Re: [Help-gsl] sparse matrices (again)

Re: [Help-gsl] sparse matrices (again)

[Brian Gough]

At Wed, 25 Apr 2007 00:30:25 +0400,
Andrew Makhorin wrote:
> I am wondering if it is planned to eventually provide GSL with general
> sparse matrix routines (e.g. like those ones which are included in NAG,
> ESSL, or Harwell Library)?

Thanks for your email.  

So far we haven't added sparse matrices because it seemed like
this would require several sparse matrix representations to be
implemented in order to be useful, each with their own functions.

Due to the potentially large amount of work involved, it seemed
like this was best left to specialised sparse matrix packages.
As I understand it the number of sparse functions in the NAG
library is pretty big.

If you have any suggestions in this area I'd be glad to hear
them.

-- 
Brian Gough

Re: [Help-gsl] sparse matrices (again)

[Andrew Makhorin]

>> I am wondering if it is planned to eventually provide GSL with general
>> sparse matrix routines (e.g. like those ones which are included in NAG,
>> ESSL, or Harwell Library)?

> So far we haven't added sparse matrices because it seemed like
> this would require several sparse matrix representations to be
> implemented in order to be useful, each with their own functions.

> Due to the potentially large amount of work involved, it seemed
> like this was best left to specialised sparse matrix packages.
> As I understand it the number of sparse functions in the NAG
> library is pretty big.

> If you have any suggestions in this area I'd be glad to hear
> them.

As it seems to me, it would be very useful to provide GSL at least
with data structures defining three most commonly used formats for
general sparse matrices, namely, coordinate format, compressed row
storage, and compressed column storage, as well as two formats for
sparse vectors based on compressed vector storage, and a minimal set
of some basic functions. This would standardize representation of
sparse objects within the GSL library and make it possible to use
such data structures on developing sparse linear algebra software
based on GSL.

Thank you,

Andrew Makhorin

Re: [Help-gsl] sparse matrices (again)

[James Bergstra]

On Sat, May 05, 2007 at 04:05:08PM +0400, Andrew Makhorin wrote:
> >> I am wondering if it is planned to eventually provide GSL with general
> >> sparse matrix routines (e.g. like those ones which are included in NAG,
> >> ESSL, or Harwell Library)?
> 
> > So far we haven't added sparse matrices because it seemed like
> > this would require several sparse matrix representations to be
> > implemented in order to be useful, each with their own functions.
[snip]
> 
> As it seems to me, it would be very useful to provide GSL at least
> with data structures defining three most commonly used formats for
> general sparse matrices, namely, coordinate format, compressed row
> storage, and compressed column storage, as well as two formats for
> sparse vectors based on compressed vector storage, and a minimal set
> of some basic functions. This would standardize representation of
> sparse objects within the GSL library and make it possible to use
> such data structures on developing sparse linear algebra software
> based on GSL.
> 
> Thank you,
> 
> Andrew Makhorin

Something like this exists as the NIST sparse BLAS:
http://math.nist.gov/spblas/

Perhaps the GSL could 
- include this code (verbatim would make 2800 lines of C), and compile it as
  libgslspblas.so, to complement libgslcblas.so (I have no idea if licensing
  would get in the way.)
- wrap it in a manner similar to gsl_matrix, to provide 
  - element access
  - simple stats (min, max, mean)
  - basic I/O (e.g.  Harwell-Boeing)
  - vector scatter / gather
  - matrix-vector multiply, (matrix-matrix multiply?)

Would these nuts and bolts make progress toward providing sparse solvers though
the GSL?  Would they be useful to many?

Cheers,

James Bergstra
-- 
http://www-etud.iro.umontreal.ca/~bergstrj

Re: [Help-gsl] sparse matrices (again)

[Brian Gough]

At Sat, 5 May 2007 16:05:08 +0400,
Andrew Makhorin wrote:
> As it seems to me, it would be very useful to provide GSL at least
> with data structures defining three most commonly used formats for
> general sparse matrices, namely, coordinate format, compressed row
> storage, and compressed column storage, as well as two formats for
> sparse vectors based on compressed vector storage, and a minimal set
> of some basic functions. This would standardize representation of
> sparse objects within the GSL library and make it possible to use
> such data structures on developing sparse linear algebra software
> based on GSL.

