madVR - high quality video renderer (GPU assisted) [Archive] - Page 504

dvzzz

18th March 2014, 23:18

@madshi you are exactly right, when I switch interlacing in MadVR to Film, blacks go to the prominent deep black but if I switch to Video via the shortcut it goes gray. Of course in the film mode picture is unwatchable but blacks are black. Anything that I can do in the Intel Graphics Control Panel to address it? Thank you.

Or more probably, when activating DXVA deinterlacing, the Intel GPU driver might do damage to the video, screwing up the black levels. Although, madVR should really detect and fix that for you. Anyway, try disabling deinterlacing (there are keyboard shortcuts for disabling/enabling deinterlacing on the fly in madVR). Does switching deinterlacing fix/unfix the black level problem?

Razoola

19th March 2014, 05:44

I think if madVR was going to change which D3D version it was going to use it would have to happen after a final version is released. Much like what NVidia are going to be doing shortly with dropping support for many older GFX cards. Once that driver is released a MadVR version could be tied with that driver and then a new build of madVR could happen on DX11 only and the newer driver.

Of course there is a big part of me is against such a thing happening but if madshi wanted to take that route the above gives a good opportunity. Of course this does not take into account AMD or INTEL users.

cyberbeing

19th March 2014, 08:08

Didn't you say that also on AMD, doing the dithering via Compute can be much slower than with OpenCL?

You are probably thinking of when madshi stated NNEDI3 via DirectCompute was ~10x slower than NNEDI3 via OpenCL for reasons unknown.

Incriminated

19th March 2014, 12:03

Much like what NVidia are going to be doing shortly with dropping support for many older GFX cards.

You got it wrong. They announced only to no more provide optimisations for those card in the driver after version 340 past 1.april 2016...this would be version 343. Still this cards will receive fixes and supported through driver 340. So driver 340 will still receive fixes in stability if necessary after april 2016. We do not know when thes cards will be finally dropped. Also it is not "shortly" but around 2 years up in the future...talking about "OLD" cards on today.... that's very far ahead. It does not change anything within the next the 2 years. And what will change in 2 years is NOT a sudden drop of supporting these cards ....

Original Source: http://nvidia.custhelp.com/app/answers/detail/a_id/3473

annovif

19th March 2014, 12:16

I' m trying my powerful :D Ati 7750 with server 2012 evaluation copy and i can set nnedi3 chroma upsampling 32 neurons with blu ray:p
It's incredible because the gpu consumption is at 96%, but in exclusive mode i have not dropped frames. In windows 7 and 8 was not possible.

Incriminated

19th March 2014, 13:56

I wonder if that energy consumption is worth for watching movies compared with very good results jinc3+ar gives at already relative high gpu-usage.

Also i cannot get these NNEDI3 chroma-upsampling to work using DXVA copy back on LAV with my nvidia gtx670, all neuro-mode have weird colors.

seiyafan

19th March 2014, 14:24

I wonder if that energy consumption is worth for watching movies compared with very good results jinc3+ar gives at already relative high gpu-usage.

It's a lot cheaper than what I pay for a movie ticket plus gas.

QBhd

19th March 2014, 14:35

And some of us don't pay for electricity, so it's a use to the fullest possible situation :)

QB

Guest

19th March 2014, 14:45

Getting OT guys. Let's not start a debate on power consumption here. Thank you.

Razoola

19th March 2014, 15:15

You got it wrong. They announced only to no more provide optimisations for those card in the driver after version 340 past 1.april 2016...this would be version 343. Still this cards will receive fixes and supported through driver 340. So driver 340 will still receive fixes in stability if necessary after april 2016. We do not know when thes cards will be finally dropped. Also it is not "shortly" but around 2 years up in the future...talking about "OLD" cards on today.... that's very far ahead. It does not change anything within the next the 2 years. And what will change in 2 years is NOT a sudden drop of supporting these cards ....

