In-place quadgk about quadgk.jl HOT 7 CLOSED

I think to keep generality, the in-place one uses in-place operations and leaves the other one alone.

Yes, but it would be good to refactor so that they can share as much code as possible. I'd hate to see the entire current implementation copy-pasted and then modified.

from quadgk.jl.

Sounds good to me. I'm not entirely sure of the API for passing f!, though. quadgk!, since presumably you'd also want to pass a pre-allocated result array?

Some tricky refactoring might be required to keep the current generality (since not all types support things like .+=) while allowing in-place operations with preallocated arrays where possible.

One approach would be to internally wrap f! in an InplaceIntegrand functor type and dispatch on that, so that we can pass it through do_quadgk etcetera, and then add various methods foo(f, ...) which use the current behavior for generic f and an optimized in-place behavior for f::InplaceIntegrand.

from quadgk.jl.

I think it would be quadgk! with a pre-allocated result array. I think to keep generality, the in-place one uses in-place operations and leaves the other one alone. quadgk would then be good for scalars and static arrays (so no broadcasting internally), while quadgk! would be more optimized for types which mutate and broadcast. Broadcast is almost a superset of indexable since GPUArrays and things like that are types which broadcast well but don't really index (they now have a fallback to avoid).

I would also like to have options to pass in the work vectors, much like it's done in the differentiation libraries, so I can avoid it allocating those each time. Or some kind of config/plan type that is made once and re-use.

from quadgk.jl.

Has there been any progress on this? Came over this issue when using quadgk in multithreaded loops on Intel KNL cpus, as the allocation made by the integrand severly limits proper scaling.

from quadgk.jl.

No, no one has worked on this yet.

from quadgk.jl.

Note that for integrating small vector integrands, you can use SVector from StaticArrays.jl and it will be fast. See also this discussion on discourse.

from quadgk.jl.

I just pushed a PR to address this issue.

Note that it still cannot operate entirely in-place, because the QuadGK algorithm must dynamically allocate a heap data structure containing integration results on subintervals determined from error estimation.

from quadgk.jl.