ruvector_solver/types.rs
1//! Core types for sparse linear solvers.
2//!
3//! Provides [`CsrMatrix`] for compressed sparse row storage and result types
4//! for solver convergence tracking.
5
6use std::time::Duration;
7
8// ---------------------------------------------------------------------------
9// CsrMatrix<T>
10// ---------------------------------------------------------------------------
11
12/// Compressed Sparse Row (CSR) matrix.
13///
14/// Stores only non-zero entries for efficient sparse matrix-vector
15/// multiplication in O(nnz) time with excellent cache locality.
16///
17/// # Layout
18///
19/// For a matrix with `m` rows and `nnz` non-zeros:
20/// - `row_ptr` has length `m + 1`
21/// - `col_indices` and `values` each have length `nnz`
22/// - Row `i` spans indices `row_ptr[i]..row_ptr[i+1]`
23#[derive(Debug, Clone)]
24pub struct CsrMatrix<T> {
25 /// Row pointers: `row_ptr[i]` is the start index in `col_indices`/`values`
26 /// for row `i`.
27 pub row_ptr: Vec<usize>,
28 /// Column indices for each non-zero entry.
29 pub col_indices: Vec<usize>,
30 /// Values for each non-zero entry.
31 pub values: Vec<T>,
32 /// Number of rows.
33 pub rows: usize,
34 /// Number of columns.
35 pub cols: usize,
36}
37
38impl<T: Copy + Default + std::ops::Mul<Output = T> + std::ops::AddAssign> CsrMatrix<T> {
39 /// Sparse matrix-vector multiply: `y = A * x`.
40 ///
41 /// # Panics
42 ///
43 /// Debug-asserts that `x.len() >= self.cols` and `y.len() >= self.rows`.
44 #[inline]
45 pub fn spmv(&self, x: &[T], y: &mut [T]) {
46 debug_assert!(
47 x.len() >= self.cols,
48 "spmv: x.len()={} < cols={}",
49 x.len(),
50 self.cols,
51 );
52 debug_assert!(
53 y.len() >= self.rows,
54 "spmv: y.len()={} < rows={}",
55 y.len(),
56 self.rows,
57 );
58
59 for i in 0..self.rows {
60 let mut sum = T::default();
61 let start = self.row_ptr[i];
62 let end = self.row_ptr[i + 1];
63
64 for idx in start..end {
65 sum += self.values[idx] * x[self.col_indices[idx]];
66 }
67 y[i] = sum;
68 }
69 }
70}
71
72impl CsrMatrix<f32> {
73 /// High-performance SpMV with bounds-check elimination.
74 ///
75 /// Identical to [`spmv`](Self::spmv) but uses `unsafe` indexing to
76 /// eliminate per-element bounds checks in the inner loop, which is the
77 /// single hottest path in all iterative solvers.
78 ///
79 /// # Safety contract
80 ///
81 /// The caller must ensure the CSR structure is valid (use
82 /// [`validate_csr_matrix`](crate::validation::validate_csr_matrix) once
83 /// before entering the solve loop). The `x` and `y` slices must have
84 /// lengths `>= cols` and `>= rows` respectively.
85 #[inline]
86 pub fn spmv_unchecked(&self, x: &[f32], y: &mut [f32]) {
87 debug_assert!(x.len() >= self.cols);
88 debug_assert!(y.len() >= self.rows);
89
90 let vals = self.values.as_ptr();
91 let cols = self.col_indices.as_ptr();
92 let rp = self.row_ptr.as_ptr();
93
94 for i in 0..self.rows {
95 // SAFETY: row_ptr has length rows+1, so i and i+1 are in bounds.
96 let start = unsafe { *rp.add(i) };
97 let end = unsafe { *rp.add(i + 1) };
98 let mut sum = 0.0f32;
99
100 for idx in start..end {
101 // SAFETY: idx < nnz (enforced by valid CSR structure),
102 // col_indices[idx] < cols <= x.len() (enforced by validation).
103 unsafe {
104 let v = *vals.add(idx);
105 let c = *cols.add(idx);
106 sum += v * *x.get_unchecked(c);
107 }
108 }
109 // SAFETY: i < rows <= y.len()
110 unsafe { *y.get_unchecked_mut(i) = sum };
111 }
112 }
113
114 /// Fused SpMV + residual computation: computes `r[j] = rhs[j] - (A*x)[j]`
115 /// and returns `||r||^2` in a single pass, avoiding a separate allocation
116 /// for `Ax`.
