On Monday, 27 April 2015 21:05:33 UTC+2, François Fayard wrote: > If one implements y = y + delta_t * dy_dt in the code with a Vector, you > create a lot of heap arrays that has a huge cost (It seems that ODE.jl is > doing that by the way). So you are forced to loop over the vector elements, > which does not look really good. How about implementing a > add_linear_combination!(y, delta_t, dy_dt) which would make things easier to > extend for any kind of "Vector". This is exactly what people from odeint have > been doing and it looks like a nice idea. If you want to go crazy you can > event implement a MPI vector whose memory is scattered around different > nodes. Then you just implement the add_linear_combination for your array, and > you are done.
The problem is that `y = y + delta_t * dy_dt` currently makes two copies, as you have observed. There is an open discussion about in-place assignment operators [1], which would help. There is also Devectorize.jl [2]. However, I agree that having something like add!(y, x, beta) would be nice and I think this also came up in some discussion. Funnily enough we just recently discussed this strategy for ODE.jl [3] ... Best, Alex. [1] https://github.com/JuliaLang/julia/issues/249 [2] https://github.com/lindahua/Devectorize.jl [3] https://github.com/JuliaLang/ODE.jl/pull/68#issuecomment-95674816