Original Source: http://nvidia.custhelp.com/app/answers/detail/a_id/3473

I did not read it like that but you could be right. I took it to mean these cards will be in the 340 build tree only. Any new build tree after this will not have the cards in. So for example a 4.x tree will not have the cards in. If a need came for an update for older cards (security related for example) they would do a new build based on the 340 build tree only.

Nvidia have stopped providing optimisations for many of the older cards for a long long time already. I can't remember the last time I have seen a driver optimisation for a 200 series card for example.

Incriminated

19th March 2014, 15:39

Yeah, but you still can download older driver that still being updated from time to time if serious bug appears for i.e. since a new windows update or whatever, seperated from the main actual branch which receives gaming optimisations. So whatever "supporting" is defined you probably will be able get these cards running for a long time, Nvidia will continue provide functioning driver download for all this card.... it's NOT like that you need to keep up community mirrors to keep the driver-files alive, or at some OS changes your driver would be broke. The 340 is far is far in the future, it will be available and stability maintained past april 2016.

Imho it is impossible to optimise so many increasing pathes for different cards with different behaviour performing architectures, so nvidia is naturally enforced to focus on there actual branches. Nvidias range in what it is actually optimising for is higher than what the actually sell this and last year. They're still optimising for 400. I appreciate that. Also it does make sense to optimise actual games for say 9600GT... whatever the performance-increase would be 50%, it still would be a lagg-show... and when you disable all effects, then theres probably less gain. xD

6233638

19th March 2014, 16:23

Madshi, if I have tagged a file with [deint=video] should I expect deinterlacing to be forced on?
I'm having to manually enable deinterlacing for these files - though it does use the correct mode once I do.

dvzzz

19th March 2014, 16:59

With "when I play Blurays in 1080i29/30" do you mean playing back 1080i29/30 content, deinterlaced to e.g. 1080p59/60? Or do you mean outputting this as 1080i29/30 to the display? Is deinterlacing turned on in madVR when the problem occurred and which display mode / refresh rate are you using in that situation?

I think it could either be a different display mode behaving differently - *if* you're using a different display mode for such kind of content.

Or more probably, when activating DXVA deinterlacing, the Intel GPU driver might do damage to the video, screwing up the black levels. Although, madVR should really detect and fix that for you. Anyway, try disabling deinterlacing (there are keyboard shortcuts for disabling/enabling deinterlacing on the fly in madVR). Does switching deinterlacing fix/unfix the black level problem?

I did more testing last night:
a) installed 0.8.7.7
b) upgraded Intel drivers to March 2014 Beta
Exactly same behavior when Interlacing is disabled via shortcut blacks are black, if Interlacing is On blacks are gray.
Not sure what else I can do to fix it?

flashmozzg

19th March 2014, 17:50

I did more testing last night:
Not sure what else I can do to fix it?

If you do not work/know someone in Intel Graphics department probably nothing. (you can try find a way to report it, but it won't be fixed anytime soon).

kiccolsd

19th March 2014, 21:03

Sorry guys, anyone noticed that switching from 0.8.4 to 0.8.7 Jinc image upscale is more power hungry?? Before with my HD4000 I could upscale 720p with antiringing option enabled now I need to deselect to avoid frame drops.

Niyawa

19th March 2014, 22:01

I'm getting my new laptop tomorrow with an 880m and 8 gigs of ram (LOL). I'll be testing it out for sure.
Thank you. I'm looking forward to it.

I did an extensive test for 270x, 290x and 770 a while ago, check them out. :)

My conclusion is based on 1080 to 1440 upscaling, 270x is the minimum card to do NNEDI3 doubling at 16 neurons, 290x can do that at 32 neurons plus error diffusion. 770 falls in between the two but it's slower with error diffusion.

