Pixel or Vertex more important, looking forward?

[MfA]

I guess I forgot a comma in that first sentence, still Chalnoth grasped what I meant

Doing occlusion tests on the fly while doing front to back rendering can do the minimum amount of rendering, you can potentially save on occlusion tests by rendering formerly visible nodes first ... on the other hand, you can still do unnecessary rendering.

Obviously doing occlusion tests on the fly isnt an option if it isnt very tightly bound with rendering ... that is why I said I wanted it on the card though. Otherwise latency is going to make it impossible, latency might still make it impossible but it is hard to make an educated guess. There is a lot of room for optimization (you can render to different parts of a screen in succession to buy you some time ... and you would of course never stall on occlusion tests, at worst they would just have used not up to date Z-buffer data).

Personally I think the actual occlusion tests are irrelevant as far as work is concerned. If so the only thing which is relevant is if their efficiency has to be compromised because of latency with front to back rendering, and if that problem is insignificant the need for temporal coherency goes out the window.

 

[GameCat]

Ah, now I get where the comma should have been. Anyway, there is no need to render objects not tested first, that is probably a bad idea actually, since it reduces the latency hiding you get by mixing occlusion tests with regular rendering. If there was no latency, like with the evaluation of occlusion test results on the card, this wouldn't matter much, agreed. But since we DO have to hide the latency, schemes exploiting temporal coherence do better. And in my testing, the cost of occlusion queries wasn't insignificant even when no stalls occurred, although this might be due to an inefficient AGP bus among other things. Anyway, this is a bit OT for this thread , feel free to check out my Master's Thesis Optimal occlusion testing for more detials on the various schemes I have tested.

 

[hovz]

oh god....dont give me this bullshit about doom 3 is held back by current video cards.

Um, nobody did.

ailuros did, thats who i was referring to.

ailuros, textures drain our peformance because developers dont take advantage of any features of newer chips. every game just has more/bigger textures while the rest of the chip sits idle. id much rather run an advanced game at 1027 res with no aa and af at 50 fps then some 3 year old looking game at 1600 res with high aa and af at 60 fps.

btw, how can doom3 by taxing our hardware when the 6800 u can run it at well over 100 fps at standard res???

 

[Ailuros]

Ever noticed that you have a way too intense obsession with just one game? Apparently I'm not the only one that notices it.

Excuse me, but many people bring up Doom3 in discussions about graphics. Not because they're obsessed, but because Doom3 is one of the latest and greatest games.
And just because people don't agree with the criticism I have on Doom3/Carmack, doesn't mean I'm wrong.

No you're not wrong, but it's just ONE game, at least for the time being. I'd prefer a far wider approach on the matter and that would include what is available on shelves today as a whole.

Yes, but I see that it's mostly about vegetation. If you take a single mesh model, and insert extra polys in it, you can get a whole lot more mileage out of the vertexcache and things, than with a set of sprites and other low-poly meshes.

Triangle count is still extremely high.

What exactly does the CPU have to do with adding polys to meshes at modeling time? Or what are you talking about?
As for Sweeney, a few years ago he believed we would be back to software rendering by now. UE3.0 is a nice product, but let's just say I don't think Sweeney is a big visionnaire.

Just because he has some weird ideas about the future, doesn't make each and every of his standpoint invalid or useless either. I usually don't agree to many of his points either, yet this one makes sense to me.

Another thing about 3DMark is that it can generally run okay (30 fps or more) if you have the fastest card available at that time. This was true for 3DMark2000 with a GF2Ultra, for 3DMark2001 with a GF3, for 3DMark03 with a 9700Pro, and I suspect this one to run okay on a 6800U or X800XTPE.

Okay is exactly the "satisfactory" point I was thinking of. It still means 1024 noAA/AF with occassional single digit minimum framerates for a 9700PRO and Mother Nature as an example.

While you will get "satisfactory" performance on a R300 in games that share the predicted analogy more or less of 3dmark03, you're still way better off with a R420 on contrary. In this case we're talking about the high end segment of the market and there high end users rarely keep hardware for too long.

 

[Ailuros]

ailuros did, thats who i was referring to.

No I didn't in that sense. There's a difference between the messenger and the receiver obviously.

ailuros, textures drain our peformance because developers dont take advantage of any features of newer chips.

