madVR - high quality video renderer (GPU assisted)

[leeperry]

Maybe you can retry tomorrow morning and compare 5 and 7 another time?
Well, I just tried again to roll 5 & 7 on many different kinds of movies as I wanted to ensure that it wasn't a bad case of teh placebo and 7 always looks noticeably fuzzier than 5 to me....which is a major bummer if you seriously plan on ditching 5 altogether :(

Actually I'm thinking about offering only 7 as the only option.

Basing this kind of most crucial decision on the experiments of half a dozen users here on d9 seems pretty drastic to me...there's no worse, there's no better........it all depends on the end-user taste and equipment. He's the one watching the movies, he's allowed to make his own decision. Again, better have too many options than not enough.

I kinda fail comprehending how you can reasonably expect a single dithering algorithm to make everyone equally happy, mostly because all displays run their own internal dithering(best case scenario being 12 bit to 8bit, otherwise plain 8bit or worse 6bit TN)....you cannot seriously expect all displays to react identically to the type of dithering noise/pattern that's being fed to them. And some ppl like a noisy picture, others like it a bit noisy when OTOH others cannot stand any kind of noise in their video...and let's not mention colorblinds, for whom I'm more than likely the only rep here.

My personal favorite would be 10
The "trade quality for performance" mVR tab currently has a section called "these should not be used, but you're the boss"...well, you don't say...I'm the boss of my PC and I like 5 far better than the others, I really don't see what the big deal would be to add another option for ED.

Can you imagine iZotope Ozone 5 Advanced (http://www.izotope.com/products/audio/ozone/features/dither.asp) coming with one single dithering algorithm?
Ozone provides the finest wordlength reduction algorithms available

You prefer 10, fine! You're the boss too :D

I'm sure there are some ppl who would prefer 2, especially on a noisy display....I'm quite sure I would hate 7 on a DLP tbh.

Anyway this wasn't a waste of my time as I'm extremely impressed by the PQ I'm getting with 5(Floyd–Steinberg FTW) and if I have to stick with this beta(hopefully not timebombed ^^) then be it, one less thing to worry about I guess :o

[peplegal]

I need something very simple and fast. It doesn't have to be encryption quality.

:goodpost:

My HD6450 really appreciates every single action you take of keeping this project lightweight.

Thank you.

[Shiandow]

Why would it shift the bits of randomHelper? It is 18000, not 0x18000. This is the well-known multiply-with-carry algorithm. It's not perfect, but it should work reasonably well. Give it a try. You might be surprised by how well it works. The main limitation comes from using only 16bit math. It would be better to use 32bit math. But unlike OpenCL, DirectCompute doesn't have a "mul_hi" instruction. So I don't know how to get the upper 32bits of the 64bit multiplication result, when multiplying two 32bit numbers.


The result of your multiplication needs more than 32bit, but your code doesn't reflect that. That's probably a simplification of the algorithm which could hurt the quality of the random numbers.

The algorithm I gave is supposed to only use the last 32bits of the answer of the multiplication, I assumed this would be done automatic. From what I understand from the explanation of multiply with carry on wikipedia, your algorithm should work as well but for some reason it doesn't seem to. It might be that multiply with carry performs badly on some seeds, but it seems more likely that the seeds aren't different for some reason.

Edit: by the way you could get the upper 32 bits of the product of two 32 bit numbers x,y by calculating ((x >> 16)*(y & 0xffff) + (x & 0xffff)*(y >>16)) >> 16 + (x >> 16)*(y >> 16) but this might be too slow.

[fairchild]

where is the link to the latest ED dither pack so I can check out this version 7 to test on my plasma. ;)

I tried searching but there were a few updates posted by madshi and not sure which has the pack in it.

Edit: found it http://madshi.net/madVRedtest.rar

[Shiandow]

@fairchild

You may also want to try http://madshi.net/madVRedtest3.rar.

[cyberbeing]

I don't like the categorizing into random and non-random. That's a technicality. The key difference should be looks, not technique. So please split the algorithms into "low noise" and "medium noise".

Random vs non-random is more than a technicality. All of the non-random ones are essentially zero noise, and their end-to-end seamless patterns is their defining look. Which behaves best in the non-random category comes down to which one has the fewest artifacts. Personally I dislike the look of all of the non-randoms, but I believe that ED2 is the most well-behaved out of them.

Into which category belong 9/10? Low-noise or medium noise? My eyes say low-noise. Do you disagree? If you agree with me, then it's unfair to expect 9/10 to be as artifact-free as the algorithms in the medium noise group.

From what I can see, ED3 has lower noise than ED7, but a little bit more noise than ED9/10. Again, please don't group algorithms for randomness, but please group them for looks alone.

ED9 & ED10, fall into the same noise level category as ED3. These are the ones which I'd consider the half-way points between low & medium noise. Personally since I dislike the appearance of the completely noiseless non-randoms, I'd classify them in the 'low-noise' category. ED7 & ED5 would then sit alone in the medium-noise category.

I personally find the appearance of ED9 superior to ED10. They both have near-identical pattern distribution and noise-level, but ED10 produces more artifacts in my tests.

