SparseSet: O(1) insert/delete/lookup with contiguous memory for efficient GPU buffer transfer

sparseset.rs

/*
* SparseSet is a data structure that most importantly keeps all its stored
* items in contiguous memory, while still allowing for O(1) insertions,
* deletions, and lookups. These are usually objects you'd like to keep
* a reference for later O(1) lookup and mutation. You may also think
* of it as a `HashMap<usize, T>` but contiguous memory holding Vec<T>
*
* The contiguous part is particularly useful to be able to map them
* into a GPU buffer, as the data transfer can only specify full blocks
* of memory. This is why this structure **must** own the data it stores.
* 
* For comparison, doing so directly with just a Vec would mean
* other objects can not hold references or pointers to the items in the Vec,
* as they would be move around in memory once the Vec is resized.
* The SparseSet has one layer of indirection to avoid this issue.
*
* A HashMap would also not be suitable, as it would not keep the items in
* contiguous memory.
*
* The key downside is really that it's more memory innefficient due to
* storing two index arrays, one of which is the size of your `data` vector,
* the other one going up to the highest index you care to use. Also, if you
* do not plan to delete any data regularly, the extra indirection is needless. 
*
* ```
* let sparseset = SparseSet::new();
*
* // This will cause the `sparse` vector to grow to a size of 2_000
* // even though there is only one item.
* sparseset.insert(2_000, "hello");
* ```
*
* As a result, this is best used with incremental indices, especially if you
* can recycle the indices.

* If sticking with a low number of items but you update/remove/add frequently,
* you may call `SparseSet<T, u8>::with_capacity(255)` which has a fixed
* memory cost over just a straight `Vec<T>` of:
*
* * `sizeof(U)*2*255` where for `u8` its `510` bytes, `u16` is `1020`
*    and `u32` is `2040`.
*
* This stabalizes the memory allocation, preventing data being moved.
*
* This implementation is inspired by the bevy implementation of a sparse set:
* * https://github.com/bevyengine/bevy/blob/f9c8f511fd50d5d2321464ef5889d33ae8eba72f/crates/bevy_ecs/src/storage/sparse_set.rs#L375
*
* Main changes are to not use the `nonmax` dependancy and just use `0` as
* sentinal value directly to say that a given slot is empty.
*
* The `dirty` flag is not required, it's simply there to flag when the data
* may need to be updated in the GPU buffer.
*/

// This stuff about `Unsigned` is to allow for the use of u8, u16, u32, usize
// as the elements of the index buffers.
// `Default` just makes accessing 0 for the type much easier
pub trait Unsigned: Copy + TryFrom<usize> + Default {
    fn as_usize(self) -> usize;
}

// u8,u16,usize support the .into::<usize>(), however u32 does not, despite `1000u32 as usize`
// being valid. Hence, we just make new methods...
impl Unsigned for usize {
    #[inline]
    fn as_usize(self) -> usize {
        self
    }
}
impl Unsigned for u8 {
    #[inline]
    fn as_usize(self) -> usize {
        self as usize
    }
}
impl Unsigned for u16 {
    #[inline]
    fn as_usize(self) -> usize {
        self as usize
    }
}
impl Unsigned for u32 {
    #[inline]
    fn as_usize(self) -> usize {
        self as usize
    }
}

#[derive(Debug, Clone)]
pub struct SparseSet<T, I: Unsigned = usize> {
    // The contiguous memory that holds all the items
    data: Vec<T>,

    // Given a sparse index, returns the index into `data`
    sparse: Vec<I>,

    // Inverse operation to go from `data` back to what
    // index in `sparse` is pointing to it.
    // This will always be modified in the same way as `data`
    dense: Vec<I>,

    /// Indicate whether the data has been modified.
    /// See [SparseSet::is_dirty], [SparseSet::mark_clean] and [SparseSet::mark_dirty]
    dirty: bool,
}

impl<T, I: Unsigned> SparseSet<T, I> {
    pub fn new() -> Self {
        SparseSet {
            data: Vec::new(),
            dense: Vec::new(),
            sparse: Vec::new(),
            dirty: false,
        }
    }

    pub fn with_capacity(capacity: I) -> Self {
        SparseSet {
            data: Vec::with_capacity(capacity.as_usize()),
            dense: Vec::with_capacity(capacity.as_usize()),
            sparse: Vec::with_capacity(capacity.as_usize()),
            dirty: false,
        }
    }

    pub fn items(&self) -> &Vec<T> {
        &self.data
    }

    pub fn items_mut(&mut self) -> &mut Vec<T> {
        &mut self.data
    }

    pub fn len(&self) -> I {
        unsafe { I::try_from(self.data.len()).unwrap_unchecked() }
    }

    pub fn is_empty(&self) -> bool {
        self.data.is_empty()
    }

    pub fn contains(&self, index: I) -> bool {
        !matches!(
            self.sparse.get(index.as_usize()).map(|x| x.as_usize()),
            Some(0) | None
        )
    }

