#texture #atlas #bin #box #packer

rectangle-pack

A general purpose, deterministic bin packer designed to conform to any two or three dimensional use case

13 unstable releases (3 breaking)

new 0.4.2 May 3, 2021
0.4.1 Apr 28, 2021
0.3.0 Mar 23, 2021
0.2.2 Mar 23, 2021
0.1.5 Mar 10, 2020

#22 in Text processing

Download history 2656/week @ 2021-01-15 2348/week @ 2021-01-22 2682/week @ 2021-01-29 2238/week @ 2021-02-05 2193/week @ 2021-02-12 2172/week @ 2021-02-19 2530/week @ 2021-02-26 3057/week @ 2021-03-05 2398/week @ 2021-03-12 2642/week @ 2021-03-19 2910/week @ 2021-03-26 3226/week @ 2021-04-02 4252/week @ 2021-04-09 3259/week @ 2021-04-16 3315/week @ 2021-04-23 5367/week @ 2021-04-30

12,465 downloads per month
Used in 48 crates (2 directly)

MIT/Apache

87KB
1.5K SLoC

rectangle-pack Actions Status docs

A general purpose, deterministic bin packer designed to conform to any two or three dimensional use case.

rectangle-pack is a library focused on laying out any number of smaller rectangles (both 2d rectangles and 3d rectangular prisms) inside any number of larger rectangles.

rectangle-pack exposes an API that gives the consumer control over how rectangles are packed - allowing them to tailor the packing to their specific use case.

While rectangle-pack was originally designed with texture atlas related use cases in mind - the library itself has no notions of images and can be used in any rectangle packing context.

Quickstart

# In your Cargo.toml
rectangle-pack = "0.4"
//! A basic example of packing rectangles into target bins

use rectangle_pack::{
    GroupedRectsToPlace,
    RectToInsert,
    pack_rects,
    TargetBin,
    volume_heuristic,
    contains_smallest_box
};
use std::collections::BTreeMap;

// A rectangle ID just needs to meet these trait bounds (ideally also Copy).
// So you could use a String, PathBuf, or any other type that meets these
// trat bounds. You do not have to use a custom enum.
#[derive(Debug, Hash, PartialEq, Eq, Clone, Ord, PartialOrd)]
enum MyCustomRectId {
    RectOne,
    RectTwo,
    RectThree,
}

// A target bin ID just needs to meet these trait bounds (ideally also Copy)
// So you could use a u32, &str, or any other type that meets these
// trat bounds. You do not have to use a custom enum.
#[derive(Debug, Hash, PartialEq, Eq, Clone, Ord, PartialOrd)]
enum MyCustomBinId {
    DestinationBinOne,
    DestinationBinTwo,
}

// A placement group just needs to meet these trait bounds (ideally also Copy).
//
// Groups allow you to ensure that a set of rectangles will be placed
// into the same bin. If this isn't possible an error is returned.
//
// Groups are optional.
//
// You could use an i32, &'static str, or any other type that meets these
// trat bounds. You do not have to use a custom enum.
#[derive(Debug, Hash, PartialEq, Eq, Clone, Ord, PartialOrd)]
enum MyCustomGroupId {
    GroupIdOne
}

let mut rects_to_place = GroupedRectsToPlace::new();
rects_to_place.push_rect(
    MyCustomRectId::RectOne,
    Some(vec![MyCustomGroupId::GroupIdOne]),
    RectToInsert::new(10, 20, 255)
);
rects_to_place.push_rect(
    MyCustomRectId::RectTwo,
    Some(vec![MyCustomGroupId::GroupIdOne]),
    RectToInsert::new(5, 50, 255)
);
rects_to_place.push_rect(
    MyCustomRectId::RectThree,
    None,
    RectToInsert::new(30, 30, 255)
);

let mut target_bins = BTreeMap::new();
target_bins.insert(MyCustomBinId::DestinationBinOne, TargetBin::new(2048, 2048, 255));
target_bins.insert(MyCustomBinId::DestinationBinTwo, TargetBin::new(4096, 4096, 1020));

// Information about where each `MyCustomRectId` was placed
let rectangle_placements = pack_rects(
    &rects_to_place,
    &mut target_bins,
    &volume_heuristic,
    &contains_smallest_box
).unwrap();

Full API Documentation

Background / Initial Motivation

Click to show the initial motivation for the library.

In my application I've switched to dynamically placing textures into atlases at runtime instead of in how I previously used an asset compilation step, so some of the problems explained here are now moot.

I still use rectangle-pack to power my runtime texture allocation, though, along with a handful of other strategies depending on the nature of the textures that need to be placed into the atlas.

rectangle-pack knows nothing about textures, so you can use it for any form of bin packing, whether at runtime, during an offline step or any other time you like.

I'm working on a game with some of the following texture atlas requirements (as of March 2020):

  • I need to be able to guarantee that certain textures are available in the same atlas.

    • For example - if I'm rendering terrain using a blend map that maps each channel to a color / metallic-roughness / normal texture I want all of those textures to be available in the same atlas. Otherwise in the worst case I might need over a dozen texture uniforms in order to render a single chunk of terrain.
  • I want to have control over which channels are used when I'm packing my atlases.

