Re: Guile numerical work and uniform arrays

[Valentin Kamyshenko <val@kamysh.materials.kiev.ua>]

>>>>> "Per" == Per Bothner <bothner@cygnus.com> writes:


>> My suggestion was, thus: 1) to give user the possibility to know
>> that a given multi-dimensional array can be treated as
>> one-dimensional; 2) to have a standard procedure to get a pointer
>> to this chunk.

Per> I cannot see how this can be mathematically meaningful, in
Per> general, nor do I see any use for it (except perhaps some
Per> optionization).

Probably, I was not clear enough.

If I have non-shared multi-dim array, it is, in agreement with your
consideration, a one-dimensional chunk of data, with additional
information about dimensions. This corresponds to the treatment of
multi-dimensional arrays in FORTRAN or C (if you define them as
A[n][m]), with the only difference that A[n][m] in C corresponds to
A(m,n) in FORTRAN. Thus, having such chunk, you can put it in the
argument of C or FORTRAN routine.

If you have shared array, you must, in general, make a new
one-dimensional chunk, and copy there values from the shared array,
before using it in arguments of FORTRAN/C procedures (although many
matrix libraries provide additional arguments in their procedures to
allow working with partially filled arrays). 

Thus, I would like to have possibility to distinguish between these
two situations, and, _if_it_is_possible_, to use the underlying
one-dimensional chunk of multi-dim. array directly, without copying
its data to new location (and back).

Few words, why I think it is important for numerical applications.

For me, using guile for such applications means, first of all, to
provide effective interplay with routines, written in C and
FORTRAN. I, personally, use BLAS and LAPACK for matrix
calculations. Let us suppose, I want to use the optimized BLAS routine
for multiplication of matrix A[n][n] and vector V[n] for 10000
different vectors V, but the same matrix A. Do you think, it is worth
to allocate temporary storage for copy of A, and to copy there
elements of A (like gh_scm2doubles does) each time I call this
subroutine (just to generate correct argument list) in the case of
n=10000 ? I want to emphasize that, for my work, this is a _typical_
example.

--
Best regards,
	V.Kamyshenko.