How to tune numerical D? (matrix multiplication is faster in g++ vs gdc)

[Lars T. Kyllingstad]

On Monday, 4 March 2013 at 12:28:25 UTC, J wrote:
> On Monday, 4 March 2013 at 08:02:46 UTC, J wrote:
>> That's a really good point. I wonder if there is a canonical 
>> matrix that would be preferred?
>
> I'm not sure if they are the recommended/best practice for 
> matrix handling in D at the moment (please advise if they are 
> not), but with a little searching, I found that the SciD 
> library has nice matrices and MattrixViews (column-major 
> storage, LAPACK compatible).
>
> Now I like MatrixViews because they let me beat the original 
> (clearly non-optimal) C matrix multiplication by a couple 
> seconds, and the D code with operator overloading in place 
> makes matrix multiplication look elegant.
>
> One shootout benchmark down, eleven to go. :-)
>
> - J
>
> p.s. Am I right in concluding there are no multimethods 
> (multiple dispatch) in D... it seemed a little awkward to have 
> to wrap the MatrixView in a new struct solely in order to 
> overload multiplication. Is there a better way that I've missed?

I'm the author of SciD.  It's great that you found it useful! :)

When I wrote scid.matrix and scid.linalg, it was because I needed 
some linear algebra functionality for a project at work.  I knew 
I wouldn't have the time to write a full-featured linear algebra 
library, so I intentionally gave it a minimalistic design, and 
only included the stuff I needed.  That's also why I used the 
name "MatrixView" rather than "Matrix"; it was only supposed to 
be a convenient way to view an array as a matrix, and not a full 
matrix type.

Later, Cristi Cobzarenco forked my library for Google Summer of 
Code 2011, with David Simcha as his mentor.  They removed almost 
everything but the linear algebra modules, and redesigned those 
from scratch.  So basically, there is pretty much nothing left of 
my code in their library. ;)  I don't think they ever completed 
the project, but I believe parts of it are usable.  You'll find 
it here:

     https://github.com/cristicbz/scid

AFAIK, their goal was to use expression templates in order to 
transform D expressions like

     x*A*B + y*C

where A, B and C are matrices, and x and y are scalars, into as 
few optimised BLAS calls as possible -- e.g., a single GEMM call 
in the example above.  I don't know how far they got on this, 
though.

Lars