At the end I settled with a MSI 270x card because the fan speed did not change a bit from idle to hours of movie playing, it was extremely quiet.
I'll check that out, not sure where it is though. Interesting, it seems that NNEDI3 is a lot more power-hungry than I thought at first. I was expecting a GTX 750 Ti to give some respectable results, seems like I was mistaken. If a 770 can take care of it in a slower fashion, I'd expect the 290 to be the sweet spot between them. Slightly off-topic but for the record, MSI makes the most silent aftermarket cards, it's no-brainer someone with silence in mind would choose them. I'm more of a silent and efficiency guy, so a GTX variant (which is relatively more efficient when power-consumption is concerned) would be my pick.

Hum, either way. You said it's a 1080p > 1440p upscale, so maybe a 280x or 770 should do just fine for 720p > 1080p. I wonder if anyone here could test a 760 for a 720p > 1080p upscale for the sake of accuracy. Regardless, this data is very valuable. Thank you very much. Should you find any other interesting bits to tell me, please do since I'm always lurking in this thread.

If anyone else has anything to add, I'd appreciate it as well. The more, the merrier.

Also, quick question for everyone: where madVR stands when SLI/Crossfire is concerned?

madshi

19th March 2014, 23:10

Hi, just one question is it safe to use madvr with Intel HD4000 drivers 15.33.0.3496 (10.18.10.3496)?? Any known issue??
No known issues, but I don't know if anybody tested that driver version yet.

@madshi you are exactly right, when I switch interlacing in MadVR to Film, blacks go to the prominent deep black but if I switch to Video via the shortcut it goes gray. Of course in the film mode picture is unwatchable but blacks are black. Anything that I can do in the Intel Graphics Control Panel to address it?
There are lots of options in the control panel. Have you tried them all? Generally I recommend to disable all funny algorithms like dynamic contrast adjustments, skin tone "optimization" etc. I'm not an expert on Intel driver control panel options, so I don't know which options exactly you should switch. But I guess maybe one of those options might fix the problem you're seeing. It's a bit weird, though, because madVR should really automatically detect the problem and fix it by itself...

Madshi, if I have tagged a file with [deint=video] should I expect deinterlacing to be forced on?
I'm having to manually enable deinterlacing for these files - though it does use the correct mode once I do.
It would make sense to have madVR enable deinterlacing in that case. Please feel free to create a bug entry for that.

Sorry guys, anyone noticed that switching from 0.8.4 to 0.8.7 Jinc image upscale is more power hungry?? Before with my HD4000 I could upscale 720p with antiringing option enabled now I need to deselect to avoid frame drops.
Those version numbers don't exist.

seiyafan

19th March 2014, 23:15

where madVR stands when SLI/Crossfire is concerned?

I read it somewhere that multiple cards provide no benefits to madVR.

ryrynz

19th March 2014, 23:22

Before with my HD4000 I could upscale 720p with antiringing option enabled now I need to deselect to avoid frame drops.

The HD4000 is borderline being able to use Jinc anyway, depending on the content you will find it drops frames on occasion so it's not a good choice.
For this reason I don't enable Jinc on my HD4000, if you like Jinc you should upgrade your GPU. Personally I use Lanzcos 4 tap AR for Luma and Bicubic 75 AR for Chroma on it.

Asmodian

20th March 2014, 01:24

Also, quick question for everyone: where madVR stands when SLI/Crossfire is concerned?

MadVR will run on SLI/Crossfire systems but at least on my SLI Titans it is much slower with SLI enabled.

The difference is extreme, 35 ms vs 85 ms, SLI off vs SLI on.

Edit: I just did a quick retest with v.87.7 from the first post and I get much better results with SLI enabled then I used to. It is still a bit slower but not so extreme.

Watching 1080p source @ 1440p with 64 neuron chroma upscaling and Jinc 3 image scaling I see 28.2 ms rendering times with SLI off and 32.8 ms rendering times with SLI on. Driver 327.23 of course.

seiyafan

20th March 2014, 01:45

Watching 1080p source @ 1440p with 64 neuron chroma upscaling and Jinc 3 image scaling I see 29.8 ms rendering times with SLI off and 34.4 ms rendering times with SLI on. Driver 327.23 of course.

