I have a code which does a lot of work on 2D/3D arrays, for which I use the 2.066 multidimensional slicing syntax through a fork of the Unstandard package [1].
Many times the order of operations doesn't matter and I thought I would give the parallelism module a try to try and get some easy speedups (I also use MPI, but that has some additional overhead).

The way I currently have my foreach loops set up, p is a size_t[2], the payload of the array v is double[9] and the array is indexed directly with a size_t[2] array and all works fine:

foreach(immutable p, ref v, arr) { double[9] stuff; arr[p] = stuff; }

If I naively try

foreach(immutable p, ref v, parallel(arr)) { ... }

I first get errors of the type "Error: foreach: cannot make v ref". I do not understand where that particular problem comes from, but I can possibly live without the ref, so I went for

foreach(immutable p, v, parallel(arr)) { ... }

Which gets me "Error: no [] operator overload for type (complicated templated type of some wrapper struct I have for arr)". I'm guessing it doesn't like that there is no such thing as a simple one-dimensional slicing operation for a multidimensional array?
Should I define an opSlice function that takes the usual two size_t arguments for the upper and lower bounds and doesn't require a dimension template argument and somehow map this to my underlying two-dimensional array? Will it need an opIndex function that takes only takes a single size_t as well?
Or is this just taking the simple parallel(...) too far and should I try to put something together myself using lower-level constructs?

Any hints would be appreciated!

[1] http://code.dlang.org/packages/unstandard

On Friday, 22 May 2015 at 10:54:36 UTC, Stefan Frijters wrote:
> I have a code which does a lot of work on 2D/3D arrays, for which I use the 2.066 multidimensional slicing syntax through a fork of the Unstandard package [1].
> Many times the order of operations doesn't matter and I thought I would give the parallelism module a try to try and get some easy speedups (I also use MPI, but that has some additional overhead).
>
> The way I currently have my foreach loops set up, p is a size_t[2], the payload of the array v is double[9] and the array is indexed directly with a size_t[2] array and all works fine:
>
> foreach(immutable p, ref v, arr) { double[9] stuff; arr[p] = stuff; }
>
> If I naively try
>
> foreach(immutable p, ref v, parallel(arr)) { ... }
>
> I first get errors of the type "Error: foreach: cannot make v ref". I do not understand where that particular problem comes from, but I can possibly live without the ref, so I went for
>
> foreach(immutable p, v, parallel(arr)) { ... }
>
> Which gets me "Error: no [] operator overload for type (complicated templated type of some wrapper struct I have for arr)". I'm guessing it doesn't like that there is no such thing as a simple one-dimensional slicing operation for a multidimensional array?
> Should I define an opSlice function that takes the usual two size_t arguments for the upper and lower bounds and doesn't require a dimension template argument and somehow map this to my underlying two-dimensional array? Will it need an opIndex function that takes only takes a single size_t as well?
> Or is this just taking the simple parallel(...) too far and should I try to put something together myself using lower-level constructs?
>
> Any hints would be appreciated!
>
> [1] http://code.dlang.org/packages/unstandard

I'd define a "flatten range" adaptor that presents n-dimensional ranges as 1d range that traverses the original array indices lexicographically, if that makes sense for your app.