117 ///
118 /// This eliminates one full memory traversal per iteration compared to
119 /// separate `spmv` + vector subtraction.
120 #[inline]
121 pub fn fused_residual_norm_sq(
122 &self,
123 x: &[f32],
124 rhs: &[f32],
125 residual: &mut [f32],
126 ) -> f64 {
127 debug_assert!(x.len() >= self.cols);
128 debug_assert!(rhs.len() >= self.rows);
129 debug_assert!(residual.len() >= self.rows);
130
131 let vals = self.values.as_ptr();
132 let cols = self.col_indices.as_ptr();
133 let rp = self.row_ptr.as_ptr();
134 let mut norm_sq = 0.0f64;
135
136 for i in 0..self.rows {
137 let start = unsafe { *rp.add(i) };
138 let end = unsafe { *rp.add(i + 1) };
139 let mut ax_i = 0.0f32;
140
141 for idx in start..end {
142 unsafe {
143 let v = *vals.add(idx);
144 let c = *cols.add(idx);
145 ax_i += v * *x.get_unchecked(c);
146 }
147 }
148
149 let r_i = rhs[i] - ax_i;
150 residual[i] = r_i;
151 norm_sq += (r_i as f64) * (r_i as f64);
152 }
153
154 norm_sq
155 }
156}
157
158impl CsrMatrix<f64> {
159 /// High-performance SpMV for f64 with bounds-check elimination.
160 #[inline]
161 pub fn spmv_unchecked(&self, x: &[f64], y: &mut [f64]) {
162 debug_assert!(x.len() >= self.cols);
163 debug_assert!(y.len() >= self.rows);
164
165 let vals = self.values.as_ptr();
166 let cols = self.col_indices.as_ptr();
167 let rp = self.row_ptr.as_ptr();
168
169 for i in 0..self.rows {
170 let start = unsafe { *rp.add(i) };
171 let end = unsafe { *rp.add(i + 1) };
172 let mut sum = 0.0f64;
173
174 for idx in start..end {
175 unsafe {
176 let v = *vals.add(idx);
177 let c = *cols.add(idx);
178 sum += v * *x.get_unchecked(c);
179 }
180 }
181 unsafe { *y.get_unchecked_mut(i) = sum };
182 }
183 }
184}
185
186impl<T> CsrMatrix<T> {
187 /// Number of non-zero entries.
188 #[inline]
189 pub fn nnz(&self) -> usize {
190 self.values.len()
191 }
192
193 /// Number of non-zeros in a specific row (i.e. the row degree for an
194 /// adjacency matrix).
195 #[inline]
196 pub fn row_degree(&self, row: usize) -> usize {
197 self.row_ptr[row + 1] - self.row_ptr[row]
198 }
199
200 /// Iterate over `(col_index, &value)` pairs for the given row.
201 #[inline]
202 pub fn row_entries(&self, row: usize) -> impl Iterator<Item = (usize, &T)> {
203 let start = self.row_ptr[row];
204 let end = self.row_ptr[row + 1];
205 self.col_indices[start..end]
206 .iter()
207 .copied()
208 .zip(self.values[start..end].iter())
209 }
210}
211
212impl<T: Copy + Default> CsrMatrix<T> {
213 /// Transpose: produces `A^T` in CSR form.
214 ///
215 /// Uses a two-pass counting sort in O(nnz + rows + cols) time and
216 /// O(nnz) extra memory. Required by backward push which operates on
217 /// the reversed adjacency structure.
218 pub fn transpose(&self) -> CsrMatrix<T> {
219 let nnz = self.nnz();
220 let t_rows = self.cols;
221 let t_cols = self.rows;
222
223 // Pass 1: count entries per new row (= old column).
224 let mut row_ptr = vec![0usize; t_rows + 1];
225 for &c in &self.col_indices {
226 row_ptr[c + 1] += 1;
227 }
228 for i in 1..=t_rows {
229 row_ptr[i] += row_ptr[i - 1];
230 }
231
232 // Pass 2: scatter entries into the transposed arrays.
