s3o vs 3do

[Argh]

Yeah, I tried it something like a year ago... and it performs worse and can lead to various draw errors.

BTW, please post the object, so that we can repeat your experiment. It really looks like bad DDS settings to me, tbh- I don't lose that much detail.

[Saktoth]

The effect is probably exaggerated by the fact that its using a 3do texture: remember its only like 64x64. Mipmaps are going to start cutting into that something chronic, the plate lines are only three pixels thick, two for black one for white. It has this 64x64 texture plastered over one whole side of the model. Nevertheless, the textures made for s3o are quite often using only 1-pixel details.

Nevertheless, thats about the texture space you'd expect on the smaller plates of the model, and the loss is visible there too. And the larger range detail is lost too: Even the gradient from dark to light has lost contrast.

Are we truly condemned to have s3o always just look crappier than 3do? This is incredibly depressing (especially after all the work everyone has put into transitioning CA into s3o, thinking it a superior technology )...

The models in that image can be found in this commit to the CA SVN/test builds

[Argh]

Well, I have 128s in World Builder, and it performs reasonably well. I think that the best solution for this is to have a specification for when we don't use a mip- OpenGL can, IIRC, be forced to keep within a certain range of mips. I know for a fact that Freelancer (albeit a DirectX title) had that rigidly defined, so that you could use low-resolution mipmaps within certain ranges- and they had a control, per object, for at what distance lower mips would be used. Anyhow, I'll take a look at the test geometry, see what can be done.

[Argh]

OK, did some tests.

1. On my hardware, in UpSpring both S3O and 3DO behave identically- they're sharp up close but it quickly degrades over distance.

2. There were no changes seen when I disabled mipmap generation for the DDS.

So, I'm wondering how much of this is Spring handling 3DO differently (I thought that the textureatlas had mipmaps auto-generated for performance purposes).

[Argh]

Did some more tests. 3DOs aren't looking a great deal different on this hardware, with 2X AA and 2X AF.

Sorry but I just don't see a striking difference that would be really noticeable to players, tbh. I'm wondering how much of this is drivers and settings, basically.

[smoth]

any difference is a bit odd though.

[Saktoth]

In Upspring?! Upspring is completely irrelevant, it handles the models very differently.

This is the same problem i've been complaining about 2 graphics cards ago. Changing settings and drivers isnt a solution, unless CA/Spring is going to come with updated drivers every time it installs.

The game looks like arse.

Anyway, if you can post your results with your hardware of those same models, i would be curious to see if you do not get the same results as i do. Unless you contend that the images i posted do not offer a 'striking difference noticeable to players', in which case i would vehemently disagree.

[smoth]

Unless you contend that the images i posted do not offer a 'striking difference noticeable to players', in which case i would vehemently disagree.

+1

[Argh]

I'm not saying that there isn't a problem. I'm just saying that I just didn't see anything really dramatic here.

I mean... meh, it's easier to show with pictures, please click to view at full resolution:





These are all using 256 skins, so it's roughly comparable to a 3DO (though not really, since it's shoving that giant texture through for every object, so it's actually inferior- the "64X64" is an illusion).

Even tiny lines are clearly visible (this is especially true when the objects are in motion or the POV changes). I don't really see some giant quality gap. I'm not saying that I wouldn't like manual control over mips, it just doesn't seem all that pressing from here.

[CarRepairer]

Argh, what brand of automobile is that? I've never seen it before.

[smoth]

those are high contrast textures. Also not what saktoth asked.

[Argh]

I have not adjusted the contrast, either with Photoshop or via driver settings.

And yeah, I think it shows my point pretty clearly, tbh. If we can perceive all those details, then where's the problem?

Oh, and the odd cars are a tribute to Syndicate.

[smoth]

you have black on yellow

that is high contrast

not a setting

it is that there is a great amount of distance between black, a color associated as negative space on your positive foreground color of yellow.

black on a mid-level grey would show more. Also those panel lines are much larger.

[Argh]

There are a lot of different colors of car there.

And the panel lines are 1 pixel wide. You're seeing specular highlights, which provide us with depth cues- that's the shader, doing what it's supposed to do.

But as I said... it's not like there isn't a problem.

If nothing else, it would be nice to set a maximum mip level that will get used, stop it from degrading into mush with stuff that's pushing the lower boundaries. with a 64 / 64 skin, for example, you probably never need to see mip1, it's just not that big of a performance boost and it's going to degrade too much.

[smoth]

so are mine...



but they are much smaller in relation to the size of the model....

[Argh]

Pixel vs. texel size is just par for the course, though. There's no way to avoid that, texture filtering and all that implies are things that can be adjusted only small amounts without rapidly hitting driver hell.

[Pressure Line]

Anyway, if you can post your results with your hardware of those same models, i would be curious to see if you do not get the same results as i do.

gratz on continuing to avoid answering the question/request Argh. Now do what the nice man says and repeat his experiment and post some pics.

[Argh]

Look... if you want to see what Spring looks like, and what happens to performance, when you set the LOD Bias low, just find TextureLODBias in SpringSettings.cfg and set it to -8 or thereabouts. Make sure your card's driver allows negative LOD bias, and that that flag is set- not every card does. Then test it for yourselves- if nothing else, it should be interesting to see who's hardware shows any real improvement, and under what circumstances.

What I suspect you will see is:

1. There isn't a massive improvement in scene quality. This isn't a magic wand.

Here's the same shot, LOD bias -8.



So, we're at mip0 for just about everything in the scene. Where's the huge difference in quality? There isn't any.

Why? Because ultimately the size of the texels means that the increased resolution is meaningless. There's more noise, that's about all. The only place I saw any improvement was transparent stuff in the very low range- 128s with very narrow transparent areas. But the amount of improvement was so slight that it certainly wasn't worth the performance hit.

2. Performance sucks. At LOD bias -8, framerate was halved, on my hardware.

3. There's noticeable shimmer and moire problems on things that use grids or lines. That's probably really noticeable if staring at 3DO, come to think, if you know what you're looking for.


So, there ya go- a test of what happens when you get rid of mip levels. Feel free to repeat it and let us see some screens of the results. This may be one of those things where, depending on the hardware / driver, it may make a difference.

[Beherith]

What tools do you use to make your dds files for your textures? Because they usually have tons of settings about how to generate mipmaps. I think the default is a box filter, which could lead to crappy results.

Saktoth, would you mind playing around with the nvdxt.exe settings on making the dds files from the sources?

[Argh]

Hey, that's a thought. Maybe what we really need is more sharpness. Lemme play with my settings, see if I can get an improved result. The settings I use as standard work well, but if that's really the issue, then setting up more aggressive sharpening might really help.