Crysis 2 Tessellation Article

[AtenRa]

http://www.youtube.com/watch?v=jQydWGlMVeM&feature=related

Watch at 1:13 (Dynamic Tessellation)

The closer we get to the model the more tessellation we get

It is possible that Crysis 2 DX-11 patch uses Dynamic Tessellation and it seams like extensive tessellation when we have a look up close.

 

[Red Hawk]

These are the biggest features of DX11

1. Tessellation
2. Compute
3. Multithreaded rendering

Yes, for gaming DX11 is about tesselation.

...no it isn't. You said it right there. The developers of BF3 have been focusing a lot on compute functionality and how it pertains to lighting, draw calls, etc. DirectX 11 isn't all about tessellation.

 

[GoodRevrnd]

It is possible that Crysis 2 DX-11 patch uses Dynamic Tessellation and it seams like extensive tessellation when we have a look up close.

It's still excessive if you're applying it to objects that don't benefit from it. Character models are obviously a fantastic use of this feature. A concrete block that looks the same whether tessellation is used or not is not.

 

[3DVagabond]

If a game is designed to run Dx9/10 then tessellation is not a big plus. The models are already hires enough to stand on their own without tessellation.

 

[SirPauly]

Ideally, one would like to use tessellation where it would make more sense. It would be nice to hear from the developers on why this is the case.

I understand what the over-all tessellation hits are with each hardware based on HardwareFR, but how much do these few examples really do to the over-all bench?

 

[BFG10K]

I was not talking about why they tessellate unnecessary objects like the concrete barrier, I was only saying that because all other models/objects are already made of high polygon count, the developer was left with little choices as to what he could be able to tessellate.

What are you trying to say here? That’s it okay for a developer to tessellate useless things if they can’t tessellate something useful?

The barriers should not be getting tessellation, plain and simple. You can argue whether it’s a conspiracy but you can’t argue that an artificial tessellation workload is being applied for no good reason.

If they would chose to tessellate everything in the screen or human/alien models, the performance hit from that would made the game unplayable even with current high end hardware, so i believe they chose to tessellate smaller objects that already had been made with lower polygon count.

The purpose of tessellation is to enhance curved surfaces and/or to enhance other surfaces that start with a low polygon count but look better with higher polygon counts.

In either case the effect should be increased image quality. In the case of the concrete barriers and the ocean, that isn’t happening. Tessellation is being misused.

I was only talking technically and saying that Z-Buffer or Z-Culling will not cull polygons. As you point out, the card clearly tessellate that mesh but the Z-buffer (Culling) not allowing the rest of the rendering to be completed and been projected in the screen (no shaders, no rasterizing and no texturing is happening, only tessellation).

I don’t understand why you keep talking about the rest of the pipeline? That isn’t relevant at all. The issue here is that tessellation resources are being wasted. The ocean should be completely culled by the game so that no polygons are generated for it in the first place.

 

[WMD]

Ideally, one would like to use tessellation where it would make more sense. It would be nice to hear from the developers on why this is the case.

I understand what the over-all tessellation hits are with each hardware based on HardwareFR, but how much do these few examples really do to the over-all bench?

That would be nice to know. Apart from that. I am really curious what would be the performance difference between rendering the extra details using tessellation and using a more detailed base mesh in the first place. (like in case of witcher 2) I have a suspicion that tessellation may not even provide that much more of a performance advantage though it definitely allows smoother LOD transitions.

These "tessellation effects" are done in DX9 the traditional way. Performance in this level on my rig is around 90fps without Ubersampling on a single OC 5850

Compare this to Crysis 2 DX11. 27 fps on 6970 staring at nothing but a tessellated wall. (image from this thread: http://www.overclock.net/amd-ati/1053900-possible-amd-fix-crysis-2-dx11-2.html)

 

[AtenRa]

What are you trying to say here? That’s it okay for a developer to tessellate useless things if they can’t tessellate something useful?

The barriers should not be getting tessellation, plain and simple. You can argue whether it’s a conspiracy but you can’t argue that an artificial tessellation workload is being applied for no good reason.

I am not the developer of the game to explain why they tessellated the concrete barrier but I will ask you this, do you actually believe that AMD cards choke with that little concrete barrier being tessellated? I dont

The purpose of tessellation is to enhance curved surfaces and/or to enhance other surfaces that start with a low polygon count but look better with higher polygon counts.

