antialiasing questions

[SonicIce]

What type of supersampling does TRSS use? I was messing with it in Live for Speed with nHancer looking at the trees, and it looks like 4xMSAA + TRSS doesn't yield trees as smooth as regular 1x2 supersampling by itself. So what kind does it use?

 

[SonicIce]

its a mystery!

 

[BFG10K]

What type of supersampling does TRSS use?

Rotated grid.

 

[ViRGE]

It has two modes, one using multisampling and the other uses supersampling. The multisampling mode is useless for the most part, supersamping offers far better results.

 

[SonicIce]

Originally posted by: BFG10K

What type of supersampling does TRSS use?

Rotated grid.

rotated grid supersampling? that doesn't make sense.

 

[Gstanfor]

The MSAA component uses a rotated grid, the SuperSampled uses Ordered Grid. Supersampling is always Ordered Grid on nvidia hardware (might change for G80, but I'm taking current and prior chipsets here).

 

[BFG10K]

rotated grid supersampling? that doesn't make sense.

If you don't understand the basics of AA then it probably won't make sense.

Supersampling is always Ordered Grid on nvidia hardware

No, this is false. When using TrAA SS the super-sampling pattern on alpha textures is rotated grid. That's one of the reasons why it looks much better than the regular xS super-sampling on alpha textures.

 

[SonicIce]

Originally posted by: BFG10K

rotated grid supersampling? that doesn't make sense.

If you don't understand the basics of AA then it probably won't make sense.

Supersampling is always Ordered Grid on nvidia hardware

No, this is false. When using TrAA SS the super-sampling pattern on alpha textures is rotated grid. That's one of the reasons why it looks much better than the regular xS super-sampling on alpha textures.

oh i didn't know supersampling could use a rotated grid.

the way i understand it, supersampling renders a multiple of the original resolution and then scales it down. like 1x2 or 2x2 or 4x4. if this is correct then im wondering which type TRSS uses. im guessing it would be 1x2 or 2x2 rotated grid for the alpha textures.

 

[BFG10K]

oh i didn't know supersampling could use a rotated grid.

It sure can. 🙂

3dfx's VSA-100 did it, AAA/TrAA does it and also the Super AA/SLI AA modes do it too.

if this is correct then im wondering which type TRSS uses.

It's actually very simple - the pattern it uses is exactly the same as the current MSAA pattern.

So if you're using 4xAA for example you will get 4xSSAA on alpha textures with the same rotated grid pattern.

ATi's quality AAA works in exactly the same way.

 

[Gstanfor]

What are you basing you rotated supersamplin grid claim on BFG10K?

If it's an AA test app, I'll bet it's actually reporting the MSAA component and NOT the SSAA component samplin positions.

 

[BFG10K]

If it's an AA test app, I'll bet it's actually reporting the MSAA component and NOT the SSAA component samplin positions.

If that were the case then it would be doing the same for TrAA MS and performance AAA, which it doesn't.

The test app is correct.

 

[Gstanfor]

Why are we discussing ATi AA methods (AAA)?

Try this, set the AA mode to SSAA 4x4 with T:SSAA stacked on top and see what sort of sample pattern you get.

 

[BFG10K]

Why are we discussing ATi AA methods

Because this proves the AA tester is not simply mimicking the number of MSAA samples (i.e. 6xMSAA with performance AAA does not show 6 super-samples).

Try this, set the AA mode to SSAA 4x4 with T:SSAA stacked on top and see what sort of sample pattern you get.

Ordered grid 16xSSAA of course.

In fact it looks exactly the same with or without TrAA and in this case that's exactly what we would expect.

 

[Gstanfor]

In fact it looks exactly the same with or without TrAA

It looks the same because all the AA happening in my scenario is SSAA, which uses Ordered Grid and not Rotated Grid...

You will only see a rotated grid pattern when you use an AA setup with an MSAA component in it.

The ATi AA is weird. They don't actually do SSAA even in T:AA modes, they mimick it somehow with MSAA. There was a good thread on it somewhere on B3D.

 

[BFG10K]

It looks the same because all the AA happening in my scenario is SSAA, which uses Ordered Grid and not Rotated Grid...

Sure but your example has absolutely nothing to do with the issue.

The issue here is that TrAA replicates the samples MSAA takes, both in pure modes and also in combined modes. If you're using a pure SSAA mode like your 4x4 example TrAA doesn't even take effect so it's meaningless to even look at that pattern.

You will only see a rotated grid pattern when you use an AA setup with an MSAA component in it.

Exactly. That's how come it can stack with the likes of 8xS and 16xS.

The ATi AA is weird. They don't actually do SSAA even in T:AA modes, they mimick it somehow with MSAA.

Not only does ATi do SSAA they do it under both quality and performance modes, unlike nVidia that only does it under the quality (SS) mode.

The AA tester shows alpha SSAA in both cases and the performance hit and IQ with ATi's quality AAA is in line with nVidia's TrAA (SS).

There was a good thread on it somewhere on B3D.

Got a link? Because I'm going to have to call BS on that claim.

 

[Gstanfor]

My example has *everything* to do with the issue. You were claiming supersampling could use a rotated grid when used as a transparency AA mode. The mode I suggested deliberately used no MSAA components. You stated there was no difference between T:SSAA on or off. Thats because none of the AA modes employed utilize a rotated grid!

To see a difference you would need to do something like 4x MSAA or 8xS AA plus T:SSAA and then the rotated grid would come from the base AA, not the stacked T:SSAA.

Alternatively you could do 4x4 SSAA and stack T:MSAA (why you would do so I have no idea though...)

 

[BFG10K]

You were claiming supersampling could use a rotated grid when used as a transparency AA mode.

That claim is correct.

