SciPy has the scipy.sparse.diags format, that stores the sparse matrix by its nonzero diagonals. In this way we collect a list of diagonals (vectors that might have different lengths) and their corresponding offsets to the center.
I think this is the same format used in packages for quantum like QuTiP or dynamiqs. Something similar seems to be implemented in SparseBandedMatrices.jl, but only for CPU arrays, and I don’t know if it is the same as the other cases.
I don’t know what is the best name for this format. GenericSparseBandedMatrix or GenericSparseDiagMatrix, or others?
It would be good to implement such format, introducing all the methods defined for the other formats, like:
mul!(y::DenseVecOrMat, A:: GenericSparseDiagMatrix, x::DenseVecOrMat)tr(A:: GenericSparseDiagMatrix)+(A::AbstractGenericSparseMatrix, B:: AbstractGenericSparseMatrix)*(A:: AbstractGenericSparseMatrix, B:: AbstractGenericSparseMatrix)+(A:: AbstractGenericSparseMatrix, B::DenseMatrix)*(A:: AbstractGenericSparseMatrix, B::DenseMatrix)kron(A:: GenericSparseDiagMatrix, B:: GenericSparseDiagMatrix)dot(x::DenseVector, A:: GenericSparseDiagMatrix, y::DenseVector)- Conversion to the other formats
- Other operations not listed because but that are already implemented for the other formats.
Of course including also the combinations for Transpose and Adjoint.
I don’t know what is the best method to store the list of diagonals. Should we store them in a simple vector? But then do the custo GPU kernels support it? Should we define them in a Tuple? But then this needs to be known at compile time. I don’t know if there are other options.
Everything should be defined without converting to CPU, as for the other formats.
SciPy has the
scipy.sparse.diagsformat, that stores the sparse matrix by its nonzero diagonals. In this way we collect a list of diagonals (vectors that might have different lengths) and their corresponding offsets to the center.I think this is the same format used in packages for quantum like QuTiP or dynamiqs. Something similar seems to be implemented in SparseBandedMatrices.jl, but only for CPU arrays, and I don’t know if it is the same as the other cases.
I don’t know what is the best name for this format.
GenericSparseBandedMatrixorGenericSparseDiagMatrix, or others?It would be good to implement such format, introducing all the methods defined for the other formats, like:
mul!(y::DenseVecOrMat, A:: GenericSparseDiagMatrix, x::DenseVecOrMat)tr(A:: GenericSparseDiagMatrix)+(A::AbstractGenericSparseMatrix, B:: AbstractGenericSparseMatrix)*(A:: AbstractGenericSparseMatrix, B:: AbstractGenericSparseMatrix)+(A:: AbstractGenericSparseMatrix, B::DenseMatrix)*(A:: AbstractGenericSparseMatrix, B::DenseMatrix)kron(A:: GenericSparseDiagMatrix, B:: GenericSparseDiagMatrix)dot(x::DenseVector, A:: GenericSparseDiagMatrix, y::DenseVector)Of course including also the combinations for
TransposeandAdjoint.I don’t know what is the best method to store the list of diagonals. Should we store them in a simple vector? But then do the custo GPU kernels support it? Should we define them in a Tuple? But then this needs to be known at compile time. I don’t know if there are other options.
Everything should be defined without converting to CPU, as for the other formats.