madVR - high quality video renderer (GPU assisted)

[vivan]

Originally Posted by James Freeman

A little early to tell but I guess that the new "Image Doubling" algorithm is more optimized than NNEDI3 (maybe even GPU optimized)

http://forum.doom9.org/showthread.php?t=169766

[6233638]

Originally Posted by vivan

http://forum.doom9.org/showthread.php?t=169766

I wish Nvidia took OpenCL performance more seriously. I don't want my next card to be an AMD one if I can help it.

Does anyone know if a Titan would perform better in this task than a 780? I know that Nvidia restricts compute performance on their "gaming class" cards.

[Vyral]

> debanding
> NNEDI3 upscale
> upscaling profiles
> release before the weekend
Aaand I have my plans for the weekend.

[iSunrise]

Originally Posted by madshi

Did you notice the "Add profile" and "Delete profile group" buttons? The whole profile logic is totally flexible. Those 4 SD and HD profiles were just example profiles I set up for testing.

Yes, I´ve noticed them and now that you say it, I think I understand now.

Ok, so what I gather from that screenshot is that you can basically setup any kind of profile "name" (which you can choose freely) and have that "referenced" in the profile rule set, with some pre-defined rules (based upon tags, is the complete madVR tagging rule set allowed?) that are fully customizable?

That would be extremely powerful, indeed. And I can´t wait to test it!

Originally Posted by 6233638

I wish Nvidia took OpenCL performance more seriously. I don't want my next card to be an AMD one if I can help it.

Does anyone know if a Titan would perform better in this task than a 780? I know that Nvidia restricts compute performance on their "gaming class" cards.

Nvidia clearly wants to become a niche player with certain aspects that make them irresistible to make money. With CUDA, they follow exactly that path, yet they miss out on deals that will rely on OpenCL more and more and increasing developer-support. It saddens me, too, because it basically means that we buy (and are also willing to buy to a certain extend) powerful GPUs, whose precious silicon is going to complete waste, since they don´t want anyone else "outside" to touch it.

Regarding Titan, it should only have a lead on double-precision performance. But since I am not a Titan or 780 owner myself, I can just go by the very few OpenCL benchmarks that are currently available on the web and my understanding of the Kepler (GK110 in this case) architecture.

[huhn]

amd is currently better than nvidia if your really need or want a nvidia card you should wait for maxwell.

[mindz]

AMD > NVidia. No thanks for the detailed reason .

[leeperry]

Maybe I missed something but I don't think I ever read madshi stating publicly that the NNEDI double scaling would be OpenCL accelerated.

[huhn]

it's opencl for sure just read: http://forum.doom9.org/showthread.php?t=169766 it never was a secret. but i wonder if it's done in 16 bit but why guessing.

edit: he even stated that the new dithering mode is opencl too

[leeperry]

Oh, spot on! I didn't find the bit about dithering and CTRL+F on "dither" in this thread doesn't give any result but OpenCL accelerated Floyd-Steinberg would be too awesome for words

I guess it'll soon take a $400 GPU if you wanna have everything maxed out in mVR

[huhn]

http://forum.doom9.org/showpost.php?p=1663079&postcount=14

and a lot more infos on about 3d on avs forum but what's the point of this X-)

[DragonQ]

Originally Posted by huhn

it's opencl for sure just read: http://forum.doom9.org/showthread.php?t=169766 it never was a secret. but i wonder if it's done in 16 bit but why guessing.

edit: he even stated that the new dithering mode is opencl too

1.85 fps on a GT 430? Well that's me out then.

[mandarinka]

Originally Posted by iSunrise

Regarding Titan, it should only have a lead on double-precision performance. But since I am not a Titan or 780 owner myself, I can just go by the very few OpenCL benchmarks that are currently available on the web and my understanding of the Kepler (GK110 in this case) architecture.

Yes, only the double-precision performance should be crippled on GeFroce 780/780Ti, and that should not really matter. Most OpenCL tasks probably run on 32-bit math, including this, because with double precision the performance would be too low - basically everything besides Tahiti, Hawaii, GK110 would be unusable.

Cape Verde/Pitcairn/Bonaire GPUs iirc only have 1/16th performance in double precision, while GK104 (and lower) give 1/24th IIRC.