233 let mut col_indices = vec![0usize; nnz];
234 let mut values = vec![T::default(); nnz];
235 let mut cursor = row_ptr.clone();
236
237 for row in 0..self.rows {
238 let start = self.row_ptr[row];
239 let end = self.row_ptr[row + 1];
240 for idx in start..end {
241 let c = self.col_indices[idx];
242 let dest = cursor[c];
243 col_indices[dest] = row;
244 values[dest] = self.values[idx];
245 cursor[c] += 1;
246 }
247 }
248
249 CsrMatrix {
250 row_ptr,
251 col_indices,
252 values,
253 rows: t_rows,
254 cols: t_cols,
255 }
256 }
257}
258
259impl<T: Copy + Default + std::ops::AddAssign> CsrMatrix<T> {
260 /// Build a CSR matrix from COO (coordinate) triplets.
261 ///
262 /// Entries are sorted by (row, col) internally. Duplicate positions at the
263 /// same (row, col) are kept as separate entries (caller should pre-merge if
264 /// needed).
265 pub fn from_coo_generic(
266 rows: usize,
267 cols: usize,
268 entries: impl IntoIterator<Item = (usize, usize, T)>,
269 ) -> Self {
270 let mut sorted: Vec<_> = entries.into_iter().collect();
271 sorted.sort_unstable_by_key(|(r, c, _)| (*r, *c));
272
273 let nnz = sorted.len();
274 let mut row_ptr = vec![0usize; rows + 1];
275 let mut col_indices = Vec::with_capacity(nnz);
276 let mut values = Vec::with_capacity(nnz);
277
278 for &(r, _, _) in &sorted {
279 assert!(r < rows, "row index {} out of bounds (rows={})", r, rows);
280 row_ptr[r + 1] += 1;
281 }
282 for i in 1..=rows {
283 row_ptr[i] += row_ptr[i - 1];
284 }
285
286 for (_, c, v) in sorted {
287 assert!(c < cols, "col index {} out of bounds (cols={})", c, cols);
288 col_indices.push(c);
289 values.push(v);
290 }
291
292 Self {
293 row_ptr,
294 col_indices,
295 values,
296 rows,
297 cols,
298 }
299 }
300}
301
302impl CsrMatrix<f32> {
303 /// Build a CSR matrix from COO (coordinate) triplets.
304 ///
305 /// Entries are sorted by (row, col) internally. Duplicate positions are
306 /// summed.
307 pub fn from_coo(
308 rows: usize,
309 cols: usize,
310 entries: impl IntoIterator<Item = (usize, usize, f32)>,
311 ) -> Self {
312 Self::from_coo_generic(rows, cols, entries)
313 }
314
315 /// Build a square identity matrix of dimension `n` in CSR format.
316 pub fn identity(n: usize) -> Self {
317 let row_ptr: Vec<usize> = (0..=n).collect();
318 let col_indices: Vec<usize> = (0..n).collect();
319 let values = vec![1.0f32; n];
320
321 Self {
322 row_ptr,
323 col_indices,
324 values,
325 rows: n,
326 cols: n,
327 }
328 }
329}
330
331impl CsrMatrix<f64> {
332 /// Build a CSR matrix from COO (coordinate) triplets (f64 variant).
333 ///
334 /// Entries are sorted by (row, col) internally.
335 pub fn from_coo(
336 rows: usize,
337 cols: usize,
338 entries: impl IntoIterator<Item = (usize, usize, f64)>,
339 ) -> Self {
340 Self::from_coo_generic(rows, cols, entries)
341 }
342
343 /// Build a square identity matrix of dimension `n` in CSR format (f64).
344 pub fn identity(n: usize) -> Self {
345 let row_ptr: Vec<usize> = (0..=n).collect();
346 let col_indices: Vec<usize> = (0..n).collect();
347 let values = vec![1.0f64; n];
348
349 Self {
350 row_ptr,
351 col_indices,
352 values,
353 rows: n,
354 cols: n,
355 }
356 }
357}
358
359// ---------------------------------------------------------------------------
360// Solver result types
361// ---------------------------------------------------------------------------
362
363/// Algorithm identifier for solver selection and routing.
364///
365/// Each variant corresponds to a solver strategy with different complexity
366/// characteristics and applicability constraints. The [`SolverRouter`] selects
367/// the best algorithm based on the matrix [`SparsityProfile`] and [`QueryType`].
368///
369/// [`SolverRouter`]: crate::router::SolverRouter
370#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, serde::Serialize, serde::Deserialize)]
371pub enum Algorithm {
372 /// Neumann series: `x = sum_{k=0}^{K} (I - A)^k * b`.
