anyhow, as of now, there is no publically available codec which can do this for an ATI/AMD.

[Daydream]

The guide I linked to at the top claims exactly that, DXVA with madVR. On whatever video card. It's true that they mentioned PotPlayer and its own decoders (which may or may not have been coded to work the needed kind of magic). I'll try to follow that guide and see if anything works, even if it's 300 pages long, especially with an AMD card, which if at all possible I would like to keep using.

[Hendrik]

That guide only works with PotPlayer. They probably modified some DXVA codec to do exactly that, without sharing their modifications with the world.

[jmone]

Personally I see DXVA as a good option to extend playback of a wider range of content on low powered PC's.  I just don't see it is a quality option that you would normally want to combine with madVR anyway....(but happy to recognise that we all have a reasons!)

[Daydream]

That guide only works with PotPlayer. They probably modified some DXVA codec to do exactly that, without sharing their modifications with the world.

Thanks nevcairiel, appreciate your insight!

I just don't see it is a quality option that you would normally want to combine with madVR anyway

You're approaching the crux of all things my friend, as far as I'm concern. Is there any solid, quantifiable proof, that the software decoding (via madvr or any other way) it's a sensible improvement so much so that you will notice it anywhere, anyway? If I play mostly 1080p where is the scaling benefit? If I play something else, how is that % improvement gonna stack up against the already proven fact that details and distance to screen and actual real resolution play a much more important part? What did we gain and what did we give up?

[jmone]

If I play something else, how is that % improvement gonna stack up against the already proven fact that details and distance to screen and actual real resolution play a much more important part?

100% agree - we are all twisted up on tech stats but gloss over the final stage where the analog light/audio waves hit our eyes and ears.  Eg, most have uncalibrated screens, with low perceived resolution (eg sit too far back on a screen that is too small to be able to resolve the detail) etc.  Anyway - like all OCD sufferers there are plenty chasing that last 1%....

What did we gain and what did we give up?

What ever you think you see!  I've never done a double blind with DXVA vs SW decoding....but the idea of the RO HQ is the best of all the bits, and comments from madshi indicate that DXVA is not a priority for him to add support for in madVR.

[BryanC]

Honestly I view DXVA as a way to save $50 a year on energy costs by using a low-voltage system.

[Hendrik]

The actual decoding is supposedly 100% the same quality, no matter if you use DXVA or a software decoder. The spec of the codecs involved does not allow for derivations. I would generally assume that all codecs produce the same quality, unless they have a bug.

What DXVA/Hardware decoding offers you:
- Possibly lower power usage, due to using a highly optimized piece of hardware, instead of the general purpose CPU
- The ability to decoder complex videos even with a slower CPU
- Hardware Deinterlacing

What Software decoding offers you:
- More flexibility (all renderers work with it, you can do software post-processing, etc, etc)
- With a fast CPU, higher performance, even on the most complex of materials

Now, with the right components in the playback chain, you can mix/match those advantages.
For example, with LAV CUVID, i get all the advantages from the Hardware Decoding section, as well as the Flexibility of a Software Codec.
Or as an alternative - if you're using EVR with a software codec, you get all the Software advantages with the ability to still do Hardware Deinterlacing.

Now we add madVR:
Pro:
- High up/downscaling quality (even on 1080p content, the Chroma needs to be upscaled from 4:2:0 to 4:4:4)
- Most fluid playback i've ever seen

Con:
- No DXVA support
- No Hardware Deinterlacing
- Higher requirements on GPU performance

Because i use LAV CUVID, the "Con"'s are not relevant for me, so the question is - why not use it?
Maybe i'll eventually get around to adding "DXVA renderless" support to my new LAV Video, for people desperately wanting it - or madVR gets DXVA support one day - we'll see what the future brings.

Conclusion:
I don't know how well the EVR in MC16 works, i've been with madVR all the way. But it might be an acceptable solution if you desperately require DXVA and Hardware Deinterlacing, and do not own a NVIDIA card. As you mentioned already, alot of people will not notice the slight improvements madVR offers, especially on 1080p content its not really that obvious.

If you want the best of the best, and you want it right now - the way to go is NVIDIA with LAV CUVID. You get madVR with Hardware Decoding and Deinterlacing. But if you already own your hardware, or want to stick to AMD/ATI for some other reason - i think alot of people are actually pretty happy with their DXVA+EVR setup as well. Now if you never watch Interlaced content, and your CPU is fast enough, you could as well just go with Software Decoding + madVR in this case.

