AMD: R8xx Speculation

[ChrisRay]

Can you please qualify the impact of this alleged subpar filtering? What would be the visible result?

Underfiltering would result in banding and visible mip map borders as you cross them. They were very heavily apparent on G7x cards. And if you recall back to the war on "Brilinear" it's essentially the same thing. Toying with LoDs and lower trilinear filtering workload.

It's really different than say "Angle Dependent" LoD calculations. Which would just not sample certain angles of the LoD field. Nvidia in their control panel has "Trilinear Optimisation" and Anistropic Mip Field Optimisations. All which adjust the level of filtering between mipmaps. But they are alot less aggressive on the G80 + Designs than they were on the Nv4x/G7x designs. ((I would argue this was largely because the G7x shared its ALU power with texturing)).

These methods of reducing workload and improving performance don't do much for G80 + cards. Which are all abundant in texture filtering as well as being decoupled from the ALUS. Most people just turn on HQ now because theres very little performance problem with it.

Due to the fact that Nvidia adjusted its aggressive texture filtering at the same time they implemented angle independent AF. Alot people assume they are connected. I'm really surprised about all the talk about texture filtering. Mostly because its been almost a non issue for Nvidia users for a long time. Seems wierd to be hyping it up at this point. Unless of course theres been a major adjustment to LoD calculation on AMD's side. For which I wouldn't know anything about.

Keep in mind that Nvidia drivers still allow you to globally control the default LoD. As well as push it back any "negative" LOD bias to zero from their control panel. ((Which is defaulted but for the most part I havent seen much use for)). Its another one of those G7x legacy functions.

 

[Silent_Buddha]

Just like sex. You spend all that time building it up only to blow your load in 5mins. Atleast it makes for comfortable sleep.

That's why you should use a...er...Review Ring to prolong that...uh...time spent...uh...Reviewing.

Regards,
SB

 

[Novum]

So, how many? (At a surface angled above max. anisotropy and at 45 degrees, which is nastiest for memory access.)

Or to put it another way, how many samples do they skip?

I don't know. There is no way that I have come up to measure it exactly. The only way is to compare it to my reference with undersampling.

I also don't think they exactly "skip" samples. They most likely take e.g. 10 instead of 16 for the coarser Mip and 14 for the more detailed one uniformly distributed with trilinear filtering when doing 16:1 AF. And you would be surprised how little samples are necessary to get a halfway good result, because most textures are low frequency.

Keep in mind that Nvidia drivers still allow you to globally control the default LoD. As well as push it back to "Zero" by using the texture filtering LoD Bias to zero.

NVIDIA does have very big problems with filtering when applying negative LOD. But D3D10/11 docs say that this is unspecified behaviour anyway.

 

[ChrisRay]

I don't know. There is no way that I have come up to measure it exactly. The only way is to compare it to my reference with undersampling.

I also don't think they exactly "skip" samples. They most likely take e.g. 10 instead of 16 for the coarser Mip and 14 for the more detailed one uniformly distributed with trilinear filtering when doing 16:1 AF. And you would be surprised how little samples are necessary to get a halfway good result, because most textures are low frequency.


NVIDIA does have very big problems with filtering when applying negative LOD. But D3D10/11 docs say that this is unspecified behaviour anyway.

Ya I adjusted the comment specifically regarding this. As there may have been some confusion. Any "negative" LoD assigned will be set to zero with the clamp enabled.

Where are you experience "major" problems with LoD's above -1?

 

[Novum]

Always. It's just a single big flickering mess on all NV GPUs since G80 with negative LOD. This is the reason for the LOD clamp option.

 

[Lux_]

So, how many? Or to put it another way, how many samples do they skip?

If I remember correctly, then Shannon theorem states that samples have to be at least twice the frequency of the original signal. Only then you can reconstruct the original signal from the samples.
When the samples are less frequent than twice the highest frequency, you start to lose details when reconstructing the original.

So, the test would need very finely detailed texture. And animation would be needed to illustrate the effect, as hypothetical undersampling is probably somewhat dynamic in nature.

 

[ChrisRay]

Always. It's just a single big flickering mess on all NV GPUs since G80 with negative LOD. This is the reason for the LOD clamp option.

Interesting. I have noticed that it can cause some textures to appear less sharp in games that use alot more default -LoDs. But most games that use -LoD are assuming you dont even know what AF is.

I mostly see this in racing sims more often than not. Have any games you might recommend which use alot of -LoD for their mipmaps which I can look at? I mean you can globally force it within the drivers. But thats not the best example either.

 

[MfA]

If I remember correctly, then Shannon theorem states that samples have to be at least twice the frequency of the original signal.

Sampling theory is neither here nor there really, anistropic filtering approximates area averaging ...

To satisfy Nyquist limits we would have to brickwall filter during mipmap creation and everything would come out like a ringing mess (you could use a more gradual filter drop off before the brick wall at, but then you'd just get a blurry ringing mess). Satisfying Nyquist limits is neither an option nor an ideal.

So, the test would need very finely detailed texture. And animation would be needed to illustrate the effect, as hypothetical undersampling is probably somewhat dynamic in nature.

A checkerboard instead of the color pattern in the tube test would be a nice.

 

[MfA]

Oh a finegrained pattern in a tube is exactly what novum used ... Novum, could you get/post a movie with slow texture movement running on the R800?

 

[rpg.314]

Just like sex. You spend all that time building it up only to blow your load in 5mins. Atleast it makes for comfortable sleep.

Well, by the looks of it, JHH will lose hi sleep in about 9 hours from now for a couple of months atleast.

 

[Davros]

@chrisray the racing game famous for showing nv in a bad light wrt filtering is ea sports f1 99-02 (aka f1 challenge 99-02)

 

[ChrisRay]

@chrisray the racing game famous for showing nv in a bad light wrt filtering is ea sports f1 99-02

Thats the one rage3d used in its review right? I never have my LoD clamped anymore. But to be fair I mostly play MMORPGs.

 

[Davros]

so what should i have my lod set to for best iq clamp ?

ps: nv profile page takes over 2 minutes to open (the first time) for people who have a lot of games installed any chance of letting the powers that be know

 

[Novum]

Zero.

 

[Ollo]

Thanks neliz for the benchmarks.

Yeah those were great!

 

[Novum]

Sampling theory is neither here nor there really, anistropic filtering approximates area averaging ...

To satisfy Nyquist limits we would have to brickwall filter during mipmap creation and everything would come out like a ringing mess (you could use a more gradual filter drop off before the brick wall at, but then you'd just get a blurry ringing mess). Satisfying Nyquist limits is neither an option nor an ideal.

A checkerboard instead of the color pattern in the tube test would be a nice.

Nyquist at least gives you a limit where more samples will not give better results. You are right that mipmapping and bilinear sampling doesn't give you the best reconstruction, but that's no reason to make it even worse.

http://www.opengl.org/registry/specs/EXT/texture_filter_anisotropic.txt
Read section 3.8.5. That's how sampling is implemented on every GPU I know.

Note that the line of anisotropy calculation that is described there is the one with maximum angle dependence (Intel style), so it's very bad

 

[Malo]

ps: nv profile page takes over 2 minutes to open (the first time) for people who have a lot of games installed any chance of letting the powers that be know

Not many users have the "Sacred Terabyte of Gaming Goodness™"

 

[Davros]

True not everyone can be the chosen one

 

[Broken Hope]

First review

http://translate.google.co.uk/trans...hp?guide=20090923&sl=it&tl=en&history_state0=

 

[Kaotik]

http://forum.beyond3d.com/showthread.php?p=1338529