On 2010-03-09 13:35:36 +0100, Fawzi Mohamed <[email protected]> said:


On 9-mar-10, at 13:21, Andrei Alexandrescu wrote:

Norbert Nemec wrote:
Multidimensional arrays are the essence of the interfaces of most numerical libraries. Having several incompatible standards is a major roadblock for acceptance in the numerical community.

I'm not so sure about that. I was very enthusiastic about defining an infrastructure of arbitary-dimensional arrays, and did a little research about it. It turns out the applicability of N-dimensional arrays falls off a cliff when N > 3. In turn, this is because high- dimensional space are weird - an N-dimensional space is not just like a 3-dimensional one, only with more dimensions; it's a downright weird beast.

yes and for dense storage, and may operations, the cost is exponential in N, so that one normally doesn't really want to go beyond 3. Still I did need up to 3, and well if you go with dense storage, up to 3 or up to N is more or less the same amount f code if you use templates...

But for non dense storage it gets messy really quick because design choices really influence the api that one can use efficiently.

Fawzi

By the way on the whole I think that D2.0 improves several things in the language, there are some design choices that I don't share so much (shared, increased use of thread local storage,...), but for sure it will make library implementations of multidimensional array better:
* immutability can be put to good use for shape and strides,
* maybe a struct implementation might be usable together with a good memory management (I have tried a couple of times to switch to a structure, but the drawbacks were more that the wins in D1.0, especially as you could not ensure deallocation of resources)
* the latest additions wrt. operator overloading are nice.

Now what is still missing for me is a good compiler for x86_64... :)

Reply via email to