PARDISO: New Version 4

[Perpf2000]

Hello!

I got word that there is a new version of PARDISO (V4). As it seems to bring about some performance-enhancements, I'd like to know whether there are plans to make this version available in MKL (and when) ?

best regards,

[Shane_S_Intel]

We are evaluating V4 currently and will decide how to proceed once that evaluation is complete. I have been told the V4 interfaceis differentwhich is unfortunate because itmaymake itincompatible with the existing MKLinterface. As you probably already know, feature wise MKL Pardiso and PARDISO have differed for awhile.

[styc]

Quoting - Shane Story (Intel)

We are evaluating V4 currently and will decide how to proceed once that evaluation is complete. I have been told the V4 interfaceis differentwhich is unfortunate because itmaymake itincompatible with the existing MKLinterface. As you probably already know, feature wise MKL Pardiso and PARDISO have differed for awhile.

What are the differences between MKL's PARDISO and Olaf Schenk's PARDISO (V3.3) besides out-of-core capability?

[Sergey_K_Intel1]

Quoting - styc

What are the differences between MKL's PARDISO and Olaf Schenk's PARDISO (V3.3) besides out-of-core capability?

There exist the number of the features been existing in MKL PARDISO for long time and these features have only recently released in Olaf Schenk's PARDISO V4. Here are some of these features: internal LP64 addressing, threaded reordering step, computation of determinant and splitting of forward and backward solves. It should be noted that all these features have been supported differently.

Among other big features supported by MKL PARDISO only and not supported by both versions of Olaf Schenk's PARDISO v4 and v3.3 , I'd like to mention the support of all data types including single and single complex data types and ILP64 support. There exist also some small differences as well. Compared to PARDISO v4, MKL PARDISO supports PGI and GNU threading libraries and MKL PARDISO provides compatibility with Intel TBB. MKL PARDISO also includes internal sparse matrix checker, MKL PARDISO allows the user to get back the permutation vector and it reports the number of equation where PARDISO detects zero or negative pivot for real symmetric positive definite and complex Hermitian positive definite matrices.

All the best

Sergey

[Alemdar__Bulent]

Quoting - Sergey Kuznetsov (Intel)

There exist the number of the features been existing in MKL PARDISO for long time and these features have only recently released in Olaf Schenk's PARDISO V4. Here are some of these features: internal LP64 addressing, threaded reordering step, computation of determinant and splitting of forward and backward solves. It should be noted that all these features have been supported differently.

Among other big features supported by MKL PARDISO only and not supported by both versions of Olaf Schenk's PARDISO v4 and v3.3 , I'd like to mention the support of all data types including single and single complex data types and ILP64 support. There exist also some small differences as well. Compared to PARDISO v4, MKL PARDISO supports PGI and GNU threading libraries and MKL PARDISO provides compatibility with Intel TBB. MKL PARDISO also includes internal sparse matrix checker, MKL PARDISO allows the user to get back the permutation vector and it reports the number of equation where PARDISO detects zero or negative pivot for real symmetric positive definite and complex Hermitian positive definite matrices.

All the best

Sergey

Sergey

Would you mind telling how to obtain this information: "MKL PARDISO allows the user to get back the permutation vector and it reports the number of equation where PARDISO detects zero or negative pivot for real symmetric positive definite and complex Hermitian positive definite matrices"? I don't remember anything about it in the release notes.

Regards
Bulent

[Sergey_K_Intel1]

Quoting - bulent

Sergey

Would you mind telling how to obtain this information: "MKL PARDISO allows the user to get back the permutation vector and it reports the number of equation where PARDISO detects zero or negative pivot for real symmetric positive definite and complex Hermitian positive definite matrices"? I don't remember anything about it in the release notes.

Regards
Bulent

Dear Bulent,

Unfortunately I don't know which MKL release you have been using. The policy is that we only list new features in release notes and it might be these features were first appeared in early MKL releases.

Soif iparm(5)=2, then PARDISO returns the permutation vector into the array perm.

And iparm(30) returns the firstnumber of equation where PARDISO detects zero or negative pivot for MTYPE=2 (real positive definite matrix) and MTYPE=4 (complex and Hermitian positive definite matrices).

Please see MKL Reference Manual for more details.

All the best
Sergey