Iris2.de

hi,

currently we have many different texture formats. I want to consolidate this, because of better handling. I thought of converting all current Textures to DDS.

Wikipedia:
http://en.wikipedia.org/wiki/S3_Texture_Compression

benefits:
- DDS is a raw surface dump and can be loaded directly into the graphiccard staying compressed in video ram
- it supports different compressions
- it supports different kinds of alpha channels (partial alpha if you use DXT2-5)
- it supports prcalculated Mipmaps in the same file
- Ogre uses a custom DDS codec and works also for OpenGL and Linux
- OpenGL + Linux Mipmap supported

drawbacks:
- real benefit only for DirectX (Windows) Users, because as far as I know it don't work with OpenGL (windows/linux). Using OpenGL, it must be decoded as all other Fileformats (png, tga, jpg).

addition: Using S3 (dxt) Texture Compression also in OpenGL - NVIDIA document




There are several types of DXT compression offered by DirectX:

DXT1
0 or 1 bit alpha
DXT2
Explicit 4-bit alpha, color data is premultiplied by alpha – UnrealEd does not use this format
DXT3
Explicit 4-bit alpha, not premultiplied
DXT4
Interpolated alpha, color data is premultiplied by alpha – UnrealEd does not use this format
DXT5
Interpolated alpha, not premultiplied

Which DXT Compression to Use?
Obviously, there are some trade-offs between the different formats which make them better or worse for different types of images. Some general rules of thumb for good use of DXT textures are as follows:

1. If your image has no alpha, use DXT1 compression. Using DXT3/5 will double your image size over DXT1 and not gain anything.
2. If your image has 1-bit (on or off) alpha information, use DXT1 with one-bit alpha. If the DXT1 image quality is too low and you don't mind doubling image size, use DXT3 or DXT5 (which one doesn't matter, they'll give the same results).
3. If your image has smooth gradations of alpha (fading in/out slowly), DXT5 is almost certainly your best bet, as it will give you the most accurate transparency representation.
4. If your image has sharp transitions between multiple alpha levels (one pixel is 100%, the next one is 50%, and another neighbor is 12%), DXT3 is probably your best bet. You may want to compare the alpha results in DXT1, DXT3 and DXT5 compression, however, to make sure.

Specification:
http://oss.sgi.com/projects/ogl-sample/ ... n_s3tc.txt

NVidia has some Tools on their Homepage to produce DDS files from Commandshell or Photoshop.

http://developer.nvidia.com/object/nv_t ... tools.html

The GIMP also has a plugin which can handle DXT-compressed DDS images.


Please tell me what you think of it.
This is a talk...so please write your opinions down.

regards

SiENcE

i would prefer sticking to standard formats (jpg,png,tga..) that can be saved, edited and viewed in all common graphic programs without installing third party plugins.
there might also be legal issues, microsoft and directx stuff tends to be proprietary.
we cannot convert the granny textures anyway so it will not improve the performance hit when a model appears, and we don't load static textures every frame, so i don't see much performance benefit either....

Why you are talking about Microsoft? Please dont think every time on Microsoft and I dont want to talk every time against or pro MS with you. ok!

It's from S3. "All" Graphiccard use this Hardware Texture DeCompression.

DXT1-5 compression is Standard for most 3D Games.

I wrote the benefit already.
-precalc Mipmaps
-better handling of Alphachannels
-i posted links to tools supporting dds (also gimp)

I dont want to convert Granny stuff. I talk about the new Terrain Engine Textures and the Static Model Textures.

We need to consolidate the texture fileformats!

I dont like the mix of tga, dds, png aso. These are all picture formats...not for graphiccard textures.

Ah and i forgot.

big benefit:
-it saves vram, because the compressed dds file its directly transfered to vram

more Infos here:

http://www.sjbrown.co.uk/?code=squish

Here is a quick comparsion between dxt compressed files and png/jpg.

I converted only the last 4 Textures (that are not dds) from our Terrainshader Engine to DDS. The converted DDS Files also include "10" pregenerated Mipmaps.

Dirt_large_pattern_1024.jpg
Grass_large_pattern_1024.jpg
Mountain_large_pattern_512.jpg
Water01.png


Memory difference:
jpg:
44,39mb vram - Ogre internal vram usage
53,38mb vram - external Vram Watcher Tool

dds:
37,17mb vram - Ogre internal vram usage
43,76mb vram - external Vram Watcher Tool

2second scene:
jpg:
54,7mb vram - Ogre internal vram usage
63,7mb vram - external Vram Watcher Tool

dds:
49,3mb vram - Ogre internal vram usage
55.24mb vram - external Vram Watcher Tool

Only 4 files with mipmaps saves 10mb of vram. I think thats a very good argument.

Ah and its the same for DirectX or OpenGL. It saves on both 3D interfaces ~10mb vram.

ps: when i also convert the skybox to dds (without mipmaps) the first scene look like this:

32,79mb vram - Ogre internal vram usage
37,92mb vram - external Vram Watcher Tool

As with any asset, source file has to be included. Using dds as such file is not a good idea (obviously). I wouldnt be too concerned about this, as long as lossless version is provided so we dont end up in similar situation like we have with current models from first version of iris.

Other than that, dds is a good choice. However, only for final stuff, not for WIP.

Yes.

Submitted in our graphic repository should be also the raw texture stuff (i prefer *.tga).


Another explanation about DXT texture compression from the "infinity" project.

found here:
http://www.gamedev.net/community/forums ... ?jn=263350