Re: [glas] multi-dimensional arrays?
Date: 2006-12-14 12:23:56
I had to developp my own lib since I could not find any libs dealing with
multi-dim vectors in the sense of Vector< Vector<..> >. Those vectors are use
with corresponding Matrix < Matrix< .. > >. (block matrices).
Such situations arise naturally when the underlying differential equations
evolves in a space that is a tensorial product of sub-spaces. For example, de
Boltzmann equation space is the phase space which can be seen as a tensorial
product of real space and momemtum space. Another example is a Cartesian 2D
space which can be seen as X \otimes Y.
In this cases, there is no problem with natural notations since a matrix vector
product A*X will leads to A(i,j)*X[j] which are again matrix vector products.
I know that the GLAS project is still in a design stage but if some of you are
interested : I made some interesting (at least for me) experiments on the
subject of Expression Template efficiency in the case of large (out of cache)
vectors (mono and multi-dim). In short my ET mecanism is specialized to use
ATLAS in the case of 1D vectors. The result is that I can type X=a*Y+b*Z (in
the multi-dim case) and I obtain results which are better that the
corresponding low level loop coding (by +/- 40 %). I also compared these
results with blitz and uBlas which (only) provides result as good as the low
level loop codes.