nalgebra/nalgebra-sparse/src/pattern.rs

403 lines
15 KiB
Rust

//! Sparsity patterns for CSR and CSC matrices.
use crate::cs::transpose_cs;
use crate::SparseFormatError;
use std::error::Error;
use std::fmt;
/// A representation of the sparsity pattern of a CSR or CSC matrix.
///
/// CSR and CSC matrices store matrices in a very similar fashion. In fact, in a certain sense,
/// they are transposed. More precisely, when reinterpreting the three data arrays of a CSR
/// matrix as a CSC matrix, we obtain the CSC representation of its transpose.
///
/// [`SparsityPattern`] is an abstraction built on this observation. Whereas CSR matrices
/// store a matrix row-by-row, and a CSC matrix stores a matrix column-by-column, a
/// `SparsityPattern` represents only the index data structure of a matrix *lane-by-lane*.
/// Here, a *lane* is a generalization of rows and columns. We further define *major lanes*
/// and *minor lanes*. The sparsity pattern of a CSR matrix is then obtained by interpreting
/// major/minor as row/column. Conversely, we obtain the sparsity pattern of a CSC matrix by
/// interpreting major/minor as column/row.
///
/// This allows us to use a common abstraction to talk about sparsity patterns of CSR and CSC
/// matrices. This is convenient, because at the abstract level, the invariants of the formats
/// are the same. Hence we may encode the invariants of the index data structure separately from
/// the scalar values of the matrix. This is especially useful in applications where the
/// sparsity pattern is built ahead of the matrix values, or the same sparsity pattern is re-used
/// between different matrices. Finally, we can use `SparsityPattern` to encode adjacency
/// information in graphs.
///
/// # Format
///
/// The format is exactly the same as for the index data structures of CSR and CSC matrices.
/// This means that the sparsity pattern of an `m x n` sparse matrix with `nnz` non-zeros,
/// where in this case `m x n` does *not* mean `rows x columns`, but rather `majors x minors`,
/// is represented by the following two arrays:
///
/// - `major_offsets`, an array of integers with length `m + 1`.
/// - `minor_indices`, an array of integers with length `nnz`.
///
/// The invariants and relationship between `major_offsets` and `minor_indices` remain the same
/// as for `row_offsets` and `col_indices` in the [CSR](`crate::csr::CsrMatrix`) format
/// specification.
#[derive(Debug, Clone, PartialEq, Eq)]
// TODO: Make SparsityPattern parametrized by index type
// (need a solid abstraction for index types though)
pub struct SparsityPattern {
major_offsets: Vec<usize>,
minor_indices: Vec<usize>,
minor_dim: usize,
}
impl SparsityPattern {
/// Create a sparsity pattern of the given dimensions without explicitly stored entries.
pub fn zeros(major_dim: usize, minor_dim: usize) -> Self {
Self {
major_offsets: vec![0; major_dim + 1],
minor_indices: vec![],
minor_dim,
}
}
/// The offsets for the major dimension.
#[inline]
#[must_use]
pub fn major_offsets(&self) -> &[usize] {
&self.major_offsets
}
/// The indices for the minor dimension.
#[inline]
#[must_use]
pub fn minor_indices(&self) -> &[usize] {
&self.minor_indices
}
/// The number of major lanes in the pattern.
#[inline]
#[must_use]
pub fn major_dim(&self) -> usize {
assert!(self.major_offsets.len() > 0);
self.major_offsets.len() - 1
}
/// The number of minor lanes in the pattern.
#[inline]
#[must_use]
pub fn minor_dim(&self) -> usize {
self.minor_dim
}
/// The number of "non-zeros", i.e. explicitly stored entries in the pattern.
#[inline]
#[must_use]
pub fn nnz(&self) -> usize {
self.minor_indices.len()
}
/// Get the lane at the given index.
///
/// Panics
/// ------
///
/// Panics if `major_index` is out of bounds.
#[inline]
#[must_use]
pub fn lane(&self, major_index: usize) -> &[usize] {
self.get_lane(major_index).unwrap()
}
/// Get the lane at the given index, or `None` if out of bounds.
#[inline]
#[must_use]
pub fn get_lane(&self, major_index: usize) -> Option<&[usize]> {
let offset_begin = *self.major_offsets().get(major_index)?;
let offset_end = *self.major_offsets().get(major_index + 1)?;
Some(&self.minor_indices()[offset_begin..offset_end])
}
/// Try to construct a sparsity pattern from the given dimensions, major offsets
/// and minor indices.
