4A Engine (Metro 2033) AA modes

brain_stew

Regular
Here's a comparison of 4A's analytical AA to 4xmsaa:

http://www.neogaf.com/forum/showpost.php?p=20297981&postcount=2389

MSAA apparently brings the framerate down by 20fps compared to AAA. Quality wise it looks pretty damn good considering the small performance cost. As expected it falls apart with fine edges (look at the wires at the top of the image) but copes OK elsewhere and there doesn't appear to be any loss in texture fidelity either. Its obvious to see why they chose to go this route over standard msaa in the 360 version. I know the character model looks blurry in the msaa shot but that's probably caused by the object based motion blur.
 
It's not just the character model being blurry. All textures are affected. The lighting is also a different strength for the orange lamp...
 
Here's a comparison of 4A's analytical AA to 4xmsaa:

http://www.neogaf.com/forum/showpost.php?p=20297981&postcount=2389

MSAA apparently brings the framerate down by 20fps compared to AAA. Quality wise it looks pretty damn good considering the small performance cost. As expected it falls apart with fine edges (look at the wires at the top of the image) but copes OK elsewhere and there doesn't appear to be any loss in texture fidelity either. Its obvious to see why they chose to go this route over standard msaa in the 360 version. I know the character model looks blurry in the msaa shot but that's probably caused by the object based motion blur.

It's not totally clear to me from that link, but are both of those taken on PC, one with traditional MSAA and the other with their post process AA? Or is one MSAA on PC and the other post process AA on 360? The sharpness difference between the two is striking, the second pic looks very blurry just about everywhere in comparison to the first so I'm wondering if it's a PC to console comparison pic.
 
According to the DF interview, Metro 2033's post process AA uses pattern recognition/edge detect to render those parts at double resolution and then downscaling. So it's edge detect 2xSSAA.
How does 2XSSAA compare to 4xMSAA?
2560x1600 is almost 2X of 1920x1080, so if we could have shots of Metro 2033 with noAA at 2560x1600 then resized to 1920x1080, it should have the same level of antialiasing on the edges that their algorithm detected, and better in places where it didn't.
 
Guys, just a question about 4A's AA method in Metro2033:

There are some statements in this forum, that AA is basically free on the Xbox360. I assume that 4A has already implemented tiling in their engine (to get even 2xMSAA@720p). So what makes it expensive? I understand that sometimes objects have to be computed twice, when they are at the tiling border...but it seemed to me (up to now) that this does not really affect the performance, as we even have games with 3 tiles.

Easy question: So why are 4A claiming that it frees the GPU when they use the new AA method although MSAA is supposed to be "nearly" free? What makes Metro2033 so special in this combination (it is just a classic FPS it seems)? How much is the potential in performance gain (I thought up to now that there is basically none) on the Xbox360 to replace the standard MSAA?
 
Its not free...nothing is ever free.
Might be that they are using some stuff in their engine which eats bandwidth & affects performance if they do MSAA over it, while if they apply their own technique its much better at conserving the performance. Not the most technical answer but I hope I am close to the point.
 
Guys, just a question about 4A's AA method in Metro2033:
First
I'm not to try to answer your post :LOL:
some decryption of the DF article would also interest me :)
There are some statements in this forum, that AA is basically free on the Xbox360. I assume that 4A has already implemented tiling in their engine (to get even 2xMSAA@720p). So what makes it expensive? I understand that sometimes objects have to be computed twice, when they are at the tiling border...but it seemed to me (up to now) that this does not really affect the performance, as we even have games with 3 tiles.
Interesting question indeed :) , I remember reading quiet some posts on the matter but the content fade away from me too. I remember reading that 2xMSAA is indeed pretty cheap, I also remember that 4xMSAA was more expensive than the extra geometry load for some reasons (may extra write back from edram to the main ram?) I would also enjoy a refresh on the matter. I also remember people not liking tiling because restriction/limitation it enforces on the renderer, a refresh here would also be appreciate too.
Easy question: So why are 4A claiming that it frees the GPU when they use the new AA method although MSAA is supposed to be "nearly" free? What makes Metro2033 so special in this combination (it is just a classic FPS it seems)? How much is the potential in performance gain (I thought up to now that there is basically none) on the Xbox360 to replace the standard MSAA?
I would also extra explanation on this part of the DF interview
The 360 was running deferred rotated grid super-sampling for the last two years, but later we switched it to use AAA. That gave us back around 11MB of memory and dropped AA GPU load from a variable 2.5-3.0 ms to constant 1.4ms. The quality is quite comparable.
 
