4 stable releases
1.2.0 | Jul 26, 2023 |
---|---|
1.1.0 | May 2, 2022 |
1.0.2 | Apr 4, 2022 |
#430 in Data structures
73KB
1K
SLoC
ZArray
Z-order indexed 2D and 3D arrays using Morton order (aka Z-order) with a convenient API for common 2D and 3D access patterns. Use of zarray in place of a Vec of Vecs often improves performance, especially for algorithms such as blurring and cellular automata.
About ZArray
The zarray crate is a lightweight Rust library that provides structs for working with 2D and 3D arrays, using internal Z-Order Morton indexing to improve data localization for better cache-line performance.
Quickstart Guide
Simply import zarray::z2d::ZArray2D and/or zarray::z3d::ZArray3D and then use the ZArray_D::new(...) function to initialize a new instance. The type will automatically be inferred from the povided default value. Note that only types which implement the Copy trait are allowed (ie not Vec or other heap-allocating types).
For example, here's a simple blur operation using ZArray2D, which generally performs better than using a Vec of Vecs by about 10-25%:
use zarray::z2d::ZArray2D;
let h: isize = 200;
let w: isize = 300;
let radius: isize = 3;
let mut src = ZArray2D::new(w as usize, h as usize, 0u8);
// set values
src.bounded_fill(100, 100, 200, 150, 255u8);
// sum neighbors values with ZArray
let mut blurred = ZArray2D::new(w as usize, h as usize, 0u16);
for y in 0..h { for x in 0..w {
let mut sum = 0;
for dy in -radius..radius+1 { for dx in -radius..radius+1 {
sum += *src.bounded_get(x+dx, y+dy).unwrap_or(&0u8) as u16;
} }
blurred.set(x as usize, y as usize, sum/((2*radius as u16+1).pow(2))).unwrap();
} }
How it works
the ZArray_D structs store data in 8x8 or 8x8x8 chuncks, using Z-order indexing to access the data within each chunk (as described here ). In so doing, the lowest 4 bits of each dimension are interdigitated to significantly improve data locality and cache-line fetch efficiency (though not as much as a Hilbert curve would do)
Why not just use Vec of Vecs (aka Vec<Vec>)?
Most of the time, using a Vec<Vec> would have great performance, so long as you remember to structure your for-loops correctly. However, when the data is not accessed in a linear fashion, such as when implementing a cellular automata or a blurring or ray tracing algorithm, then the performance of a Vec<Vec> can be significantly impaired by frequent RAM access and cache-line misses. This is when data locality matters most for performance.
Why not Z-Order the entire data array?
Two reasons: Firstly, Z-Order indexing only works on square/cube shaped data, so a pure Z-Order index would waste huge amount of memory for 2D and 3D arrays that are long and thin. Second, on most CPU architectures (Intel, AMD, and Arm), memory is accessed in 64-byte cache-lines, thus the performance gains from Z-order indexing are less significant above 6 bits of linear addressing space (ie 8x8 or 4x4x4)
Note
Only types with the Copy trait can be stored in ZArrady_D structs. Thus zarray works for all numerical types and most simple data structs, but not for heap-allocating data types such as Vec. This limitation arises from complexities in filling the data patches with initial values. In addition, you would not see a performance improvement over a simple Vec<Vec> if the data resided on the heap.
License
This library is provided under the MIT license. In other words: free to use as you wish.
Contributing
If you'd like to contribute, go ahead and fork the GitHub repo and/or submit a pull request
Dependencies
~165KB