README.md

cuSOLVER — Dense Solver Examples

5 examples covering direct linear system solvers: LU, QR, Cholesky, SVD, and eigenvalue decomposition. Enable with -Dcusolver=true.

Build & Run

zig build run-cusolver-<name> -Dcusolver=true

zig build run-cusolver-getrf  -Dcusolver=true
zig build run-cusolver-gesvd  -Dcusolver=true
zig build run-cusolver-potrf  -Dcusolver=true
zig build run-cusolver-syevd  -Dcusolver=true
zig build run-cusolver-geqrf  -Dcusolver=true

---

Examples

Example	File	Factorization	Description
getrf	getrf.zig	LU (PA = LU)	LU factorization + linear solve via backsubstitution
gesvd	gesvd.zig	SVD (A = UΣVᵀ)	Full singular value decomposition
potrf	potrf.zig	Cholesky (A = LLᵀ)	Cholesky factorization for SPD matrices + solve
syevd	syevd.zig	Eigen (Av = λv)	Symmetric eigenvalue decomposition
geqrf	geqrf.zig	QR (A = QR)	QR factorization with explicit Q extraction

---

Key API

const cusolver = @import("zcuda").cusolver;

const solver = try cusolver.CusolverDnContext.init(ctx);
defer solver.deinit();
const ext = try cusolver.CusolverDnExt.init(ctx);
defer ext.deinit();

// LU factorization (PA = LU)
const buf_size = try solver.sgetrf_bufferSize(m, n, d_a, lda);
const workspace = try stream.alloc(f32, @intCast(buf_size));
const d_piv  = try stream.alloc(i32, @min(m, n));
const d_info = try stream.alloc(i32, 1);
try solver.sgetrf(m, n, d_a, lda, workspace, d_piv, d_info, stream);

// Solve Ax = b (after LU)
try solver.sgetrs(.no_transpose, n, 1, d_a, lda, d_piv, d_b, ldb, d_info, stream);

// SVD (A = U Σ Vᵀ)
try ext.gesvdj(m, n, d_a, d_s, d_u, d_vt, d_info, .{
    .tol       = 1e-7,
    .max_sweeps = 100,
}, stream);

Note

devInfo (d_info) is a GPU-side CudaSlice(i32) per cuSOLVER API contract. Use stream.memcpyDtoH after ctx.synchronize() to read factorization status on the host.

→ Full API reference: docs/cusolver/README.md