I need to think about end users. They won't like one option to be called "error diffusion with some random" and "error diffusion without random". That won't help them choose. They want "error diffusion, low noise, mild artifacts" and "error diffusion, medium noise, no artifacts" as options. So that's what we need to classify the algorithms in.

If I was really going to narrow it down, my top 3 combinations for a medium noise and low noise category would be the following from best to worse:

Medium Noise: ED7 | Low Noise: ED3

Medium Noise: ED7 | Low Noise: ED9

Medium Noise: ED7 | Low Noise: ED2

Edit: Though if leeperry, who is colorblind, really finds ED7 more fuzzy than ED5, I'd be okay with swapping that in instead.

[kerimcem]

my old video card 3650 ati dont work all test version,How to shut down d.c option?

[fairchild]

I just tried ED7 and ED10 and I honestly can't tell much of a difference even when looking at a very close distance 4-6 feet. I will leave it to madshi and the other's which seem to have a better grasp on things when it comes to dithering methods. ;)

I will just be happy if the version that ends up being picked is a low noise version which is what will end up looking best on my plasma which inherently does dithering at different levels. Out of curiousity, where does the standard random dithering which has been used by madVR for years until it finally switched to the new ED version fares as far as noise level? Is it considered low-noise?

[Werewolfy]

Out of curiousity, where does the standard random dithering which has been used by madVR for years until it finally switched to the new ED version fares as far as noise level? Is it considered low-noise?

Very high noise level ;)


So it seems build 7 is a favorite of many people. So I'm wondering: Does anybody still need a lower-noise algorithm than 7? Or maybe I could use build 7 as the only error diffusion algorithm?

If a lower noise algorithm is still in demand, there doesn't seem to be any consensus about that yet. Earlier many people liked build 2. But, personally, I think build 10 is better than 2 (same noise level, less patterns).

Thoughts?

My favorite one is 7 too and I don't mind if you only use this one because I don't like low noise algorithm, it seems there is more banding when there is less noise and the picture appears to be less sharp. But if I had to pick one between 2 and 10, I'd choose 10.

[noee]

madshi, I (and my 7 year old with amazing eyes and ears) prefer 10 and then 7 in my viewing scenario. I tested (for what seems now like days on my older 1080p IPS panel) with *only* SD and 1080p24 material (no test patterns at all), so take it fwiw. I now get what leeperry and others say about "pop effect" now that I have spent so much time rolling these things.

[kasper93]

I take a look at the screenshots posted in this thread and I think Build 7 looks best. I personally would go only with that option. No need for lower noise option it's already low... But if you really want it build 10 and 9 is fine. (In fact I can't see the difference between those, but again I'm looking at the screenshots)

[XMonarchY]

Bah, no matter what content I try - I keep on coming back to DirectCompute testbuild 3 because it just makes the picture really sharp, even though I own an LCD, which produces a sharp image anyway. I also get ramp banding, so it would seem I could benefit from something like #7, but I don't like it one bit...

All these super-contrast zoomed-in tests are not very representative of the overall perceptions from HD video playback. Even when it comes to image doubling and etc. - sharpness outweighs smoothness.

[har3inger]

I have a problem with error diffusion and nnedi3 on ATI 13.12, 8870m and intel 4000 switchable graphics, win 8.0.

Anything related to OpenCL (nnedi, the old ED) caused my computer to immediately bsod. With the new test builds, this is still the case.

Now, with direct compute ED, I get a black screen with audio when I turn it on. This is similar to what happens when I check "use separate device for presentation". For now, I have no idea why.

Running opencl or directcompute benchmarks shows that both are at least working on the 8870, even though GPU-z says that opencl, and not directcompute is installed for the discreet gpu. (I'm guessing GPU-z is just wrong here).

Are there any tips for how I should begin troubleshooting the problem? I could live without nnedi since I usually watch 1080p content at native, but the error diffusion is intriguing, especially since I have so much horsepower at hand.

[Asmodian]

All these super-contrast zoomed-in tests are not very representative of the overall perceptions from HD video playback. Even when it comes to image doubling and etc. - sharpness outweighs smoothness.

I really like the sharpness of the new NNEDI3 options but I cannot say that I agree sharpness outweighs smoothness all the time. I like all the ED builds that offer lowish noise - and I have a lot of trouble telling them apart with "real" content even from two feet away (27" 2560x1440 glossy IPS).

Are there any tips for how I should begin troubleshooting the problem? I could live without nnedi since I usually watch 1080p content at native, but the error diffusion is intriguing, especially since I have so much horsepower at hand.

I do not believe NNEDI will be usable on a HD8870M (a 25% declocked HD7770) even if you do get it to work, while not a bad card it isn't what I would call fast.

It sounds like directcompute isn't working correctly on your system and that is what is used by ED dither now.

[huhn]

I really like the sharpness of the new NNEDI3 options but I cannot say that I agree sharpness outweighs smoothness all the time. I like all the ED builds that offer lowish noise - and I have a lot of trouble telling them apart with "real" content even from two feet away (27" 2560x1440 glossy IPS).

i just tried DC 3 and took a screen with random dither, ED and no dither
the difference between no dither and dither is totally obvious because the black level is risen and less bending when dithering is used but i can't see a real difference between ED and random dithering on a real source.

and i looked at it from 20-60 cm on a 42° calibrated lcd tv screen.

ed dither compare: http://screenshotcomparison.com/comparison/61913

screen without dither with really deep black: http://abload.de/img/nodithershin5gkij.png

is the content to see this so special or my eyes/tv are terrible bad... ?