Is that with doubling disabled?

Asmodian

20th March 2014, 01:55

Yes, doubling 1080p luma with 32 neurons (even with lanczos chroma scaling) is a bit too hard for my Titan with 30fps content and turning other settings down to enable it looks worse to me.

I am also using ED2, no debanding, and no trade quality for performance.

seiyafan

20th March 2014, 01:59

I see, it's faster than the 290x I had.

Asmodian

20th March 2014, 02:06

Well the Titan has had its power throttling disabled (up to 350W) and is water cooled, so sort of an overclock though the clock settings are at stock (running at 1097 MHz for these tests).

Edit: I edited my post with corrected rendering times, the original times were with Jinc4 image scaling for some reason. :o

seiyafan

20th March 2014, 02:30

Delerue

20th March 2014, 03:22

Asmodian

20th March 2014, 04:40

Can Titan utilize it's strong double precision floating point in madVR?

I understand madVR does everything in single precision floating point (32 bit) so the double precision performance is probably meaningless. I don't know if madVR's OpenCL or DirectCompute code uses any double precision FP but I doubt it does. Video work really doesn't need the extra precision.

Mangix

20th March 2014, 05:20

I understand madVR does everything in single precision floating point (32 bit) so the double precision performance is probably meaningless. I don't know if madVR's OpenCL or DirectCompute code uses any double precision FP but I doubt it does. Video work really doesn't need the extra precision.

I thought madVR uses 16-bit arithmetic to do its processing.

In any case, AMD tends to have faster integer performance.

edit: after reading this again, I'm confused. In C, a float (single-precision) is 16-bit. A double (double-precision) is 32-bit.

Niyawa

20th March 2014, 05:40

I read it somewhere that multiple cards provide no benefits to madVR.

Watching 1080p source @ 1440p with 64 neuron chroma upscaling and Jinc 3 image scaling I see 28.2 ms rendering times with SLI off and 32.8 ms rendering times with SLI on. Driver 327.23 of course.
My knowledge with madVR in SLI/Crossfire always leaves me with 'tearing' or downright 'not working' issues. If that is in the past, then it's all good.

shimaflarex

20th March 2014, 07:52

edit: after reading this again, I'm confused. In C, a float (single-precision) is 16-bit. A double (double-precision) is 32-bit.
You are mistaken. Single-precision is 32 bits, double-precision is 64 bits.

leeperry

20th March 2014, 10:04

Well, I think this link already made the AMD/Titan GP-GPU situation rather clear (http://www.extremetech.com/computing/153467-amd-destroys-nvidia-bitcoin-mining):
AMD cards outperform their Nvidia counterparts [..] because the SHA-256 algorithm utilizes a 32-bit integer right rotate operation.[..] AMD GPUs can do this operation in a single step. Prior to the launch of the GTX Titan, Nvidia GPUs required three steps — two shifts and an add.

We say “prior to Titan,” because one of the features Nvidia introduced with Compute Capability 3.5 (only supported on the GTX Titan and the Tesla K20/K20X) is a funnel shifter. The funnel shifter can combine operations, shrinking the 3-cycle penalty Nvidia significantly.

290x was nice and much faster but it ran hot and fans were noisy
Throttle its fans until they almost reach TJ Max? All the silent AMD boards I've seen use this trick and I do too on my 7850......I cap its fans to 20% et voilà ))

nevcairiel

20th March 2014, 10:11

Well, I think this link already made the AMD/Titan GP-GPU situation rather clear (http://www.extremetech.com/computing/153467-amd-destroys-nvidia-bitcoin-mining):

Because of performance differences due to one instruction being present on one architecture but not the other? (and needing 3 instructions to replicate the same result?)

How does that make anything clear?
In fact, its extremely specific to this one algorithm compared there.

