matlisp-users

[Matlisp-users] help with gross array hack needed

[rif <ri...@MI...>]

Is there any way, in CMUCL, using some implementation-specific
construct, to create an "aliased array"?  Suppose I create a
double-float array of size 10:

(make-array 10 :element-type 'double-float)

I know I can take it's address, with #'system:vector-sap.

What I want to do is manipulate that address and somehow "invert"
vector-sap to create a new double-float array whose first element is
(for example) the third element of the original array.

I need this because I need to pass matlisp a subset of an array.  More
specifically, I want to allocate an n by n+1 array, so that I can
store half of each of two n by n symmetric arrays in n(n+1) space
rather than 2n^2 space.  This is a common technique for using
BLAS/LAPACK, but to get at it, I need to be able to give a Fortran
matlisp routine a way to get at the square n by n matrix that "starts
with the second column" of the n by n+1 matrix, and the way the
interface is defined, it wants actual arrays, not pointers.  If I can
somehow create the appropriate array, I don't have to go mucking about
in the internals of matlisp's def-fortran-routine and so forth.

(Note that I only need this "aliased array" temporarily, and I'd of
course turn gc off while I needed it.)

Any suggestions or advice are appreciated.  If you respond, please cc
to rif directly --- I think I'm on both these lists, but I'm not 100%
sure.

rif

[Matlisp-users] Re: help with gross array hack needed

[Marcin T. <mm...@ze...>]

http://www.lisp.org/HyperSpec/Body/glo_d.html#displaced_array

Re: [Matlisp-users] help with gross array hack needed

[Raymond T. <ray...@er...>]

>>>>> "rif" == rif  <ri...@MI...> writes:

    rif> Is there any way, in CMUCL, using some implementation-specific
    rif> construct, to create an "aliased array"?  Suppose I create a
    rif> double-float array of size 10:

    rif> (make-array 10 :element-type 'double-float)

    rif> I know I can take it's address, with #'system:vector-sap.

    rif> What I want to do is manipulate that address and somehow "invert"
    rif> vector-sap to create a new double-float array whose first element is
    rif> (for example) the third element of the original array.

    rif> I need this because I need to pass matlisp a subset of an array.  More
    rif> specifically, I want to allocate an n by n+1 array, so that I can
    rif> store half of each of two n by n symmetric arrays in n(n+1) space
    rif> rather than 2n^2 space.  This is a common technique for using
    rif> BLAS/LAPACK, but to get at it, I need to be able to give a Fortran
    rif> matlisp routine a way to get at the square n by n matrix that "starts
    rif> with the second column" of the n by n+1 matrix, and the way the
    rif> interface is defined, it wants actual arrays, not pointers.  If I can
    rif> somehow create the appropriate array, I don't have to go mucking about
    rif> in the internals of matlisp's def-fortran-routine and so forth.

Marc has already mentioned displaced arrays.  This should work, but
depends on what you are really trying to do.  What matlisp function
are you trying to call?

Ray

Re: [Matlisp-users] help with gross array hack needed

[Marco A. <ma...@cs...>]

On Dec 8, 2004, at 7:43 AM, Raymond Toy wrote:

>>>>>> "rif" == rif  <ri...@MI...> writes:
>
>     rif> Is there any way, in CMUCL, using some implementation-specific
>     rif> construct, to create an "aliased array"?  Suppose I create a
>     rif> double-float array of size 10:
>
>     rif> (make-array 10 :element-type 'double-float)
>
>     rif> I know I can take it's address, with #'system:vector-sap.
>
>     rif> What I want to do is manipulate that address and somehow 
> "invert"
>     rif> vector-sap to create a new double-float array whose first 
> element is
>     rif> (for example) the third element of the original array.
>
>     rif> I need this because I need to pass matlisp a subset of an 
> array.  More
>     rif> specifically, I want to allocate an n by n+1 array, so that I 
> can
>     rif> store half of each of two n by n symmetric arrays in n(n+1) 
> space
>     rif> rather than 2n^2 space.  This is a common technique for using
>     rif> BLAS/LAPACK, but to get at it, I need to be able to give a 
> Fortran
>     rif> matlisp routine a way to get at the square n by n matrix that 
> "starts
>     rif> with the second column" of the n by n+1 matrix, and the way 
> the
>     rif> interface is defined, it wants actual arrays, not pointers.  
> If I can
>     rif> somehow create the appropriate array, I don't have to go 
> mucking about
>     rif> in the internals of matlisp's def-fortran-routine and so 
> forth.
>
> Marc has already mentioned displaced arrays.  This should work, but
> depends on what you are really trying to do.

Not necessarily.  Displaced arrays are good for row-major slices.  So 
if the matrix where (N + 1 x N), then you could "displace away" the 
first row.  Unfortunately, FORTRAN is column-major.  The only way I see 
around this is to transpose the array first, and then displace.  
However, that may cause problems for the algorithm at hand.

Isn't this fun? :)

Cheers



--
Marco Antoniotti					http://bioinformatics.nyu.edu
NYU Courant Bioinformatics Group		tel. +1 - 212 - 998 3488
715 Broadway 10th FL				fax. +1 - 212 - 998 3484
New York, NY, 10003, U.S.A.