Initial port from nalgebra::CsCholesky factorization to CscCholesky

This commit is contained in:
Andreas Longva 2021-01-11 15:14:54 +01:00
parent 6e34c23d05
commit aad2216c56
7 changed files with 395 additions and 1 deletions

View File

@ -17,3 +17,5 @@ proptest = { version = "0.10", optional = true }
[dev-dependencies]
itertools = "0.9"
matrixcompare = "0.1.3"
nalgebra = { version="0.23", path = "../", features = ["compare"] }

View File

@ -115,7 +115,6 @@ impl<T> CscMatrix<T> {
}
}
/// An iterator over non-zero triplets (i, j, v).
///
/// The iteration happens in column-major fashion, meaning that j increases monotonically,

View File

@ -0,0 +1,294 @@
// TODO: Remove this allowance
#![allow(missing_docs)]
use crate::pattern::SparsityPattern;
use crate::csc::CscMatrix;
use core::{mem, iter};
use nalgebra::{U1, VectorN, Dynamic, Scalar, RealField};
use num_traits::Zero;
use std::sync::Arc;
use std::ops::Add;
pub struct CscSymbolicCholesky {
// Pattern of the original matrix that was decomposed
m_pattern: Arc<SparsityPattern>,
l_pattern: SparsityPattern,
// u in this context is L^T, so that M = L L^T
u_pattern: SparsityPattern
}
impl CscSymbolicCholesky {
pub fn factor(pattern: &Arc<SparsityPattern>) -> Self {
assert_eq!(pattern.major_dim(), pattern.minor_dim(),
"Major and minor dimensions must be the same (square matrix).");
// TODO: Temporary stopgap solution to make things work until we can refactor
#[derive(Copy, Clone, PartialEq, Eq, Debug)]
struct DummyVal;
impl Zero for DummyVal {
fn zero() -> Self {
DummyVal
}
fn is_zero(&self) -> bool {
true
}
}
impl Add<DummyVal> for DummyVal {
type Output = Self;
fn add(self, rhs: DummyVal) -> Self::Output {
rhs
}
}
let dummy_vals = vec![DummyVal; pattern.nnz()];
let dummy_csc = CscMatrix::try_from_pattern_and_values(Arc::clone(pattern), dummy_vals)
.unwrap();
let (l, u) = nonzero_pattern(&dummy_csc);
// TODO: Don't clone unnecessarily
Self {
m_pattern: Arc::clone(pattern),
l_pattern: l.pattern().as_ref().clone(),
u_pattern: u.pattern().as_ref().clone()
}
}
pub fn l_pattern(&self) -> &SparsityPattern {
&self.l_pattern
}
}
pub struct CscCholesky<T> {
// Pattern of the original matrix
m_pattern: Arc<SparsityPattern>,
l_factor: CscMatrix<T>,
u_pattern: SparsityPattern,
work_x: Vec<T>,
work_c: Vec<usize>
}
#[derive(Debug, PartialEq, Eq, Clone)]
pub enum CholeskyError {
}
impl<T: RealField> CscCholesky<T> {
pub fn factor(matrix: &CscMatrix<T>) -> Result<Self, CholeskyError> {
let symbolic = CscSymbolicCholesky::factor(&*matrix.pattern());
assert_eq!(symbolic.l_pattern.nnz(), symbolic.u_pattern.nnz(),
"u is just the transpose of l, so should have the same nnz");
let l_nnz = symbolic.l_pattern.nnz();
let l_values = vec![T::zero(); l_nnz];
let l_factor = CscMatrix::try_from_pattern_and_values(Arc::new(symbolic.l_pattern), l_values)
.unwrap();
let mut factorization = CscCholesky {
m_pattern: symbolic.m_pattern,
l_factor,
u_pattern: symbolic.u_pattern,
work_x: vec![T::zero(); matrix.nrows()],
// Fill with MAX so that things hopefully totally fail if values are not
// overwritten. Might be easier to debug this way
work_c: vec![usize::MAX, matrix.ncols()],
};
factorization.refactor(matrix.values())?;
Ok(factorization)
}
pub fn refactor(&mut self, values: &[T]) -> Result<(), CholeskyError> {
self.decompose_left_looking(values)
}
pub fn l(&self) -> &CscMatrix<T> {
&self.l_factor
}
pub fn take_l(self) -> CscMatrix<T> {
self.l_factor
}
/// Perform a numerical left-looking cholesky decomposition of a matrix with the same structure as the
/// one used to initialize `self`, but with different non-zero values provided by `values`.
fn decompose_left_looking(&mut self, values: &[T]) -> Result<(), CholeskyError> {
assert!(
values.len() >= self.m_pattern.nnz(),
// TODO: Improve error message
"The set of values is too small."
);
let n = self.l_factor.nrows();