[ikarad]

Originally Posted by Vyral

> NNEDI3 upscale
.

What is nnedi3 upscale? Is to replace jinc,lanczos and similar?

Originally Posted by DarkSpace

If image doubling does what I think it does, will it be possible to, say, perform image doubling from 480p to 960p and then perform rest of the scaling to 1080p using the normal upscaling algorithms?

Is it a new version of smooth motion?

[Mangix]

Originally Posted by mandarinka

Yes, only the double-precision performance should be crippled on GeFroce 780/780Ti, and that should not really matter. Most OpenCL tasks probably run on 32-bit math, including this, because with double precision the performance would be too low - basically everything besides Tahiti, Hawaii, GK110 would be unusable.

Cape Verde/Pitcairn/Bonaire GPUs iirc only have 1/16th performance in double precision, while GK104 (and lower) give 1/24th IIRC.

In general nvidia hardware does very well with 64-bit operations whereas AMD does very well with 32-bit ones.

AMD also does better on ints than nvidia. This is why oclHashcat is faster on 32-bit hashes with amd hardware. 64-bit hashes are faster on nvidia hardware.

oclHashcat also has the nvidia kernels compiled(it is not open source) using CUDA and NOT OpenCL. There was a section in the wiki where the author explained why this was but I cannot find it. If memory serves me correctly, he mentioned driver issues, a few CUDA-only features, and general stability.

He wanted to switch to OpenCL for nVidia but the cross-compiler seems to produce slower code.

[6233638]

Originally Posted by ikarad

What is nnedi3 upscale? Is to replace jinc,lanczos and similar?

Yes, it should be a higher quality option than Jinc scaling.

Originally Posted by ikarad

Is it a new version of smooth motion?

No, it would seem that "image doubling" is going to be an option to use a high quality algorithm to scale the image 2x, and another to scale that to your display resolution.

nnedi3 is a very demanding algorithm, so few systems are probably going to be able to use it.
However, it might be possible for them to use nnedi3 to scale 2x (which should be faster) and then use another algorithm to scale the resulting image to your display resolution, while still benefiting from some of nnedi3's quality improvements.

[ikarad]

Originally Posted by 6233638

Yes, it should be a higher quality option than Jinc scaling.

No, it would seem that "image doubling" is going to be an option to use a high quality algorithm to scale the image 2x, and another to scale that to your display resolution.

nnedi3 is a very demanding algorithm, so few systems are probably going to be able to use it.
However, it might be possible for them to use nnedi3 to scale 2x (which should be faster) and then use another algorithm to scale the resulting image to your display resolution, while still benefiting from some of nnedi3's quality improvements.

Thanks.

If I well understand, image doubling make upscale (multiply by 2 the resolution of image) and downscale (at the display resolution) after, doesn't it?

nnedi3 is a very demanding algorithm, so few systems are probably going to be able to use it.

Do you think that a i5 3570k at 4.5ghz with a gtx780 is more than enough?

[6233638]

Originally Posted by ikarad

If I well understand, image doubling make upscale (multiply by 2 the resolution of image) and downscale (at the display resolution) after, doesn't it?

More likely something like 360p > 720p (nnedi3) > 1080p (lanczos), or 480p > 960p (needi3) > 1080p (lanczos)

It would probably be faster to go from 720p > 1080p with nnedi3 directly, than 720p > 1440p (nnedi3) > 1080p (catmull-rom)

Originally Posted by ikarad

Do you think that a i5 3570k at 4.5ghz with a gtx780 is more than enough?

While AMD seems to be about twice as quick as Nvidia cards (I'm concerned that my GTX 570 may not cut it) I would hope that a 780 is fast enough to handle it.

We'll see soon enough. I'd rather not speculate.

[vivan]

Originally Posted by 6233638

It would probably be faster to go from 720p > 1080p with nnedi3 directly, than 720p > 1440p (nnedi3) > 1080p (catmull-rom)

but nnedi3 can upscale only by factor of 2x. Or not?

[truexfan81]

Originally Posted by vivan

but nnedi3 can upscale only by factor of 2x. Or not?

nnedi3 only upscales by powers of 2 (ie, 2x, 4x, 8x...).

[6233638]

Oh - I was not aware of that. I assumed it was a performance-saving feature, not a requirement.