Sparse matrix - vector multiplication with threads.

thread_sparse.jl

# Gives roughly a 2x speedup using JULIA_NUM_THREADS=2
# requires 0.7, due to closure type inference bug in 0.6

using Random, SparseArrays

srand(1); X = sprand(20_000,100_000,0.1); v = ones(100_000);

function mul(X,v)
    m,n = size(X)
    U = zeros(size(X,1), Threads.nthreads())
    @assert n == length(v)
    Threads.@threads for col = 1:n
        thread = Threads.threadid()
        @inbounds a = v[col]
        @inbounds for i = X.colptr[col]:(X.colptr[col + 1] - 1)
            row = X.rowval[i]
            @inbounds U[row, thread] = muladd(X.nzval[i], a, U[row, thread])
        end
    end
    sum(U,2)
end

@time X*v;
@time X*v;
@time X*v;

@time mul(X,v);
@time mul(X,v);
@time mul(X,v);