Thanks for the suggestions.  Sorry for the delay in replying.

It might be feasible to add sparse matrix and vector datatypes and a
sparse blas, without adding any specific solver algorithms (which is
what I'd like to avoid).

If this was added to GSL, would you want to use it in GLPK, or is it
just something you thought would be generally useful to users of GSL?

-- 
Brian Gough

Re: [Help-gsl] sparse matrices (again)

[Brian Gough]

At Sun, 6 May 2007 19:18:09 -0400,
James Bergstra wrote:
> Something like this exists as the NIST sparse BLAS:
> http://math.nist.gov/spblas/
> 
> Perhaps the GSL could 
> - include this code (verbatim would make 2800 lines of C), and compile it as
>   libgslspblas.so, to complement libgslcblas.so (I have no idea if licensing
>   would get in the way.)
> - wrap it in a manner similar to gsl_matrix, to provide 
>   - element access
>   - simple stats (min, max, mean)
>   - basic I/O (e.g.  Harwell-Boeing)
>   - vector scatter / gather
>   - matrix-vector multiply, (matrix-matrix multiply?)
> 
> Would these nuts and bolts make progress toward providing sparse solvers though
> the GSL?  Would they be useful to many?

I think the sparse blas is something we could reasonably implement.
It it not too complicated, and it is a standard.  I don't know how
useful it would be in practice, in terms of number of users.

-- 
Brian Gough

RE: [Help-gsl] sparse matrices (again)

[Konkle, Daniel]

I apologize for this email but I've tried the web interface several
times now unsuccessfully.  I have been trying to discontinue receiving
the emails from the list server.  I ask it to unsubscribe or send me my
password to login but I never get the email.

Can anyone help me?

Danny

Re: [Help-gsl] sparse matrices (again)

[Brian Gough]

At Mon, 14 May 2007 13:12:53 -0500,
Konkle, Daniel wrote:
> 
> I apologize for this email but I've tried the web interface several
> times now unsuccessfully.  I have been trying to discontinue receiving
> the emails from the list server.  I ask it to unsubscribe or send me my
> password to login but I never get the email.
> 
> Can anyone help me?

Try the method here -- the address in the 'List-Unsubscribe' header
mentioned in the second link should work:

http://sourceware.org/lists.html#unsubscribe-full
http://sourceware.org/lists.html#unsubscribe-full-addr-unknown

-- 
Brian Gough

Re: [Help-gsl] sparse matrices (again)

[Andrew Makhorin]

Sorry for a long time delay in my response.

> Andrew Makhorin wrote:
>> As it seems to me, it would be very useful to provide GSL at least
>> with data structures defining three most commonly used formats for
>> general sparse matrices, namely, coordinate format, compressed row
>> storage, and compressed column storage, as well as two formats for
>> sparse vectors based on compressed vector storage, and a minimal set
>> of some basic functions. This would standardize representation of
>> sparse objects within the GSL library and make it possible to use
>> such data structures on developing sparse linear algebra software
>> based on GSL.

> Thanks for the suggestions.  Sorry for the delay in replying.

> It might be feasible to add sparse matrix and vector datatypes and a
> sparse blas, without adding any specific solver algorithms (which is
> what I'd like to avoid).

> If this was added to GSL, would you want to use it in GLPK, or is it
> just something you thought would be generally useful to users of GSL?

Probably the latter rather than the former. Sparse matrix algorithms
used in GLPK are very specific and probably are not interesting for
most of the users of GSL. On the other hand, some standardization of
sparse calculations, in particular, representation of sparse objects
seems to me extremely useful. I cannot judge, however, it seems to me
that GSL might include standard data types for representing sparse
matrix, sparse graphs, as well as basic algorithms for such objects.

Thank you,

Andrew Makhorin