373 ///
374 /// Requires spectral radius < 1. Best for diagonally dominant, very sparse
375 /// systems where the series converges in O(log(1/eps)) terms.
376 Neumann,
377 /// Jacobi iterative solver.
378 Jacobi,
379 /// Gauss-Seidel iterative solver.
380 GaussSeidel,
381 /// Forward Push (Andersen-Chung-Lang) for Personalized PageRank.
382 ///
383 /// Computes an approximate PPR vector by pushing residual mass forward
384 /// along edges. Sublinear in graph size for single-source queries.
385 ForwardPush,
386 /// Backward Push for target-centric PPR.
387 ///
388 /// Dual of Forward Push: propagates contributions backward from a target
389 /// node.
390 BackwardPush,
391 /// Conjugate Gradient (CG) iterative solver.
392 ///
393 /// Optimal for symmetric positive-definite systems. Converges in at most
394 /// `n` steps; practical convergence depends on the condition number.
395 CG,
396 /// Hybrid random-walk approach combining push with Monte Carlo sampling.
397 ///
398 /// For large graphs where pure push is too expensive, this approach uses
399 /// random walks to estimate the tail of the PageRank distribution.
400 HybridRandomWalk,
401 /// TRUE (Topology-aware Reduction for Updating Equations) batch solver.
402 ///
403 /// Exploits shared sparsity structure across a batch of right-hand sides
404 /// to amortise factorisation cost. Best when `batch_size` is large.
405 TRUE,
406 /// Block Maximum Spanning Subgraph Preconditioned solver.
407 ///
408 /// Uses a maximum spanning tree preconditioner for ill-conditioned systems
409 /// where CG and Neumann both struggle.
410 BMSSP,
411 /// Dense direct solver (LU/Cholesky fallback).
412 ///
413 /// Last-resort O(n^3) solver used when iterative methods fail. Only
414 /// practical for small matrices.
415 Dense,
416}
417
418impl std::fmt::Display for Algorithm {
419 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
420 match self {
421 Algorithm::Neumann => write!(f, "neumann"),
422 Algorithm::Jacobi => write!(f, "jacobi"),
423 Algorithm::GaussSeidel => write!(f, "gauss-seidel"),
424 Algorithm::ForwardPush => write!(f, "forward-push"),
425 Algorithm::BackwardPush => write!(f, "backward-push"),
426 Algorithm::CG => write!(f, "cg"),
427 Algorithm::HybridRandomWalk => write!(f, "hybrid-random-walk"),
428 Algorithm::TRUE => write!(f, "true-solver"),
429 Algorithm::BMSSP => write!(f, "bmssp"),
430 Algorithm::Dense => write!(f, "dense"),
431 }
432 }
433}
434
435// ---------------------------------------------------------------------------
436// Query & profile types for routing
437// ---------------------------------------------------------------------------
438
439/// Query type describing what the caller wants to solve.
440///
441/// The [`SolverRouter`] inspects this together with the [`SparsityProfile`] to
442/// select the most appropriate [`Algorithm`].
443///
444/// [`SolverRouter`]: crate::router::SolverRouter
445#[derive(Debug, Clone, Copy, PartialEq, Eq)]
446pub enum QueryType {
447 /// Standard sparse linear system `Ax = b`.
448 LinearSystem,
449
450 /// Single-source Personalized PageRank.
451 PageRankSingle {
452 /// Source node index.
453 source: usize,
454 },
455
456 /// Pairwise Personalized PageRank between two nodes.
457 PageRankPairwise {
458 /// Source node index.
459 source: usize,
460 /// Target node index.
461 target: usize,
462 },
463
464 /// Spectral graph filter using polynomial expansion.
465 SpectralFilter {
466 /// Degree of the Chebyshev/polynomial expansion.
467 polynomial_degree: usize,
468 },
469
470 /// Batch of linear systems sharing the same matrix `A` but different
471 /// right-hand sides.
472 BatchLinearSystem {
473 /// Number of right-hand sides in the batch.
474 batch_size: usize,
475 },
476}
477
478/// Sparsity profile summarising the structural and numerical properties
479/// of a matrix that are relevant for algorithm selection.
480///
481/// Computed once by [`SolverOrchestrator::analyze_sparsity`] and reused
482/// across multiple solves on the same matrix.