In either case the effect should be increased image quality. In the case of the concrete barriers and the ocean, that isn’t happening. Tessellation is being misused.

That is what im trying to say, they only tessellated low polygon objects like the concrete wall in order to keep higher frame rates and I don’t argue that tessellating the concrete barrier was necessary but I believe they dint have more choices as to what objects to tessellate in the first place.

I don’t understand why you keep talking about the rest of the pipeline? That isn’t relevant at all. The issue here is that tessellation resources are being wasted. The ocean should be completely culled by the game so that no polygons are generated for it in the first place.

I was trying to show that technically if you have the wireframe and the developer apply tessellation in that mesh the card will not cull it(in that instance the water mesh) unless the developer completely remove it from there, the Z-buffer will only cull the rest of the pipe-line.

In the end, what I really trying to say is, creating a DX-9 game and applying tessellation in a few objects just to say that you have tessellation is not what developers should do and all those DX-11 patches they do more damage than produce a better outcome.

 

[BFG10K]

I am not the developer of the game to explain why they tessellated the concrete barrier but I will ask you this, do you actually believe that AMD cards choke with that little concrete barrier being tessellated? I dont

So it’s okay to waste resources as long as AMD cards don’t choke?

That is what im trying to say, they only tessellated low polygon objects like the concrete wall in order to keep higher frame rates and I don’t argue that tessellating the concrete barrier was necessary but I believe they dint have more choices as to what objects to tessellate in the first place.

Again, why tessellate something that reaps absolutely no benefit from tessellation? If they can’t tessellate something useful then they shouldn’t do it.

I was trying to show that technically if you have the wireframe and the developer apply tessellation in that mesh the card will not cull it(in that instance the water mesh) unless the developer completely remove it from there, the Z-buffer will only cull the rest of the pipe-line.

The game should be culling it in the first place. That’s my point. There’s absolutely no reason to tessellate a mesh that will never be seen.

 

[AtenRa]

So it’s okay to waste resources as long as AMD cards don’t choke?


Again, why tessellate something that reaps absolutely no benefit from tessellation? If they can’t tessellate something useful then they shouldn’t do it.


The game should be culling it in the first place. That’s my point. There’s absolutely no reason to tessellate a mesh that will never be seen.

In the end, what I really trying to say is, creating a DX-9 game and applying tessellation in a few objects just to say that you have tessellation is not what developers should do and all those DX-11 patches they do more damage than produce a better outcome.

 

[3DVagabond]

As the model gets more complex the more work it is to apply tessellation. Especially after the fact. That's why they added it to flat models with few edges. It can be done with POM. The Ocean would have been a new model. Not a very complex bit of work.

 

[WMD]

As the model gets more complex the more work it is to apply tessellation. Especially after the fact. That's why they added it to flat models with few edges. It can be done with POM. The Ocean would have been a new model. Not a very complex bit of work.

Its funny that most of the visual details on the ground and some of the walls are actually done with POM which doesn't even need DX11. While most of the tessellated objects looks the same without tessellation. Turning off tessellation by lowering objects from ultra turns off POM as well.

 

[Red Hawk]

That would be nice to know. Apart from that. I am really curious what would be the performance difference between rendering the extra details using tessellation and using a more detailed base mesh in the first place. (like in case of witcher 2) I have a suspicion that tessellation may not even provide that much more of a performance advantage though it definitely allows smoother LOD transitions.

These "tessellation effects" are done in DX9 the traditional way. Performance in this level on my rig is around 90fps without Ubersampling on a single OC 5850

Compare this to Crysis 2 DX11. 27 fps on 6970 staring at nothing but a tessellated wall. (image from this thread: http://www.overclock.net/amd-ati/1053900-possible-amd-fix-crysis-2-dx11-2.html)

Bad comparison. First of all, the Crysis 2 screenshot is at a higher resolution than the TW2 screenshot. Secondly, the Crysis 2 screenshot has a closer viewpoint than the TW2 screenshot. It might be generating more detail because it's that close.

