Resizing multi-dimensional `DiffCache`s

[ranocha]

We use some multi-dimensional caches in Trixi.jl that we have to resize! for some operations. Since Julia only supports to resize! vectors, we construct the plain storage as vectors and unsafe_wrap them in multi-dimensional arrays, e.g.,

# create multi-dimensional cache
_cache = zeros(10)
cache = unsafe_wrap(Array, pointer(_cache), (2, 5))

# resize it
resize!(_cache, 12)
cache = unsafe_wrap(Array, pointer(_cache), (2, 6))

To use the same pattern with DiffCaches, we have to use knowledge about the interval structure of them at the moment, e.g.,

# create multi-dimensional cache
_cache = DiffCache(zeros(10))
s = (2, 5)
cache = DiffCache(unsafe_wrap(Array, pointer(_cache.du), s),
                  unsafe_wrap(Array, pointer(_cache.dual_du), s),
                  Any[])

# resize it
resize!(_cache, 12)
s = (2, 6)
cache = DiffCache(unsafe_wrap(Array, pointer(_cache.du), s),
                  unsafe_wrap(Array, pointer(_cache.dual_du), s),
                  Any[])

• Can we assume that this is safe to do?
• If not, can we add something to the public API of PreallocationTools.jl that allows us to obtain multi-dimensional DiffCaches that can be resize!d like this without allocating a lot of memory every time?

[thomvet]

Probably I am missing something, but why is a reshape based workflow not appropriate here?

using PreallocationTools

_cache = DiffCache(zeros(10))
s = (2, 5)
x = 1.0 #Or ForwardDiff.Dual for the other cache, etc.
cache = reshape(get_tmp(_cache, x), s)

# resize it
resize!(_cache, 12)
s = (2, 6)
cache = reshape(get_tmp(_cache, x), s)

a = cache[2,6] #acts like a multidimensional array

[ranocha]

The most important reason is that we use caches at several places in the code and would have to make more involved changes, adding additional logic where it is not required.

[ChrisRackauckas]

I'm not sure you can even consider that public API for Julia, since `resize!` on a `Matrix` fails with an error, and so you're using an `unsafe` function to get around it. It's all kind of fishy... but we're happy to try to support it as much as we can. I don't think what you're doing there is too prone to breaking. It's definitely not public API as to what the names of the caches are an how many there are, but this package isn't changing all that much so it should be relatively safe. It would be better if there was a public API that handled this... but I can't easily think of a nice one. Internally in PreallocationTools, DiffCache is just using a vector and resizing, so maybe the `resize!` function could just do this, but it would take some exploration. On 9/3/2025 11:24:56 AM, Hendrik Ranocha ***@***.***> wrote: *ranocha* left a comment (SciML/PreallocationTools.jl#140) <#140 (comment)> The most important reason is that we use caches at several places in the code and would have to make more involved changes, adding additional logic where it is not required. — Reply to this email directly, view it on GitHub <#140 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAN25HQ3KZ6RV6JE6CP3FI33Q4W7RAVCNFSM6AAAAACFESSRWWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTENJQGMYDCMZUGA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[ranocha]

Thanks!