Struct BatchedSolver 

pub struct BatchedSolver { /* private fields */ }

Batched matrix factorization engine.

Each thread block handles one matrix. For very small matrices (n <= 16), multiple matrices per thread block. All computation in registers/shared memory.

Implementations

impl BatchedSolver

pub fn new(handle: SolverHandle) -> Self

Creates a new batched solver.

pub fn handle(&self) -> &SolverHandle

Returns a reference to the underlying solver handle.

pub fn handle_mut(&mut self) -> &mut SolverHandle

Returns a mutable reference to the underlying solver handle.

pub fn batched_lu<T: GpuFloat>( &mut self, matrices: &mut DeviceBuffer<T>, pivots: &mut DeviceBuffer<i32>, n: usize, batch_count: usize, ) -> SolverResult<BatchedResult>

Batched LU factorization: factorize batch_count matrices of size n x n.

Input: matrices[batch_count * n * n] (column-major, contiguous). Output: in-place LU factors, pivots[batch_count * n].

Errors

Returns SolverError::DimensionMismatch if buffer sizes are incorrect or matrix size is out of range.

pub fn batched_qr<T: GpuFloat>( &mut self, matrices: &mut DeviceBuffer<T>, tau: &mut DeviceBuffer<T>, m: usize, n: usize, batch_count: usize, ) -> SolverResult<BatchedResult>

Batched QR factorization.

Output: in-place QR factors, tau[batch_count * min(m, n)] (Householder scalars).

Arguments

• matrices — contiguous buffer of batch_count matrices, each m x n, column-major.
• tau — output Householder scalars, length batch_count * min(m, n).
• m — number of rows per matrix.
• n — number of columns per matrix.
• batch_count — number of matrices.

Errors

Returns SolverError::DimensionMismatch if buffer sizes are incorrect.

pub fn batched_cholesky<T: GpuFloat>( &mut self, matrices: &mut DeviceBuffer<T>, n: usize, batch_count: usize, ) -> SolverResult<BatchedResult>

Batched Cholesky factorization (for SPD matrices).

Output: in-place lower triangular Cholesky factors.

Arguments

• matrices — contiguous buffer of batch_count SPD matrices, each n x n.
• n — matrix dimension.
• batch_count — number of matrices.

Errors

Returns SolverError::DimensionMismatch if buffer sizes are incorrect.

pub fn batched_solve<T: GpuFloat>( &mut self, a_matrices: &mut DeviceBuffer<T>, b_matrices: &mut DeviceBuffer<T>, n: usize, nrhs: usize, batch_count: usize, ) -> SolverResult<BatchedResult>

Batched linear solve using LU: solve A_i * X_i = B_i for each i.

First performs batched LU factorization on a_matrices, then uses the factors to solve the system. Both a_matrices and b_matrices are modified in-place.

Arguments

• a_matrices — contiguous batch_count coefficient matrices (n x n each).
• b_matrices — contiguous batch_count RHS matrices (n x nrhs each).
• n — system dimension.
• nrhs — number of right-hand sides.
• batch_count — number of systems to solve.

Errors

Returns SolverError::DimensionMismatch if dimensions are invalid.