madVR - high quality video renderer (GPU assisted)

[robl45]

Probably just missing a colon - scale to 16:9.

Yes, I assumed everyone would know without the colon as I was typing from my phone.

I'm looking for whatever option will just stretch everything to fit my 16:9 screen. Unfortunately stretch to 16:9 stretches everything including actual 16:9 films.

[Damien147]

When I drag the mouse on the seekbar the seekbar blinks,I restored settings but nothing,any idea?


off topic:I've seen some people here using blueskyfrc.Any idea which one is smoothest?BlueskyFrc in mode 2 or svp 4 free?

[ashlar42]

Which basic performance item of an GPU is more important for madVR: Texel Rate, Pixel Rate, Memory bandwidth or Single precision performance?

I would like to know if there's enough reason to upgrade my laptop, from my current 880M to the current generation of 1060 for laptop, the two GPU have their own pros and cons in the aforementioned 4 basic performance test items, 1060 outperforms 880M in 3 of them except for Texel Rate.I would say performance rated in Gflops. But someone more knowledgeable than me will surely give a more precise answer soon.

[huhn]

Yes, I assumed everyone would know without the colon as I was typing from my phone.

I'm looking for whatever option will just stretch everything to fit my 16:9 screen. Unfortunately stretch to 16:9 stretches everything including actual 16:9 films.

if you stretch a 16/9 movie to 16/9 nothing bad happens.

[ryrynz]

I would say performance rated in Gflops. But someone more knowledgeable than me will surely give a more precise answer soon.

I've never seen any in depth comparison between cards besides what was posted recently vs the 1060 and 480. What's been deduced so far is this

Shader performance determines 99% of madVR performance

and this

If you want to know which GPU to get exactly, my advice has always been the same: Get the faster GPU you can afford and that fits your thermal requirements.

[Warner306]

I've never seen any in depth comparison between cards besides what was posted recently vs the 1060 and 480. What's been deduced so far is this



and this

How is shader performance measured? What technical spec outlines this?

[robl45]

if you stretch a 16/9 movie to 16/9 nothing bad happens.

You'd think it wouldn't, but the scale to 16:9 option that is in mpc-be and mpc-hc scales the 16:9 movie off the screen.

[Asmodian]

How is shader performance measured? What technical spec outlines this?

FLOPS, and I agree that it seems to indicate madVR performance quite accurately.

[CarlosCaco]

FLOPS, and I agree that it seems to indicate madVR performance quite accurately.

A little offtopic but maybe you guys can help me... I'm planning to buy a new gpu
Seens the best based on price/performance here in Brasil is a GTX 1060 (I have a GTX 570)
So my questions is:

I m on a i5 2500 and really don t use for game, only for madVR, will cpu bottleneck this card using madVR?

There two versions that have a good price a standard pny version with the clocks from factory and 6gb memory and 1280 Cuda cores and msi gaming x 3gb memory with higher clocks in OC and Gaming mode but with 1152 Cuda cores
On game debate the 3gb shows 96% shader performance vs 100% of 6gb version

In yours opinions what is the best bang for madVR: msi 3gb less memory less Cuda cores but higher clocks or pny standard 6gb more Cuda cores but standar clocks only...

[huhn]

for madVr only.

if the card with higher Gflops is still the better card for madVR like in the past than the RX 470/480 4G version with around 5-6 gflops.

if nnedi3 plays a huge rule. i can recommend anything the openCL performance is hard to judge...

for UHD i would avoid the 3g 1060 version. 4Gb should be totally fine.

[ashlar42]

FLOPS, and I agree that it seems to indicate madVR performance quite accurately.Ok, so my idea was correct.

If this is the case, there's a huge gap between 1050Ti and 1060.
2.1 Tflops vs 4.4.

RX480 should offer 5.8 Tflops. Is this confirmed by real testing (it being faster than GTX1060 in madVR)?

Note: the above values are calculated from specifications, according to Tech Report.

Edit: I searched for the comparison spreadsheet that was posted here, based on real testing, and no, GTX480 does not seem to confirm the theoretical advantage that it has based on specs.

