Hello everyone, I need some help on optimization of a function, which looks like this # re-edit SpMatrixListUnion = Union{Vector{SparseMatrixCSC{Float64,Int64}},Vector{SparseMatrixCSC{Complex128,Int64}}} function avg_V_on_ψ!( res::Array{Float64,2}, # result to w…

Your code doesn’t run here. I’m assuming that SpMatrixUnion should be SpMatrixListUnion, but then I still get julia> avg_V_on_ψ!(res2,vlist,oplist) ERROR: MethodError: no method matching avg_V_on_ψ!(::Array{Float64,2}, ::Array{Array{Float64,2},1}, ::Array{SparseMatrixCSC{Float64,Int64},1}) Closest …

Using views give a meagre 20% improvement: function avg_V_on_ψ!2( res::Array{Float64,2}, vec_list::Vector{Matrix{Float64}}, OP_list ) for i=1:length(vec_list) I,J,V = findnz(OP_list[i]) v = vec_list[i] vI = view(v, I, :) vJ = view(v, J, :) …

Sorry I’ve got a paranoid way of managing types. I have replaced the SpMatrixUnion by SpMatrixListUnion. This time it should work …

Excellent ! Thanks a lot. It seems that the broadcast function prefers views of matrices, instead of matrices themselves.

I think the performance gain is just by avoiding having to copy the sub-matrices.

BTW kristofer.carlsson just removed the type after OP_list, that makes the code more transparent …

I just realized that real.(sum(vI .* broadcast(*, V, vJ), 1)) is equivalent to res[:,i] = real.(RowVector(V)*(vI.*vJ)). The latter leads to 20% gain

Or V' * (vI .* vJ) (a bit more compact)

[image] kristoffer.carlsson: for i=1:length(vec_list) I,J,V = findnz(OP_list[i]) v = vec_list[i] vI = view(v, I, :slight_smile: vJ = view(v, J, :slight_smile: res[:,i] = real.(sum(vI .* broadcast(*, V, vJ), 1)) end yes, only if the O in op_list has all real entries

Help on optimization

General Usage

tkoolen April 9, 2018, 12:26am 2

Would you mind quoting your code to make it more readable?

Topic		Replies	Views
Matrix computation Performance	6	838	July 27, 2020
Speeding up elementwise Vector-SparseMatrixCSC multiplication broadcasting Performance sparse	11	667	December 12, 2024
A lot of big sparse matrices Performance	8	554	December 12, 2020
Matrix-vector multiplication slower than a 'naive' for loop? Performance vector	7	1777	July 30, 2020
Weighted sum of sparse matrices New to Julia	3	1646	August 27, 2017

Help on optimization

Related topics