TXAA Anti-Aliasing Makes Its Debut In Latest Update For The Secret World

[BallaTheFeared]

This statement really is quite laughable. Perhaps you should see the stills that Sirpauly has of TXAA at 44 fps @ 1080p, and thats a still - combat performance is signifigantly lower still with TXAA. Or you can just believe what you want. Whatever pal. I'm done with you.

Actually it's your inablity to actually understand your hardware that is causing the issues here.

Pauly is seeing a 15% performance loss from 2xTXAA over FXAA HQ.

I don't see pauly reporting his gpu usage dropping to 50% and a subsequent over 50% loss of frame rates to the point where before your clocks were fluctuating because it wasn't demanding to being unable to get 30 fps, let alone 60.


It's hard talking to someone so opinionated who lacks commonsense when it comes to hardware.

 

[TimothyLottes]

Apparently, it does not scale properly with SLI.

As for TXAA and SLI scaling, TXAA actually forced the driver team to fix some SLI performance bugs, and those changes got into the first driver which provided TXAA support. The changes should help any game using a temporal effect. This includes CryEngine with their temporal AA effect, and anyone doing temporal feedback for screen-space ambient occlusion effects. That being said SLI scaling is a challenging problem, and likely you could still see problems with any number of other things, not related to TXAA.

TXAA requires a GPU to GPU copy with SLI which is proportional to the resolution of the framebuffer. This copy doesn't cause stalling problems with SLI because the data for the copy is generated at the time TXAA runs, and isn't needed until the start of the next time TXAA runs (which with SLI will be on another GPU). Now the copy itself can load up the PCIe bus which can limit scaling somewhat. On systems which have crappy PCIe buses, this will be a problem.

Even then, DICE didn't exactly implement MSAA correctly in BF3. The amount of stuff it doesn't touch is incredibly sad.

I wouldn't hammer on DICE for this, those guys are actually quite awesome. The problem you are seeing with MSAA not removing aliasing on trees and such is caused because MSAA aims to shade only one sample per pixel. The outline on tree leaves is generated in the shader ("alpha test") not from actual geometry (triangle edges are the only things which get multiple samples by default). This is a common problem in just about every game with MSAA.

There is one technique developers can use to increase the quality with "alpha test" which involves selectively super-sampling at critical parts of the engine. For instance one can render with super-sampling into depth only with "alpha-test" on to get the good outline via multiple samples, then do a second pass to shade without super-sampling to avoid the extra shading cost.

lazy developers do not implement proper AA

MLAA/FXAA/SMAA definitely didn't help here. But the real trend started with the popularity of deferred rendering and a mix of consoles not having enough perf to do deferred with AA, and the techniques to reduce the cost of deferred with MSAA still not being widely understood.

Actually yes, they did. I don't know what the underlying technical reason is, but it's considered to be a bad thing to have the drivers overriding the developer's intentions with LOD.

The problem with screwing with LOD is that you need to know which textures to apply this to. Biasing LOD on the wrong textures causes problems. The best thing here would be the developers building this bias option into their games, touching only the correct textures.

With this statement being in reference to TXAA having a similar cost to MSAA, I don't buy it.

I forget if I posted some numbers here, but 4xTXAA on a mid-range mobile GT650M at 1280x800 takes 1.3 ms/frame extra over just turning on 4xMSAA. On BF3's highly optimized MSAA implementation the cost of just turning on 4xMSAA is about 9.1 ms/frame (on the same resolution and same GPU). So yeah, most of the cost of TXAA is in just the MSAA part, and the cost of the MSAA part will greatly vary per title. Also there are some optimizations going into the fixed cost of the TXAA bit, which will hit later titles.

SSAA is already capable of delivering what TXAA does

The difference is that with TXAA the shading cost is the same as with MSAA. 4xMSAA might only have a 50% hit in a title, with 4xRGSSAA this is going to be around a 400% hit for the pixel shading part of the frame (which at high resolutions is the dominate cost).

 

[SirPauly]

It's so refreshing to receive information!

 

[Akantus]

Have you even tried TXAA or are another speculative poster? SLI has better performance than single card, and the unideal scaling is likely due to the shortcomings of TXAA. I could be wrong and if I am, great. It has been documented that the SLI scaling of TXAA is not ideal - although the performance is still greater than single card. I never directly said anything regarding Timothy Lottes lying, its nice that you put words in my mouth pal. I did say this:

I don't see 2x MSAA causing anywhere NEAR the performance hit that TXAA 2x-4x causes for me in TSW. // In other words, words are cheap, nvidia needs to prove it by refining the performance of TXAA.

