New 3Dc demo

[Humus]

It's been a while since the last time, but now I got another demo ready.

The main features are the usage of 3Dc and the use of detail normal mapping. The rest is standard stuff I've done before.

Available at the usual place:
http://esprit.campus.luth.se/~humus/

It will run even if you don't have 3Dc, but then DXT5 and uncompressed is the only options. On X800 cards you can compare quality the quality between 3Dc and DXT5.

The performance difference is interesting:

With shadows:
No compression: 125fps
3Dc: 146fps (+17%)
DXT: 136fps (+9%)
3Dc & DXT: 158fps (+26%)

Without shadows:
No compression: 164fps
3Dc: 210fps (+28%)
DXT: 195fps (+19%)
3Dc & DXT: 239fps (+46%)

 

[jvd]

u sure it takes off 3dc . I can't tell any diffrence haha . Anyway I"m getting about 300fps with all set to true with a 2500+ athlon xp and a x800xt at 1027x768 full screen mode

 

[991060]

Why there is "3Dc & DXT", you used them together? Isn't it that only method can be used at one time?
And, 17+9 = 26(26), 28+19 = 47(46), coincidence?

 

[Mr. Travis]

excellent work detail normal maps, great idea!

 

[Humus]

u sure it takes off 3dc . I can't tell any diffrence haha .

Look again!

Notice the angular artifacts on DXT5, while 3Dc is all smooth.

 

[Humus]

Why there is "3Dc & DXT", you used them together? Isn't it that only method can be used at one time?

DXT5 for base texture and 3Dc for normal maps. DXT5 as opposed to DXT1 because I put the heightmap into alpha for offset mapping.

 

[Kurtisl8]

The normal maps are blockier and not as crisp with the DXT5, while it looks perfect with 3DC

 

[AlNom]

I'm getting a higher fps with compressed normal maps set to to false....

no lights, base maps compressed

 

[Ostsol]

I'm getting a higher fps with compressed normal maps set to to false....

no lights, base maps compressed

There's probably some additional latency for sampling compressed textures. The data does have to be decompressed, after all. Just a guess, though. . .

EDIT: Maybe nevermind. . . I didn't look at Humus' results. . .

 

[AlNom]

oh....... .........

so... there's no point to having compressed normal maps I don't get why I'm getting the opposite results of what everyone else is getting.

 

[SuperBORG]

I can confirm DXT5 brings some artifacts/banding. 3Dc does not. But I've only gained 5% more with everything on as pose to off.

 

[Ante P]

I just get a black screen with a white dot moving around on my 6800.

 

[OpenGL guy]

I'm getting a higher fps with compressed normal maps set to to false....

no lights, base maps compressed

There's probably some additional latency for sampling compressed textures. The data does have to be decompressed, after all. Just a guess, though. . .

EDIT: Maybe nevermind. . . I didn't look at Humus' results. . .

Decompression for 3Dc textures is free, just like it is for DXT1-5.

 

[lowdog]

I just get a black screen with a white dot moving around on my 6800.

Same here.

 

[euan]

I just get a black screen with a white dot moving around on my 6800.

sounds like an illegal optimisation.

 

[Humus]

I'm getting a higher fps with compressed normal maps set to to false....

I have noticed that myself, a couple of fps difference from some points of view, and no difference in other. Not sure why that it. With my first tests I got results the other way around, 3Dc a few fps faster than DXT5. I guess it could have something to do with the optimizer in the driver that maybe for some reason does a better job when DXT5 is used, though I'm not so sure about that either. DXT5 requires that you flip component order with the .wzyx swizzle, so I must use a different shader for DXT5. I've tried looking at the assembly code generated with fxc. I noticed earlier that it screwed up the sqrt(1-dot(bump.xy, bump.xy)) expression generating a lot more code than needed, but it seems this is not the case for the final version. The shaders are equally long and seems to be "done right" with the dp2add instruction for what's under the squareroot. I assume the compiler in fxc is identical to the one that gets linked into the application.

Anyway, the difference is quite small so it could be anything.

 

[Humus]

I can confirm DXT5 brings some artifacts/banding. 3Dc does not. But I've only gained 5% more with everything on as pose to off.

Well, if you want to compare you shouldn't toggle shadows on/off, because that's a whole different thing you're benchmarking then. What's interesting to compare performancewise is the first two, compressed base and bump on/off.

 

[Humus]

oh....... .........

so... there's no point to having compressed normal maps I don't get why I'm getting the opposite results of what everyone else is getting.

Exactly what results are you getting?

 

[SuperBORG]

I can confirm DXT5 brings some artifacts/banding. 3Dc does not. But I've only gained 5% more with everything on as pose to off.

Well, if you want to compare you shouldn't toggle shadows on/off, because that's a whole different thing you're benchmarking then. What's interesting to compare performancewise is the first two, compressed base and bump on/off.

Without the shadows, the gain is very significant. With it, however, not as much.

 

[AlNom]

oh....... .........

so... there's no point to having compressed normal maps I don't get why I'm getting the opposite results of what everyone else is getting.

Exactly what results are you getting?

specs:

2600+
R9800p

looking at the wall across from where the camera view starts (up close)
Fullscreen, shadows set to false, both compressions to true, cat 4.6

~171fps w/4xAA
~191 w/o AA
same thing, but bump maps compression set to false (same view)

~179 fps w/4xAA
~201 w/o AA

Turning off AA, if I look at the bumpmaps (rectangular tiles) directly, my fps goes quite low (~100 w/ BM compression)....Looking at this particular section, the gain in fps w/o BM compression is much smaller (~2fps)

Even so, it's lower than if I go into a corner and get a view of the entire room (~165), some other thing I thought was weird.