We have a (time critical function) with the following prototype:
heapinsert!{N}(xs::Array{Tuple{NTuple{N, UInt}, heap_s}, 1}, exp::NTuple{N,
UInt}, x::heap_s)
However, one line of code in the implementation is allocating gigabytes of
memory, which is causing a massive performance deficit in our program:
xs[1] = (exp, x)
Note heap_s is just an ordinary Julia type here, which contains a couple of
Ints and another heap_s. As far as I can see it is irrelevant. It should
just be a pointer at the machine level.
It seems that Julia allocates space for the tuple (exp, x) and then writes
it into the array xs.
Does anyone have any ideas how I can stop it from making this unnecessary
allocation? It's slowing down multivariate polynomial multiplication by a
factor of at least 2 (and possibly 10) and causing a massive blowout in
memory allocation (this is really performance critical code).
I've already tried making using an immutable rather than a tuple. Same
problem. I'm tracing this using julia --track-allocation=all.
Bill.
