Adaptive scalable texture compression

Adaptive scalable texture compression (ASTC) is a lossy block-based texture compression algorithm developed by Jørn Nystad et al. of ARM Ltd. and AMD. Full details of ASTC were first presented publicly at the High Performance Graphics 2012 conference, in a paper by Olson et al. entitled "Adaptive Scalable Texture Compression". ASTC was adopted as an official extension for both OpenGL and OpenGL ES by the Khronos Group on 6 August 2012. On Linux, all Gallium 3D drivers have a software fallback since 2018, so ASTC can be used on any AMD Radeon GPU. The method of compression is an evolution of Color Cell Compression with features including numerous closely spaced fractional bit rates, multiple color formats, support for high-dynamic-range (HDR) textures, and real 3D texture support. The stated primary design goal for ASTC is to enable content developers to have better control over the space/quality tradeoff inherent in any lossy compression scheme. With ASTC, the ratio between adjacent bit rates is of the order of 25%, making it less expensive to increase quality for a given texture. Encoding different assets often requires different color formats. ASTC allows a wide choice of input formats, including luminance-only, luminance-alpha, RGB, RGBA, and modes optimized for surface normals. The designer can thus choose the optimal format without having to support multiple different compression schemes. The choices of bit rate and color format do not constrain each other, so that it's possible to choose from a large number of combinations. Despite this flexibility, ASTC achieves better peak signal-to-noise ratios than PVRTC, S3TC, and ETC2 when measured at 2 and 3.56 bits per texel. For HDR textures, it produces results comparable to BC6H at 8 bits per texel. ASTC supports anywhere from 1 to 4 channels. In modes with 2–4 channels, one of the channels can be treated as "uncorrelated" and be given a separate gradient for prediction. In any case, the data is decoded as RGBA. Each of these may be encoded as low or high dynamic range.