Textures, compression, filtering

1
Textures, compression, filtering

• One way to reduce the overhead of using very many
  textures per scene is to compress the textures
• Compressed textures use less bandwidth in
  transmission from hard disc to system RAM to
  graphics card texture buffer
• If textures are kept compressed, they use less
  graphics card texture buffer

2
Textures, compression, filtering

• This requires hardware to decode textures during
  rendering
• S3 and 3dfx had texture compression systems built
  into their hardware, with support in DirectX 7
  onwards (DXTC)
• Surviving graphics companies have embraced and
  extended this approach

3
Texture Filtering

• Point sampling is used in software rendering when
  3D acceleration is unavailable
• Bi-linear filtering is the method that is
  standard today, often accompanied by something
  called mip-mapping
• Tri-linear filtering is a "high-quality" mode
  used to enhance visual realism yet another notch
  but it usually comes with a slight speed penalty

4
Three levels of mip-map textures
5
Mip-mapping (level-of-detail)

• Several versions of the same texture are used to
  represent a given texture on screen at different
  distances
• Further away a smaller version of the texture is
  used and closer a bigger, more detailed one
• Mip-mapping can be used regardless of filtering
  method

6
Mip-mapping (level-of-detail)

• The advantages of using Mip-mapping lies mainly
  in that it reduces memory bandwidth requirements
  of rendering
• How? think about it
• It also offers better quality in the rendered
  image

7
Bilinear vs Pixel resize
8
Bi-linear texture filtering

• Bi-linear filtering is more or less the standard
  today
• Supported by just about every 3D accelerator on
  the market
• The accelerator uses an interpolation method to
  produce smooth transitions between different
  pixels in the source texture
• This is done by sampling the four closest pixels
  of the source texture (or most suitable mip-map)
  and interpolating these values before rendering
  each single texel on screen

9
Tri-linear texture filtering

• Tri-linear filtering is similar to bi-linear
  filtering
• But tri-linear filtering also uses the four
  closest pixels in the second most suitable
  mip-map to produce smoother transitions between
  mip-map levels
• It samples eight pixels and interpolates these
  before rendering, twice as much as bi-linear does
• Tri-linear filtering always uses mip-mapping

10
Anisotropic texture filtering

• Anisotropic filtering is the latest filtering
  type to be implemented in 3D accelerators and is
  even more advanced than tri-linear filtering
• Anisotropic (which means non-uniform shape)
  samples from more than 8 pixels, depending on the
  degree of surface tilting in the X,Y and Z
  directions
• The sample area is based on an area of visibility
• It requires even more processing than tri-linear
  filtering, which means it can be much slower

11
Antialiasing
12
Antialiasing
13
Antialiasing

• All modern 3D video accelerators support some
  sort of anti-aliasing feature
• Polygon edge anti-aliasing was popular for a
  year or so
• When using PEAA the accelerator or the driver or
  the application determines which triangle edges
  to anti-alias and then renders these edges again
  with a slight offset and lower opacity to produce
  a smoothing effect
• The end result is a smoother image and no ugly
  jagged polygon edges

14
Antialiasing

• Anti-aliasing by down-sampling is a very
  different approach from polygon edge
  anti-aliasing
• The accelerator renders the scene at an increased
  resolution and then scales it down with bi-linear
  or bi-cubic filtering to the end screen
  resolution
• Visual quality is very high
• We may be quadrupling the load on the GPU
• Some newer accelerators have a high enough fill
  rate to consider this form of anti-aliasing

15
Hardware Transform Lighting

• Hardware TL moves a lot of work away from the
  CPU and onto the GPU
• With hardware TL, the CPU is no longer expected
  to break an object down into constituent
  triangles, then transform and illuminate them and
  send the data to the GPU
• Rather we send the objects (including lights) to
  the GPUs object buffer, then tell the GPU where
  to put the objects within the scene

16
References

• For current coverage
• www.sharkyextreme.com
• www.firingsquad.com
• www.tomshardware.com
• History
• S3 Savage 2000 review, PC Pro magazine, March
  2000, page 185
• Graphics Card Shoot-Out, Personal Computer World,
  March 2000, page 80