That's a very interesting point; mind if I don't agree with it? Find an application that can measure trilinear fill-rates with different amounts of texturing stages and run a couple of tests with filtering optimisations on and off.

I'll give it another shot:

FarCry

sprites --> no sprites = 2x the poly count = about -15%

4xAF optimized --> 4xAF full trilinear = about -30%

every game just has more/bigger textures while the rest of the chip sits idle. id much rather run an advanced game at 1027 res with no aa and af at 50 fps then some 3 year old looking game at 1600 res with high aa and af at 60 fps.

Ok so you don't like image quality improving features and prefer aliasing fests and blurry textures. It's a matter of taste. But that's not what some people buy high end GPUs for and that for the past couple of years.

btw, how can doom3 by taxing our hardware when the 6800 u can run it at well over 100 fps at standard res???

Where did I say that it is in fact taxing the current hardware to the absolute maximum possible? If it would have the polycounts of let's say Far Cry it most likely would come very close to that.

However I will still insist on my former point concerning IQ improving features.

The 6800U gets ~50 fps in 1280*1024 with 4xAA/high quality. If it would reach the polycounts of other games not only would achieving the former score be fairly impossible, yet you couldn't also run it on a GF4MX in 640/low detail with the same ease as today.

 

[hovz]

ailuros did, thats who i was referring to.

No I didn't in that sense. There's a difference between the messenger and the receiver obviously.

ailuros, textures drain our peformance because developers dont take advantage of any features of newer chips.

That's a very interesting point; mind if I don't agree with it? Find an application that can measure trilinear fill-rates with different amounts of texturing stages and run a couple of tests with filtering optimisations on and off.

I'll give it another shot:

FarCry

sprites --> no sprites = 2x the poly count = about -15%

4xAF optimized --> 4xAF full trilinear = about -30%

every game just has more/bigger textures while the rest of the chip sits idle. id much rather run an advanced game at 1027 res with no aa and af at 50 fps then some 3 year old looking game at 1600 res with high aa and af at 60 fps.

Ok so you don't like image quality improving features and prefer aliasing fests and blurry textures. It's a matter of taste. But that's not what some people buy high end GPUs for and that for the past couple of years.

btw, how can doom3 by taxing our hardware when the 6800 u can run it at well over 100 fps at standard res???

Where did I say that it is in fact taxing the current hardware to the absolute maximum possible? If it would have the polycounts of let's say Far Cry it most likely would come very close to that.

However I will still insist on my former point concerning IQ improving features.

The 6800U gets ~50 fps in 1280*1024 with 4xAA/high quality. If it would reach the polycounts of other games not only would achieving the former score be fairly impossible, yet you couldn't also run it on a GF4MX in 640/low detail with the same ease as today.

running tilinear fillrate tests proves that our current cards are memory bandwidth limited. instead of lazily adding more textures they could render using shaders, and actually take advantage of the extreme processing power these cards have, rather than lazily adding an ugly texture to every surface and calling it an advanced engine.

your comment about different number of texture stages is my point. its 2004 and developers are still desiging game engines around flat walls with textures on them. thats why the graphics of newer games for the most part, havent improved much at all in the last few years. far cry was the biggest leap visually imo, but that was only because they somewhat pushed the cards with actual gemoetry, not just a bunch of layered textures. to achieve lifelike graphics you NEED extremely high polygon counts. rather then putting a texture they could use material shaders. that would take all the strain off the memory bus and fillrate and put some of it to use in the pixel shaders which go idle or work at 10% load most of the time. i wont even mention the vertex shaders which basically sit idle during almost every game.

 

[KimB]

running tilinear fillrate tests proves that our current cards are memory bandwidth limited.

Except trilinear filtering takes more fillrate than memory bandwidth.

 

[hovz]

http://www.digit-life.com/articles2/radeon/r420.html

r420 is bandwidtt limited unless processing many textures, even with tri

 

[KimB]

And, if you'll notice, the texel rate increases when trilinear filtering is used. This means that trilinear filtering takes more fillrate than it does memory bandwidth.

 

[hovz]

and no game uses anywhere near the fillrate that the xtpe has unless your in an extremely high res with high levels of aa and af so whats ur point? at the fillrate levels games use at normal resolutions and no aa and af its extremely memory limited.

