3. Required concepts not contained in this format

This specification encodes how atomic data should be interpreted in a manner which is independent of the layout and dimensionality of the collective data. Collections of data may have a “compatible format” in that their format descriptor may be identical, yet be different sizes. Some additional information is therefore expected to be recorded alongside the “format description”.

The API which controls the bulk data is responsible for controlling which memory location corresponds to the indexing mechanism chosen. A texel block has the concept of a coordinate offset within the block, which implies that if the data is accessed in terms of spatial coordinates, a texel block has spatial locality as well as referring to contiguous memory (per plane). For texel blocks which represent only a single spatial location, this is irrelevant; for block-based compression, for formats with downsampled channels, or for Bayer-like formats, the texel block represents a finite extent in up to four dimensions. However, the mapping from coordinate system to the memory location containing a texel block is beyond the control of this API.

The minimum requirements for accessing a linearly-addressed buffer is to store the start address and a stride (typically in bytes) between texels in each dimension of the buffer, for each plane contributing to the texel block. For the first dimension, the memory stride between texels may simply be the byte size of texel block in that plane — this implies that there are no gaps between texel blocks. For other dimensions, the stride is a function of the size of the data structure being represented — for example, in a compact representation of a two-dimensional buffer, the texel block at coordinate (x,y+1) might be found at the address of coordinate (x,y) plus the buffer width multiplied by the texel size in bytes. Similarly in a three-dimensional buffer, the address of the pixel at (x,y,z+1) may be at the address of (x,y,z) plus the byte size of a two-dimensional slice of the texture. In practice, even linear layouts may have padding, and often more complex relationships between coordinates and memory location are used to encourage locality of reference. The details of all of these data structures are beyond the remit of this specification.

Most simple formats contain a single plane of data. Those formats which require additional planes compared with a conventional representation are typically downsampled YUV formats, which already have the concept of separate storage for different color channels. While this specification uses multiple planes to describe texel blocks that span multiple scan lines if the data is disjoint, there is no expectation that the API using the data formats needs to maintain this representation — interleaved planes should be easy to identify and coalesce if the API requires a more conventional representation of downsampled formats.

Some image representations are composed of tiles of texels which are held contiguously in memory, with the texels within the tile stored in some order that improves locality of reference for multi-dimensional access. This is a common approach to improve memory efficiency when texturing. While it is possible to represent such a tile as a large texel block (up to the maximum representable texel block size in this specification), this is unlikely to be an efficient approach, since a large number of samples will be needed and the layout of a tile usually has a very limited number of possibilities. In most cases, the layout of texels within the tile should be described by whatever interface is aware of image-specific information such as size and stride, and only the format of the texels should be described by a format descriptor.

The complication to this is where texel blocks larger than a single pixel are themselves encoded using proprietary tiling. The spatial layout of samples within a texel block is required to be fixed in the basic format descriptor — for example, if the texel block size is 2×2 pixels, the top left pixel might always be expected to be in the first byte in that texel block. In some proprietary memory tiling formats, such as ones that store small rectangular blocks in raster order in consecutive bytes or in Morton order, this relationship may be preserved, and the only proprietary operation is finding the start of the texel block. In other proprietary layouts such as Hilbert curver order, or when the texel block size does not divide the tiling size, a direct representation of memory may be impossible. In these cases, it is likely that this data format standard would be used to describe the data as it would be seen in a linear format, and the mapping from coordinates to memory would have to be hidden in proprietary translation. As a logical format description, this is unlikely to be critical, since any software which accesses such a layout will necessarily need proprietary knowledge anyway.