A simple @nogc matrix math library for the D Programming language.
Numd is currently just a matrix math library intended to support EbbCFD. It can perform basic matrix math operations in an intuitive API. It is not particularly fast or efficient but it does get the job done. More work on performance is intended (quite possibly by utilizing mir)
One of the primary goals of Numd's matrix math library was ease of use. If one is familliar with blas and lapack, the matrix routines are very cumbersom to use. On the flip side, a package like Matlab has very intuitive matrix math support. I aimed for the later when designing this library.
I also wanted to use D's type system to its fullest when performing matrix operations. This means emmiting compile errors when operating on incompatible matricies, not allowing certain functions to be called on matricies that cannot support them (e.g. inverse on non-square matricies, dot on non-vectors, and cross on non length 3 vectors). The effect of this is that the matrix struct is a template that takes in its layout at compile time. This does have limitations, but is acceptable in many cases.
Further, I wanted to be able to use this library in a @nogc context.
All matrix operations are @nogc and the matrix object uses a statically
allocated array for its underlying data. For a heap allocated matrix
the aliases RMatrix and RVector provide easy heap matricies
that use ALMOST the same semantics. However, because the
RefCounted wrapper has an opAssign, when you assign one
matrix to another, it only assigns a reference and does not copy the
underlying data.
Some basic usage:
auto m1 = Matrix!(3, 2)(1, 2,
3, 4,
5, 6);
auto m2 = Matrix!(2, 2)(7, 8,
9, 10);
auto m3 = m1 * m2;
writeln(m3);
/+
[ 25 28 ]
[ 57 64 ]
[ 89 100 ]
+/Originally Numd was to be a complete numerical methods library. This original vision has not been lost but in order to reduce dependancies that made it somewhat difficult to get running on a few clusters, it has been stripped down to only include the matrix math library for now.
The plan going forward is to move things from EbbCFD into Numd when they reach a stable level with an interface that I am happy with. The first things that will likely be brought over are the time integrators that EbbCFD uses. They currently have a specific interface to EbbCFD but this should be changing soon.