CUDA extension

CSR(A::CuSparseMatrixCSC)

Wrap A (m × n) without copying as the CSR of transpose(A) (n × m, transposed = true), sharing colPtr, rowVal and nzVal.

CSR(A::CuSparseMatrixCSR)

Wrap A without copying: the CSR shares rowPtr, colVal and nzVal (one-based, transposed = false).

DirectSolver(A::CuSparseMatrixCSR, structure::String, view::Char; index = 'O')
DirectSolver(A::CuSparseMatrixCSC, structure::String, view::Char; index = 'O')

Solver on A's arrays, without copies (≅ CudssSolver(A, structure, view)). A CuSparseMatrixCSC is read as the CSR of its transpose with the view flipped (MadNLP's lower-triangle CSC becomes an upper-triangle CSR), see DirectSolver.

CuSparseMatrixCSR(A::CSR)

The stored CSR matrix of A (the transposed flag is not applied) as a CuSparseMatrixCSR. One-based arrays are shared; zero-based index arrays are copied and rebased. Batches (nbatch(A) > 1) are rejected.

cholesky(A::CuSparseMatrixCSR, NoPivot(); view = 'F', check = false) -> DirectSolver
cholesky(Symmetric(A::CuSparseMatrixCSR)) / cholesky(Hermitian(A::CuSparseMatrixCSR))

LLᵀ/LLᴴ factorization of A on the GPU (≅ CUDSS.jl's cholesky); the wrappers pass their uplo as the view. See cholesky(::CSR).

ldlt(A::CuSparseMatrixCSR; view = 'F', check = false) -> DirectSolver
ldlt(Symmetric(A::CuSparseMatrixCSR)) / ldlt(Hermitian(A::CuSparseMatrixCSR))

LDLᵀ/LDLᴴ factorization of A on the GPU (≅ CUDSS.jl's ldlt); the wrappers pass their uplo as the view. See ldlt(::CSR).

lu(A::CuSparseMatrixCSR; check = false) -> DirectSolver

L D U factorization (structure "G") of A on the GPU (≅ CUDSS.jl's lu). See lu(::CSR).

max_local_bytes(::CUDABackend) -> Int

48 KiB: @localmem is static shared memory, and ptxas rejects more than 48 KiB of static shared data per kernel on every CUDA device (larger amounts need dynamic shared memory with an opt-in attribute). The regime-A kernels therefore stop at the 48 KiB class on CUDA (issue #60).

max_threads_per_multiprocessor(::CUDABackend) -> Int

CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR of the current device.

to_backend(A::SparseMatrixCSC, ::CUDABackend; index = 'O') -> CSR

CSR arrays of A built on the host, then uploaded as CuVectors.

update!(solver::DirectSolver, A::CuSparseMatrixCSR)
update!(solver::DirectSolver, A::CuSparseMatrixCSC)

Point solver at the arrays of A (≅ cudss_update(solver, A)), see update!.