S3 Texture Compression
S3 Texture Compression or S3TC is a compression scheme for three or four color channel textures. This functionality is exposed by the ubiquitous extension EXT_s3_texture_compression. There are 3 forms of S3TC allowed by OpenGL.
S3TC is a technique for compressing images for use as textures. Standard image compression techniques like JPEG and PNG can achieve greater compression ratios than S3TC. However, S3TC is designed to be implemented in high-performance hardware. JPEG and PNG decompress images all-at-once, while S3TC allows specific sections of the image to be decompressed independently.
S3TC is a block-based format. The image is broken up into 4x4 blocks. For non-power-of-two images that aren't a multiple of 4 in size, the other colors of the 4x4 block are taken to be black. Each 4x4 block is independent of any other, so it can be decompressed independently.
There are 3 forms of S3TC accepted by OpenGL. These forms are named after the equivalent forms in Direct3D: DXT1, DXT3 and DXT5.
A DXT1-compressed image is an RGB image format. As such, the alpha of any color is assumed to be 1. Each 4x4 block takes up 64-bits of data, so compared to a 24-bit RGB format, it provides 6:1 compression. You can get a DXT1 image by using the
GL_COMPRESSED_RGB_S3TC_DXT1_EXT as the internal format of the image.
Each 4x4 block stores color data as follows. There are 2 16-bit color values, color0 followed by color1. Following this is a 32-bit unsigned integer containing values that describe how the two colors are combined to determine the color for a given pixel.
The 2 16-bit color values are stored in little-endian format, so the low byte of the 16-bit color comes first in each case. The color values are stored in RGB order (from high bit to low bit) in 5_6_5 bits.
The 32-bit unsigned integer is also stored in little-endian format. Every 2 bits of the integer represent a pixel; the 2 bits are a code that defines how to combine color0 and color1 to produce the color of that pixel. In order from highest bit to lowest bit (after the little-endian conversion), the pixels are stored in row-major order. Every 8 bits, 4 2-bit codes, is a single row of the image.
The interpretation of the 2-bit values depends on how color0 and color1 compare to each other. If the integer value of color0 is greater than color1, then the 2-bit values mean something different than if color0 is less than or equal to color1. The meaning of the 2-bit values is as follows:
|code||color0 > color1||color0 <= color1|
|2||(2*color0 + color1) / 3||(color0 + color1) / 2|
|3||(color0 + 2*color1) / 3||Black|
The arithmetic operations are done per-component, not on the integer value of the colors. And the value "Black" is simply R=G=B=0.
DXT1 with 1-bit Alpha
There is a form of DXT1 available that provides a simple on/off alpha value. This format therefore uses an RGBA base format. To get this format, use the
GL_COMPRESSED_RGBA_S3TC_DXT1_EXT internal format.
The format of the data is identical to the above case, which is why this is still DXT1 compression. The interpretation differs slightly. You always get an alpha value of 1 unless the pixel uses the code for Black in the above table. In that case, you get an alpha value of 0.
Note that this means that the RGB colors will also be 0 on any pixel with a 0 alpha. This also means that bilinear filtering between neighboring texels will result in colors combined with black. If you are using premultipled alpha blending, this is what you want. If you aren't, then it almost certainly is not what you want.
When using OpenGL to compress a texture, the GL implementation will assume any pixel with an alpha value < 0.5 should have an alpha of 0. This is another reason to manually compress images.
The DXT3 format is an RGBA format. Each 4x4 block takes up 128 bits of data. Thus, compared to a 32-bit RGBA texture, it offers 4:1 compression.
Each block of 128 bits is broken into 2 64-bit chunks. The first chunk contains the color information, compressed almost as in the DXT1 case; the difference being that color0 is always assumed to be less than color1 in terms of determining how to use the codes to extract the color value. The second chunk contains the alpha information.
The alpha 64-bit chunk is stored as a little-endian 64-bit unsigned integer. The alpha values are stored as 4-bit-per-pixel alpha values. The alpha values are stored in row-major order, from the highest bit of the 64-bit unsigned integer.
The DXT5 format is an alternate RGBA format. As in the DXT3 case, each 4x4 block takes up 128 bits. So it provides the same 4:1 compression as in the DXT3 case.
Just as for the DXT3 format, there are two 64-bit chunks of data per block: an RGB chunk compressed as for DXT1 (with the same caveat as for DXT3), and an alpha chunk.
Where DXT3 and DXT5 differ is how the alpha chunk is compressed. DXT5 compresses the alpha using a compression scheme similar to DXT1.
The alpha data is stored as a 2 8-bit alpha values, alpha0 and alpha1, followed by a 48-bit unsigned integer that describes how to combine these two reference alpha values to achieve the final alpha value.
The 48-bit unsigned integer contains 3-bit codes that describe how to compute the final alpha value. These codes are stored in the identical order as the codes in DXT1; they simply are 3 bits in size rather than 2.
Just as in the DXT1 case, the codes have different meanings depending on how alpha0 and alpha1 compare to one another. Here is the table of codes and computations:
|code||alpha0 > alpha1||alpha0 <= alpha1|
|2||(6*alpha0 + 1*alpha1)/7||(4*alpha0 + 1*alpha1)/5|
|3||(5*alpha0 + 2*alpha1)/7||(3*alpha0 + 2*alpha1)/5|
|4||(4*alpha0 + 3*alpha1)/7||(2*alpha0 + 3*alpha1)/5|
|5||(3*alpha0 + 4*alpha1)/7||(1*alpha0 + 4*alpha1)/5|
|6||(2*alpha0 + 5*alpha1)/7||0.0|
|7||(1*alpha0 + 6*alpha1)/7||1.0|
sRGB and S3TC
Images compressed with S3 texture compression can also be in the sRGB colorspace. This extension does not qualify as a ubiquitous extension, but it is widely supported nevertheless.
As with any sRGB operation, the question arises as to whether the two key colors (which are in the sRGB colorspace) are converted to a linear colorspace before the decompression step or after. While converting before decompression is preferred, the specification allows implementations to do the colorspace conversion afterwards the decompression. The decompression requires linearly interpolating color values, which generally should not be done in a non-linear colorspace like sRGB.
In general, hardware that is incapable of GL 3.0 or above will do the colorspace conversion after decompression, while 3.0 or better hardware (aka: Direct3D 10-capable) will do the conversion before decompression.