On 2009-04-16 09:12:33 +0200, Lars Kyllingstad
<pub...@kyllingen.nospamnet> said:
Robert Jacques wrote:
On Wed, 15 Apr 2009 10:07:38 -0400, Lars Kyllingstad
<pub...@kyllingen.nospamnet> wrote:
I am writing a D library based some of the stuff in SLATEC, and I've
come to a point where I need to decide on a way to manipulate vectors
and matrices. To that end, I have some ideas and questions I would like
comments on from the community.
Ideally, I want to restrict the user as little as possible, so I'm
writing heavily templated code in which one can use both
library-defined vector/matrix types and built-in arrays (both static
and dynamic). My reasons for this are:
a) Different problems may benefit from different types. Sparse
matrices, dense matrices, triangular matrices, etc. can all be
represented differently based on efficiency and/or memory requirements.
yes the problem is that depending on the representation some operation
might be not efficient or maybe even not worth being implemented,
especially for sparse matrixes.
This field is quite difficult, I think Bill idea of having your wrapper
is probably the easiest, second possibility would be the have template
arguments for the operations you use.
You could even have a template parameter that acts just a way to
"select" at compiletime the correct function to call. Doable but I
think that it would need some thinking to get it right.
b) I hope that, at some point, my library will be of such a quality
that it may be useful to others, and in that event I will release it.
Interoperability with other libraries is therefore a goal for me, and a
part of this is to let the user choose other vector/matrix types than
the ones provided by me.
Good, Bill sparse matrixes lib and Don Glade lib apart (that it seem
you already found) maybe you can also take a look at my blip lib
http://dsource.org/projects/blip
It gives dense multidimensional arrays that can call lapack through
nice to use wrappers eigv for eigenvalues, dot for dot product, easy
subslicing, basically never copying unless really needed...
It needs the svn version of tango though...
c) Often, for reasons of both efficiency and simplicity, it is
desirable to use arrays directly.
My first question goes to those among you who do a lot of linear
algebra in D: Do you think supporting both library types and arrays is
worth the trouble? Or should I just go with one and be done with it?
I'd say its worth the trouble.
if for example you need just matrix,vector multiplication I would try
to abstract that away.
If your needs are more complex then if you manage to abstract away
nicely then go for it, but beware of tying to abstract away the things
before beginning, as the problem is common and well known, and there
are different approaches non compatible with each other... Start with
what you use, and try to abstract that away.
A user-defined matrix type would have opIndex(i,j) defined, and to
retrieve elements one would write m[i,j]. However, the syntax for
two-dimensional arrays is m[i][j], and this means I have to put a lot
of static ifs around my code, in order to check the type every time I
access a matrix. This leads me to my second question, which is a
suggestion for a language change, so I expect a lot of resistance. :)
I consider m[i][j] to be a jagged array, which is logically different
from matrix types. (i.e. its not square, etc.)
Would it be problematic to define m[i,j,...] to be equivalent to
m[i][j][...] for built-in arrays, so that arrays and user-defined types
could be used interchangeably?
Actually, I'd prefer actual dense arrays over syntactic sugar for
jagged arrays.
Me too, but that's a bigger language change. At least there ought to be
a matrix type in the standard library.
This is a prime example of something that sparse matrixes might not
define so efficiently, looping over the indexes, maybe row-wise or
column-wise this is more likely to be always efficient, indexing... mhh
not so sure.
So think about your needs and start from there, but yes try to do
algorithms that need just matrix vector multiplication abstract.
(And, importantly, are there anyone but me who think they would benefit
from this?)
I would appreciate it :)
ciao
Fawzi
-Lars
Good numerics and linear algebra is always appreciated. To that end
there's a nice performance speedup in storing machine/byte strides
instead of logical/element strides. (See:
http://dobbscodetalk.com/index.php?option=com_content&task=view&id=502&Itemid=52
Also, my lab maintains a vector/numerics/robotics package that might be
of interest https://trac.lcsr.jhu.edu/cisst)
Thanks for the tip!
-Lars
