On Saturday, 25 February 2017 at 11:06:28 UTC, data pulverizer
wrote:
I have noticed that some numerical packages written in D use
pointer semantics heavily (not referring to packages that link
to C libraries). I am in the process of writing code for a
numerical computing library and would like to know whether
there times when addressing an array using pointers conveys
performance benefits over using D's array or functional
semantics?
Pointers can behave like iterators, while to have an iterator on
top of array you need to store array and index. This is slower
and may disable vectorisation. Iterators semantic is required to
build multidimensional random access ranges (ndslice). ndslice
uses iterators heavily in its internals. Iterators semantic
helped to create many multidimensional Phobos analogs in few
dozens LOC, while Phobos has few hundreds LOC for the same
functionality.
Phobos functional semantics disable vectorisation and other
optimisations. mir.ndslice.topology [1], mir.ndslice.algorithm[1]
and mir.functional[1] in combination with other Mir modules and
packages can be used instead of Phobos functional utilities from
std.algorithm, std.range, std.functional.
You may want to use the new ndslice [1]. Slice!(Contiguous, [1],
T*) can replace T[].
[1] https://github.com/libmir/mir-algorithm