NVIDIAs real problem is that they have fewer but stronger computing units (SMX processors), while AMD has more but weaker units. For GP-GPU processing, its easier to utilize more units with smaller work loads, and harder to keep strong, but fewer units busy 100% of the time.
(Note: this changes a bit with Maxwell, as the new "SMM" units are internally structured differently to be used more efficiently in parallel, as long as you don't need FP64 processing)

dread

20th March 2014, 11:05

Mangix

20th March 2014, 11:18

Well, I think this link already made the AMD/Titan GP-GPU situation rather clear (http://www.extremetech.com/computing/153467-amd-destroys-nvidia-bitcoin-mining)

Rotate and bit select operations are highly specific to cryptography and not used much elsewhere. SSEx and AVXx for example do not accelerate either whereas XOP does(go figure).

And what nev said.

Asmodian

20th March 2014, 11:41

My knowledge with madVR in SLI/Crossfire always leaves me with 'tearing' or downright 'not working' issues. If that is in the past, then it's all good.

That is in the distant past. :)

I thought madVR uses 16-bit arithmetic to do its processing.

It does the math in 32 bit but stores the internal buffers in 16 bit as more is not needed.

Anatasia

20th March 2014, 13:20

Well, I dumped 290x and got a 270x, thought I was done. Tomorrow I am going to dump the 270x because I ordered a 280x, lol!

270x was nice and quiet but has limitations. 290x was nice and much faster but it ran hot and fans were noisy. Hope the 280x can strike a balance between the two. So I think this time I am really done.

Can Titan utilize it's strong double precision floating point in madVR?

I'm using 290x too and I admit that it's too noisy. Then I've replaced the default thermal unit and used this http://www.tomshardware.com/reviews/r9-290-accelero-xtreme-290,3671.html (IMHO, much quieter, cooler and dirty cheap). In my country it's almost the same price with gtx 680 and the performance in madVR is really good.

Ceremony

20th March 2014, 14:16

Throttle its fans until they almost reach TJ Max? All the silent AMD boards I've seen use this trick and I do too on my 7850......I cap its fans to 20% et voilà ))Do NOT do this on any card ever!

The solder points are more likely to deteriorate under high temperatures than under normal to warm temps.

if you simply try to keep it at TJ Max, i doubt you get more than 1 1/2 years out of that card!

Instead, if you dont need all that raw power and want a quieter version, underclock it and lower its voltage.

I did the same for my GTX480, now running at 500MHz (0.900V) instead of 700MHz (1.013V). Cooler, more efficient and plenty for games and madVR

6233638

20th March 2014, 15:30

Do NOT do this on any card ever!All the current high-end Nvidia cards target 80C and AMD cards target 95C.

They basically run as fast as the cooling system allows. (unless you're using water cooling, at which point voltage generally becomes the limit)

leeperry

20th March 2014, 15:45

How does that make anything clear?
In fact, its extremely specific to this one algorithm compared there.
OK fair enough, but even in non-BTC specific OpenCL benchmarks, it seems rather clear that AMD's have quite an edge when it comes to handling 32-bit integer: http://www.anandtech.com/show/7764/the-nvidia-geforce-gtx-750-ti-and-gtx-750-review-maxwell/21

Anyway, in mVR's case there's that interop lag that seems completely random....I can do 1080p 64x NNEDI on 720p25 but not 720p30 dammit, hopefully someone at AMD will pity us.

Do NOT do this on any card ever!

The solder points are more likely to deteriorate under high temperatures than under normal to warm temps.
The story is rather simple, many of the newest AMD boards shoot for slightly lower than TJ Max(95°C on the current chips FWIR): http://www.guru3d.com/articles_pages/msi_radeon_r9_290x_gaming_oc_review,10.html

My 1GHz 7850 (http://www.vtx3d.com/products_features.asp?id=160) comes with a pretty serious heatsink but the fans gets unbearably loud over 25%...and it goes up to +40% under load, cooling down the GPU to 50°C, what do I care? When I read reviewers claiming the board to be fairly silent they lose all credibility.

