Module sparse_gpu 

GPU-ready sparse matrix formats and iterative solvers with data-oriented layouts.

Provides CSR, ELLPACK, Hybrid (ELL+COO), and Block-CSR formats along with iterative solvers (CG, BiCGSTAB, preconditioned CG) suitable for GPU offload.

Structs

BlockCsrMatrix

Block-CSR matrix where every stored entry is a block_size × block_size dense tile.

CsrMatrix

Compressed Sparse Row (CSR) matrix stored as plain f64 arrays.

EllMatrix

ELLPACK-format sparse matrix: rows padded to max_nnz_per_row.

HybridMatrix

Hybrid ELL+COO matrix: regular rows stored in ELL, overflow in COO.

SparseTriplet

Coordinate (COO) format sparse matrix for incremental assembly.

Functions

assemble_1d_laplacian

Assemble a 1D Laplacian matrix of size n × n (tridiagonal: 2 on diag, -1 off-diag).

axpy

y + alpha * x (AXPY).

bicgstab_solve

BiCGSTAB solver for general (possibly non-symmetric) systems A x = b.

cg_solve

Conjugate Gradient solver for symmetric positive-definite systems A x = b.

compute_nnz_per_row

Compute the number of non-zeros per row for a CSR matrix.

csr_to_ell

Convert a CSR matrix to ELLPACK format (convenience wrapper).

dot

Dot product of two vectors.

extract_diagonal

Extract the main diagonal of a CSR matrix.

frobenius_norm

Compute the Frobenius norm of a sparse matrix: sqrt(sum(a_ij^2)).

jacobi_preconditioned_cg

Conjugate Gradient with Jacobi (diagonal) preconditioner for A x = b (SPD).

norm2

Euclidean norm of a vector.

optimal_ell_row_width

Choose the ELLPACK row width (max non-zeros per row) to minimize padding waste.

scale_vec

Scale every element of x by s.

simulate_spmv_throughput

Estimate SpMV throughput in GFLOPS given matrix dimensions and nnz count.

sparse_lower_triangular_solve

Forward-substitution solve L x = b where L is lower-triangular (CSR).

sparse_upper_triangular_solve

Back-substitution solve U x = b where U is upper-triangular (CSR).

spmv_segmented

Segmented SpMV: processes each row independently to prepare for GPU-style parallel execution. Functionally identical to CsrMatrix::spmv but structured for row-parallel dispatch.