[turbojet]

huhn: I don't notice much either between ED and RD. ED is a little sharper and a little less noise but it's very costly. It's only usable for me (650) on a few things so I don't even use it. A very long profile could rule out 60p, 1080i to ivtc, <960p (nnedi3) which leaves maybe 10% of videos, figured it wasn't worth the time to make the profile.

This isn't saying its bad, just its price/performance ratio is very high with current gpus.

[cyberbeing]

ed dither compare: http://screenshotcomparison.com/comparison/61913
is the content to see this so special or my eyes/tv are terrible bad... ?


Not a great comparison, since your source shows signs of being dithered with error diffusion before encoding, most likely to 10bit. The most obvious effects which can be seen by madVR dithering are on those left three archways. With madVR error diffusion they appear smoother with a checkerboard pattern. With madVR random dither they are noisy.

[huhn]

Not a great comparison, since your source shows signs of being dithered with error diffusion before encoding, most likely to 10bit. The most obvious effects which can be seen by madVR dithering are on those left three archways. With madVR error diffusion they appear smoother with a checkerboard pattern. With madVR random dither they are noisy.
maybe... but from normal viewing conditions not really? i have to say it's not zero but nearly nothing.
this is a 8 bit untouched blu ray, so this is what we get from the disc.

of cause i don't know what the studio did to it.

[cyberbeing]

maybe... but from normal viewing conditions not really? i have to say it's not zero but nearly nothing.
this is a 8 bit untouched blu ray, so this is what we get from the disc.

of cause i don't know what the studio did to it.

Ah, okay. Well it either had to be have an insane bitrate with 8bit (like a Blu-ray) or with 10bit (re-encoded with filtering) to retain dithering patterns after encoding like that.

Honestly, in your example I prefer 'random dither'. As you'll find from me earlier in the thread, I was the first to express displeasure with the quality of that original Direct Compute Build 3 you used, and didn't think it was worth the performance cost. madshi listened and now we have things like the ED7 build which are quite nice. Either way the entire point of dithering is to do its job, but be subtle enough to be undetectable. Error diffusion algorithms also have the benefit of offering more accurate representations of the source. The current debates back and forth are nitpicking objective differences through exaggerated examples as well as subjective quality on various display setups.

Like many settings in madVR, it's about squeezing another 1% of quality out of a source at orders of magnitude higher performance costs. Add up enough of those 1% quality increases, and you may begin to notice significant differences compared to other video renderers. The availability of such settings are what makes madVR unique, but if they are worth it is up to each individual person to decide. In most cases GPU power is the limiting factor here.

[ryrynz]

Find these builds aren't working on my HD3000, the video just launches paused and when I close MPC it crashes. 0.84.4 and The Directcompute 3 build work fine. This is with ED even set to random.

[madshi]

On 0.87.4:
Playing SD content at 200% (720x406 -> 1440x812)

Case 1: Using NNEDI3 color upscaling
* observed result: video plays, but missing chroma (green screen).

Case 2: Using NNEDI3 image doubling
* observed result: video freezes, 0% GPU usage, GPU down-clocks into fallback mode

setup: Quadro K2000M (~660M)
driver version 320.27 WHQL/ 331.82 WHQL

(So probably the same D3D9 interop issue, but downgrading drivers doesn't solve the issue for those GPUs)
Don't know why older drivers don't work for you. I guess we can only hope for a driver fix from NVidia.

You prefer 10, fine! You're the boss too :D
No, I don't prefer 10 over 7. I prefer 10 over 2 and 4, if we have to have a low-noise alternative. I like 7 and 5 most.

The algorithm I gave is supposed to only use the last 32bits of the answer of the multiplication
You're right, thanks, I misinterpreted the way the algorithm work. Will switch to your suggested algorithm.

I still believe the algorithm I was using was alright, though, although limited to only 16bit, so not as good as yours.

Random vs non-random is more than a technicality. All of the non-random ones are essentially zero noise, and their end-to-end seamless patterns is their defining look.
I agree so far.

ED9 & ED10, fall into the same noise level category as ED3.
I don't agree with this. From all I can see, ED9&10 have similar noise levels to ED2, and lower noise levels than ED3. This is not visible in all test images, though. If you look at only one test image, ED3 can sometimes appear to have noise levels as low as the non-random algorithms, but on some gray levels it's higher. From all the test images I've seen, ED9&10 have very similar noise levels to ED2. So IMHO ED9&10 are valid competitors to ED2, while ED3 has slightly higher noise levels (but lower than ED5&7).

If I was really going to narrow it down, my top 3 combinations for a medium noise and low noise category would be the following from best to worse:

Medium Noise: ED7 | Low Noise: ED3

Medium Noise: ED7 | Low Noise: ED9

Medium Noise: ED7 | Low Noise: ED2
I can agree with this. However, ED3 is almost identical to the old test build 5 which many users rejected, due to having higher noise levels than the old test build 3. I fear ED3 is not a valid choice for the Low Noise algorithm cause the noise level is still a bit too high, compared to ED2&9&10.

