Intel® oneAPI Math Kernel Library
Ask questions and share information with other developers who use Intel® Math Kernel Library.
6956 Discussions

Ineffective implementation Relatively Robust Representations algorithm in Intel MKL

yuriisig
Beginner
540 Views
On my pagehttp://www.thesa-store.com/products/ (it is not currently available: see below) (for processor P4, item 2.2) was seen compared my algorithm and the algorithm proposed in the late last century, Inderjit S. Dhillon and sold in a package Interl MKL, both in speed and accuracy, and the results of comparison were not in favor of this algorithm.

And what we have now? With regard to the orthogonality of vectors, then the implementation in recent releases Intel encouraging. A parallelization in dstegr Intel MKL is not implemented, and speed problems.
For the tridiagonal matrix from paragraph 2.2 of the size n = 30001 is my result - with 56.6 sec (hardware configuration: i7 860 processor (Speed: 2.80 GHz), Motherboard DP55KG,DDR31333 MHz (8 GB), OS Windows XP Professional x64 Edition SP2,Intel MKL 10.2 Update 4, EM64T,HT off). And dstegr Intel MKL provides 19 min. 37 sec. (result is given to the frequency of 2.80 GHz to compensate for the turbo boost, because parallelization in dstegr Intel MKL is not implemented). The difference in more than 20 times!


P.S.
The results presented here relates to an improved algorithm, on which information is published on my page. I also want to note that the parallelization of my algorithm is not complete (work on a full parallelization is), making it an advantage over the RRR algorithm even more impressive.Regarding the accuracy of the eigenvectors, it is not inferior RRR algorithm. My web page (it is not currently available) and publications, which used my diagonalization, can be downloaded here: http://depositfiles.com/files/fmy2ueaad
0 Kudos
10 Replies
Gennady_F_Intel
Moderator
540 Views

Hello Jury,

1) Yes, you are right. This function is not threadedat all. Our implementation of this algorithm is the same as netlib has.

2)how can we verify this? Can you give us the binaries to check the problem on our side?

--Gennady

0 Kudos
yuriisig
Beginner
540 Views

2)how can we verify this? Can you give us the binaries to check the problem on our side?

--Gennady

Hello Gennady,

mine C-file for testing dstegr Intel MKL? Program implementation of my algorithm, I would not want to give: let Intel representatives come to my house and see that.

--Yurii
0 Kudos
Gennady_F_Intel
Moderator
540 Views
Yurii,
we didn't asked your source files, but binaries files would be better...
--Gennady

0 Kudos
yuriisig
Beginner
540 Views
Gennady,
The fact that my algorithms are not published, but only links to the articles (9-11 http://www.thesa-storre.com/products/english.php ), in which they are used. Specialist easy to disassemble the translated code: ie I risk, and you sent me a specialist home, no risk. Previously I skip the tests on my computer, which you have given me.
--Yurii
0 Kudos
yuriisig
Beginner
540 Views
Gennady,

While I am prepared to provide the source code and the idea of the algorithm, referred to in paragraph 5.1 on my web page http://www.thesa-store.com/products/english.php .

--Yurii
0 Kudos
barragan_villanueva_
Valued Contributor I
540 Views
Hi,

This site containscomparison with Intel very old MKL 8.0 and 8.1 :(
What about the latest MKL 10.2.4?
0 Kudos
yuriisig
Beginner
540 Views
What about the latest MKL 10.2.4?

Yes. But the idea of my algorithms have not changed. And Intel in that time has moved only in addressing the matrix multiplication for IA32 by prof. Granovsky (for 65 nm. processors). And Intel has also significantly improved the RRR algorithm. Regarding the algorithm, which I counterpoise RRR, he is much improved.

0 Kudos
yuriisig
Beginner
540 Views
Quoting yuriisig
I also want to note that the parallelization of my algorithm is not complete (work on a full parallelization is)...
Now the algorithm is fully parallel.
0 Kudos
yuriisig
Beginner
540 Views
I also want to note, except for very low speedof the RRR algorithm and the impossibility of parallelism, rather low accuracy of the calculated eigenvectors, that does not allow many of the calculations correctly. But the algorithm for finding the eigenvectors of tridiagonal matrix is the key to finding the eigenvectors of real symmetric matrix. Thus, the RRR algorithm makes sense to use if you do not need high accuracy of calculations.
0 Kudos
yuriisig
Beginner
540 Views
we didn't asked your source files, but binaries files would be better...

Gennady,
can use the source code: this is my gift. Id like to notice that dlarfb has a lot of applications so the importance of changes submitted by me is quite high.
http://software.intel.com/en-us/forums/showthread.php?t=77331&o=d&s=lr
--Yurii

0 Kudos
Reply