[Daydream]

Excellent description!

As Matt points above Ro + additional filters can help towards DXVA.

It's just that I didn't see much discussion about it, like it was relegated to an afterthought. In my opinion it's not that insignificant, and maybe a more up front 'DXVA' labeled solution in MC would be warranted. There's no question that madvr+LAV (CUVID) is the top solution, it's just that people with HTPCs that do not resemble a desktop (less power used, less heat, less noise, smaller form factor) and with laptops and the likes not made with Nvidia hardware (or by... Alienware) would benefit greatly from some more attention.

Same goes towards LAV. If there was not enough demand for tricks and twists to make it work with DXVA, well nev, you got me as +1 . Regardless, keep up the good work, what you do is amazing.

[JustinChase]

a very well presented explanation

Thank you for that!

I recently purchased an ATI  HD5550 video card, mainly because I (thought I) wanted to bitstream to my receiver, and just stuck with ATI.  Not so important to me now.  I'm running a several year old, cheaper CPU* in this HTPC also.

LAV and madVR will not play all videos without stuttering, which I assume to be caused by my insufficient hardware and/or lack of DXVA decoding.

Is it reasonable to assume that simply installing a current nVidia card would resolve this limitation, or will the CPU need upgraded also to overcome this situation?



ATI  HD5550 - [320 Stream Processing Units, 16 Texture Units, DDR3/GDDR5 memory interface]
*Intel dual core E5200 - [2M Cache, 2.50 GHz, 800 MHz FSB]

[Matt]

I did a little power consumption testing because of this thread:
http://yabb.jriver.com/interact/index.php?topic=65585.0

[jmone]

I agree that we should also have a profile if possible of RO with DXVA.  It will extend the use of RO to lower spec PC's, is GPU agnostic + we already have the filters as part of FFDSHOW (apart from that I'm of little help as I've never had to use DXVA)!

[Daydream]

I did a little power consumption testing because of this thread:

In DXVA land this will be a no-contest battle. Using the shaders means using the GPU; using DXVA - it uses that piece of ASIC on every board that kicks in and decodes everything on single digits W power. I like your idea, I'll grab my kill-o-watt too and throw in some numbers.

[BryanC]

I did a little power consumption testing because of this thread:
http://yabb.jriver.com/interact/index.php?topic=65585.0

Interesting. In my earlier post I was referring to running something like an ION versus a full-fledged desktop processor 24/7/365 (like my HTPC/Library Server).

[roolark]

RO + ffdshow dxva decoder + EVR is a combination that I finally got setup and discovered works perfectly for playing blu-rays on my low power ati/amd system.

I know it's not the madvr, but at least the setup works well.

Would ffdshow dxva decoder work with madvr as a renderer, as it does with EVR?

[jmone]

This is a good example of where DXVA helps, well done getting it working and her is hoping JR can build on this for a "RedOctober - DXVA" profile! 

Also madVR has no compatibility (at least at this stage) with DXVA anyway.....You don't say what GPU you have but I'd not worry about it as all you are likely to add is more issues as madVR likes a GPU with high speed DDR5 memory and lots of shaders.

[roolark]

The GPU is Ati Radeon 6310 HD. It's actually built into the cpu, as AMD's new offering of combines cpu/gpu. The cpu's power is between 18-30 watts or something near there. Of course, the HDD sucks more than that power, lol.

[Daydream]

The GPU is Ati Radeon 6310 HD. It's actually built into the cpu, as AMD's new offering of combines cpu/gpu. The cpu's power is between 18-30 watts or something near there. Of course, the HDD sucks more than that power, lol.

So you have a Zacate platform. Nice! These things are the ones benefiting the most from DXVA. And no, a standard 7200rpm HDD consumes between 6 to 11W (idle - stress), a green one - half of that, an SSD - half of a green one (at least on paper). So, unless we're talking multiple HDDs, one alone is under the power req. for the APU.

[Daydream]

Some power consumption numbers for DXVA.

System used was my laptop since it's closer in config to an HTPC than a desktop build (actually the laptop IS my HTPC right now).
Intel Core i5 - 450M
ATI Mobility Radeon HD 5650, 1GB DDR3, 400 unified shaders.
Screen resolution is 1600x900 (so there is always some scaling).
MPEG-2 is always decoded in software since I don't have a MPEG-2 HW decoder installed and was too lazy to install one.