My favorite one is 7 too and I don't mind if you only use this one because I don't like low noise algorithm, it seems there is more banding when there is less noise and the picture appears to be less sharp. But if I had to pick one between 2 and 10, I'd choose 10.
K, thx.

madshi, I (and my 7 year old with amazing eyes and ears) prefer 10 and then 7 in my viewing scenario. I tested (for what seems now like days on my older 1080p IPS panel) with *only* SD and 1080p24 material (no test patterns at all), so take it fwiw. I now get what leeperry and others say about "pop effect" now that I have spent so much time rolling these things.
Thanks, appreciated!

I take a look at the screenshots posted in this thread and I think Build 7 looks best. I personally would go only with that option. No need for lower noise option it's already low... But if you really want it build 10 and 9 is fine. (In fact I can't see the difference between those, but again I'm looking at the screenshots)
Thx.

Bah, no matter what content I try - I keep on coming back to DirectCompute testbuild 3
The old test build 3 is almost identical to the new build 4. I believe the new build 2 is an improvement over both. And I believe the new builds 9&10 are another improvement over the new build 2. Of course that's only my personal opinion, but the majority of users seems to agree with that.

I have a problem with error diffusion and nnedi3 on ATI 13.12, 8870m and intel 4000 switchable graphics, win 8.0.

Anything related to OpenCL (nnedi, the old ED) caused my computer to immediately bsod.
Ouch.

Now, with direct compute ED, I get a black screen with audio when I turn it on. This is similar to what happens when I check "use separate device for presentation". For now, I have no idea why.

Running opencl or directcompute benchmarks shows that both are at least working on the 8870, even though GPU-z says that opencl, and not directcompute is installed for the discreet gpu. (I'm guessing GPU-z is just wrong here).

Are there any tips for how I should begin troubleshooting the problem? I could live without nnedi since I usually watch 1080p content at native, but the error diffusion is intriguing, especially since I have so much horsepower at hand.
I've got no idea, to be honest. My recommendation would be to fully uninstall &clean the AMD drivers, and then reinstall the latest drivers, just to be safe. Other than that I don't know what to suggest. Maybe a full Windows reinstall, but something like that is very time consuming, and there's no guarantee that it will work afterwards...

-------

I'm sorry about this, but could we please do one final test for the higher noise algorithm? I've implemented Shiandow's improved random number algorithm. And instead of just builds 5 and 7 there are now 2 variants of both:

- Floyd-Steinberg, 1.0 weight sum (old build 5)
- Floyd-Steinberg, 0.97 weight sum
- Filter-Lite, 1.0 weight sum
- Filter-Lite, 0.97 weight sum (old build 7)

Which of those 4 do you prefer? It's perfectly alright to say that they all look the same to you. If that's what your eyes tell you, please do say that. The number of "looks the same" votes is helpful, too. Personally, I don't see much difference between those 4. I'm sure leeperry will pick one algorithm to be the best, of course... :D If the rest of you don't care, I've no problem letting him choose. But if there are multiple votes with clear favorites, I'll go with the majority vote.

http://madshi.net/madVRednoisy.rar

[cyberbeing]

I don't agree with this. From all I can see, ED9&10 have similar noise levels to ED2, and lower noise levels than ED3. This is not visible in all test images, though. If you look at only one test image, ED3 can sometimes appear to have noise levels as low as the non-random algorithms, but on some gray levels it's higher. From all the test images I've seen, ED9&10 have very similar noise levels to ED2. So IMHO ED9&10 are valid competitors to ED2, while ED3 has slightly higher noise levels (but lower than ED5&7).
I probably should have been a bit more clear. ED3 does have more noise than ED9 & ED10, but the type of noise, error, and patterns it produces are in the same category. The non-random ED2 is in a different category, and ED5 & ED7 are together in the same category. You can see what I mean by this if you boost the saturation way up on a gray bars pattern.

Edit: And yes, I realize that earlier you said you disliked classifying them like this, but hopefully these images exhibit why I do so...
__
ED2 (Noiseless Patterns)
http://i.imgbox.com/ctIBzTnP.png

ED3 | ED9 | ED10 (Random Patterns)
http://i.imgbox.com/53E2xzCm.png http://i.imgbox.com/76TCqBaW.png http://i.imgbox.com/ymIaXtws.png

ED5 | ED7 (Patternless)
http://i.imgbox.com/6ODKJRSj.png http://i.imgbox.com/x51CP9vw.png

I can agree with this. However, ED3 is almost identical to the old test build 5 which many users rejected, due to having higher noise levels than the old test build 3. I fear ED3 is not a valid choice for the Low Noise algorithm cause the noise level is still a bit too high, compared to ED2&9&10.

Yeah I know, though it seems through your blind testing a few more people began to like ED3 (old test build 5). Overall I do agree that it would probably make sense to go with something even lower noise than ED3 for the low noise option. You should probably weight your choice for the low noise option from the opinions of those who preferred the old directcompute test build 3, which was a noiseless non-random build. It's ultimately up to them how much randomness & noise they find acceptable.

[Shiandow]

You're right, thanks, I misinterpreted the way the algorithm work. Will switch to your suggested algorithm.