And that statement stands. Sorry, I expect 680 sli to be up to the task performance wise. And for that reason, I am not impressed. If that can happen anytime in the near future i'll revise my opinion, as it is now even 1080p has less than 60 fps with TXAA - I don't think thats a good tradeoff for whatever image benefit it may provide. If any of these experiences change with a future driver or future TSW build, i'm sure my opinion of TXAA will be revised.

Sorry if I offended you, that wasn't my goal, but I can't see other way how to understand what you wrote than, that you don't "believe" (because you don't have proof, since traditional MSAA isn't implemented to compare to) that 2xTXAA perf hit is 2xMSAA perf hit + little extra. Timothy Lottes even provided his measurements.

And no I haven't tried it since I don't have neither GTX600 series card nor TSW, so my only option is to go by videos, screens and what posters like you provide. It would be pretty empty "discussion" if only posters who had tried it were allowed to post...

And about SLI, now you have posted that your fps is better with SLI than single card with TXAA, I get it. I was just trying to provide reasonable explanation why you are seeing such performance hit.. sorry about that :\

Again, I'm sorry if my post offended you, my only desire is to engage in reasonable conversation about topic that interests me.

 

[Akantus]

TXAA requires a GPU to GPU copy with SLI which is proportional to the resolution of the framebuffer. This copy doesn't cause stalling problems with SLI because the data for the copy is generated at the time TXAA runs, and isn't needed until the start of the next time TXAA runs (which with SLI will be on another GPU). Now the copy itself can load up the PCIe bus which can limit scaling somewhat. On systems which have crappy PCIe buses, this will be a problem.

That's the solution I thought of.
Since we now have PCIe 3.0 with loads of bandwidth the performance hit should be quite small. (Guessing here, don't have any proof)

One thing that interests me, does the copying become a problem with 3/4-way SLI? Especially at surround resolutions?

 

[Gordon Freemen]

That's the solution I thought of.
Since we now have PCIe 3.0 with loads of bandwidth the performance hit should be quite small. (Guessing here, don't have any proof)

One thing that interests me, does the copying become a problem with 3/4-way SLI?

I was under the impression that PCI-E 2.0 does not even bottle neck a GTX 690.

 

[ViRGE]

I wouldn't hammer on DICE for this, those guys are actually quite awesome. The problem you are seeing with MSAA not removing aliasing on trees and such is caused because MSAA aims to shade only one sample per pixel. The outline on tree leaves is generated in the shader ("alpha test") not from actual geometry (triangle edges are the only things which get multiple samples by default). This is a common problem in just about every game with MSAA.

There is one technique developers can use to increase the quality with "alpha test" which involves selectively super-sampling at critical parts of the engine. For instance one can render with super-sampling into depth only with "alpha-test" on to get the good outline via multiple samples, then do a second pass to shade without super-sampling to avoid the extra shading cost.

I know, and I'm not. It's just disappointing; if anyone could have (and should have) gotten MSAA right, it would be them.

As it stands the aliasing even with MSAA is atrocious at times (Thunder Run, anyone?). They're not just failing to anti-alias fake geometry like trees, but somehow they're only anti-aliasing a limited selection of polygons too, which is just weird.

 

[TimothyLottes]

That's the solution I thought of.
Since we now have PCIe 3.0 with loads of bandwidth the performance hit should be quite small. (Guessing here, don't have any proof)

One thing that interests me, does the copying become a problem with 3/4-way SLI? Especially at surround resolutions?

Adding more GPUs doesn't change the amount of framebuffer which needs a copy with TXAA. It does reduce the window of time which the copy can run in before you end up with a stall, or maybe could hit some other exotic unknown bug in the driver with internal resource limits.

FPS and resolution however definitely changes SLI overhead. If surround means 1920x1080 times 3 monitors, this is only 50% more pixels than 2560x1600. Going from 2560x1600 at 30 Hz to 60 Hz is a much higher change in SLI overhead for TXAA.

I'd suggest on the next game with TXAA if it has MSAA (or if a TSW patch adds MSAA support), that if you end up with scaling problems with SLI only with TXAA and not with MSAA, that you complain through the proper channels at NVIDIA (any NVIDIA forums, etc). Because I can use this as leverage to get the priority needed to fix things.

