Hellbinder says full trilinear would cost 50% preformance?

[Hellbinder]

And hence, Hellbinder is a rabid fanboy who hears a few very vague rumors, spices them up with his own insane ideas without mentioning that, and then capitalizes random words.

Please...

That is so not true its not even funny. But thats ok if you all want to hack on me i dont care. It does not change the truth about me at all.

 

[bloodbob]

No,,, I dont think so,,,

I Think if you look back at hardware that did Actualy Full trilinear with no performance hacks at all it was nearly a 50% hit. Furthermore as i mentioned as scene complexity increases Full trilinear will get more and more expensive.

So laugh if you want. I dont mind. 8)

I'll do a benchmark on my TNT2 later Hell.

 

[Hellbinder]

Re: Hellbinder says full trilinear would cost 50% preformanc

Hellbinder is the one in the know for ATI he always has the scoop

Haha. Hahaha. HAHAHAHAHAHA!

Hellbinder knows very little. His "technical details" are often nothing more than FUD. If HB says something, believe the opposite.

I am really puzzled why you are attacking me like this. Especially after i got permanently banned rom NVnews becuase i stuck up for you guys who "left" there.

As an example it is on record months and months ago that I said the R420 was a Quad souped up version of the 9600. Which is not a perfect description but pretty close. I also pointed out MONTHS ago that the R420 would have a new AF method that got rid of Mip-maps even in performance mode. AND that it was in the 9600 but not fully exposed in drivers so not to cause confusion with the rest of the line.

Yeah, I made some mistakes along the way.. But you guys in this Thread are not being fair or accurate (or very nice for that matter).

 

[Eronarn]

Re: Hellbinder says full trilinear would cost 50% preformanc

or very nice for that matter

People not being nice??! ON THE INTERNET!?!?

OH NOES! CALL THE POLICE!



(sorry, it had to be said)

 

[Ruined]

From the benchmarks I've seen, like CoD, ATI seems to take a 15-20% hit with 8x AF at 1600x1200 compared to full trilinear (mipmap colorson)

 

[SiliconAbyss]

Re: Hellbinder says full trilinear would cost 50% preformanc

or very nice for that matter

People not being nice??! ON THE INTERNET!?!?

OH NOES! CALL THE POLICE!



(sorry, it had to be said)

Very intelligent, thanks for that.

 

[keegdsb]

No,,, I dont think so,,,

I Think if you look back at hardware that did Actualy Full trilinear with no performance hacks at all it was nearly a 50% hit. Furthermore as i mentioned as scene complexity increases Full trilinear will get more and more expensive.

So laugh if you want. I dont mind. 8)

Please delete this post, it serves no useful purpose.

Ha, I just realized that right before Hellbinder said the below

 

[Hellbinder]

btw, i may not have been very clear. I was thinking full trillinear filtered AF. Not stand alone Trillinear.

 

[Hellbinder]

From the benchmarks I've seen, like CoD, ATI seems to take a 15-20% hit with 8x AF at 1600x1200 compared to full trilinear (mipmap colorson)

Thats Angle dependant AF. What do you think it would be without that? and wihout any optimizations at all.?

 

[Eronarn]

Re: Hellbinder says full trilinear would cost 50% preformanc

Very intelligent, thanks for that.

It's true, though- not everyone is going to treat everyone else perfectly nicely. Just try to ignore people like that Hellbinder, no need to feel personally offended.

 

[Hellbinder]

Re: Hellbinder says full trilinear would cost 50% preformanc

Very intelligent, thanks for that.

It's true, though- not everyone is going to treat everyone else perfectly nicely. Just try to ignore people like that Hellbinder, no need to feel personally offended.

I am not really offended. Just puzzled. i hold Baron in pretty high esteem and actually feel a little blindsided by his comments.

as for my post being fud. My trillinear comments were directed at performance in general and not at ATi or Nvidia. Just telling people that it is unreasonable IMO to want full everything with the crazy performance hit it would cost and that we should expect even more optimizations as Game complexity goes through the roof.

That is optimizations from BOTH companies.