And I'm not really buying that you get that good performance on The Witcher 2 with everything but ubersampling on, it's one of the most infamously taxing games.:hmm:

 

[Carfax83]

You'd have to ask CryTek. Personally i think it is stupid to assume that Nvidia would cripple their own midrange cards' performance just to make "AMD look bad". That is a conspiracy mindset and ignores that it is asinine for business - and Nvidia is profit-oriented.

People are assuming that this patch is the last one. According to interviews, evidently it is not. i expect further optimizations.

This. People just love a good conspiracy theory though, and have already decided that Nvidia and Crytek just wanted to fuck over anyone with AMD or mid range Nvidia hardware.

 

[WMD]

Bad comparison. First of all, the Crysis 2 screenshot is at a higher resolution than the TW2 screenshot. Secondly, the Crysis 2 screenshot has a closer viewpoint than the TW2 screenshot. It might be generating more detail because it's that close.

And I'm not really buying that you get that good performance on The Witcher 2 with everything but ubersampling on, it's one of the most infamously taxing games.:hmm:

Closer viewpoint does not mean less fps, infact it is usually the opposite. All the gpu does is mostly render the wall and nothing else. Open areas usually have much lower fps in games. That witcher 2 image is from the internet as I do not have a screenshot of that level. I run the game at 1680x1050 while the crysis 2 screenshot is done at 1920x1080. But keep in mind that a 6970 @ 1920x1080 is still much much faster than a 5850 @ 1680x1050.

It does run 90-110fps in those dungeons at my settings (ubersampling, motion blur and DOF off at 1680x1050). With everything ultra except UBS, it still runs over 60fps in every level except the biggest fight. The game isnt very taxing on the gpu actually and performance is alot better with the latest 11.8 drivers.

Some of fraps captures from my PC. This is from a much more taxing level.

 

[BenSkywalker]

http://techreport.com/r.x/crysis2/barrier-dx9-full.jpg

http://techreport.com/r.x/crysis2/barrier-dx11-full.jpg

This is what people are using as an example of what isn't needed? Seriously?

Where to start- Look at the top of the barrier where the hooks are to move it. On the shader setup the near hook is washed out and flat looking, the second hook almost vanishes completely due to gamma impacting the shader routine too much. The geometry example is nice and clear on both. The plate on the side that has marking on it on the shader example blurs to be one blob with the concrete within a short distance off from camera, the geometry example offers a crisp differentation from near field all the way through. These aren't something you need to do a pixel compare on, these aren't even remotely close. If you want to get into finer levels of detail you can still see the farthest hook on the second barrier in the geometry example.

A lot of people on these forums gripe about the differences betwee then uber high AA settings and the AF optimizations used by either company. If you can not instantly see a very clear difference in the barriers in these shots then you seriously need to shut your mouths on fine IQ differences. You can do a per pixel breakdown if you'd like, but these tesselation shots show significantly greater differences then moving from 4x to a hybrid 8x AA mode or using a better optimized AF routine- it isn't even remotely close.

Anyone who thinks that this level of detail improvement isn't worth the performance impact needs to leave the video forum if they care about anything other then cheerleading their team of choice. We are deep into diminishing returns, IQ improvements are going to keep getting smaller and are going to continue to get even more demanding on hardware. Wait until you see the kind of performance hit radiosity has versus shader based lighting compared to the IQ difference. Seriously people, if you can't clearly see the by comparison huge differences between the barriers, it's time for you to get a new hobby.

The game should be culling it in the first place. That’s my point. There’s absolutely no reason to tessellate a mesh that will never be seen.

Come on BFG, the kids may not know but I know you know that it is inate on the render sort order of the engine if they are capable of doing that or not. If the render sort isn't set for front to back on a given item then they can't cull it. That is outside of the codepath of a render routine, that is a core element of the engine.

 

[Red Hawk]

The DX 11 screenshot may look a teensy bit better because it's closer to the concrete wall than the other...that's about it. Nothing to get excited about and nothing that's worth the performance drop.

 

[Lonbjerg]

http://techreport.com/r.x/crysis2/barrier-dx9-full.jpg

http://techreport.com/r.x/crysis2/barrier-dx11-full.jpg

This is what people are using as an example of what isn't needed? Seriously?

