#buffer #layout #interface #allocated #data-structures #texel #rearranging

image-texel

A texel type and allocated buffers suitable for image data

7 releases

0.3.1 Nov 2, 2022
0.3.0 Jun 9, 2022
0.2.1 May 26, 2022
0.1.2 May 22, 2022

#912 in Images

38 downloads per month
Used in image-canvas

MIT license

175KB
3K SLoC

image-texel docs.rs: image-texel Build Status

An allocated buffer suitable for image data. Compared to a Vec of all samples, it restricts the possible data types but on the other hand focusses on a more efficient interface to casting the data or rearranging the contents.

What it is

The buffer types provided provide several utility operations out of the box, or they can be used as internal structures on which a more concrete interface is built.

  • As the inner type of opaque image struct that offers conversion.
  • As standard forms of particularly shaped image data in interfaces.
  • As references to image data with their layout attached.
  • As a basis for optimized transform libraries.
  • As a mechanism for an FFI interface.

Why

After some discussion it was concluded that there is likely no safe way to reinterpret the allocation of a Vec<T> for a different Vec<U>, except in a very restricted set of cases¹. This includes grouping samples in logic pixel structs, even if those are annotated #[repr(C)] or alike. This is hardly the fault of Vec<_>, as images and the necessary transmutations of the representative binary data are far from general but Vec must be.

¹This was slightly relaxed in Rust 1.61 where few casts such as Vec<[u8; 3]> to Vec<u8> are now permitted, and the reverse under a restricted set of circumstances. This may provide us a good 'out' to perform zero-allocation conversion into and from a standard vector. It's still far from flexible enough for most high-performance needs.

What it is not

It is not optimized to provide nd-algebra or other matrix style operations. This firstly keeps the implementation complexity low, allows other implementation details, and separates concerns. If you nevertheless find it useful for such purposes, rest assured that we find this incredibly cool even though we may not accept PRs to introduce additional complexity motivated solely by such use. Rather, use the code and free license to shape it to those other needs.

Todo

  • Safely wrap SIMD iteration/transmutation/map-operation that are sound under the alignment guarantees of the buffer and slice types provided.

Dependencies

~140KB