You stated there was no difference between T:SSAA on or off. Thats because none of the AA modes employed utilize a rotated grid!

You providing an irrelevant AA mode has no bearing on the fact that my original claim was correct.

It's like me claiming nVidia supports MSAA and then you ask me to use a pure SSAA mode to "prove" I'm wrong.

Your arguments are nonsensical.

To see a difference you would need to do something like 4x MSAA or 8xS AA plus T:SSAA and then the rotated grid would come from the base AA, not the stacked T:SSAA.

I still don't think you understand how TrAA actually works so let me explain it to you again as clearly as possible.

The TrAA sample patten mimicks the current MSAA pattern when it detects alpha textures.

The stacking in combined modes comes from the fact that the MSAA patterns are used on alpha textures while the regular combined pattern is used on the rest of the scene.

So in the case of 8xS you'll get 8xSSAA on alpha textures and 4xMSAA + 2xSSAA for the rest of the scene.

Alternatively you could do 4x4 SSAA and stack T:MSAA (why you would do so I have no idea though...)

That wouldn't do anything because as I explained TrAA derives its sampling from the MSAA pattern currently in use.

If you're using a pure SSAA mode like 4x4 TrAA won't have any effect which is why your example is totally meaningless and proves nothing.

 

[Gstanfor]

Oh, I understand how the mode works BFG10K - I have the T:AA whitepaper and nvidia's GPU programming guide to tell me how it works...

 

[BFG10K]

Then I would suggest you consult your documentation again as you've clearly missed some very important details.

 

[Gstanfor]

No I haven't... I suggest you read the documentation since the behaviour you are trying to describe applies to the T:MSAA mode, not the T:SSAA mode!

 

[BFG10K]

the behaviour you are trying to describe applies to the T:MSAA mode, not the T:SSAA mode!

Uh, no.

While it's true that the pattern is the same for both MS and SS, TrAA MS isn't actually visible in the AA app because the samples are multi-samples.

 

[Gstanfor]

T:SSAA simply resamples the geometry n times to get its samples. T:MSAA has to compute the alpha values. Read the damn documentation!

There are no MSAA samples for T:SSAA to modify, since MSAA *IGNORES* alpha (which is what T:AA works on, *not* the rest of the scene)!

 

[BFG10K]

This is getting comical beyond belief. Lets read nVidia's documentation, shall we?

I'll even bold the relevant bits since in the past you have demonstrated an alarming lack of reading comprehension skills even when presented with clear English sentences.

On GeForce 7800 GTX, NVIDIA has exposed a Supersample renderstate allowing individual primitives to be supersampled (in combination with a multisampled backbuffer). This is especially effective for textures with hard edges like a chain link fence or lattice work.

Transparency supersampling provides optimal flexibility: because transparency supersampling can be controlled on a per-primitive basis, you get the advantage of extremely high quality supersampling for alpha-tested textures but you can avoid this cost for other geometry.

In other words your get SSAA for alpha textures and MSAA for the rest. If you can't understand even this basic scenario then you don't have any hope of possibly understanding how other forms of TrAA and AAA work.

Not that this has anything to do with your original incorrect claim that you can't get rotated grid SSAA on nVidia cards, but responding with irrelevant and random comments has always been your modus operandi.

So instead of spamming up the forum with your blatant misinformation and irrelevant rhetoric I suggest you actually do some research before being totally proven wrong as per usual.

 

[Gstanfor]

How has what any of what you just posted contradicted any of what I've said in previous posts?

It doesn't matter how you try to twist it BFG10K, the T:SSAA is done with an ordered grid. The MSAA backbuffer is totally irrelevant since T:SSAA ****only applies to alpha textured primitives**** In other words an MSAA'd primitve will never receive T:SSAA filtering.

Heres another quote from that document for you:

Multisampling, on the other hand, is based on geometric pixel coverage, and will not work with fragments where the underlying geometry does not match the actual pixel?s geometry.

The only comical person in this thread is you. Why don't you try reading my signature once you've had someone read you nvidia's documentation?

 

[BFG10K]

How has what any of what you just posted contradicted any of what I've said in previous posts?

That depends which of your posts you're talking about given you tend to post randomly, almost like a "spin the bootle" approach.

It doesn't matter how you try to twist it BFG10K, the T:SSAA is done with an ordered grid.

Total and utter BS.

If you think that's ordered grid SS I suggest you get your vision checked.

Furthermore I've already debunked your false claim that the AA tester simply outputs the same amount of SSAA samples by pointing out the example of performance AAA but you're too obtuse and obstinate to accept anything outside of the world of nVidia.

Even without the application it's trivially easy to tell SS TrAA is doing something different than regular OGSS as the alpha textures have vastly better image quality.

Stop making up crap Gstanfor, your trolling is getting tiresome.

T:SSAA ****only applies to alpha textured primitives**** In other words an MSAA'd primitve will never receive T:SSAA filtering.

No sh!t sherlock, thanks for stating the obvious. The point of the quote was to demonstrate that the TrAA SS pattern operates in lock-step with the regular MS pattern, unlike your claim that some kind of fictional ordered grid is being used.

Where is your evidence of this magical ordered grid being used in TrAA Gstanfor? So far you have providing nothing that backs your claim, as per usual. You simply make up steaming piles and then spam up the forum with useless rhetoric when actual evidence disproving you is presented.

Multisampling, on the other hand, is based on geometric pixel coverage, and will not work with fragments where the underlying geometry does not match the actual pixel?s geometry.

Um, that's from the section labelled "transparency multi-sampling".

You claimed super-sample TrAA can't do rotated grid SSAA and then you point me to the section for multi-sample TrAA to prove you're right?

Seriously, are you for real?

Reading comprehension again and all that.