 

[Tim Murray]

HB, I don't consider any rumor to you to be particularly credible, especially after the Overdrive bit. No offense, just based on track record. (PS--the filtering has been exposed in RV3x0 cards since Cat 3.4. and the capitalization comment was a joke. I'm a joking kind of guy.)

And, how are you going to estimate the performance hit of non-angle-dependent AF when R420 isn't capable of it? You could get some goofy theoretical fillrate hit numbers, but I doubt that those would be anywhere close to accurate or could be used to interpolate performance hits in games. I think that estimating the performance hit of anything is kind of silly. Sure, you could compare R3x0 and NV3x performance hits, but that seems very silly.

 

[KimB]

btw, i may not have been very clear. I was thinking full trillinear filtered AF. Not stand alone Trillinear.

It doesn't matter. The performance hit for enabling trilinear filtering cannot be more than 50%. It doesn't matter how complex scenes get. The worst it can do is approach 50%.

And going into the next couple of years, I expect trilinear filtering to have less and less of a performance hit, as shaders get longer and so it becomes easier to hide the latency for texture filtering.

 

[bloodbob]

btw, i may not have been very clear. I was thinking full trillinear filtered AF. Not stand alone Trillinear.

It doesn't matter. The performance hit for enabling trilinear filtering cannot be more than 50%. It doesn't matter how complex scenes get. The worst it can do is approach 50%.

And going into the next couple of years, I expect trilinear filtering to have less and less of a performance hit, as shaders get longer and so it becomes easier to hide the latency for texture filtering.

Actual figures from a memory limited situation IE my 9500 pro over clocked but with the memory significantly UNDER clocked ( 200ish if I went much below I got artifacts ).

2631 3dmark2k3 16x AF with Tri
2743 3dmark2k3 16x AF with Bi

Yes and thats on all stages.

What Chalnoth says is a mathematical fact. Bilinear with mipmaps ( remeber we talk about bilinear filter we generally mean bilinear filter with mipmaps ) = 1 bilinear look up and Trilinear filter = 2 bilinear look ups. If have no memory cache then we have to look them up every time so if we NEVER write to frame buffer the speed difference is 1/2 as soon as we write to the frame buffer or do anything else interesting it become (1+x)/(2+x). Just to make things more interesting as X tends towards infinity the preformance loss tends to 0.

Oh and if we move to AF + Tri you are basicly just taking more samples so it becomes (Ax1)/(Ax2) which becomes 1/2. Very boring.

 

[ZenOps]

Umm, you can't do one pass trilinear... It would be impossible, as the image memory framebuffer must be passed through once to get bilinear and once again to get trilinear. There is no two ways around it, and the performance hit would be exactly 50 percent.

It is possible to do a quick second pass though (Nvidias so called brilinear or optimized trilinear) that only takes quick samples of the memory buffer and applies the pass only to textures that need it.

I think ATi is doing the right thing here, makes perfect sense to me really... Sort of like Z-culling, where noone ever sees the areas of the screen that are not being rendered - why even bother to put another rendering pass on non-coloured (or even transparent wireframe) pixels?

Honestly though... Full trilinear is quality overkill for both NVidia and ATi, bilinear isn't really all *that* bad. Somewhere in between lies good quality without the heavy performance hit of full trilinear.

IE: would you rather have full trilinear and 16XAF or an almost indistinguishable optimized trilinear and 256xAF (Theoretically would be about the same peformance) Would you rather have noAA with full trilinear or the next level of AA with an optimized trilinear.

True trilinear on everthing (including non-visible areas and areas that do not show a visual difference with even bilinear) is wayyyyy overhyped IMO.

 

[bloodbob]

Umm, you can't do one pass trilinear... It would be impossible, as the image memory framebuffer must be passed through once to get bilinear and once again to get trilinear. There is no two ways around it, and the performance hit would be exactly 50 percent.

Umm if your not doing an alpha blend or anything why do you have to read the frame buffer? ( WTF IS ALL THIS FUD COMING FROM ??!???!?!? )

Calculate Z
Calculate LOD
Do 2 bilnear sample for the texture
Blend the two samples according to the LOD
write to the frame buffer.