483///
484/// [`SolverOrchestrator::analyze_sparsity`]: crate::router::SolverOrchestrator::analyze_sparsity
485#[derive(Debug, Clone)]
486pub struct SparsityProfile {
487 /// Number of rows.
488 pub rows: usize,
489 /// Number of columns.
490 pub cols: usize,
491 /// Total number of non-zero entries.
492 pub nnz: usize,
493 /// Fraction of non-zeros: `nnz / (rows * cols)`.
494 pub density: f64,
495 /// `true` if `|a_ii| > sum_{j != i} |a_ij|` for every row.
496 pub is_diag_dominant: bool,
497 /// Estimated spectral radius of the Jacobi iteration matrix `D^{-1}(L+U)`.
498 pub estimated_spectral_radius: f64,
499 /// Rough estimate of the 2-norm condition number.
500 pub estimated_condition: f64,
501 /// `true` if the matrix appears to be symmetric (checked on structure only).
502 pub is_symmetric_structure: bool,
503 /// Average number of non-zeros per row.
504 pub avg_nnz_per_row: f64,
505 /// Maximum number of non-zeros in any single row.
506 pub max_nnz_per_row: usize,
507}
508
509/// Estimated computational complexity for a solve.
510///
511/// Returned by [`SolverOrchestrator::estimate_complexity`] to let callers
512/// decide whether to proceed, batch, or reject a query.
513///
514/// [`SolverOrchestrator::estimate_complexity`]: crate::router::SolverOrchestrator::estimate_complexity
515#[derive(Debug, Clone)]
516pub struct ComplexityEstimate {
517 /// Algorithm that would be selected.
518 pub algorithm: Algorithm,
519 /// Estimated number of floating-point operations.
520 pub estimated_flops: u64,
521 /// Estimated number of iterations (for iterative methods).
522 pub estimated_iterations: usize,
523 /// Estimated peak memory usage in bytes.
524 pub estimated_memory_bytes: usize,
525 /// A qualitative complexity class label.
526 pub complexity_class: ComplexityClass,
527}
528
529/// Qualitative complexity class.
530#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
531pub enum ComplexityClass {
532 /// O(nnz * log(1/eps)) -- sublinear in matrix dimension.
533 SublinearNnz,
534 /// O(n * sqrt(kappa)) -- CG-like.
535 SqrtCondition,
536 /// O(n * nnz_per_row) -- linear scan.
537 Linear,
538 /// O(n^2) or worse -- superlinear.
539 Quadratic,
540 /// O(n^3) -- dense factorisation.
541 Cubic,
542}
543
544/// Compute lane priority for solver scheduling.
545#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
546pub enum ComputeLane {
547 /// Low-latency lane for small problems.
548 Fast,
549 /// Default throughput lane.
550 Normal,
551 /// Batch lane for large problems.
552 Batch,
553}
554
555/// Budget constraints for solver execution.
556#[derive(Debug, Clone)]
557pub struct ComputeBudget {
558 /// Maximum wall-clock time allowed.
559 pub max_time: Duration,
560 /// Maximum number of iterations.
561 pub max_iterations: usize,
562 /// Target residual tolerance.
563 pub tolerance: f64,
564}
565
566impl Default for ComputeBudget {
567 fn default() -> Self {
568 Self {
569 max_time: Duration::from_secs(30),
570 max_iterations: 1000,
571 tolerance: 1e-6,
572 }
573 }
574}
575
576/// Per-iteration convergence snapshot.
577#[derive(Debug, Clone)]
578pub struct ConvergenceInfo {
579 /// Iteration index (0-based).
580 pub iteration: usize,
581 /// Residual L2 norm at this iteration.
582 pub residual_norm: f64,
583}
584
585/// Result returned by a successful solver invocation.
586#[derive(Debug, Clone)]
587pub struct SolverResult {
588 /// Solution vector x.
589 pub solution: Vec<f32>,
590 /// Number of iterations performed.
591 pub iterations: usize,
592 /// Final residual L2 norm.
593 pub residual_norm: f64,
594 /// Wall-clock time taken.
595 pub wall_time: Duration,
596 /// Per-iteration convergence history.
597 pub convergence_history: Vec<ConvergenceInfo>,
598 /// Algorithm used.
599 pub algorithm: Algorithm,
600}