I still believe the algorithm I was using was alright, though, although limited to only 16bit, so not as good as yours.

After reading up on Multiply-with-carry random number generators I also believe that your original algorithm should be alright. The main advantages of the algorithm I provided is that it has a slightly larger period (4294967296 instead of 589823999) and it also works if the seed is 0.

But I now no longer have an explanation for why the algorithm you used made the output periodic. It would be nice if it's solved now but I have no idea why.

[leeperry]

The old test build 3 is almost identical to the new build 4.
Interesting, I rolled 4 and 5 quite extensively but ultimately 5 won by quite a long run.

I'm sure leeperry will pick one algorithm to be the best, of course... :D If the rest of you don't care, I've no problem letting him choose. But if there are multiple votes with clear favorites, I'll go with the majority vote.
Ah, lemme see :D

I might just do it the russian way and bribe everybody to sing along, you guys take paypal? :devil:

[Ver Greeneyes]

@madshi, I was wondering about some things regarding dithering and monitor calibration.

One of the things that caught my eye when this discussion first started is that 6233638 reported being able to better differentiate between patches of similar brightness, and I was wondering if this might improve my dispcal calibration near black. With that in mind, I have the following questions:
1) My calibrator takes the average of a group of pixels, so it probably won't care about worm patterns - but if the pattern is the same every frame, could the average vary between one location and the next?
2) For the purposes of differentiating between patches, you'd want the least amount of randomness possible right? But I'd also probably want to measure under the same conditions as I use for viewing later on. Assuming we end up with two algorithms to choose from in the end, should I calibrate with the one I prefer or with the one that has the lowest amount of noise? (assuming they're not one and the same)
3) Very near black, is it possible that the dithering will be clipped (causing a higher average)? If so, would this be worse for an algorithm with more randomness?

Note that these questions are for the purposes of calibration, not viewing. I want calibration to leave my black level as low as possible while still accurately adjusting dark shades of grey (relative to the black and white level).

[leeperry]

BTW, would anyone know about a dead silent graphic card that's got the muscles for serious NNEDI goodness?

I got this o/c 7850 (http://www.vtx3d.com/products_features.asp?id=160) for fairly cheap but its thermal paste is prolly junk as heat instantly raises, it's no DirectCU and its fans get incredibly loud >20%...I guess I could replace the paste and slap an 800rpm 120mm onto the card but then I could kiss the 2 years warranty goodbye and it gave me the dreaded 7850 black screen on XP a few times.

Ideally I'd like to get an Asus DirectCU 660 (http://www.techpowerup.com/reviews/ASUS/GeForce_GTX_660_Direct_Cu_II/32.html), but apparently nvidia missed the NNEDI train and I especially don't want a Sapphire. I guess a DIY kludge is in good order.

[James Freeman]

I probably said it already, but any opinion is a good opinion.

I can see clear difference between Random Dithering vs Error Diffusion.
I can't see the difference between any of the ED versions.

*I'm on a 27" HD Monitor, maybe on projector or larger sized TV's the difference is more noticeable from closer range.
*Calibrated monitor with i1 Display Pro,


What will be chosen is fine by me.
Thanks.

[6233638]

One of the things that caught my eye when this discussion first started is that 6233638 reported being able to better differentiate between patches of similar brightness, and I was wondering if this might improve my dispcal calibration near black.Sorry, this was an error on my part.

[YxP]

BTW, would anyone know about a dead silent graphic card that's got the muscles for serious NNEDI goodness?


I got a basic 1GB model of 7850 and it can't do image doubling, so for "serious" NNEDI you probably need to go higher than that card.

[cyberbeing]

- Floyd-Steinberg, 1.0 weight sum (old build 5)
- Floyd-Steinberg, 0.97 weight sum
- Filter-Lite, 1.0 weight sum
- Filter-Lite, 0.97 weight sum (old build 7)

Which of those 4 do you prefer? It's perfectly alright to say that they all look the same to you. If that's what your eyes tell you, please do say that.

That new random algorithm changed the behavior of these quite a bit.

My initial impressions via test patterns:

First place "madVR - noisy 2". (This build appears to have the most uniform noise distribution)

Second Place "madVR ED7"

Third Place "madVR ED5"

Forth place "madVR - noisy 4.

Fifth place "madVR - noisy 3"

Sixth place "madVR - noisy 1"


I'll need to do some more tests later, but so far "noisy 2" seems to be my preferred choice.

[Werewolfy]

BTW, would anyone know about a dead silent graphic card that's got the muscles for serious NNEDI goodness?

I got this o/c 7850 (http://www.vtx3d.com/products_features.asp?id=160) for fairly cheap but its thermal paste is prolly junk as heat instantly raises, it's no DirectCU and its fans get incredibly loud >20%...I guess I could replace the paste and slap an 800rpm 120mm onto the card but then I could kiss the 2 years warranty goodbye and it gave me the dreaded 7850 black screen on XP a few times.

Ideally I'd like to get an Asus DirectCU 660 (http://www.techpowerup.com/reviews/ASUS/GeForce_GTX_660_Direct_Cu_II/32.html), but apparently nvidia missed the NNEDI train and I especially don't want a Sapphire. I guess a DIY kludge is in good order.

