sum(sparse) -> dense? about sparsearrays.jl

[Sacha0]

👍 for @abraunst's comments; optimal methods for hypersparse arrays often differ from those for sparse arrays. As an alternative to the keyword idea, we could extend sum! to accept a destination array, and dispatch on the destination type. Best!

[abraunst]

Just to simplify the example, your slice looks very similar to s=sprand(10^7,150,10^-7,i->rand(1:10,i)). It is a bit extreme, with only ~ 150 nonzeros in 1.5*10^9 entries. The fastest way to sum I found in this case was:

function sum2(x::SparseMatrixCSC)
    o = spzeros(eltype(x), size(x,1))
    @inbounds for i=1:nnz(x)
        o[x.rowval[i]] += x.nzval[i]
    end
    o
end

I think that the problem with this kind of approaches is that it could fail badly when the density is a bit higher. For example, replacing the second 10^-7 with 10^-4 or so, which still gives a really sparse matrix (most entries in the output vector are still 0), then sum(s, dims=2) is about 40 times faster. The difference becomes much larger for higher densities of course. I think it's really hard to optimize for all possible scenarios.

[StefanKarpinski]

In general, the assumption for a SparseMatrix{R,C}SC, S, is that the number of non-zeros is roughly O(size(S,1) + size(S,2)). Since the result size of sum(S, dims=1) and sum(S, dims=2) are both of that size, it seems like the reasonable choice is to return a dense vector in both cases. For CSC the storage it makes no sense to return a sparse vector when summing along the columns, for example. For summing along the rows, the result might happen to be sparse if all the row values happen to be in a small subset of rows, but that's not something that should generally be expected, and there's no real reason to expect that any more than the result of summing the rows of a dense matrix might happen to have lots of zero values. In general, if size(S,1) ≈ size(S,2) then we would expect both row and column sums to be dense, so producing a dense vector for these operations seems sane to me.

[ViralBShah]

Yes, this is good to do - which is make the sum of SparseMatrixCSC along either dimension a SparseVector.

[ViralBShah]

I think the right thing to do is to always have a sparse output, and the user can explicitly convert it to dense if required.

[abraunst]

Note that

• the output 1xN SparseMatrixCSC would not be really sparse (i.e. requires still O(N) storage even if it's completely zero)
• empty columns in the input MxN SparseMatrixCSC still take space (because of the colptr vector) and need to be iterated over.

so summing in the dims=1 dimension would probably be slower with a sparse output. The dims=2 could maybe benefit some, but not sure how this "really really sparse" case is relevant. Note also that both A*ones(N) and ones(1,N)*A give dense arrays (i.e. alternative ways of achieving the same results).

[yurivish]

The dims=2 could maybe benefit some, but not sure how this "really really sparse" case is relevant.

If it helps to illustrate my concrete use case with some size numbers, I worked with a square matrix of 11 million x 11 million census blocks, and the sum was returning a dense vector of length 11 million x 100 or so.

I used findnz to find the nonzero elements in my columns of interest and then did a group sum by the row index, which was faster than using the built-in sum.

[abraunst]

The dims=2 could maybe benefit some, but not sure how this "really really sparse" case is relevant.

If it helps to illustrate my concrete use case with some size numbers, I worked with a square matrix of 11 million x 11 million census blocks, and the sum was returning a dense vector of length 11 million x 100 or so.

I used findnz to find the nonzero elements in my columns of interest and then did a group sum by the row index, which was faster than using the built-in sum.

Could you clarify the example? What do you mean by "columns of interest"? How many nonzeros in the output row column vector? What were the respective times (i.e. how much faster)?

[yurivish]

Sorry, here's a quick example where a hacky sparse sum using Dicts outperforms the naive sum. Perhaps I'm missing something and this isn't a rigorous benchmark by any means, but I think it replicates the essence of the situation I found myself in where I was slicing a big square matrix and summing some of the rows or columns together.

using SparseArrays, SplitApplyCombine
function test()
    s = let
        I = []; J = []; V = Int[]
        sz = 10_000_000
        for i in 1:sz
            j = rand(1:sz)
            v = rand(1:10)
            push!(I, i)
            push!(J, j)
            push!(V, v)
        end
        sparse(I, J, V)
    end;
    # my actual use had noncontiguous indexes in the second dimension
    slice = s[:, 100_000:100_150]; 
    @show typeof(slice)
    @time a = sum(slice, dims=2);
    @time b = let
        nt = ((i=i, j=j, v=v) for (i, j, v) in zip(findnz(slice)...))
        groupsum(x -> x.i, x -> x.v, nt)
    end
    a, b
end
a, b = test();

Timings on my laptop (2012 Macbook Pro) after compilation are

  0.465381 seconds (10 allocations: 152.588 MiB, 82.09% gc time)
  0.028346 seconds (29.54 k allocations: 1.358 MiB)

[SobhanMP]

@Wimmerer do you think this would be a good feature to add?

[rayegun]

I imagine this SparseVector will very often be dense but that's the only correct route.