 

[Grooveriding]

Adding more GPUs doesn't change the amount of framebuffer which needs a copy with TXAA. It does reduce the window of time which the copy can run in before you end up with a stall, or maybe could hit some other exotic unknown bug in the driver with internal resource limits.

FPS and resolution however definitely changes SLI overhead. If surround means 1920x1080 times 3 monitors, this is only 50% more pixels than 2560x1600. Going from 2560x1600 at 30 Hz to 60 Hz is a much higher change in SLI overhead for TXAA.

I'd suggest on the next game with TXAA if it has MSAA (or if a TSW patch adds MSAA support), that if you end up with scaling problems with SLI only with TXAA and not with MSAA, that you complain through the proper channels at NVIDIA (any NVIDIA forums, etc). Because I can use this as leverage to get the priority needed to fix things.

This is very cool :thumbsup: We'll see how Borderlands 2 turns out.

 

[blackened23]

Adding more GPUs doesn't change the amount of framebuffer which needs a copy with TXAA. It does reduce the window of time which the copy can run in before you end up with a stall, or maybe could hit some other exotic unknown bug in the driver with internal resource limits.

FPS and resolution however definitely changes SLI overhead. If surround means 1920x1080 times 3 monitors, this is only 50% more pixels than 2560x1600. Going from 2560x1600 at 30 Hz to 60 Hz is a much higher change in SLI overhead for TXAA.

I'd suggest on the next game with TXAA if it has MSAA (or if a TSW patch adds MSAA support), that if you end up with scaling problems with SLI only with TXAA and not with MSAA, that you complain through the proper channels at NVIDIA (any NVIDIA forums, etc). Because I can use this as leverage to get the priority needed to fix things.

Awesome :thumbsup:

I won't lie -- I'm pretty hard headed and opinionated and i'll freely admit it, Hopefully none of the things stated regarding TXAA bother you. Anyway - its pretty awesome that you can take the time out of your busy schedule to post tidbits regarding TXAA - We all certainly appreciate your posting your thoughts here, thanks. :thumbsup:

 

[lopri]

SLI has better performance than single card, and the unideal scaling is likely due to the shortcomings of TXAA.

I think a better interpretation is that TXAA amplifies the shortcomings of SLI.

And to those who compare AA in games with films: These two are very different media of expression/experience as BFK10K noted, and as far as their similarity reaches let me remind you there is an entire industry dedicated to bitch and moan about the quality of DVD/Blu-Ray products. (such as this) It is not difficult to spot poor visuals and fail effects on (TV) screens.

 

[Grooveriding]

I think a better interpretation is that TXAA amplifies the shortcomings of SLI.

And to those who compare AA in games with films: These two are very different media of expression/experience as BFK10K noted, and as far as their similarity reaches let me remind you there is an entire industry dedicated to bitch and moan about the quality of DVD/Blu-Ray products. (such as this) It is not difficult to see poor visuals and fail effects on (TV) screens.

True. Also worth mentioning high quality film/blu-ray is running at 24fps, whereas the gaming ideal is 60fps, in some cases 120fps. Gaming is also a very high speed and rapidly changing environment being presented in most cases, movies spend a lot of time on static scenes.

Two very different mediums imo.

 

[TimothyLottes]

Hopefully none of the things stated regarding TXAA bother you

No problem, certainly we all have different personal preferences here, but we actually have some very important common ground. For instance I'd like to help developers get the most out of MSAA with selective super-sampling, etc. The aim here would be that in-game you'd have the option to run with standard MSAA on, then crank up the super-sampling in various ways on a slider which ends at full RGSSAA (with a separate slider for LOD bias). Then for those who like the really sharp view, you just use the in game MSAA-to-RGSSAA with the slider, for those who want the film look, you just toggle on TXAA on top of that.

The real challenge here is getting critical mass so that the developer thinks things like selective super-sampling are important to users. That takes some work, because you need to sell it to the average joe too, who probably has a $100 video card. That average joe running with 1080p or soon retina resolution displays simply isn't going to have the perf at native resolution to super-sample on anything but a simple 2.5D game with no depth complexity or triangle density.