///
/// Returns an error if the data does not conform to the requirements.
pub fn try_from_offsets_and_indices(
major_dim: usize,
minor_dim: usize,
major_offsets: Vec<usize>,
minor_indices: Vec<usize>,
) -> Result<Self, SparsityPatternFormatError> {
use SparsityPatternFormatError::*;
if major_offsets.len() != major_dim + 1 {
return Err(InvalidOffsetArrayLength);
}
// Check that the first and last offsets conform to the specification
{
let first_offset_ok = *major_offsets.first().unwrap() == 0;
let last_offset_ok = *major_offsets.last().unwrap() == minor_indices.len();
if !first_offset_ok || !last_offset_ok {
return Err(InvalidOffsetFirstLast);
}
}
// Test that each lane has strictly monotonically increasing minor indices, i.e.
// minor indices within a lane are sorted, unique. In addition, each minor index
// must be in bounds with respect to the minor dimension.
{
for lane_idx in 0..major_dim {
let range_start = major_offsets[lane_idx];
let range_end = major_offsets[lane_idx + 1];
// Test that major offsets are monotonically increasing
if range_start > range_end {
return Err(NonmonotonicOffsets);
}
let minor_indices = &minor_indices[range_start..range_end];
// We test for in-bounds, uniqueness and monotonicity at the same time
// to ensure that we only visit each minor index once
let mut iter = minor_indices.iter();
let mut prev = None;
while let Some(next) = iter.next().copied() {
if next >= minor_dim {
return Err(MinorIndexOutOfBounds);
}
if let Some(prev) = prev {
if prev > next {
return Err(NonmonotonicMinorIndices);
} else if prev == next {
return Err(DuplicateEntry);
}
}
prev = Some(next);
}
}
}
Ok(Self {
major_offsets,
minor_indices,
minor_dim,
})
}
/// An iterator over the explicitly stored "non-zero" entries (i, j).
///
/// The iteration happens in a lane-major fashion, meaning that the lane index i
/// increases monotonically, and the minor index j increases monotonically within each
/// lane i.
///
/// Examples
/// --------
///
/// ```
/// # use nalgebra_sparse::pattern::SparsityPattern;
/// let offsets = vec![0, 2, 3, 4];
/// let minor_indices = vec![0, 2, 1, 0];
/// let pattern = SparsityPattern::try_from_offsets_and_indices(3, 4, offsets, minor_indices)
/// .unwrap();
///
/// let entries: Vec<_> = pattern.entries().collect();
/// assert_eq!(entries, vec![(0, 0), (0, 2), (1, 1), (2, 0)]);
/// ```
///
#[must_use]
pub fn entries(&self) -> SparsityPatternIter<'_> {
SparsityPatternIter::from_pattern(self)
}
/// Returns the raw offset and index data for the sparsity pattern.
///
/// Examples
/// --------
///
/// ```
/// # use nalgebra_sparse::pattern::SparsityPattern;
/// let offsets = vec![0, 2, 3, 4];
/// let minor_indices = vec![0, 2, 1, 0];
/// let pattern = SparsityPattern::try_from_offsets_and_indices(
/// 3,
/// 4,
/// offsets.clone(),
/// minor_indices.clone())
/// .unwrap();
/// let (offsets2, minor_indices2) = pattern.disassemble();
/// assert_eq!(offsets2, offsets);
/// assert_eq!(minor_indices2, minor_indices);
/// ```
pub fn disassemble(self) -> (Vec<usize>, Vec<usize>) {
(self.major_offsets, self.minor_indices)
}
/// Computes the transpose of the sparsity pattern.
///
/// This is analogous to matrix transposition, i.e. an entry `(i, j)` becomes `(j, i)` in the
/// new pattern.
#[must_use]
pub fn transpose(&self) -> Self {
// By using unit () values, we can use the same routines as for CSR/CSC matrices
let values = vec![(); self.nnz()];
let (new_offsets, new_indices, _) = transpose_cs(
self.major_dim(),
self.minor_dim(),
self.major_offsets(),
self.minor_indices(),
&values,
);
// TODO: Skip checks
Self::try_from_offsets_and_indices(
self.minor_dim(),
self.major_dim(),
new_offsets,
new_indices,
)
.expect("Internal error: Transpose should never fail.")
}
}
/// Error type for `SparsityPattern` format errors.
#[non_exhaustive]
#[derive(Copy, Clone, Debug, PartialEq, Eq)]
pub enum SparsityPatternFormatError {
/// Indicates an invalid number of offsets.
///
/// The number of offsets must be equal to (major_dim + 1).
InvalidOffsetArrayLength,
/// Indicates that the first or last entry in the offset array did not conform to
/// specifications.
///
/// The first entry must be 0, and the last entry must be exactly one greater than the
/// major dimension.
InvalidOffsetFirstLast,
/// Indicates that the major offsets are not monotonically increasing.