// Reset `work_c` to the column pointers of `l`.
self.work_c.clear();
self.work_c.extend_from_slice(self.l_factor.col_offsets());
unsafe {
for k in 0..n {
// Scatter the k-th column of the original matrix with the values provided.
let range_begin = *self.m_pattern.major_offsets().get_unchecked(k);
let range_end = *self.m_pattern.major_offsets().get_unchecked(k + 1);
let range_k = range_begin..range_end;
*self.work_x.get_unchecked_mut(k) = T::zero();
for p in range_k.clone() {
let irow = *self.m_pattern.minor_indices().get_unchecked(p);
if irow >= k {
*self.work_x.get_unchecked_mut(irow) = *values.get_unchecked(p);
}
}
for &j in self.u_pattern.lane(k) {
let factor = -*self
.l_factor
.values()
.get_unchecked(*self.work_c.get_unchecked(j));
*self.work_c.get_unchecked_mut(j) += 1;
if j < k {
let col_j = self.l_factor.col(j);
let col_j_entries = col_j.row_indices().iter().zip(col_j.values());
for (&z, val) in col_j_entries {
if z >= k {
*self.work_x.get_unchecked_mut(z) += val.inlined_clone() * factor;
}
}
}
}
let diag = *self.work_x.get_unchecked(k);
if diag > T::zero() {
let denom = diag.sqrt();
{
let (offsets, _, values) = self.l_factor.csc_data_mut();
*values
.get_unchecked_mut(*offsets.get_unchecked(k)) = denom;
}
let mut col_k = self.l_factor.col_mut(k);
let (col_k_rows, col_k_values) = col_k.rows_and_values_mut();
let col_k_entries = col_k_rows.iter().zip(col_k_values);
for (&p, val) in col_k_entries {
*val = *self.work_x.get_unchecked(p) / denom;
*self.work_x.get_unchecked_mut(p) = T::zero();
}
} else {
// self.ok = false;
// TODO: Return indefinite error (i.e. encountered non-positive diagonal
unimplemented!()
}
}
}
Ok(())
}
}
fn reach<T>(
m: &CscMatrix<T>,
j: usize,
max_j: usize,
tree: &[usize],
marks: &mut Vec<bool>,
out: &mut Vec<usize>,
) {
marks.clear();
marks.resize(tree.len(), false);
// TODO: avoid all those allocations.
let mut tmp = Vec::new();
let mut res = Vec::new();
for &irow in m.col(j).row_indices() {
let mut curr = irow;
while curr != usize::max_value() && curr <= max_j && !marks[curr] {
marks[curr] = true;
tmp.push(curr);
curr = tree[curr];
}
tmp.append(&mut res);
mem::swap(&mut tmp, &mut res);
}
// TODO: Is this right?
res.sort_unstable();
out.append(&mut res);
}
fn nonzero_pattern<T: Scalar + Zero>(
m: &CscMatrix<T>
) -> (CscMatrix<T>, CscMatrix<T>) {
// TODO: In order to stay as faithful as possible to the original implementation,
// we here return full matrices, whereas we actually only need to construct sparsity patterns
let etree = elimination_tree(m);
let (nrows, ncols) = (m.nrows(), m.ncols());
let mut rows = Vec::with_capacity(m.nnz());
// TODO: Use a Vec here instead
let mut cols = unsafe { VectorN::new_uninitialized_generic(Dynamic::new(nrows), U1) };
let mut marks = Vec::new();
// NOTE: the following will actually compute the non-zero pattern of
// the transpose of l.
for i in 0..nrows {
cols[i] = rows.len();
reach(m, i, i, &etree, &mut marks, &mut rows);
}
// TODO: Get rid of this in particular
let mut vals = Vec::with_capacity(rows.len());
unsafe {
vals.set_len(rows.len());
}
vals.shrink_to_fit();
// TODO: Remove this unnecessary conversion by using Vec throughout
let mut cols: Vec<_> = cols.iter().cloned().collect();
cols.push(rows.len());
let u = CscMatrix::try_from_csc_data(nrows, ncols, cols, rows, vals).unwrap();
// TODO: Avoid this transpose
let l = u.transpose();
(l, u)
}
fn elimination_tree<T>(m: &CscMatrix<T>) -> Vec<usize> {
let nrows = m.nrows();
let mut forest: Vec<_> = iter::repeat(usize::max_value()).take(nrows).collect();
let mut ancestor: Vec<_> = iter::repeat(usize::max_value()).take(nrows).collect();
for k in 0..nrows {
for &irow in m.col(k).row_indices() {
let mut i = irow;
while i < k {
let i_ancestor = ancestor[i];
ancestor[i] = k;
if i_ancestor == usize::max_value() {
forest[i] = k;
break;
}
i = i_ancestor;
}
}
}
forest
}

