Sparse Matrix with CUDA.jl

[boutor2]

Thank you very much for your comprehensive answer. However, it does not seem trivial to convert CartesianIndex in a CSC format. But may be I am wrong?

Here is a proposal to do so, that uses the sparse method from SparseArrays

Cartesian indices, linear indices, and sparse matrices

Thanks for writing this up as a comprehensive question! It’d probably be nice to add a sparse method that takes arrays of CartesianIndexes. Here’s a simple implementation that could serve you in the mean-time: function SparseArrays.sparse(IJ::Vector{<:CartesianIndex}, v, m, n) IJ′ = reinterpret(Int, reshape(IJ, 1, :)) return sparse(view(IJ′, 1, :), view(IJ′, 2, :), v, m, n) end In the above I used reinterpret and lazy views to prevent needlessly allocating new vectors that contai…

but as you mentioned in an other post we miss the sparse method for CuSparseMatricCSC

CuSparseMatrixCSC constructor

This only works for sparse from SparseArrays: julia> using SparseArrays julia> sparse(Array(I), Array(J), Array(V), 3, 3) 3×3 SparseMatrixCSC{Float64,Int64} with 2 stored entries: [1, 1] = 0.2 [1, 2] = 0.3 You need to convert your coordinate lists (I and J) to CSC format (colPtr and rowVal). julia> colPtr = CuArray([1, 2, 3, 3]); julia> rowVal = CuArray([1, 1]); julia> CUSPARSE.CuSparseMatrixCSC(colPtr, rowVal, V, (3,3)) 3×3 CuSparseMatrixCSC{Float64} with 2 stored entries: [1, 1…