NonmonotonicOffsets,
/// One or more minor indices are out of bounds.
MinorIndexOutOfBounds,
/// One or more duplicate entries were detected.
///
/// Two entries are considered duplicates if they are part of the same major lane and have
/// the same minor index.
DuplicateEntry,
/// Indicates that minor indices are not monotonically increasing within each lane.
NonmonotonicMinorIndices,
}
impl From<SparsityPatternFormatError> for SparseFormatError {
fn from(err: SparsityPatternFormatError) -> Self {
use crate::SparseFormatErrorKind;
use crate::SparseFormatErrorKind::*;
use SparsityPatternFormatError::DuplicateEntry as PatternDuplicateEntry;
use SparsityPatternFormatError::*;
match err {
InvalidOffsetArrayLength
| InvalidOffsetFirstLast
| NonmonotonicOffsets
| NonmonotonicMinorIndices => {
SparseFormatError::from_kind_and_error(InvalidStructure, Box::from(err))
}
MinorIndexOutOfBounds => {
SparseFormatError::from_kind_and_error(IndexOutOfBounds, Box::from(err))
}
PatternDuplicateEntry => SparseFormatError::from_kind_and_error(
#[allow(unused_qualifications)]
SparseFormatErrorKind::DuplicateEntry,
Box::from(err),
),
}
}
}
impl fmt::Display for SparsityPatternFormatError {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
match self {
SparsityPatternFormatError::InvalidOffsetArrayLength => {
write!(f, "Length of offset array is not equal to (major_dim + 1).")
}
SparsityPatternFormatError::InvalidOffsetFirstLast => {
write!(f, "First or last offset is incompatible with format.")
}
SparsityPatternFormatError::NonmonotonicOffsets => {
write!(f, "Offsets are not monotonically increasing.")
}
SparsityPatternFormatError::MinorIndexOutOfBounds => {
write!(f, "A minor index is out of bounds.")
}
SparsityPatternFormatError::DuplicateEntry => {
write!(f, "Input data contains duplicate entries.")
}
SparsityPatternFormatError::NonmonotonicMinorIndices => {
write!(
f,
"Minor indices are not monotonically increasing within each lane."
)
}
}
}
}
impl Error for SparsityPatternFormatError {}
/// Iterator type for iterating over entries in a sparsity pattern.
#[derive(Debug, Clone)]
pub struct SparsityPatternIter<'a> {
// See implementation of Iterator::next for an explanation of how these members are used
major_offsets: &'a [usize],
minor_indices: &'a [usize],
current_lane_idx: usize,
remaining_minors_in_lane: &'a [usize],
}
impl<'a> SparsityPatternIter<'a> {
fn from_pattern(pattern: &'a SparsityPattern) -> Self {
let first_lane_end = pattern.major_offsets().get(1).unwrap_or(&0);
let minors_in_first_lane = &pattern.minor_indices()[0..*first_lane_end];
Self {
major_offsets: pattern.major_offsets(),
minor_indices: pattern.minor_indices(),
current_lane_idx: 0,
remaining_minors_in_lane: minors_in_first_lane,
}
}
}
impl<'a> Iterator for SparsityPatternIter<'a> {
type Item = (usize, usize);
#[inline]
fn next(&mut self) -> Option<Self::Item> {
// We ensure fast iteration across each lane by iteratively "draining" a slice
// corresponding to the remaining column indices in the particular lane.
// When we reach the end of this slice, we are at the end of a lane,
// and we must do some bookkeeping for preparing the iteration of the next lane
// (or stop iteration if we're through all lanes).
// This way we can avoid doing unnecessary bookkeeping on every iteration,
// instead paying a small price whenever we jump to a new lane.
if let Some(minor_idx) = self.remaining_minors_in_lane.first() {
let item = Some((self.current_lane_idx, *minor_idx));
self.remaining_minors_in_lane = &self.remaining_minors_in_lane[1..];
item
} else {
loop {
// Keep skipping lanes until we found a non-empty lane or there are no more lanes
if self.current_lane_idx + 2 >= self.major_offsets.len() {
// We've processed all lanes, so we're at the end of the iterator
// (note: keep in mind that offsets.len() == major_dim() + 1, hence we need +2)
return None;
} else {
// Bump lane index and check if the lane is non-empty
self.current_lane_idx += 1;
let lower = self.major_offsets[self.current_lane_idx];
let upper = self.major_offsets[self.current_lane_idx + 1];
if upper > lower {
self.remaining_minors_in_lane = &self.minor_indices[(lower + 1)..upper];
return Some((self.current_lane_idx, self.minor_indices[lower]));
}
}
}
}
}
}