Software used to read GPU load: GPU-Z

The system defaults to 26W at idle.

Notations:
- /24p, /30i means there are 24 or 30 frames, progressive and interlaced, respectively. The other kind of notations, like 1080i60, which refer to fields, are not used.
- WBM = Windows Based Merit (with EVR)

Test file 1 - Kylie Minogue - Better Than Today (Children in Need - UK TV) - 1440x1080/25 MBAFF H264
Test file 2 - birds scene, Planet Earth, cut from Blu-Ray - 1920x1080/24p VC-1
Test file 3 - file 024 Anantech media test suite - 1920x816/24p
Test file 4 - file 036 Anantech media test suite - MPEG-2 1280x720/60p
Test file 5 - Kylie Minogue - Get Outta My Way (on Leno - NBC) - 1920x1080/30i
Test file 6 - file 005 Anantech media test suite - 1440x1080/30i


                     WBM                     RO Standard                  RO HQ (with GPU load)
Test 1             33W                           40W                                 57W / 61%
Test 2             30W                           43W                                 51W / 38%
Test 3             33W                           38W                                 47W / 31%
Test 4             36W                           36W                                 43W / 54%
Test 5             34W                           35W                                 62W / 79%
Test 6             33W                           39W                                 60W / 68%

Average power consumption WBM / RO ST. / ROHQ is

                     33.17W                  38.5W                          53.33W
                    (100%)                 (~116%)                       (~161%)

(in $$$ that's change money; at 53.3W/h and 0.12$/kWh you can run a device like that 24/7/365 for $56)

Note: as with any statistic, one has to be careful with details and numbers, otherwise the results can be used to prove anything. My average power consumption above can be heavily distorted by what you play (mostly Blu-rays, mostly 1080/30i files, etc; hell, I don't have any SD files on my test). Second thing the tests were run in fullscreen windowed mode, not exclusive. I have a feeling exclusive mode is not so taxing.

But those are not the problems. Here's what's troubling me:
- there were dropped frames with RO HQ (madVR). Not so many with exclusive mode but still there were. Even as I didn't hit 100% on either the CPU or the GPU. Which highlights that shaders numbers are not everything and the rest of the internal architecture of the video card matters (in general, or memory speed in special). Which doesn't help one bit in figuring out where we draw the line - from this line of cards up it works smooth like butter; from it down is garbage. And I want that demarcation line.
- the other thing that's killing us is the deinterlacing. In software is of high-impact, in hardware it's the CUVID way or the highway. I don't like this one bit, above all because it's limiting, and I don't like any limitations to my toys, even if it's out of principle.

Other conclusions:
- the $$$ impact is not that much - as proven by Matt somewhere else, now the numbers extended to other architectures.
- the impact is from a different angle -> can you take a GTX 480 (which I'm sure with all its 1.5Gigs of 384bit DDR5 memory doesn't drop frames) and put it in a M350 Universal Mini-ITX?
- as said, a device that uses 50-something W/h is nothing but how many can you build like that when you have to put top dollar video cards in them? What if you have more than one system like that in the house? Would a user still run them all 24/7 as desktop builds? The difference between 35W to 60W it's just the same as 160W to 185W, but which one would you prefer to keep running all the time?

So because of all these, please imagine a blinking purple with bells (to stand out) request, for a DXVA with EVR automatic profile.

[MrC]

'Round these parts, we're in a tiered energy usage scenario, and we're always in the top tier at .40/kWh.  So we use that rate for luxury items calculations.  That's $187/year.

[Matt]

Using DXVA will be easy with a coming build:
NEW: Added Options > Video > Hardware accelerate video decoding when possible.  It will try LAV CUVID (requires nVidia card), and also DXVA (requires EVR renderer instead of madVR).

For now, the hardware acceleration option will be off by default.  LAV CUVID has been very stable for me, but I haven't tested DXVA much.  Once we gather enough feedback, we can reevaluate the best default.

[JustinChase]

Using DXVA will be easy with a coming build:

It doesn't look like it made it into .145; or am I just missing it?  I'm anxious to see if it runs okay on my slower/old CPU, faster GPU setup.

[BryanC]

Ignore. Wrong thread.