Jump to content

  • Log In with Google      Sign In   
  • Create Account


DXT for terrain normals?

  • You cannot reply to this topic
8 replies to this topic

#1 mark ds   Members   -  Reputation: 1089

Like
0Likes
Like

Posted 10 July 2014 - 03:09 PM

This is just an exploratory question really, but here's my situation.

 

I have a very large terrain, with all the textures stuffed into a texture array, addressable by their array index. Eventually there will be quite a lot of textures (I envisage around 30 or more in total), but any given 'chunk' will only have a subset of the total - no more than say 16, which conveniently fits into 4 bits. I can then use a base offset for each chunk to determine which of the consecutive textures I can access in the array. The simplest would be to use an R8 texture, but that requires 8 bits per pixel, when 4 bits would do.

 

DXT3 textures handily provide 4 bits per pixel in the alpha, leaving RGB empty, which would be ideal for terrain vertex normals. Now, I understand that DXT is generally a crappy way to store normals, but was wondering if anyone knows how bad it would be for terrains, bearing in mind that most normal point more or less up, rather than every-which-way.

 

Maybe there is a way of using RG & B to better encode XY normals? Or a better compression method that allows 4 bits per pixel.

 

I suspect however, that I'll just have to use bytes and bit shift to get at the data. A 'built in' method (like DXT) would be nicer/easier though, and would avoid "if( mod(...) )" type statements.

 

Any suggestions?


Edited by mark ds, 10 July 2014 - 03:58 PM.


Sponsor:

#2 kalle_h   Members   -  Reputation: 1280

Like
0Likes
Like

Posted 10 July 2014 - 03:33 PM

So you don't want albedo texture and normal maps to match?

If you could use second texture array for normal maps and then use that compressed normal as base normal that is then blended with detail normal map using reoriented normal mapping found here.

http://blog.selfshadow.com/publications/blending-in-detail/



#3 mark ds   Members   -  Reputation: 1089

Like
0Likes
Like

Posted 10 July 2014 - 03:40 PM

Sorry, I should have been clearer (I've edited the OP) - I meant the terrain vertex normals, not the texture normals.

 

1  2

-----

|    |

-----

3  4

 

If my terrain consists of one quad (above), I'd have 4 heightmap values, 4 vertex normals , and 4 texture array indices - and then blend between the textures. I'm trying to combine the vertex normal and the texture array indices into the smallest space possible. I'm ignoring the texture normal for now.


Edited by mark ds, 10 July 2014 - 03:58 PM.


#4 phil_t   Crossbones+   -  Reputation: 3213

Like
0Likes
Like

Posted 10 July 2014 - 03:59 PM

Keep in mind older GPUs may only support vertex texture sampling from floating point formats (i.e. not DXT).



#5 swiftcoder   Senior Moderators   -  Reputation: 9672

Like
0Likes
Like

Posted 10 July 2014 - 04:02 PM

You don't need vertex normals for terrain, strictly speaking, if you are going to normal map it.

 

Since terrain tends to face statically upwards, and not be affected by skeletal animation, object-space normal maps work just as well, and remove the need for vertex normals.


Tristam MacDonald - Software Engineer @Amazon - [swiftcoding]


#6 mark ds   Members   -  Reputation: 1089

Like
0Likes
Like

Posted 10 July 2014 - 04:33 PM

You don't need vertex normals for terrain, strictly speaking, if you are going to normal map it.

 

Since terrain tends to face statically upwards, and not be affected by skeletal animation, object-space normal maps work just as well, and remove the need for vertex normals.

 

Yeah - but you need to use the vertex data to derive the normal map, and then 'stuff it' somewhere, hence the question about the appropriateness of using the 'spare' RGB slots in a DXT texture, assuming I use the Alpha for storing the texture array indices. I'm simply trying to avoid using 8 bits when I can get away with four - unfortunately there are no native 4 bit single channel formats available.



#7 MJP   Moderators   -  Reputation: 10282

Like
2Likes
Like

Posted 10 July 2014 - 04:49 PM

The old trick for DXT5 (AKA BC3) is to stuff the X and Y into the A and G channels of the texture, and then reconstruct Z. If your hardware supports it, BC5 (AKA 3Dc) is a much better option (it's basically 2 DXT5 alpha channels put together). BC7 can also work for normals, but that requires even newer hardware and API's.



#8 Hodgman   Moderators   -  Reputation: 27930

Like
3Likes
Like

Posted 10 July 2014 - 05:18 PM

DXT3 is basically DXT1+4bpp alpha, and DXT1 really sucks for storing normals. The problem is that it uses a palette of RGB values, and blend weights for that palette, which means that R, G an B are inter-dependent; a gradient in the R channel will impact the quality of the G channel.

As MJP mentions above, a common solution was to use DXT5, where RGB and A are seperately encoded, an then store zeroes in R/B, and the normal data in G/A...
So I would not recommend the scheme you're thinking of.

Anyway, do you need texture blending at each terrain point? I would think you would want to store at least two 4-bit indices and an 8-bit blend weight ;-)

#9 mark ds   Members   -  Reputation: 1089

Like
1Likes
Like

Posted 10 July 2014 - 08:05 PM

So I would not recommend the scheme you're thinking of.


I kind of figured that would be the case!
 

Anyway, do you need texture blending at each terrain point? I would think you would want to store at least two 4-bit indices and an 8-bit blend weight ;-)

 

Believe it or not, I specifically want a 'grid' of texture points. It's the way Skyrim does it, although I'm doing for very different reasons - basically, I'm investing heavily in 2-colour, corner based Wang Tiles for the textures. I'm extending this to include Wang Tile 'blend textures' to avoid the default (read hideous) blending that would otherwise occur. This will enable me to create more interesting blend edges, in turn removing the 'gridiness' of the terrain texturing.







PARTNERS