[huhn]

the first test show it was not the case with polaris.

i just got my RX 480 back and it can easily do nnedi3 64 neuron 24p FHD-> UHD.
but that is mostly openCL.

just wait for the new algo i'm pretty sure people with different hardware will test it.

[Shiandow]

FLOPS, and I agree that it seems to indicate madVR performance quite accurately.

Keep in mind that FLOPS can only do so much to make things faster; running shaders always has some overhead (reading / writing memory etc.) and for modern GPUs this overhead seems to dominate the computation time for all but the most computation intensive shaders (e.g. NNEDI3).

Case in point, a while back I upgraded my GTX 560Ti to a GTX 960, and while the GTX 960 is definitely faster the difference in FLOPS is far greater than the difference in render times.

[Asmodian]

Keep in mind that FLOPS can only do so much to make things faster; running shaders always has some overhead (reading / writing memory etc.) and for modern GPUs this overhead seems to dominate the computation time for all but the most computation intensive shaders (e.g. NNEDI3).

Case in point, a while back I upgraded my GTX 560Ti to a GTX 960, and while the GTX 960 is definitely faster the difference in FLOPS is far greater than the difference in render times.

I have noticed this too. All the easy shader passes take 0.49ms on my Titan X and this is independent of clock speed and resolution. This means an easy six pass shader, like thin edges, is the same 3 ms on my Titan X as it was on my 980Ti while NNEDI3 256 is much faster on my Titan X.

More FLOPS = Faster in madVR, it isn't 1 to 1 but other aspects don't do much to help madVR at all.

[Shiandow]

I have noticed this too. All the easy shader passes take 0.49ms on my Titan X and this is independent of clock speed and resolution. This means an easy six pass shader, like thin edges, is the same 3 ms on my Titan X as it was on my 980Ti while NNEDI3 256 is much faster on my Titan X.

Interesting... the 0.5ms per shader pass seems to be pretty universal, I've had similar results on my GTX 560Ti and GTX 960. Not sure what's causing it, it doesn't seem to be just memory bandwidth as reading in more pixels doesn't slow things down much (in a rather extreme case I've had a shader read 1920 pixels per output pixel, that definitely slowed things down but it still worked). I think the number of output pixels has a bigger effect but I haven't checked it thoroughly.

At any rate if it's 0.5ms for every GPU then there's indeed no point in trying to make it faster, so FLOPS are pretty much the only benchmark that will have some effect. Although 2x more FLOPS won't necessarily mean 2x faster rendering.

[AngelGraves13]

VRAM plays a role too when upscaling. My 1080 GTX is a huge upgrade over my 980 GTX.

NNEDI3 at 256 for Chroma still isn't possible, but maybe one day...

[Asmodian]

VRAM plays a role too when upscaling. My 1080 GTX is a huge upgrade over my 980 GTX.

NNEDI3 at 256 for Chroma still isn't possible, but maybe one day...

Do you mean VRAM speed or amount? I haven't noticed a change in performance due to VRAM speed myself, and I never need more than 3GB of VRAM even with 4K and large buffer sizes.

Are you sure the improvement isn't simply due to the large increase in FLOPS (4981 v.s. 8873)?

edit: Why would you want NNEDI3 256 for chroma? I can do it but I do not.

[huhn]

with a GPU buffer of 16 i get more than 3Gb VRAM usages.

[Asmodian]

with a GPU buffer of 16 i get more than 3Gb VRAM usages.

Relatively larger buffers then, or maybe simply not small buffers, the default of 8 stays under 3GB. :p

The 980 has 4GB of VRAM anyway, I don't see how the extra 4GB of VRAM on the 1080 helps.

[AngelGraves13]

Do you mean VRAM speed or amount? I haven't noticed a change in performance due to VRAM speed myself, and I never need more than 3GB of VRAM even with 4K and large buffer sizes.

Are you sure the improvement isn't simply due to the large increase in FLOPS (4981 v.s. 8873)?

edit: Why would you want NNEDI3 256 for chroma? I can do it but I do not.

I'd say VRAM usage. Doubt speed matters much, if at all. It's just better to have more buffer, especially for 4K video, though it will be a few years until we can play back 4K Ultra HD on PC, assuming the copy protection will ever be cracked :(

I have 128 for CPU, 24 for GPU and 16 for Present.