So part of my goal with TXAA was finding a way, like with DVDs or BluRays, that the output content could be up-sampled well so that these higher end techniques could be used on lower end cards if the gamer wanted to trade sharpness. Or alternatively on a GPU which isn't fast enough to enable 4xMSAA at native resolution and hit the 120 Hz you want for input latency in game, TXAA could be used with up-sampling, to get that 120 Hz frame rate .... by now you should realize that I'm a latency snob who doesn't care about sharpness, but hates temporal aliasing.

And one other option, which applies to those who like maximum sharpness with RGSSAA: if 4xRGSSAA has a 400% hit on frame rate, and 4xTXAA in an optimized title has say a 50% hit on frame rate, then in theory you'd be able to run the game with 4xTXAA at 2x the resolution (which would be an effective 300% hit on frame rate), and down-sample to have a performance win with TXAA compared to RGSSAA, and still get the sharpness you are looking for thanks to a new quality down-sampling (and I'm not referring to the relatively poor quality you currently get with the NVIDIA GPU-does-the-scaling driver hack).

Anyway there is a lot of things to do...

 

[boxleitnerb]

And one other option, which applies to those who like maximum sharpness with RGSSAA: if 4xRGSSAA has a 400% hit on frame rate, and 4xTXAA in an optimized title has say a 50% hit on frame rate, then in theory you'd be able to run the game with 4xTXAA at 2x the resolution (which would be an effective 300% hit on frame rate), and down-sample to have a performance win with TXAA compared to RGSSAA, and still get the sharpness you are looking for thanks to a new quality down-sampling (and I'm not referring to the relatively poor quality you currently get with the NVIDIA GPU-does-the-scaling driver hack).

Anyway there is a lot of things to do...

I'm a bit confused by those calculations. I very often play with 4xSGSSAA and it usually carries a performance hit of about 50% vs. no AA. Meaning, if I have 100fps without any AA, I have 50fps with 4xSG.

Regarding the bold part:
It is VERY interesting that you know about this hack!
I agree, for the performance hit, quality is not very good. The undeniable benefit I see with this hack is that in principle it works with absolutely every game, regardless of TXAA/MSAA/FXAA support, API etc. In combination with FXAA it actually looks very decent.

Do you think it would be possible to get a more effective down-sampling filter for this hack? Also, there is a bug that prevents 3840x2160 on Kepler cards (red screen bug). Maybe there could be something done about that? Sorry to be a little OT, but it was too tempting after your very appreciated comment

 

[SirPauly]

For instance I'd like to help developers get the most out of MSAA with selective super-sampling, etc. The aim here would be that in-game you'd have the option to run with standard MSAA on, then crank up the super-sampling in various ways on a slider which ends at full RGSSAA (with a separate slider for LOD bias). Then for those who like the really sharp view, you just use the in game MSAA-to-RGSSAA with the slider, for those who want the film look, you just toggle on TXAA on top of that.

The holy grail of flexibility enhancements, one may imagine. That would be so very welcomed considering the vast amount of different subjective tastes and tolerances.

 

[Silverforce11]

This could also explain microstutter during movement with TXAA, it feels like the game is skipping frames and is not smooth at all. I can play other games at 30 fps and they feel okay, but this one feels like its stuttering constantly with TXAA enabled.

Pure fail right there, massive blur, huge perf hit, screwed SLI scaling AND microstutter..

Yeah lets see the usual bunch defend this. As it is, it's lame, needs to be improved massively before being a valid option.

I'm saddened a new PC game with dx11 features add this lame AA without even trying MSAA support. <- this is the problem. I wouldn't care if they included MSAA but add new trial AA modes which can be optimized down the road etc.

 

[BFG10K]

No it's not.
Not in theory (temporal aliasing is not removed),

You seem to be throwing around that term a lot without actually understanding what it means.

Temporal aliasing is nothing more than a fancy name for aliasing that’s apparent during movement. This happens because areas where edge based schemes are weak (or non-existent) are under-sampled.

You can AA these areas by either by generating additional data (SSSA), or by blurring your existing data to remove high frequency noise (TXAA). The problem with the latter is that it removes too much other useful information at the same time.

and not in practice (performance penalty is huge)

2xSSAA has very similar performance to 8xMSAA in many cases. If you’re worried about performance then use FXAA anyway. It’s not as good as removing aliasing but it doesn’t have significant blur, and it works everywhere, not just on a per-game basis.

Again not true. Negative LOD bias is what's usually employed to sharpen the image which is otherwise softened by supersampling.

Your response makes absolutely no sense in the context of what I posted.

