Non inlining of function arguments causing slow creation of sparse matrices from IJV format

[KristofferC]

The default functions being passed into the sparse matrix constructors from IJV vectors, for example:

sparse(I,J,V::AbstractVector,m,n) = sparse(I, J, V, Int(m), Int(n), +)

are not inlined, which means that significant time is spent to do type inference when the combine function is called many times.

In my code, for a medium sized FEM problem, the difference between inlining the plus or not is about 4x:

julia> @time sparse_inline(dof_rows, dof_cols, k_values, fp.n_eqs, fp.n_eqs)
elapsed time: 0.096662539 seconds (61 MB allocated, 3.20% gc time in 2 pauses with 0 full sweep)
80400x80400 sparse matrix with 1119192 Float64 entries:
.
.

julia> @time Base.sparse(dof_rows, dof_cols, k_values, fp.n_eqs, fp.n_eqs)
elapsed time: 0.393187843 seconds (257 MB allocated, 27.98% gc time in 11 pauses with 1 full sweep)
80400x80400 sparse matrix with 1119192 Float64 entries:
.
.

Since plus is the most common, maybe a special version should be made for that.

I realize that inlining of function arguments is coming so then this will be fixed by itself, however, if this it far away, then it might be worth looking at this.

[JeffBezanson]

Yes, the default should probably be special-cased. My sense is that wanting a function other than + is quite rare; is that right?

[KristofferC]

I can't answer how common it is to use other functions than + but for assembling the stiffness matrix in finite elements you want to use + so it would be nice to have a fast version for that.

[simonster]

Maybe just replace + with AddFun() (and modify the method signatures)?

[andreasnoack]

Or make + a type and overload call(::Type{+},x,y).

[mauro3]

Matlab only implements + for (fast) accumulation, so presumably that is what is used the most by far. Certainly the case for me.

[mlubin]

+ is certainly the most common. I've also used funky combine functions like vcat, but in that case there's a diminished expectation of performance.

[ViralBShah]

+ is the certainly most common. So far the performance was ok and nobody complained, and I didn't want to have a totally different implementation just for +. The functor approach is perfect.

I often use other combine functions when using sparse matrices as graph data structures and doing graph algorithms. Often, just overwriting rather than combining is useful to have. Matlab created accumarray as a generalization, but it doesn't work for sparse, even though it is incredibly useful.

[jiahao]

I think sparse accumarray exists now.

[ViralBShah]

Perhaps it was not there 5 years back (or maybe it was), when I used matlab a lot. We've come a long way. :-)

[ViralBShah]

@KristofferC Can you try and see how the latest master performs now? I merged @mauro3 's PR.