Actually you tried to do it in two passes and their alphablending its gonna be very ugly possible impossible I would have to sit down and think about it ( well with the current frame buffers you could have linked lists ect but lets forget about that ).

IE: would you rather have full trilinear and 16XAF or an almost indistinguishable optimized trilinear and 256xAF (Theoretically would be about the same peformance) Would you rather have noAA with full trilinear or the next level of AA with an optimized trilinear.

Next how do you work out that 256xAF is only a 50% preformance hit compated to 16xAF????????

Or better yet how about the companies give me a choice on wether or not I choose to use the optimistation then I can have whatever I want when I want.

 

[ZenOps]

AF is very efficient.

Its possible to get only a 5 percent speed drop going from 2x to 16x. If you extend it out, 16x to 256x should still stay well within the 50 percent peformance range.

Y'know... I just thought of something:

Would it be possible to do a temporal bilinear filtering to approximate a trilinear without any peformance hit? (sort of like Temporal AA)

Shift the bilinear stages ever so slightly across the horizon, and it should "blur" into an equivalent trilinear image.

 

[KimB]

Umm, you can't do one pass trilinear... It would be impossible, as the image memory framebuffer must be passed through once to get bilinear and once again to get trilinear. There is no two ways around it, and the performance hit would be exactly 50 percent.

No. All trilinear is single-pass. It may require more than one clock to get the texture samples required, but it does not require an extra write to the framebuffer.

It is possible to do a quick second pass though (Nvidias so called brilinear or optimized trilinear) that only takes quick samples of the memory buffer and applies the pass only to textures that need it.

This is not the case at all. The "brilinear" technique basically reduces the area of the screen to which trilinear filtering is applied, increasing performance by reducing the average number of texture samples taken per pixel.

Honestly though... Full trilinear is quality overkill for both NVidia and ATi, bilinear isn't really all *that* bad. Somewhere in between lies good quality without the heavy performance hit of full trilinear.

Oh, come on. Trilinear filtering has been pretty standard since about the time of the original GeForce.

IE: would you rather have full trilinear and 16XAF or an almost indistinguishable optimized trilinear and 256xAF (Theoretically would be about the same peformance) Would you rather have noAA with full trilinear or the next level of AA with an optimized trilinear.

Higher degrees of anisotropy both have lesser performance hits and are applied to smaller portions of the screen. 256-degree anisotropic filtering would be absolutely pointless. Maybe one day we'll get as high as 32-degree, but that's about where it'll stop. Why spend 256+ clocks on one pixel in a million, anyway?

 

[bloodbob]

AF is very efficient.

Its possible to get only a 5 percent speed drop going from 2x to 16x. If you extend it out, 16x to 256x should still stay well within the 50 percent peformance range.

Well distance adpative AF is reasonable efficent but I think in most of those cases your not really that GPU limited. Try doing it at 2048xwhatever and I think you will find the drops are much bigger. The oringinal incarnations of AF used the full number of samples everwhere but really you only need a high number of samples at anistropic sampling areas ( upclose its fairly isotropic ). Anyway the really savings ATI is getting atm is close to 2% as Dave said earlier so you might get 32x for free.

On a side not angle AF should be shot NOW yeah it wasn't too bad but now we are getting more and more vertexs which allows developers to have slopped surfaces at every possible angle rather then just vertical/horizonta surfacesl of the wolfenstien days. If the R500 or the NV50 have angle dependant AF I won't be happy.

Y'know... I just thought of something:

Would it be possible to do a temporal bilinear filtering to approximate a trilinear without any peformance hit? (sort of like Temporal AA)

Shift the bilinear stages ever so slightly across the horizon, and it should "blur" into an equivalent trilinear image.

Umm not really that wouldn't work becuase you would need atlest 64 different patterns which means you would need like 3800+ fps.

 

[MikeC]

How much of a performance hit will ATi take with adaptive AF disabled. Inform us all with your superior knowledge.

I don't think there is any proper way to nail down a number, it will depend totally on the situation and what game is tested.

I'm conducting some off the wall experimentation using AquaMark3, but it's a work-in-progress

http://www.nvnews.net/vbulletin/showthread.php?t=29077