Memory leak when calling associate_rhs!() many times

[johnomotani]

I have a program where I factorize a matrix once, and apply it many times to different rhs vectors (they are generated sequentially, in a time-stepping loop). I noticed that the memory usage grows linearly.

I noticed that every time I call associate_rhs!(), another entry is added to the mumps._gc_haven(), but I think the real culprit is that these two lines make copies of the rhs

MUMPS.jl/src/interface.jl

Line 248 in 09313ba

associate_rhs!(mumps::Mumps, rhs::AbstractVector) = associate_rhs!(mumps, repeat(rhs, 1, 1))

MUMPS.jl/src/interface.jl

Line 240 in 09313ba

y = convert(Matrix{T}, rhs)[:]

Is there a reason for this? Am I using the wrong calls for this use case?

Is an easy fix to replace those lines with:

associate_rhs!(mumps::Mumps, rhs::AbstractVector) = associate_rhs!(mumps, reshape(rhs, length(rhs), 1))

and

y = convert(Matrix{T}, rhs)

I think that way at least if you pass a Vector or Matrix, there would be no copy - you'd end up with y pointing to the same memory as the original rhs that the user passed in.

[amontoison]

Is it not possible to replace y = convert(Matrix{T}, rhs)[:] by y = vec(convert(Matrix{T}, rhs))?
We developed the interface a few years ago and didn’t optimize it.
We have not been careful with the allocations.

Can you open a PR?
I can review it before the end of the week.

[johnomotani]

yes, I think I can make a PR - will look into it today