NNEDI3 128 I can do, but 256 isn't possible on a 1080 GTX. I use Jinc w/ SuperRes 4 for Chroma and Upscaling.

[huhn]

Relatively larger buffers then, or maybe simply not small buffers, the default of 8 stays under 3GB. :p

The 980 has 4GB of VRAM anyway, I don't see how the extra 4GB of VRAM on the 1080 helps.

even with GPU buffer of 8 you get close to 3 GB Vram usage.
of cause 8 GB doesn't help but 3 GB is still very limited for UHD. with larger buffer size you simply have no chance with 3GB.

[Betroz]

With 720p, 23.976fps content upscaled to 1440p, and then down to 1080p on my 1080p TV I can use the following settings in MadVr :

- Chroma upscaling : NNEDI3 64
- image downscaling : SSIM2D 100%, AR, LL
- image doubling : double Luma NNEDI3 256 + double Chroma NNEDI3 32 (quadruple off)
- image upscaling : Jinc AR
- upscaling refinement : only SuperRes 1X (2X is possible, but a bit higher render times)

This gives me a rendertime of 32-35ms. 480p content is upscaled with Jinc AR from 960p -> 1080p (after NNEDI3 settings has done it's job up to 960).

Or I can use these settings with the same content :

- Chroma upscaling : NNEDI3 256
- image downscaling : SSIM2D 100%, AR, LL
- image doubling : double Luma NNEDI3 64 + double Chroma NNEDI3 64 (quadruple off)
- image upscaling : Jinc AR
- upscaling refinement : only SuperRes 1X (2X is possible, but a bit higher render times)

Very little difference in IQ between the two setups, although I like NNEDI3 256 for Chroma Upscaling with native 1080p content on my 1080p TV that is not using any image upscaling/doubling or SuperRes refinement.

[ryrynz]

I like NNEDI3 256 for Chroma Upscaling

Spot the difference (http://screenshotcomparison.com/comparison/188367). :sly:

[Betroz]

Spot the difference (http://screenshotcomparison.com/comparison/188367). :sly:

Hehe, I can't ;)
I use the first settings from my previous post anyway (the one with NNEDI3 256 for Luma).

[ryrynz]

Hehe, I can't ;)


:)

I'd almost suggest removing 128 and 256 neuron NNEDI3 for chroma but I don't think that'll happen, IMO it's best to avoid them.

[Shiandow]

Seems I may need to hurry along my latest experiment (http://i.imgur.com/M2wklSI.jpg) a bit.

[Ver Greeneyes]

Seems I may need to hurry along my latest experiment (http://i.imgur.com/M2wklSI.jpg) a bit.
Compared to NNEDI3, seems to sharpen the sleeve and instrument of the guy on the left a little more, but at the expense of some aliasing (or clipping) that makes it look a little rough. Had to switch between the images several times to even notice though! (I somehow noticed the aliasing before the increase in sharpness)

[aufkrawall]

even with GPU buffer of 8 you get close to 3 GB Vram usage.
of cause 8 GB doesn't help but 3 GB is still very limited for UHD. with larger buffer size you simply have no chance with 3GB.
Considering that I'm getting presentation glitches with default queues, which I don't get with longer queues, this is a real dealbreaker for 3GB cards.

[Shiandow]

Compared to NNEDI3, seems to sharpen the sleeve and instrument of the guy on the left a little more, but at the expense of some aliasing (or clipping) that makes it look a little rough. Had to switch between the images several times to even notice though! (I somehow noticed the aliasing before the increase in sharpness)

True, although softer images almost automatically look less aliased. And compared to NNEDI3 it's quite a bit faster. Still somewhat of a work in progress though.

[sauma144]

True, although softer images almost automatically look less aliased. And compared to NNEDI3 it's quite a bit faster. Still somewhat of a work in progress though.

Is it SSimSuperRes, its successor or a totally new algo?
What's the performance difference between NNEDI3 (64n) and your algo?

[Shiandow]

Is it SSimSuperRes, its successor or a totally new algo?
What's the performance difference between NNEDI3 (64n) and your algo?