I have a Asus DirectCU 660. It's very quiet most of the time and even with NNEDI the fans don't make too much noise. I can use NNEDI64 for luma channel on SD 24 fps movies and NNEDI32 on 720p 24 fps movies.

[leeperry]

I got a basic 1GB model of 7850 and it can't do image doubling, so for "serious" NNEDI you probably need to go higher than that card.
Mine is factory overclocked to 1Ghz and it's got all the muscles I need tbh, it's almost as fast as a 7870 FWIR.

I have a Asus DirectCU 660. It's very quiet most of the time and even with NNEDI the fans don't make too much noise. I can use NNEDI64 for luma channel on SD 24 fps movies and NNEDI32 on 720p 24 fps movies.

Yes, I briefly owned this card (https://www.asus.com/Graphics_Cards/GTX660DC2OCPH2GD5/) and it was so amazingly quiet, every review raves about this and it comes with a 3y warranty too: http://www.techpowerup.com/reviews/ASUS/GeForce_GTX_660_Direct_Cu_II/32.html
extremely quiet in both idle and load

I don't understand why Asus can't get AMD GPU's to quiet down under load: http://www.techpowerup.com/reviews/ASUS/R9_270X_Direct_Cu_II_TOP/31.html
Noisy during gaming

The NNEDI OpenCL benchmarks make rather clear that AMD leads the way IIRC, and madshi seemed to agree.

[YxP]

Mine is factory overclocked to 1Ghz and it's got all the muscles I need tbh, it's almost as fast as a 7870 FWIR.

Ahh, my bad, I mistakenly read this as if you were thinking about buying that card :)

[QBhd]

My DirectCU II TOP 270x is very quiet... not sure why they would say noisy, even their own results don't warrant a Noisy statement... it's not the quietest on the list, but 38 dB under full load is far from noisy.

QB

[tschi]

I have a Asus DirectCU 660. It's very quiet most of the time and even with NNEDI the fans don't make too much noise. I can use NNEDI64 for luma channel on SD 24 fps movies and NNEDI32 on 720p 24 fps movies.
Seriously ?
What is yours rendering times and yours settings ?
I have a amd R9 280x (7970 gpu) and I can't use NEEDI for luma on 720p movies (24fps and other), I think my rendering times are around 50ms (only needi for luma).
if you refers to benchmark on SEt http://forum.doom9.org/showthread.php?t=169766 The AMD cards are faster than nvidia cards with opencl needi. I think with madvr opencl impletementation the result is not so different.
Maybe, we can make a benchmark of GPU with the same video and same settings, in order to clarify what 's really possible to do in realtime with madVR and the differents graphic cards.

[Werewolfy]

Chroma upscaling : Jinc3 AR
Image downscaling : Catmull-Rom AR LL
Image doubling : NNEDI3 32 neurons for luma
Image upscaling : Jinc3 AR
Debanding : low

My rendering times with 720p 24fps movies : 31.59ms without Error Diffusion and 36.10ms with Error Diffusion (DC3).

[TheLion]

Sorry to interrupt the very interesting ongoing dithering discussion, but I hope somebody can help me with the following issue (all 0.87.4 builds).

I am using a AMD 5870 board with latest 13.12 drivers. To get best possible Blu-Ray playback I wanted to experiment with the new NNEDI3 chroma upsampling option (no other scaling involved as output is native 1080p).
Looking at some chroma test pics I realized that no-matter the neuron setting it is actually always JINC3 been used. So I can select NNEDI, it doesn't give me any alert message, but internally seems to fallback to JINC3.
Is this to be expected with the 5870 (openCL support)? Would it be possible to give feedback to the user that NNEDI is in fact disabled and something else used?

Thanks!

[Procrastinating]

Madshi, have you looked into this? Someone was developing a dithering algorithm mased on magic numbers rather than "pure random" numbers. The intended application is slightly different, but I figured it could be taken into account.

It could be faster to create psuedorandom outputs simply from a list of "random-looking" numbers rather than random-number generation, though I'm just throwing ideas around here.

http://pippin.gimp.org/a_dither/

[6233638]

I don't understand why Asus can't get AMD GPU's to quiet down under load: http://www.techpowerup.com/reviews/ASUS/R9_270X_Direct_Cu_II_TOP/31.htmlAMD's current GPUs run very hot compared to Nvidia's.

I would recommend buying one of MSI's Twin Frozr IV cards - they seem to be the quietest of all the current generation GPU coolers.

[pie1394]

Seriously ?
What is yours rendering times and yours settings ?
I have a amd R9 280x (7970 gpu) and I can't use NEEDI for luma on 720p movies (24fps and other), I think my rendering times are around 50ms (only needi for luma).


There could be something wrong with your system. Your card is a rebranded HD7970GHz, which is 10% faster than regular HD7970. A regular HD7970 can do it with less than 32ms timing for 720p24 + following settings to 1920x1080p24 output:

- Chroma upscaling NNEDI3_16
- Image Doubling: Luma - NNEDI3_32 , Chroma - NNEDI3_16
- Image Upscaling: Jinc3AR
- Image Downscaling: Spline3AR_LL
- Debanding_with_angle_detection.



if you refers to benchmark on SEt http://forum.doom9.org/showthread.php?t=169766 The AMD cards are faster than nvidia cards with opencl needi. I think with madvr opencl impletementation the result is not so different.
Maybe, we can make a benchmark of GPU with the same video and same settings, in order to clarify what 's really possible to do in realtime with madVR and the differents graphic cards.