 

[hovz]

or i should say cpu limited cuz developers still use software rendered low poly low detail everything and leave the gpu idle

 

[Ailuros]

hovz,

running tilinear fillrate tests proves that our current cards are memory bandwidth limited. instead of lazily adding more textures they could render using shaders, and actually take advantage of the extreme processing power these cards have, rather than lazily adding an ugly texture to every surface and calling it an advanced engine.

Chalnoth is right it's more about fill-rate in that case. I'm not sure what you're proposing exactly. Textures might eventually get replaced (if that's what you meant) through shaders (math tables), yet it'll still take time. I can't imagine how you could run such a game on a DX7 or DX8.1 accelerator today.

In that case either the developer would have to sit down and code two versions of the same game, which isn't much of an option or a very high persentage of users out there would get excluded right now. I doubt that the majority out there has dx9.0 class accelerators even today.

your comment about different number of texture stages is my point.

I don't think we meant the same thing. If I'm to mention fill-rate today then I'll have to concentrate on multitexturing or more specifically on quad-texturing at least.

far cry was the biggest leap visually imo, but that was only because they somewhat pushed the cards with actual gemoetry, not just a bunch of layered textures.

Yet they could have done a way better job after all and it hasn't all that much to do with texturing either. If Crytek would have optimised the game a bit better, it wouldn't waste as much GPU and CPU resources after all.

to achieve lifelike graphics you NEED extremely high polygon counts.

Not only. Give me all the geometry you want, if I get entirely unfiltered and aliased output in the end, it's anything but lifelike to me. I'd rather have it all step by step.

i wont even mention the vertex shaders which basically sit idle during almost every game.

They most certainly don't sit idle in games like UT2k3/4 or any game that's using the U2 engine as a simple example. Disable in the game's ini the hardware T&L and VS settings and you'll end up with half the performance.

By the way it's my understanding that it's way easier to discard pixel than vertex shader data overall.

rather then putting a texture they could use material shaders. that would take all the strain off the memory bus and fillrate and put some of it to use in the pixel shaders which go idle or work at 10% load most of the time.

The sollution isn't when a head hurts to cut it off. Transitions take time.

 

[Ailuros]

and no game uses anywhere near the fillrate that the xtpe has unless your in an extremely high res with high levels of aa and af so whats ur point? at the fillrate levels games use at normal resolutions and no aa and af its extremely memory limited.

First of all games that request trilinear should get trilinear and that goes for both IHVs. Just because an accelerator ends up doing half the work doesn't mean that it's got fill-rate to waste. If that would be the case there wouldn't be a single reason for optimisations to start with.

Multisampling is in relative terms fill-rate "free", while AF isn't. It gets almost there with the filtering related optimisations (including angle-dependancy).

I'm not sure what you want lower resolutions for either. High resolution gaming is a totally different experience, otherwise I could easily fetch a console and call it a day. There's a reason why I consciously use the PC for gaming and to that with a 21" CRT.

 

[hovz]

god this is going to be a test of patience...who gives a fuck about trilinear ops...that has nothing to do with this.

look. would u rather play a game like doom 3 at 1076 res with no aa or af? or a game like quake 3 at 1600 with high aa and af? does that make any sense to you?

obviously i didnt mean sit completely idle, but if you think any games on the old ass unreal tournament 2003/4 engine are stressing these cards at all u need help.

i realize we dont have the power to totally phase out textures, but we have power to do alot more than we are now.

your point about dx9 class cards is exactly what i mean. developers code with the lowest common demonimator in mind as the base of their engine, and just add higher res textures and a few effects to scale up. that is NOT a scalable engine. a scalable engine would scale down the shaders, polygon complexity, etc. not just blur the textures and lower the res. how can you expect graphics to ever advance with the mentality of current developers?

my comment about high poly was directed at everyone who says high poly isnt the wave of the future. totally retarded train of thought.

yes change is gradual, but there has been almost zero change in the last 4 years.

 

[KimB]

It doesn't help when you're wrong on pretty much all of your talking points. For example:

rather then putting a texture they could use material shaders. that would take all the strain off the memory bus and fillrate and put some of it to use in the pixel shaders which go idle or work at 10% load most of the time.

