KristofferC
commented
9 years ago There is apparently a type instability for sparse when it is called with Int32s.
julia> @code_warntype Base.SparseMatrix.sparse(r,c,v,N,N, Base.AddFun());
# .
# pdest::Any
# .Edit:
pdest starts of as an Int32 but gets promoted to an Int64 at https://github.com/JuliaLang/julia/blob/master/base/sparse/csparse.jl#L89
bdqp
Constructing a large sparse matrix, using Int32 indices is a lot slower. Is this a performance bug or expected behavior? I kind of thought it would be quicker due to more elements being available in cache memory.
julia> N=1000000; NNZ=20*N; julia> r = rand(int32(1:N),NNZ); julia> c = rand(int32(1:N),NNZ); julia> v = rand(Float32,NNZ); julia> tic(); A = sparse(r,c,v,N,N); toc() elapsed time: 11.829277923 seconds 11.829277923
julia> r = rand(int64(1:N),NNZ); julia> c = rand(int64(1:N),NNZ); julia> tic(); B = sparse(r,c,v,N,N); toc() elapsed time: 6.594016314 seconds 6.594016314
julia> whos() A 1000000x1000000 SparseMatrixCSC{Float32,Int32} B 1000000x1000000 SparseMatrixCSC{Float32,Int64}