I think GCN could be much utilized with complex OpenCL code due to its architecture design. It is more compiler-optimization-friendly to schedule/combine different thread task jobs if compared to previous VLIW5 / VIEW4 processor's.

But I am curious why DirectCompute5 HLSL compiler results in 10x slower code than OpenCL one of madshi's NNEDI3 shader function. It is quite difficult to imagine that unless any of following reason:

- A lot of DRAM re-access per thread group's calculation processing.
- some NNEDI3 heavy math calculations are not directly translated to processor's FPU instruction -- but by emulation
- 8 UAV count per cycle restriction has very big impact.

Compiler -O0 vs -O2/-O3 does not often produce such big difference even on heavy pure math calculation code.

[Vyral]

Sorry to interrupt the very interesting ongoing dithering discussion, but I hope somebody can help me with the following issue (all 0.87.4 builds).

I am using a AMD 5870 board with latest 13.12 drivers. To get best possible Blu-Ray playback I wanted to experiment with the new NNEDI3 chroma upsampling option (no other scaling involved as output is native 1080p).
Looking at some chroma test pics I realized that no-matter the neuron setting it is actually always JINC3 been used. So I can select NNEDI, it doesn't give me any alert message, but internally seems to fallback to JINC3.
Is this to be expected with the 5870 (openCL support)? Would it be possible to give feedback to the user that NNEDI is in fact disabled and something else used?

Thanks!
AMD Radeon HD 5870 does support OpenCL.

Try the following to confirm the problem :
1. select JINC3 in the settings
2. play a video and show the OSD (Ctrl + J)
3. Note the rendering time
4. stop playback and close the media player
5. select NNEDI in the settings
6. repeat step 2 and 3 with the same video

The rendering time should be higher with NNEDI than Jinc3.

Still, I think it would be a good idea to show in the OSD which chroma upsampling and image upscaling/downscaling algorithms are being used.

[cyberbeing]

I'll need to do some more tests later, but so far "noisy 2" seems to be my preferred choice.

Did some more tests on real content as well as a couple gradient debanding tests, and came to the same conclusion. "madVR - noisy 2" is superior.

But I now no longer have an explanation for why the algorithm you used made the output periodic. It would be nice if it's solved now but I have no idea why.

On solid colors, I'm seeing still seeing periodic repetitions exactly every 256 pixels, just like the previous builds. It unfortunately does not appear this issue was resolved with the new random number method.

[Qaq]

The rendering time should be higher with NNEDI than Jinc3.
With my 7750 I have same rendering time, but present queue goes to zero.
BTW, as far as I remember 87.1b (or was it 87.3b?) gave me NNEDI chroma w/o drops, but only for 16.

[6233638]

Please don't concentrate on "near black" too much. The key thing is overall looks. We'll have to look for a solution for your 3dlut black level problem elsewhere. This should not be the deciding factor for choosing the error diffusion algorithm.Well I meant how values near black looked (e.g. shades in the 0-5% range) not how noisy black itself was, which is another issue.

But I think I will defer to Cyberbeing's judgement - he seems to be doing much more thorough testing of the dither patterns than I have had the time to, and while in my limited testing #2 looked to be the least noisy without many patterns showing up, he seems to be able to find situations where the patterns become very obvious.

Considering that any of these are significantly less noisy than random dithering, I'll go with whatever he deems to be the best choice.

[leeperry]

ED5 | ED7 (Patternless)
http://i1.someimage.com/Rslmaai.png http://i1.someimage.com/Z4dEIMX.png
Very nice, I guess that would explain why 7 looks more "popping" than 5...it's like the good ole days when I was playing around with Grainfactory3() or how noisy SmoothLevels()'s ED looks, with the right noise settings you can actually process some sort of noise/grain-based EE. I guess that's why very noisy movies such as 21 grams look so "deep".

Anyway, I've always agreed that 7 gives the best depth impression and I'm getting used to it now :)

I guess the subjective pop effect is what this hobby is all about(Samsung have mastered it with their DNIE thingie), so I'll also +1 for 7. No more e-drama on my end.

I would recommend buying one of MSI's Twin Frozr IV cards - they seem to be the quietest of all the current generation GPU coolers.
I read a lot of bad feedback about their twin frozr 660's, if they can't cool down a 660 quietly then my hopes are low...and to me MSI is much like Sapphire but fair enough, maybe they managed to improve.

Oh well, I don't currently need anything faster than a 1Ghz 7850 and I'm not buying a +150€ card for silence. My 7850 is prolly worth 70€ on the used market, either I add +80€ for silence or I lose 70€ if the 7850 goes belly up(due to the black screen problem for instance).

I've recently seen 3GB 7950's selling for 120€ but their fans looked pretty low grade and these were Sapphire...I guess the days when you had to buy a TOTL GPU fan and install it yourself for silence aren't quite history. I still got a Zalman VF1000 that's serious business with a 120mm slapped onto it, I wonder if it'd fit.