Material shaders don't get rid of, or even reduce, the number of textures typically used.

 

[hovz]

SIGH

 

[Ailuros]

god this is going to be a test of patience...who gives a fuck about trilinear ops...that has nothing to do with this.

I'm showing already more patience than I usually do and no I don't agree with you in many departments. Do you really need that attitude? You're not going to force your personal opinion on me at the end of the day.

look. would u rather play a game like doom 3 at 1076 res with no aa or af? or a game like quake 3 at 1600 with high aa and af? does that make any sense to you?

Yes it does make sense. I play today Doom3 in 1280*1024 with AA/AF and Q3a in 2048*1536 with AA/AF. In fact there isn't a single game I haven't played in the past years in a lower resolution than 1152*864*32 and apart from some corner cases where MSAA wouldn't work, there was always a pinch of AA/AF added to the mix.

obviously i didnt mean sit completely idle, but if you think any games on the old ass unreal tournament 2003/4 engine are stressing these cards at all u need help.

It still remains the best multiplayer FPS I have on my system right now, with a very live and active 3rd party mapping community. It was a perfectly feasable point, since according to you VS units sit entirely idle; actually they don't because many games out there have T&L optimized code these days.

If anything else you have 4 VS units @ 380MHz on your R350, delivering far more than a 3.0 GHz P4.

i realize we dont have the power to totally phase out textures, but we have power to do alot more than we are now.

I still don't see the hardware available for that right now. I'd love too to have far more demanding games available, but I'd also love to have the according hardware too.

your point about dx9 class cards is exactly what i mean. developers code with the lowest common demonimator in mind as the base of their engine, and just add higher res textures and a few effects to scale up. that is NOT a scalable engine. a scalable engine would scale down the shaders, polygon complexity, etc. not just blur the textures and lower the res. how can you expect graphics to ever advance with the mentality of current developers?

You won't get any shader functionalities with a GF4MX today, if you'd play FarCry or any other game containing shaders out there. As for textures, here's another point why we need both more advanced hardware and underlying API in the form of dynamic on chip LOD, to tax the CPU even less. That's another advantage of a Geometry Shader and/or PPP in WGF.

my comment about high poly was directed at everyone who says high poly isnt the wave of the future. totally retarded train of thought.

I didn't see anyone even hinting anything close to that; rather the contrary.

yes change is gradual, but there has been almost zero change in the last 4 years.

Almost zero is a good one, especially considering how many average polys games had in 2000 and how they overall looked like.

If I'd have a reason to complaint, then it would be the marginal advancements in gameplay or original ideas for that department, but that's entirely OT.

 

[Laa-Yosh]

Material shaders don't get rid of, or even reduce, the number of textures typically used.

It's safe to say that shaders increase texture data. Instead of a single color map, you'll get a color, a normal and aspecular map - and that's for an average material, but you could easily end up with 4-5 or more textures per surface.
And procedural fractals won't look good enough in 95% of games, so don't count on them.

 

[KimB]

And procedural fractals won't look good enough in 95% of games, so don't count on them.

Well, I don't think it could be broken down into which games procedural materials look good in, but procedural materials only do look good for a select few types of materials (typically relatively chaotic materials with an underlying pattern, such as wood and marble).

 

[Scali]

Well, I don't want to destroy the nice flamewar here... but I think both sides are saying pretty much the same, but approaching it from a different angle.

I believe one side stresses on the quality of textures and AA/AF to increase realism... And the other side stresses on geometric detail.

However, if I understand hovz correctly, he is saying that we already have high-quality textures and good AA/AF levels, so adding more geometry would be nice for now. But it seems that the others interpret it as trading texture/image quality for geometry.

I agree with hovz. Personally I don't need more than 1024x768, as long as I have decent textures and something 4xAA/8xAF. My Radeon 9600Pro already handles this quite nicely in most games.
However, if you look at the polycount today, and say, 4 years ago, it has pretty much stagnated, and that's rather sad. Agreed, the same polycount looks better today, because of the better textures and normalmaps and pixelshading, but still, the silhouettes are blocky, and animation is a tad limited. So yes, I would be happy if this area were explored, and texture quality and resolution remained at the current level for now.