Where to start- Look at the top of the barrier where the hooks are to move it. On the shader setup the near hook is washed out and flat looking, the second hook almost vanishes completely due to gamma impacting the shader routine too much. The geometry example is nice and clear on both. The plate on the side that has marking on it on the shader example blurs to be one blob with the concrete within a short distance off from camera, the geometry example offers a crisp differentation from near field all the way through. These aren't something you need to do a pixel compare on, these aren't even remotely close. If you want to get into finer levels of detail you can still see the farthest hook on the second barrier in the geometry example.

A lot of people on these forums gripe about the differences betwee then uber high AA settings and the AF optimizations used by either company. If you can not instantly see a very clear difference in the barriers in these shots then you seriously need to shut your mouths on fine IQ differences. You can do a per pixel breakdown if you'd like, but these tesselation shots show significantly greater differences then moving from 4x to a hybrid 8x AA mode or using a better optimized AF routine- it isn't even remotely close.

Anyone who thinks that this level of detail improvement isn't worth the performance impact needs to leave the video forum if they care about anything other then cheerleading their team of choice. We are deep into diminishing returns, IQ improvements are going to keep getting smaller and are going to continue to get even more demanding on hardware. Wait until you see the kind of performance hit radiosity has versus shader based lighting compared to the IQ difference. Seriously people, if you can't clearly see the by comparison huge differences between the barriers, it's time for you to get a new hobby.



Come on BFG, the kids may not know but I know you know that it is inate on the render sort order of the engine if they are capable of doing that or not. If the render sort isn't set for front to back on a given item then they can't cull it. That is outside of the codepath of a render routine, that is a core element of the engine.

+1

 

[Arkadrel]

Look at the top of the barrier where the hooks are to move it.

But isnt it possible to get the same effect, if the only part of that concrete that uses tesselation is the hooks?

Because useing tessellation on a FLAT object, doesnt make anysense (doesnt improve the image quality). Even the cracks in the object is just the texture coloured in. I could understand if they wanted to use tessellation on the concrete slab, so they could make the cracks in it more detailed useing tessellation (BUT THEY DONT), so what is the point of haveing the intire thing tessellated to that degree?

Useing very high tessellation levels on a flat surface of a concrete block makes no sense. Useing it only on the "hook's" you want to stand out, makes more sense.

Its pretter with it (dx11) than without (dx9), no question about it, but is it wrong to ask game developers to think about haveing their game's be optimised to run the best they possibly can? and not needlessly waste GPU resources? But instead to use them only in the area's where the biggest return in image quality is?

Easy to make a pretty game, that no one has the hardware to run.
Harder to make a pretty game, optimised so more people can run it.

being lazy with tessellation is a bad thing.

If this is what happends when you dump some money on a game developer, its a bad way to go about things. The fault lies with the lazy game developers though.

 

[3DVagabond]

The fault lies with the lazy game developers though.

Yeppers! :thumbsup:


Actually, cheap game devs. They cheaped out on the patch. They needed to pay for some new models. Looks like they just reworked some of the scene props and the ocean.

 

[Lonbjerg]

If this is what happends when you dump some money on a game developer, its a bad way to go about things. The fault lies with the lazy game developers though.

No, this is what happens when AMD drop the ball on tesselation, making a fixed function hardware solution that is slower performing than NVIDIA's unified approch.

Before "fermi" AMD hyped tessellation as the sizzle...now they really to to down play it and even talks about to much tesselation.

It's also what happens when one company really spend time and money on developer relations...AMD need to take a cue...and not only from a hardware perspective.

They have realized now, that VILW isn't the best solution to GPGPU...or physics.

Those things really matters in the future...hence AMD is going SIMD...just like NVIDIA, NVIDA just realized it sooner.

That is the fact.

Your "claim" is just undocumented fanboy speculation...I would like to see you document your claim...but my guess is that will never happen.

 

[3DVagabond]

No, this is what happens when AMD drop the ball on tesselation, making a fixed function hardware solution that is slower performing than NVIDIA's unified approch.

Before "fermi" AMD hyped tessellation as the sizzle...now they really to to down play it and even talks about to much tesselation.

It's also what happens when one company really spend time and money on developer relations...AMD need to take a cue...and not only from a hardware perspective.

They have realized now, that VILW isn't the best solution to GPGPU...or physics.

