Why does my devectorized code keep allocating memory?

[ChrisRackauckas]

@WouterJRV

du.=1im*(1 .-1im*η).*(NBmat*u.-4*u)/gridsize^2 +(.-(1im).*V+α.-(σ.+1im*g).*abs2.(u)).*u;

Why not just @. du = ...? You’re not fusing the whole operation: you forgot a few dots so it’s allocating a few steps. That RHS function is not written in a very fast way, so that will impact the speed you see.

Edit: I see, matmul. Everything else should be dotted those, and it’s not.

@WouterJRV

Evolving it with Vern7() is maybe 20% faster than MATLAB’s ode113 result and takes about 600 sec on my desktop to evolve t from 0 until 100.

Is that the right solver? Sounds like more of a problem for VCABM, QNDF, CVODE_BDF, LSODA, ROCK2, or ROCK4. Did you do a comparative study of methods?

@WouterJRV

In a very analogous way to this example, I also tried to rewrite my PDE in a devectorized way. Note that I’ve modified the original complex vector U of length N into a real Nx times Nx times 2 array, where the real and imaginary parts are separated. This gives in full

Not sharing code makes it impossible to really debug, but from experience, I would guess that your function is not type-stable because you made p3 a Vector when it should be a Tuple since it’s filled with different typed objects. Did you look at @code_warntype to see?