    • For example - I need to be able to easily pack my metallic and roughness textures into one channel each, while packing color and normal channels into three channels.
      • This means that my rectangle packer needs to expose configuration on the number of layers/channels available in our target bins.
  • I need to be able to ensure that uncommon textures aren't taking up space in commonly used atlases

    • For example - if a set of textures is only used in one specific region of the game - they shouldn't take up space in an atlas that contains a texture that is used for very common game elements.
      • This means that the packer needs to cater to some notion of groups or priority so that uncommon textures can be placed separately from common ones.
    • This allows us to minimize the number of textures in GPU memory at any time since atlases with uncommon texture atlases can be removed after not being in use for some time.
      • Without meeting this requirement - a large texture might be sitting on the GPU wasting space indefinitely since it shares an atlas with very common textures that will never be evicted.
    • Note that we might not end up achieving this at the API level. This could potentially be achieved by just having the consumer call the library multiple times using whichever input rectangles they determine to be of similar priority.
      • Or some other solution.
  • I need to be able to pack individual bits within a channel. For example - if I have a texture mask that encodes whether or not a fragment is metallic I want to be able to pack that into a single bit, perhaps within my alpha channel.

    • This means that our layers concept needs to support multi-dimensional needs. A layer within a layer.
      • For example - In color space one might be thinking of RGBA channels / layers or be thinking about within the Alpha channel having 255 different sub-layers. Or even a smaller number of variable sized sub-layers. Our API needs to make this simple to represent and pack.
    • We don't necessarily need to model things that way internally or even expose a multi-layered notion in the API - we just need to enable those use cases - even if we still think of things as one dimension of layers at the API level.
      • In fact .. as I type this .. one dimensions of layers at the API level both internally and externally sounds much simpler. Let the consumer worry about whether a channel is considered one layer (i.e. alpha) or 255 layers (i.e. every bit in the alpha channel).
  • I need to be able to allow one texture to be present in multiple atlases.

    • For example - say there is a grass texture that is used in every grassy region of the game. Say each of those regions has some textures that are only used in that region and thus relegated to their own atlas. We want to make sure our grass texture is copied into each of those textures so that one texture can support the needs of that region instead of two.

These requirements are the initial guiding pillars to design the rectangle-pack API.

The API shouldn't know about the specifics of any of these requirements - it should just provide the bare minimum required to make them possible. We're trying to push as much into user-land as possible and leave rectangle-packs responsibility to not much more than answering:

Given these rectangles that need to be placed, the maximum sizes of the target bins to place them in and some criteria about how to place and how not to place them, where can I put all of these rectangles?

no_std

rectangle-pack supports no_std by disabling the std feature.

rectangle-pack = {version = "0.4", default-features = false}

Disabling the std feature does the following.

  • BTreeMaps are used internally in places where HashMaps would have been used.

Features

  • Place any number of 2d / 3d rectangles into any number of 2d / 3d target bins.

    • Supports three dimensional rectangles through a width + height + depth based API.
  • Generic API that pushes as much as possible into user-land for maximum flexibility.

  • Group rectangles using generic group id's when you need to ensure that certain rectangles will always end up sharing a bin with each other.

  • Supports two dimensional rectangles (depth = 1).

  • User provided heuristics to grant full control over the packing algorithm.

  • Zero dependencies, making it easier to embed it inside of a more use case specific library without introducing bloat.

  • Deterministic packing.

    • Packing of the same inputs using the same heuristics and the same sized target bins will always lead to the same layout.
      • This is useful anywhere that reproducible builds are useful, such as when generating a texture atlas that is meant to be cached based on the hash of the contents.
  • Ability to remove placed rectangles and coalesce neighboring free space.

Future Work

The first version of rectangle-pack was designed to meet my own needs.

As such there is functionality that could be useful that was not explored since I did not need it.

Here are some things that could be useful in the future.

Three-Dimensional Incoming Rectangle Rotation

When attempting to place a Rectangle into the smallest available bin section we might want to rotate the rectangle in order to see which orientation produces the best fit.

This could be accomplished by:

  1. The API exposes three booleans for every incoming rectangles, allow_global_x_axis_rotation, allow_global_y_axis_rotation, allow_global_z_axis_rotation.

  2. Let's say all three are enabled. When attempting to place the rectangle/box we should attempt it in all 6 possible orientations and then select the best placement (based on the ComparePotentialContainersFn heuristic).

  3. Return information to the caller about which axis ended up being rotated.

Mutually exclusive groups

An example of this is the ability to ensure that certain rectqngle groups are not placed in the same bins.

Perhaps you have two plates (bins) and two groups of cheese (rectangles), one for Alice and one for Bob.

When packing you want to ensure that these groups of cheese each end up in a different bin since Alice and Bob don't like to share.

Stats on how the bins were packed

Things such as the amount of wasted space - or anything else that would allow the caller to compare the results of different combinations of target bin sizes and heuristics to see which packed the most efficiently.


If you have a use case that isn't supported, go right ahead and open an issue or submit a pull request.

Packing Algorithm

We started with the algorithm described in rectpack2D and then made some adjustments in order to support our goal of flexibly supporting all use cases.

  • The heuristic is provided by the caller instead of having rectangle-pack decide on a user provided heuristic.

  • When splitting an available section of a bin into two new sections of a bin - we do not decide on how the split should occur arbitrarily. Instead, we base it on the user provided more_suitable_containers heuristic function.

  • There is a third dimension.

In The Wild

Here are some known production users of rectangle-pack.

  • Akigi uses rectangle-pack to power parts of its runtime texture allocation strategy.

  • Bevy uses rectangle-pack to create texture atlases.

Contributing

If you have a use case that isn't supported, a question, a patch, or anything else, go right ahead and open an issue or submit a pull request.

To Test

To run the test suite.

# Clone the repository
git clone git@github.com:chinedufn/rectangle-pack.git
cd rectangle-pack

# Run tests
cargo test

See Also

  • rectpack2D
    • Inspired parts of our initial implementation

No runtime deps

Features

  • std