Subj : Re: Matrix Operation
To   : borland.public.cpp.borlandcpp
From : Mohsen
Date : Sun Apr 24 2005 01:30 pm


maeder@glue.ch (Thomas Maeder [TeamB]) wrote:
>maeder@glue.ch (Thomas Maeder [TeamB]) writes:
>
>> One obvious alternative is an array with rows*columns elements. This
>> will certainly speed up both allocation and deallocation. I'm not an
>> expert, but the contiguity of all the elements also might have an
>> impact on how the matrix can be cached.
>
>Oh, and you'll also dereference one less pointer per element access.

But this has the drawback that in order to access the elements
of farthest rows great number of multiplication is necessary.
I've wrapped the matrix in a template. I've noticed that the
compiler generates a code such as this one

mov esi, [esi+ebx*4]

for an indexed access to individual elements of an array such 
as this one:

for (unsigned i = 0; i < n; i++)
  a = array[i] * 2;

but this one produces a better code:

double* ptr1 = array;
double* ptr2 = ptr1 + n;

for (; ptr1 < ptr2; ptr1++)
  a = *ptr1 * 2;

This eliminates the multiplication by 4 in previous assembly
code and could have a great influence when the dimensions are
large. But in matrix case it is inevitable to bypass the
problem mentioned above and I thought it could be solved by
changing the way we save the matrix. For example accessing the
elements in each row is OK, but elements in columns still have
the same problem. I don't know whether I could clarify my 
problem.

Mohsen

.