For a large part it's completely new, but it's somewhat similar to the Bilateral chroma scaler. On my GTX 960 it takes about 2~3 times as long as bicubic, or about an additional 0.75ms. Still need to optimise it some more though. At any rate it should be significantly less performance intensive than NNEDI3.

[CarlosCaco]

For a large part it's completely new, but it's somewhat similar to the Bilateral chroma scaler. On my GTX 960 it takes about 2~3 times as long as bicubic, or about an additional 0.75ms. Still need to optimise it some more though. At any rate it should be significantly less performance intensive than NNEDI3.

Is this the new algo madshi talked about?

[sauma144]

For a large part it's completely new, but it's somewhat similar to the Bilateral chroma scaler. On my GTX 960 it takes about 2~3 times as long as bicubic, or about an additional 0.75ms. Still need to optimise it some more though. At any rate it should be significantly less performance intensive than NNEDI3.

So it is a chroma only scaler?
I hope I am wrong but it seems that the MPDN development is in standby.
Your previous algorithms/shaders were initially made for MPDN.
Do you still create them for MPDN?
Is there an hidden option in madVR to use external shaders?

[Ver Greeneyes]

True, although softer images almost automatically look less aliased.Yeah, and I do think the sharpening looks good (though I can't judge whether it's accurate to the original). It's mostly the topmost fold of cloth on the sleeves of both men that looks a bit jagged where NNEDI3 makes them look like smooth gradients. If you could fix that I think it would improve upon NNEDI3 in that image, assuming the sharpness is 'correct' (for some sense of the word).

[Shiandow]

Is this the new algo madshi talked about?

Probably not.

So it is a chroma only scaler?
[...]
Do you still create them for MPDN?
Is there an hidden option in madVR to use external shaders?

Yeah, it's meant just for the chroma. It's basically another attempt at trying to recover chroma information from luma. So far it seems to be more successful than other methods I tried, although very small details can be a bit problematic. To avoid going too far off topic I won't go into depth on the current state of MPDN but I still use it to try out new algorithms. No hidden options for injecting shaders in MadVR, unfortunately.

[adhara]

Hi all,

This is about RGB output levels with a 10 bits display (so dithering 10bits and more activated in MadVr).
I know this is not a new question but I would like a clear explanation as we can read everything and its opposite arround this subject.

Basically there are 3 options:

Setup 1 :
GPU outputs 0-255
MadVr outputs 0-255
Display expects 0-255

Setup 2:
GPU outputs 16-235
MadVr outputs 16-235
Display expects 16-235

Setup 3:
GPU outputs 0-255
MadVr outputs 16-235
Display expects 16-235

Problems:
-There is no way on my UHD (10 bits) TV to select limited or full range.
- Graphic Card GTX 1070: When output is set to 2160p(4k) - RGB Full range, bitdepth is limited to 8 bits / To have 10/12 bits outptut on the GPU, I have to lower outptu level to limited range (16-235).

So what is the best setting for my own use (4k) ?

Madvr (0-255 / 10 bits) --> GPU (0-255 / 8 bits) --> Display

OR

Madvr (16-235 / 10 bits) --> GPU (16-235 / 10 bits) --> Display

OR

Madvr (0-255 / 10 bits) --> GPU (16-235 / 10 bits) --> Display

OR (other ?)

Thanks

[ryrynz]

Seems I may need to hurry along my latest experiment (http://i.imgur.com/M2wklSI.jpg) a bit.

I think it looks too sharp. NNEDI3 isn't the one to beat anyway, it's super-xbr AR. Considering you seem to like working with bilateral for chroma I can only assume it has benefits but I haven't seen anything that makes me consider it yet. I've wondered if it would be an interesting idea to use this or something similar as a second pass for chroma on top of super-xbr and averaging the result.

[Shiandow]

I think it looks too sharp. NNEDI3 isn't the one to beat anyway, it's super-xbr AR. Considering you seem to like working with bilateral for chroma I can only assume it has benefits but I haven't seen anything that makes me consider it yet. I've wondered if it would be an interesting idea to use this or something similar as a second pass for chroma on top of super-xbr and averaging the result.

