Module cusolver 

cuSOLVER-backed GPU linear algebra: SVD, Cholesky, QR, Solve.

Each operation follows the cuSOLVER pattern:

1. Query workspace size via *_bufferSize.
2. Allocate workspace + output buffers on the device.
3. Call the cuSOLVER routine.
4. Check devInfo — non-zero means the operation failed (singular matrix, etc.).

All functions operate on column-major data because cuSOLVER (LAPACK-style) uses column-major layout. The caller is responsible for transposing row-major tensors before calling and transposing outputs back.

Functions

gpu_cholesky_f32

Compute the Cholesky decomposition of an n-by-n SPD matrix (row-major f32).

gpu_cholesky_f64

Compute the Cholesky decomposition of an n-by-n SPD matrix (row-major f64).

gpu_qr_f32

Compute the reduced QR decomposition of an m-by-n matrix (row-major f32).

gpu_qr_f64

Compute the reduced QR decomposition of an m-by-n matrix (row-major f64).

gpu_solve_f32

Solve A * X = B for X where A is n-by-n and B is n-by-nrhs (row-major f32).

gpu_solve_f64

Solve A * X = B for X where A is n-by-n and B is n-by-nrhs (row-major f64).

gpu_svd_f32

Compute the thin SVD of an m-by-n matrix (row-major f32).

gpu_svd_f64

Compute the thin SVD of an m-by-n matrix (row-major f64).