I’ll repeat what I said: the vast majority of games I’ve tested with SSAA show no significant blur and for the rare title that does, a LOD adjustment more than fixes it. So you can sharpen the game while still retaining less aliasing than edge based schemes.

Please, do show us some crisp SSAA images of BF3, Crysis 2, Shogun 2, Metro 2033, Sniper Elite V2.

Crysis 2 can’t have real MSAA and Metro 2033 is too slow to even use 4xMSAA, so neither TXAA or SSAA would be fast enough. I don’t have the other games.

How about you show us TXAA in those games? Oh wait, it currently only works in one game, and looks like it’ll never be driver enforceable. Whoops.

 

[BFG10K]

The difference is that with TXAA the shading cost is the same as with MSAA. 4xMSAA might only have a 50% hit in a title, with 4xRGSSAA this is going to be around a 400% hit for the pixel shading part of the frame (which at high resolutions is the dominate cost).

I accept the cost is higher, but the quality is also far better. It also works anywhere where MSAA works and is driver forceable, unlike TXAA which only works in one game.

I can’t help but feel a better option would be to simply make a new version of FXAA which is less aggressive about detecting edges, and let those that want more blurring use that.

 

[railven]

Here are shots with a meager GTX 670 in the same area as Blackened:

FXAAHQ

http://img843.imageshack.us/img843/8950/thesecretworlddx1120120l.jpg

TXAA 2

http://img225.imageshack.us/img225/7104/thesecretworlddx1120120s.jpg

Thanks for the pics Pauly, I have a question about them - not sure if it's the engine or the TXAA options:

But looking at the in-game time, they seem to be 7minutes apart, but why is there such a huge difference in lighting?

It's almost like one is done in the day (FXAAHQ) and the other at night (TXAAx2) since almost all the shadows are gone.

The blurring is definitely visible, but then I noticed how different the light sources are. Just curious if this is the game itself (ie drastic shifts in lighting from "evening" to "night" - 7 minute time diff) or TXAA (I recall reading in this thread lighting is affected, but that is a huge affect.)

 

[SirPauly]

Lighting is very dynamic in this title over-all.

TXAA x2

http://img4.imageshack.us/img4/5708/thesecretworlddx1120120g.jpg

FXAA HQ

http://img502.imageshack.us/img502/1683/thesecretworlddx1120120o.jpg

 

[railven]

Lighting is very dynamic in this title over-all.

TXAA x2

http://img4.imageshack.us/img4/5708/thesecretworlddx1120120g.jpg

FXAA HQ

http://img502.imageshack.us/img502/1683/thesecretworlddx1120120o.jpg

Thanks for some follow-up pics, and here you can see the "time of day" change, but in the last set of pics you posted, the left side of the two pics are relatively similar (ie you can see some transition of time), and you can see the stars are no longer in the exact same spot, but on the right side all the shadows are gone.

There were shadows casted by the trees on the building, the fances casted shadows, and there was a clear light source on the building, in the TXAA screen all that is gone, as if the sun set.

Gonna guess from these set of pics, it was a time shift, literally went from "evening" to "night" in 7 in-game minutes, haha.

EDIT: Tracing a circular light source - is that a spotlight illuminating the buildings on the right? (For your FXAAHQ first Pic)

 

[SirPauly]

Lighting transitions move pretty quick during sun-sets and sun-rises. Very dynamic lighting in this title.

 

[njdevilsfan87]

TXAA is making FXAA look really good right now.

 

[SirPauly]

Based on static images, can't disagree but in movement TXAA, really shines in this context. Believe the feature holds great potential based on solving the aliasing problems with movement, while being efficient is the hard part to me -- offering flexibility to improve clarity in the future would be very, very welcomed.

Right now, this ability is proprietary, not a forced feature, offers a noticeable blur, but it's a start and a lot of good here. The title had no AA, FXAA, FXAA HQ, and the title suffered from so much aliasing; it was ridiculous -- now it is virtually gone. Very welcomed for the title for me.

 

[f1sherman]

Please, do show us some crisp SSAA images of BF3, Crysis 2, Shogun 2, Metro 2033, Sniper Elite V2.

can&#8217;t [...] too slow [...] Whoops.

Don't be hard on yourself. And thanks for trying.

Selective editing of quotes in this manner is a good way to get yourself in trouble for trolling. I would suggest you not do that.
-ViRGE