Quantcast

ublas+multiprecision+lapack

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

ublas+multiprecision+lapack

Carsten Burgard
Hi everyone!

I've been using the UBLAS matrix library for about a year now, mostly
for very simple operations (matrix inversion, condition number). Due to
very special requirements, we're using the types from boost
multiprecision to achieve higher precision numerics, which works perfectly.

Now, we're trying to go one step further and also calculate some
eigenvectors (and eigenvalues). As far as I understand, UBLAS does not
natively implement any numerics algorithms for these types of
operations, but instead relies on bindings to existing matrix libraries
like LAPACK.

However, it's not clear to me whether this will still work when
templated to multiprecision types. So far, I've been unable to come up
with a way to make it work (see, e.g., here.
http://stackoverflow.com/questions/37705633/boost-multiprecision-numeric-bindings-lapack).

Is there some way to make it work, possibly an example showing how it
can be done floating around somewhere on the web? Is there some special
backend that needs to be used instead of standard LAPACK? Or is this
whole endeavour just futile?

Regards,

Carsten


_______________________________________________
ublas mailing list
[hidden email]
http://lists.boost.org/mailman/listinfo.cgi/ublas
Sent to: [hidden email]
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: ublas+multiprecision+lapack

Oswin Krause
Hi,

its futile. LAPACK does not know multiprecision types. This would need a
templated implementation, luckily there are LAWNs showing that it is not
too hard to implement in an "efficient" manner (as EVs can not be
computed efficiently at all...)

On 2016-06-11 11:55, Carsten Burgard wrote:

> Hi everyone!
>
> I've been using the UBLAS matrix library for about a year now, mostly
> for very simple operations (matrix inversion, condition number). Due to
> very special requirements, we're using the types from boost
> multiprecision to achieve higher precision numerics, which works
> perfectly.
>
> Now, we're trying to go one step further and also calculate some
> eigenvectors (and eigenvalues). As far as I understand, UBLAS does not
> natively implement any numerics algorithms for these types of
> operations, but instead relies on bindings to existing matrix libraries
> like LAPACK.
>
> However, it's not clear to me whether this will still work when
> templated to multiprecision types. So far, I've been unable to come up
> with a way to make it work (see, e.g., here.
> http://stackoverflow.com/questions/37705633/boost-multiprecision-numeric-bindings-lapack).
>
> Is there some way to make it work, possibly an example showing how it
> can be done floating around somewhere on the web? Is there some special
> backend that needs to be used instead of standard LAPACK? Or is this
> whole endeavour just futile?
>
> Regards,
>
> Carsten
>
>
> _______________________________________________
> ublas mailing list
> [hidden email]
> http://lists.boost.org/mailman/listinfo.cgi/ublas
> Sent to: [hidden email]
_______________________________________________
ublas mailing list
[hidden email]
http://lists.boost.org/mailman/listinfo.cgi/ublas
Sent to: [hidden email]
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: ublas+multiprecision+lapack

Carsten Burgard
Hi Oswin!

Thanks for this quick response.

In that case, we need to investigate other backend options. Efficiency
is not terribly important to us, as our matrices are comparably small
and runtime is not a concern for this part of the code.

If you happen to know of a backend for UBLAS that we could use in
connection with multiprecision types, ideally one that comes with a bit
of documentation and/or usage examples, that would be awesome!

In any case, thanks again for your input!

Regards,

Carsten


On 06/11/2016 12:37 PM, Oswin Krause wrote:

> Hi,
>
> its futile. LAPACK does not know multiprecision types. This would need
> a templated implementation, luckily there are LAWNs showing that it is
> not too hard to implement in an "efficient" manner (as EVs can not be
> computed efficiently at all...)
>
> On 2016-06-11 11:55, Carsten Burgard wrote:
>> Hi everyone!
>>
>> I've been using the UBLAS matrix library for about a year now, mostly
>> for very simple operations (matrix inversion, condition number). Due to
>> very special requirements, we're using the types from boost
>> multiprecision to achieve higher precision numerics, which works
>> perfectly.
>>
>> Now, we're trying to go one step further and also calculate some
>> eigenvectors (and eigenvalues). As far as I understand, UBLAS does not
>> natively implement any numerics algorithms for these types of
>> operations, but instead relies on bindings to existing matrix libraries
>> like LAPACK.
>>
>> However, it's not clear to me whether this will still work when
>> templated to multiprecision types. So far, I've been unable to come up
>> with a way to make it work (see, e.g., here.
>> http://stackoverflow.com/questions/37705633/boost-multiprecision-numeric-bindings-lapack).
>>
>>
>> Is there some way to make it work, possibly an example showing how it
>> can be done floating around somewhere on the web? Is there some special
>> backend that needs to be used instead of standard LAPACK? Or is this
>> whole endeavour just futile?
>>
>> Regards,
>>
>> Carsten
>>
>>
>> _______________________________________________
>> ublas mailing list
>> [hidden email]
>> http://lists.boost.org/mailman/listinfo.cgi/ublas
>> Sent to: [hidden email]

_______________________________________________
ublas mailing list
[hidden email]
http://lists.boost.org/mailman/listinfo.cgi/ublas
Sent to: [hidden email]
Loading...