Both are native 1080p and PC shots. I dunno what's causing the subtle blurriness in the MSAA shot. 4A did talk about implementing some form of (adaptive?) supersampling so perhaps that is being used when you enable 4xmsaa and without a negative LOD bias, texture fidelity is lost? I really have no idea.

Here's another comparison, this time from PCgames.de:

http://www.pcgameshardware.de/aid,7...t-DirectX-11-und-GPU-PhysX/Action-Spiel/Test/

The subtle blurring is again apparent when 4xmsaa is used. Regardless of what's messing up the 4xmsaa shots, I think their AAA holds up really well.
 
Last edited by a moderator:
I would also extra explanation on this part of the DF interview

It sounds like initially they were using a different (but still custom) AA method and they recently switched to a new method. So the memory and performance savings they describe are probably not in comparison to hardware msaa, but in comparison to their original AA method. The clue to that is the "11mb" savings they describe, because using hardware msaa on 360 doesn't use any extra memory. So it implies that their original custom method was more memory heavy. As for what their original method did, hard to know without more info from them.
 
Easy question: So why are 4A claiming that it frees the GPU when they use the new AA method although MSAA is supposed to be "nearly" free? What makes Metro2033 so special in this combination (it is just a classic FPS it seems)? How much is the potential in performance gain (I thought up to now that there is basically none) on the Xbox360 to replace the standard MSAA?

Generally speaking, there is no reason to go with a custom AA approach on 360 unless your engine is somehow incompatible with the hardware msaa. The msaa part on the 360 is free, the titling part has a cost but its relatively small. Still, if you replace hardware msaa with some other approach then you are just swapping one cost for another (tiling cost vs. post process cost), so either way there is some gpu cost. It's better to just stick with the hardware msaa if you can, which also helps preserve the sharpness in the scene. I don't know about Gow3's approach, but Saboteurs and Metro's custom AA definitely softens the image, way too much for my liking personally. It's possible that tiling simply created too many issues with their engine so they passed on it. Or perhaps they were willing to live with some blur in exchange for a fixed performance cost (tiling cost is variable whereas a post process AA can be consistent cost). I'm guessing that tiling caused some issues with their engine so they tossed it and went custom, which is a totally fine reason to do so.
 
Generally speaking, there is no reason to go with a custom AA approach on 360 unless your engine is somehow incompatible with the hardware msaa. The msaa part on the 360 is free, the titling part has a cost but its relatively small. Still, if you replace hardware msaa with some other approach then you are just swapping one cost for another (tiling cost vs. post process cost), so either way there is some gpu cost. It's better to just stick with the hardware msaa if you can, which also helps preserve the sharpness in the scene. I don't know about Gow3's approach, but Saboteurs and Metro's custom AA definitely softens the image, way too much for my liking personally. It's possible that tiling simply created too many issues with their engine so they passed on it. Or perhaps they were willing to live with some blur in exchange for a fixed performance cost (tiling cost is variable whereas a post process AA can be consistent cost). I'm guessing that tiling caused some issues with their engine so they tossed it and went custom, which is a totally fine reason to do so.

You know in the Metro 2033 PC shots, the sharper one was the MLAA and the blurry one was the 4xMSAA right? The MLAA doesn't smooth all the edges but textures remained sharper based on those two 1080p shots linked in this thread.
 