Those things really matters in the future...hence AMD is going SIMD...just like NVIDIA, NVIDA just realized it sooner.

That is the fact.

Your "claim" is just undocumented fanboy speculation...I would like to see you document your claim...but my guess is that will never happen.

AMD didn't write the articles. It was independent review sites. They are the ones pointing out too much tessellation.

As far as AMD switching architectures goes, it's a new process. Because they've changed for the upcoming process doesn't mean that the VLIW5/4 were wrong for the older processes. It was far more efficient both perf/watt and perf/mm^2 than nVidia's. On 28nm they can have far more GCN type "cores" than they could have on previous processes. That's why it makes sense to change now, where it didn't before. Consider they went from VLIW5 to VLIW4 @ 40nm because it was getting too difficult to feed all those shader units. Now imagine 28nm and trying to feed +3000 SPU. Now is the correct time to make the change.

Really no reason to get personal and call people names just because you disagree with them.

 

[Jacky60]

This. People just love a good conspiracy theory though, and have already decided that Nvidia and Crytek just wanted to fuck over anyone with AMD or mid range Nvidia hardware.

Why do you call it a conspiracy as if businesses don't do underhand things to adversely affect their competition. The point is (as you appear to have completely missed it) that Nvidia hardware suffers less of a performance hit from all this completely unnecessary tessellation than does amd hardware thus artificially benefitting Nvidia and attendant benchmarks disproportionately. It's not that difficult to understand for someone over eight years old without resorting to calling whistleblowers conspiracy theorists. Believe it or not some companies and even governments say things in public and do the opposite in private. I guess you call that a conspiracy theory.

 

[Jacky60]

http://techreport.com/r.x/crysis2/barrier-dx9-full.jpg

http://techreport.com/r.x/crysis2/barrier-dx11-full.jpg

This is what people are using as an example of what isn't needed? Seriously?

Where to start- Look at the top of the barrier where the hooks are to move it. On the shader setup the near hook is washed out and flat looking, the second hook almost vanishes completely due to gamma impacting the shader routine too much. The geometry example is nice and clear on both. The plate on the side that has marking on it on the shader example blurs to be one blob with the concrete within a short distance off from camera, the geometry example offers a crisp differentation from near field all the way through. These aren't something you need to do a pixel compare on, these aren't even remotely close. If you want to get into finer levels of detail you can still see the farthest hook on the second barrier in the geometry example.

A lot of people on these forums gripe about the differences betwee then uber high AA settings and the AF optimizations used by either company. If you can not instantly see a very clear difference in the barriers in these shots then you seriously need to shut your mouths on fine IQ differences. You can do a per pixel breakdown if you'd like, but these tesselation shots show significantly greater differences then moving from 4x to a hybrid 8x AA mode or using a better optimized AF routine- it isn't even remotely close.

Anyone who thinks that this level of detail improvement isn't worth the performance impact needs to leave the video forum if they care about anything other then cheerleading their team of choice. We are deep into diminishing returns, IQ improvements are going to keep getting smaller and are going to continue to get even more demanding on hardware. Wait until you see the kind of performance hit radiosity has versus shader based lighting compared to the IQ difference. Seriously people, if you can't clearly see the by comparison huge differences between the barriers, it's time for you to get a new hobby.



Come on BFG, the kids may not know but I know you know that it is inate on the render sort order of the engine if they are capable of doing that or not. If the render sort isn't set for front to back on a given item then they can't cull it. That is outside of the codepath of a render routine, that is a core element of the engine.

QFT you're talking bollocks. Either that or you have super high resolution eyes because my girlfriend could see no difference until I pointed out the handles on top. If you really believe there's a significant difference YOU should leave the video forums and go and work for NASA as a replacement for the hubble space telescope.

 

[SirPauly]

imho,

All it takes is more impressive tessellation performance from AMD to knock off conspiracy theories. It has been AMD's tessellation that is the competitive disadvantage.

edit: Just to reinforce that idea, a guy at Hardocp sent a 6970 to XFX for RMA, he got a 6770 in return, AND a coupon for a "soon to be released next generation graphics product of similar performance (to his 6970). This was because XFX said they had no 6970's on hand.

http://www.rage3d.com/board/showpost.php?p=1336675956&postcount=30

Hopefully, the newer generation from AMD will perform better.