There is a core limitation in XQuery 3.0 data types that make 2+D matrixes
painful.
That is sequences cannot nest, they automatically flatten.
e.g.
( ( 1 , 2 , 3 ) , (4 , 5 , 6) )
becomes
( 1,2,3,4,5,6)
So there is no natural datatype for matrix (> 1D)
You can implement them with elements or functions but both impose an overhead
that is harder for the optimizer to get rid of.
e.g. as elements (arbitrary name chosen)
let $a := <c>
<c>
<c>1</c>
<c>2</c>
<c>3</c>
</c>
<c>
<c>4</c>
<c>5</c>
<c>6</c>
</c>
</c>
return $a/c[2]/c[2]
Or as functions
let $a := ( function () { (1,2,3) } , function () { (4,5,6) } )
return $a[2]()[2]
Both return 5
The syntax is a bit cumbersome but it works ... its also hard on the optimizer
This among other reasons (such as supporting json arrays) is why native array
types that can nest are often supported as vendor extensions,
and I belive are being considered by W3C as XQuery native types.
From: [email protected] [mailto:[email protected]] On Behalf Of
Michael Sokolov
Sent: Tuesday, December 31, 2013 10:49 AM
To: jean-marc Mercier; Andrew Welch
Cc: xquery-discuss; ihe onwuka
Subject: Re: [xquery-talk] Matrix Multiplication
I would love to see some tests of pure XQuery implementations of both sparse
and dense operations. I suspect performance of matrix multiply, inversion,
etc, will be poorer than in C++ or Java, but I would expect performance
comparable to Perl or Python (w/o its numpy extension) - just a wild guess.
I'd also expect it might be easier to get good sparse performance. I don't
know why, just a hunch.
But the more interesting question for me is whether language changes are really
needed to support this. I would have thought that proper optimization of
operations on sequences would be enough? So for example, an extension module
using sequences as matrix datatypes could possibly optimize performance using a
lower-level implementation. Does anyone see any reason why that wouldn't be
possible?
-Mike
PS:
I reviewed the discussion you referred to, jean-marc, but it seems to have more
to do with using functions as map keys, and I don't see any direct connection
to linear algebra?
On 12/31/2013 9:55 AM, jean-marc Mercier wrote:
It is not due to the spec. It is rather due to the common usage of XQUERY,
forcing vendor solutions (as BaseX) to focus primarily on XML Data Base
requests more than algorithmic performances.
There are actually some threads that are discussing these performance issues in
the context of maps (maps are for instance used for sparse matrix
representations) : look for instance to ""map module for XQUERY ?" that I
initiated or "Higher-order XQuery Modules" by Leo from BaseX, on
[email protected]<http://x-query.com> mailing list.
Anyhow, to write a serious linear algebra modulus, the basic need is to have a
vector containers. Unfortunately, XQUERY does not provide any performant vector
containers at present time, and it is not possible to code them in pure XQUERY
: I have tried, and more experienced xquery developpers than me have also
tried, without success.
We have to wait for the XQUERY version that will give us these containers, a
decision to be taken by the W3C.
2013/12/31 Andrew Welch
<[email protected]<mailto:[email protected]>>
Are you saying the spec as it stands somehow forces all implementations to be
1000x slower, or just what you have observed in some particular implementation?
On 31 Dec 2013 14:27, "jean-marc Mercier"
<[email protected]<mailto:[email protected]>> wrote:
>
> As far as I understand, you want to write a linear algebra module using
> XQUERY ?
> If so, I opened a thread some months ago about this idea. My opinion today is
> that this is a false good idea at present time.
>
> 1) XQUERY would be really good for writing concise, efficient linear algebra
> modulus.
> 2) However, I strongly recommend to wait a little bit for starting coding :
> the current version of XQUERY (3.0) suffers from performance issues. A linear
> algebra modulus written in XQUERY is expected to have performances
> performances 1000 X slower than its corresponding C++ or JAVA (you can
> measure it precisely). Any mathematician linear algebra modulus would
> probably trashed your modulus after the first test.
>
> Hope this helps
>
>
>
> 2013/12/31 Ihe Onwuka <[email protected]<mailto:[email protected]>>
>>
>> Assuming a sparse representation it is about 4 lines of SQL. This is known
>> not least because you can read enough articles and papers that discuss it
>> and it optimises well. The obvious google search does not reveal any
>> corresponding XQuery discussion, neither does it appear to have surfaced on
>> this or the eXist mailing list (allowing for my deficient search skills).
>> For something so "trivial" I thought that was rather strange. Hence I
>> thought it would be prudent to ask before naively embarking on a 600k X
>> 40k matrix multiplication.
>>
>>
>>
>> On Tue, Dec 31, 2013 at 11:31 AM, Andrew Welch
>> <[email protected]<mailto:[email protected]>> wrote:
>>>
>>>
>>> It should be pretty trivial...
>>>
>>> On 31 Dec 2013 11:07, "Ihe Onwuka"
>>> <[email protected]<mailto:[email protected]>> wrote:
>>> >
>>> > Has anybody tried this in XQuery or if I am so foolish (not yet but give
>>> > me time) would I be the courageous
>>> > <culturalReference>http://www.youtube.com/watch?v=ik8JT2S-kBE</culturalReference>
>>> > early adopter.
>>> >
>>> >
>>> > _______________________________________________
>>> > [email protected]<mailto:[email protected]>
>>> > http://x-query.com/mailman/listinfo/talk
>>
>>
>>
>> _______________________________________________
>> [email protected]<mailto:[email protected]>
>> http://x-query.com/mailman/listinfo/talk
>
>
_______________________________________________
[email protected]<mailto:[email protected]>
http://x-query.com/mailman/listinfo/talk
_______________________________________________
[email protected]
http://x-query.com/mailman/listinfo/talk