The board comes with a 2y warranty, with a 20% fan speed cap it barely reaches 75°C with mVR, if it ever dies I'll send it back and play dumb....not my fault if they majorly failed on the fan speed curve. The sales pitch of this board is "extremely freezing power", better deliver...especially when the Asus 660's are so amazingly silent.

nevcairiel

20th March 2014, 15:47

OK fair enough, but even in non-BTC specific OpenCL benchmarks, it seems rather clear that AMD's have quite an edge when it comes to handling 32-bit integer: http://www.anandtech.com/show/7764/the-nvidia-geforce-gtx-750-ti-and-gtx-750-review-maxwell/21

Too bad madVR doesn't use 32-bit integer then! :)

flashmozzg

20th March 2014, 16:55

Question to madshi: would it be too hard to implement sli/crossfire support for computations such as nnedi?

markanini

20th March 2014, 17:17

Do NOT do this on any card ever!

The solder points are more likely to deteriorate under high temperatures than under normal to warm temps.

if you simply try to keep it at TJ Max, i doubt you get more than 1 1/2 years out of that card!

EVGA will suffer massive losses due to the 3 year warranty, right? Consider me sceptical.

Guest

20th March 2014, 17:23

Please stay on topic in this thread. Thank you.

madshi

20th March 2014, 18:44

1- Open a DVD (disc or folder) with MPC-HC;
2- Drag a .srt file to MPC-HC window;
Result: Subtitles are blinking fast and the timing is wrong;

This problem doesn't happen with VMR9 Renderless, for example.

Can anyone confirm this?

Seven x64 with GTX 650 Ti @332.21 and MPC-HC (20/03/2014)
Which subtitle renderer are you using? Is smooth motion FRC on/off? If it's on, does the problem go away if you turn it off?

I understand madVR does everything in single precision floating point (32 bit) so the double precision performance is probably meaningless.
It does the math in 32 bit but stores the internal buffers in 16 bit as more is not needed.
Correct.

Question to madshi: would it be too hard to implement sli/crossfire support for computations such as nnedi?
It would be hard without being able to test it. I don't have an SLI/CrossFire setup here. SLI/CrossFire support is not planned for any time soon. I'm not doing anything specific to stop SLI/CrossFire from working, but I'm also not doing anything to make it work. Maybe I'll look into this at some point in the future, but not any time soon.

GTPVHD

20th March 2014, 20:01

http://blogs.nvidia.com/blog/2014/03/20/directx-12/

NVIDIA will support the DX12 API on all the DX11-class GPUs it has shipped; these belong to the Fermi, Kepler and Maxwell architectural families. With more than 50% market share (65% for discrete graphics) among DX11-based gaming systems, NVIDIA alone will provide game developers the majority of the potential installed base.

burfadel

20th March 2014, 20:12

Current AMD hardware fully supports DirectX 12 as well, Nvidia is blowing the smoke up their own a**e a bit there!

In reality it's probably AMD that is at the greater advantage since DirectX 12 will also be on the Xbox One (AMD hardware).

MistahBonzai

20th March 2014, 20:22

I'm having a problem getting XySubFilter to load with MadVR when ffdshow raw video filter is enabled. Works fine without ffdshow in the stream. My ffdshow raw video filter option is as follows:

Nothing selected but avisynth.

Same behaviour even with avisynth disabled - just ffdshow raw video filter enabled in MPC-HC external filters.

Entering XySubFilter as an external filter and setting to preferred has no effect.

Again XySubFilter works fine (Shows as XySubFilter (=> madVr) under filters) with ffdshow disabled in external filters.

This is most likely not an MadVR issue but I'm outta leads having googled extensively and combing through the XySubFilter open issues list. Anyone? TIA :)

madshi

20th March 2014, 20:39