    pub fn get(&self, index: I) -> Option<&T> {
        match self.sparse.get(index.as_usize()).map(|x| x.as_usize()) {
            Some(0) | None => None,
            Some(dense_index) => {
                // -1 because the indices contained in the sparse set are 1+
                let real_index = dense_index - 1;
                self.data.get(real_index)
            }
        }
    }

    pub fn get_mut(&mut self, index: I) -> Option<&mut T> {
        match self.sparse.get(index.as_usize()).map(|x| x.as_usize()) {
            Some(0) | None => None,
            Some(dense_index) => {
                // -1 because the indices contained in the sparse set are 1+
                let real_index = dense_index - 1;
                self.data.get_mut(real_index)
            }
        }
    }

    pub fn insert(&mut self, index: I, value: T) -> &T {
        let uindex = index.as_usize();
        match self.sparse.get(uindex).map(|x| x.as_usize()) {
            Some(0) => {
                // In bounds, but nothing in that slot
                let dense_index = self.data.len();
                let sparse_index = dense_index + 1;

                // Unsafe here as it's possible we fail to convert from usize above
                // into `I`  (e.g. u8). However we grow according to the `I` type so
                // this should not happen.
                // TODO: Test this with u8 up to boundary and see if it panics
                self.sparse[uindex] = unsafe { I::try_from(sparse_index).unwrap_unchecked() };

                self.dense.push(index);
                self.data.push(value);

                &self.data[dense_index]
            }
            Some(dense_index) => {
                // In bounds, something in the slot, we just update
                let real_index = dense_index - 1;
                self.data[real_index] = value;
                &self.data[real_index]
            }
            None => {
                // Out of bounds, we need to resize the sparse array
                self.sparse.resize(uindex + 1, I::default());
                let dense_index = self.data.len();
                let sparse_index = dense_index + 1;
                self.sparse[uindex] = unsafe { I::try_from(sparse_index).unwrap_unchecked() };

                self.data.push(value);
                self.dense.push(index);

                &self.data[dense_index]
            }
        }
    }

    pub fn remove(&mut self, index: I) -> Option<T> {
        let uindex = index.as_usize();
        match self.sparse.get(uindex).map(|x| x.as_usize()) {
            Some(0) | None => None,
            // In bounds and not empty slot
            Some(dense_index) => {
                let item_index = dense_index - 1;
                let last_item_index = self.data.len() - 1;

                // We are removing the last element
                if item_index == last_item_index {
                    self.sparse[uindex] = I::default();
                    self.dense.pop();
                    return self.data.pop();
                }

                let value = self.data.swap_remove(item_index);
                let last_index = self.dense.last().unwrap().as_usize();
                let sparse_index = self.dense.swap_remove(item_index);

                self.sparse[last_index] = unsafe { I::try_from(item_index + 1).unwrap_unchecked() };
                self.sparse[uindex] = I::default();

                Some(value)
            }
        }
    }

    pub fn is_dirty(&self) -> bool {
        self.dirty
    }

    pub fn mark_dirty(&mut self) {
        self.dirty = true;
    }

    pub fn mark_clean(&mut self) {
        self.dirty = false;
    }
}

impl<T> Default for SparseSet<T> {
    fn default() -> Self {
        Self::new()
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_sparse_set_insert() {
        let mut sparse_set = SparseSet::new();

        sparse_set.insert(0_u8, "a");
        sparse_set.insert(1, "b");
        sparse_set.insert(2, "c");
        assert_eq!(sparse_set.len(), 3);

        assert_eq!(sparse_set.get(0), Some(&"a"));
        assert_eq!(sparse_set.get(1), Some(&"b"));
        assert_eq!(sparse_set.get(2), Some(&"c"));

        assert!(sparse_set.contains(0));
        assert!(sparse_set.contains(1));
        assert!(sparse_set.contains(2));

        assert_eq!(sparse_set.items(), &["a", "b", "c"]);

        assert_eq!(sparse_set.get(3), None);
        assert!(!sparse_set.contains(3));
    }

    #[test]
    fn test_sparse_remove() {
        let mut sparse_set = SparseSet::new();

        sparse_set.insert(0_u8, "a");
        sparse_set.insert(1, "b");
        sparse_set.insert(2, "c");
        sparse_set.insert(4, "e");

        assert_eq!(sparse_set.remove(1), Some("b"));
        assert_eq!(sparse_set.len(), 3);
        assert_eq!(sparse_set.items(), &["a", "e", "c"]);

        assert_eq!(sparse_set.remove(2), Some("c"));
        assert_eq!(sparse_set.len(), 2);
        assert_eq!(sparse_set.items(), &["a", "e"]);

        assert_eq!(sparse_set.remove(0), Some("a"));
        assert_eq!(sparse_set.len(), 1);
        assert_eq!(sparse_set.items(), &["e"]);

        assert_eq!(sparse_set.remove(4), Some("e"));
        assert_eq!(sparse_set.len(), 0);
        assert!(sparse_set.is_empty());
    }
}