Generally speaking, there is no reason to go with a custom AA approach on 360 unless your engine is somehow incompatible with the hardware msaa. The msaa part on the 360 is free, the titling part has a cost but its relatively small. Still, if you replace hardware msaa with some other approach then you are just swapping one cost for another (tiling cost vs. post process cost), so either way there is some gpu cost. It's better to just stick with the hardware msaa if you can, which also helps preserve the sharpness in the scene.
From what they said they saved some memory by going with post processing AA, so I would belive that they somehow stored G-Buffer in full resolution in a main ram.
Perhaps their problems were inherited from that.
You know in the Metro 2033 PC shots, the sharper one was the MLAA and the blurry one was the 4xMSAA right? The MLAA doesn't smooth all the edges but textures remained sharper based on those two 1080p shots linked in this thread.
Sadly the sharpness seems to be more of the shader aliasing, also the new method seems to fail in most obvious cases like thin/small polygons and long edges.
 
Last edited by a moderator:
You know in the Metro 2033 PC shots, the sharper one was the MLAA and the blurry one was the 4xMSAA right? The MLAA doesn't smooth all the edges but textures remained sharper based on those two 1080p shots linked in this thread.

Yeah you're right, the broken up thin lines on the sharper shot indicate no hardware msaa. However, hardware msaa doesn't blur textures at all, it only does it's mojo on edges. So where is all that extra sharpness on their post process version coming from? Is it artificial, ie, does it look all shimmery in motion? I haven't seen the game in person yet.
 
First
I'm not to try to answer your post :LOL:
some decryption of the DF article would also interest me :)

Interesting question indeed :) , I remember reading quiet some posts on the matter but the content fade away from me too. I remember reading that 2xMSAA is indeed pretty cheap, I also remember that 4xMSAA was more expensive than the extra geometry load for some reasons (may extra write back from edram to the main ram?) I would also enjoy a refresh on the matter. I also remember people not liking tiling because restriction/limitation it enforces on the renderer, a refresh here would also be appreciate too.

I would also extra explanation on this part of the DF interview

I think this time, the DF interview was not so good. A lot of questions and contradicting stuff...grandmaster should have asked him more stuff.

How is it possible that in the same interview he first state that they implented the new AA method to spare some memory and GPU time....and at the end of the interview he claims that the Xbox360 is quite underused and that they often have up to 100 MB left (okay, why then even consider the new AA method) and that the game runs with 40-50 fps (why save GPU?) - this is contradicting and GM should have asked him about this in more detail!


They only explanation I can find is somewhat a conspiracy theory ;-)
I know you guys want to hear it, so here we go:

The dev clearly states that no PS3 version is in production. What if there is one? What if Metro2033 is only time exclusive? That would explain why they want to spare memory, spare GPU time and maybe would even explain why the Xbox360 is underused
:oops:
 
How is it possible that in the same interview he first state that they implented the new AA method to spare some memory and GPU time....and at the end of the interview he claims that the Xbox360 is quite underused and that they often have up to 100 MB left (okay, why then even consider the new AA method) and that the game runs with 40-50 fps (why save GPU?) - this is contradicting and GM should have asked him about this in more detail!

Its a conspiracy !!
Seriously speaking I re-read the interview today & those are the EXACT questions I had. ;)

1) Why bother saving 11MB in favor of a worse looking AA when you already have 100MB unused..??

2) Why not have the game 100% V-synced with no Tearing if your game runs at 40-50FPS ?

3) Why not utilize the left over juice and bring down the FPS to 30 by using extra graphical features ?
 
1) Why bother saving 11MB in favor of a worse looking AA when you already have 100MB unused..??

They already explained that with the performance metrics.

2) Why not have the game 100% V-synced with no Tearing if your game runs at 40-50FPS ?
Instead, ask them why they don't do it the Bioshock way... i.e. they may have preferred the response of higher framerate. Perhaps they "simply" just were not thinking some would prefer the v-sync.
*Don't have the game to check.
3) Why not utilize the left over juice and bring down the FPS to 30 by using extra graphical features ?
Because then you risk going under 30fps... Were it so easy. :|
 
Back
Top