[TheLion]

AMD Radeon HD 5870 does support OpenCL.

Try the following to confirm the problem :
1. select JINC3 in the settings
2. play a video and show the OSD (Ctrl + J)
3. Note the rendering time
4. stop playback and close the media player
5. select NNEDI in the settings
6. repeat step 2 and 3 with the same video

The rendering time should be higher with NNEDI than Jinc3.

Still, I think it would be a good idea to show in the OSD which chroma upsampling and image upscaling/downscaling algorithms are being used.

Thanks. Same thing - With Jinc3 I get about 8.4ms rendering time, when I switch to ANY of the NNEDI options it stays the same. Switching in general works - eg. Jinc8 results in 41ms.

I agree about the OSD - on the other hand it shouldn't get overloaded.

[6233638]

Very nice, I guess that would explain why 7 looks more "popping" than 5...
Actually, when I look at them again, 5 seems to have a much more defined edge than 7.
7 seems to blur the edges together a bit.

As the goal of dithering is not to deband the image, that seems worse.

I read a lot of bad feedback about their twin frozr 660's, if they can't cool down a 660 quietly then my hopes are low...and to me MSI is much like Sapphire but fair enough, maybe they managed to improve.I don't know if the 660 is using the latest Twin Frozr IV design. (probably not?)

You mentioned ASUS' DCUII 270X which is 38dBA under load (http://www.techpowerup.com/reviews/ASUS/R9_270X_Direct_Cu_II_TOP/25.html), compared to MSI's which is only 27dBA under load (http://www.techpowerup.com/reviews/MSI/R9_270X_Gaming/23.html).
Same situation with the 700 series cards from Nvidia.

[TheLion]

After switching back and forth a couple of times - always while looking at the rendering times - I think there may be a bug.

What I did is starting a bdmv - only scaling is chroma upsampling. Switching between all other chroma upscaling options is fine and works as expected (longer rendering with more complex algorithms and more taps. When switching to one of the NNEDI the behavior is strange and pretty random.

E.g. I start a video with NNEDI64 - it shows 8.4ms (same as Jinc3). I switch to NNEDI16 - suddenly rendering time increases to 31ms. I press stop. Restart the video - now rendering time (average) again goes back to 8.4ms. Switching to any of the NNEDI neuron options doesn't change that - it stays at 8.4ms even with 256. But switching to any other algorithm works as expected again - eg. Jinc 3->4 goes from 8.4ms to 34.5ms.

I am not sure what's going on - but I guess chroma upsampling falls back to Jinc3 when selecting NNEDI in my setup - the strange thing is that "sometimes" rendering times suggest that NNEDI works after switching to it. Till the next restart.

I am using Windowed Mode. Full Screen Exclusive shows the same behavior.

[GREG1292]

Very nice, I guess that would explain why 7 looks more "popping" than 5...it's like the good ole days when I was playing around with Grainfactory3() or how noisy SmoothLevels()'s ED looks, with the right noise settings you can actually process some sort of noise/grain-based EE. I guess that's why very noisy movies such as 21 grams look so "deep".

Anyway, I've always agreed that 7 gives the best depth impression and I'm getting used to it now :)

I guess the subjective pop effect is what this hobby is all about(Samsung have mastered it with their DNIE thingie), so I'll also +1 for 7. No more e-drama on my end.


I read a lot of bad feedback about their twin frozr 660's, if they can't cool down a 660 quietly then my hopes are low...and to me MSI is much like Sapphire but fair enough, maybe they managed to improve.

Oh well, I don't currently need anything faster than a 1Ghz 7850 and I'm not buying a +150&euro; card for silence. My 7850 is prolly worth 70&euro; on the used market, either I add +80&euro; for silence or I lose 70&euro; if the 7850 goes belly up(due to the black screen problem for instance).

I've recently seen 3GB 7950's selling for 120&euro; but their fans looked pretty low grade and these were Sapphire...I guess the days when you had to buy a TOTL GPU fan and install it yourself for silence aren't quite history. I still got a Zalman VF1000 that's serious business with a 120mm slapped onto it, I wonder if it'd fit.


Last night looked at 1-10 all over again with a fresh install of 8.1 windows and 7 keeps coming up as the clear favorite on my system. Also
then compared 7 with a sapphire 7870xt to the gtx-770 and preferred
the GTX-770. Guess I will be sticking
with 7 and Nvidia. Maxwell is coming
soon so who knows what is in store
For MadVR in the future. Great work
everyone PQ is amazing.


Sent from my iPhone using Tapatalk

[Shiandow]

On solid colors, I'm seeing still seeing periodic repetitions exactly every 256 pixels, just like the previous builds. It unfortunately does not appear this issue was resolved with the new random number method.

That's what I'm seeing as well. So since we've now completely ruled out the possibility that the algorithm was wrong I guess that the problem must lie with the seeds that are used by the algorithm.

For some reason only the last 8 bits of the position of the pixel are used as seed, this is made worse by the fact that the last 4 bits are always the same... If madshi can find out why this happens then it should be trivial to solve, if not then I suppose it's worth a try to just divide the position by 17 (or some similar number) this should at least help a bit.