View File

@ -0,0 +1,4 @@
//! Matrix factorization for sparse matrices.
mod cholesky;
pub use cholesky::*;

View File

@ -89,6 +89,7 @@ pub mod csr;
pub mod pattern;
pub mod ops;
pub mod convert;
pub mod factorization;
pub(crate) mod cs;

View File

@ -0,0 +1,93 @@
#![cfg_attr(rustfmt, rustfmt_skip)]
use crate::common::{value_strategy, PROPTEST_MATRIX_DIM, PROPTEST_MAX_NNZ};
use nalgebra_sparse::csc::CscMatrix;
use nalgebra_sparse::factorization::{CscCholesky};
use nalgebra_sparse::proptest::csc;
use nalgebra::{Matrix5, Vector5, Cholesky, DMatrix};
use proptest::prelude::*;
use matrixcompare::assert_matrix_eq;
fn positive_definite() -> impl Strategy<Value=CscMatrix<f64>> {
let csc_f64 = csc(value_strategy::<f64>(),
PROPTEST_MATRIX_DIM,
PROPTEST_MATRIX_DIM,
PROPTEST_MAX_NNZ);
csc_f64
.prop_map(|x| {
// Add a small multiple of the identity to ensure positive definiteness
x.transpose() * &x + CscMatrix::identity(x.ncols())
})
}
proptest! {
#[test]
fn cholesky_correct_for_positive_definite_matrices(
matrix in positive_definite()
) {
let cholesky = CscCholesky::factor(&matrix).unwrap();
let l = cholesky.take_l();
let matrix_reconstructed = &l * l.transpose();
// TODO: Use matrixcompare instead
let diff = DMatrix::from(&(matrix_reconstructed - matrix));
prop_assert!(diff.abs().max() < 1e-8);
// TODO: Check that L is in fact lower triangular
}
}
// This is a test ported from nalgebra's "sparse" module, for the original CsCholesky impl
#[test]
fn cs_cholesky() {
let mut a = Matrix5::new(
40.0, 0.0, 0.0, 0.0, 0.0,
2.0, 60.0, 0.0, 0.0, 0.0,
1.0, 0.0, 11.0, 0.0, 0.0,
0.0, 0.0, 0.0, 50.0, 0.0,
1.0, 0.0, 0.0, 4.0, 10.0
);
a.fill_upper_triangle_with_lower_triangle();
test_cholesky(a);
let a = Matrix5::from_diagonal(&Vector5::new(40.0, 60.0, 11.0, 50.0, 10.0));
test_cholesky(a);
let mut a = Matrix5::new(
40.0, 0.0, 0.0, 0.0, 0.0,
2.0, 60.0, 0.0, 0.0, 0.0,
1.0, 0.0, 11.0, 0.0, 0.0,
1.0, 0.0, 0.0, 50.0, 0.0,
0.0, 0.0, 0.0, 4.0, 10.0
);
a.fill_upper_triangle_with_lower_triangle();
test_cholesky(a);
let mut a = Matrix5::new(
2.0, 0.0, 0.0, 0.0, 0.0,
0.0, 2.0, 0.0, 0.0, 0.0,
1.0, 1.0, 2.0, 0.0, 0.0,
0.0, 0.0, 0.0, 2.0, 0.0,
1.0, 1.0, 0.0, 0.0, 2.0
);
a.fill_upper_triangle_with_lower_triangle();
// Test crate::new, left_looking, and up_looking implementations.
test_cholesky(a);
}
fn test_cholesky(a: Matrix5<f64>) {
// TODO: Test "refactor"
let cs_a = CscMatrix::from(&a);
let chol_a = Cholesky::new(a).unwrap();
let chol_cs_a = CscCholesky::factor(&cs_a).unwrap();
let l = chol_a.l();
let cs_l = chol_cs_a.take_l();
let l = DMatrix::from_iterator(l.nrows(), l.ncols(), l.iter().cloned());
let cs_l_mat = DMatrix::from(&cs_l);
assert_matrix_eq!(l, cs_l_mat, comp = abs, tol = 1e-12);
}

View File

@ -1,4 +1,5 @@
mod coo;
mod cholesky;
mod ops;
mod pattern;
mod csr;