Blending the images would probably get you the worst of both worlds. Anyway, here (http://screenshotcomparison.com/comparison/188509) is a rather extreme example that shows why I like 'bilateral' methods. Sure in most cases the difference is more subtle, but it's the cases where it's not that matter. The difference also becomes more noticeable if you use upscaling. By the way, if I recall correctly Chroma-SuperXbr also uses some luma information, just not as extensively.

[ryrynz]

here (http://screenshotcomparison.com/comparison/188509) is a rather extreme example that shows why I like 'bilateral' methods.

Those jaggies tho.. :(

[Shiandow]

Those jaggies tho.. :(

Yeah. most are the result of not using the full luma information. As far as I can tell Superxbr isn't doing much better. I managed to beef things up a bit and get this (http://i.imgur.com/7Ak18WH.jpg) which is somewhat better, but I need to check if I can raise the strength that much without making other images worse.

[Stereodude]

FLOPS, and I agree that it seems to indicate madVR performance quite accurately.
I don't think you can use FLOPS to compare across hardware generations.

[Asmodian]

I don't think you can use FLOPS to compare across hardware generations.

I don't know for sure but it seems like you can, it is not super precise but significantly higher FLOPS gives higher performance. FLOPS seems to be a better indicator for madVR performance than for most other types of loads.

Like we were discussing earlier, FLOPS doesn't give a 1:1 performance boost, but for NNEDI3 the card with more FLOPS runs NNEDI3 faster.

[Georgel]

Spot the difference (http://screenshotcomparison.com/comparison/188367). :sly:

I can't see any difference either...

Yeah. most are the result of not using the full luma information. As far as I can tell Superxbr isn't doing much better. I managed to beef things up a bit and get this (http://i.imgur.com/7Ak18WH.jpg) which is somewhat better, but I need to check if I can raise the strength that much without making other images worse.

SuperXBR has a bit of jagged edges, but still, those results do show a bit too...

At that point NNEDI, bilateral or reconstruction or even Jinc look better. At least if you're alergic to jagged edges.


By the way, anyone knows if there is a big difference in image downscaling between SSIM 2D and 1D, or rather what is the difference?

[leeperry]

SSIM 2D and 1D (..) what is the difference?
jinc/bicubic

[har3inger]

I'm a big fan of bilateral with some superchromares, so I'm looking forward to your new scaler, shiandow. Generally, every other non-reconstruction chroma scaler will look thick and fuzzy due having half the resolution, but bilateral cleans things up nicely, as it basically works with built in line thinning. Artifacts are cleaned up a bit with SR at the cost of chroma cleanliness.

[Stereodude]

I don't know for sure but it seems like you can, it is not super precise but significantly higher FLOPS gives higher performance. FLOPS seems to be a better indicator for madVR performance than for most other types of loads.

Like we were discussing earlier, FLOPS doesn't give a 1:1 performance boost, but for NNEDI3 the card with more FLOPS runs NNEDI3 faster.
The data seems to suggest that's very much not true. The GTX 1060 outperforms the RX480 (just barely) in the madVR data posted despite having 33% less FLOPS. The same data shows the GTX 1060 outperforming the GTX 970 by >20% despite having only a 10% FLOPS advantage.

I wish we had data comparing an R9 card with basically the same FLOPS as the RX 480. I'm fairly confident it would show the RX 480 inferior despite comparable FLOPS.

[seiyafan]

Two years I had a 270x and GTX 770 and they were about the same in madVR even though the 770 had more FLOPS. So now it seems the opposite is true.

[huhn]

the 1060 has less gflops at the stock boost clock but every 1060 clocks way hjigher than the "max" boost clock that makes normal Gflop calculation pretty much worthless.

than we have openCL efficiency in the past AMD was way way better at this than nvidia even with copyback bug.

[madshi]

Sorry guys, my new algo needs some more tweaking time. Maybe 1-2 weeks more, I don't know. Quality is constantly improving, though. Performance seems ok, but will require a decent GPU.

Will be back to answer all open questions/comments when the new algo is completed.

[ryrynz]

Will be back to answer all open questions/comments when the new algo is completed.

At least this.. upscaler or post processing?