https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Tridiagonal-system-and-PARDISO/m-p/882539#M9724 Hi, <BR /><BR />I need to solve two tridiagonal (non symmetric though) systems at each time step (coming from a split finite difference scheme). My strategy was as follow:<BR />- define two context handles pt1 and pt2<BR />- perform symbolic factorization (phase 11) on each system<BR />- then at each time step, use a preconditioned CGS and solve the systems (phase 23)<BR /><BR />Unfortunately, I found no speed improvment wether I used preconditioned CGS or not. Also, solving this tridiagonal systems with PARDISO is like 10 times slower than solving it with LAPACK dgtsv()...<BR /><BR />Am I missing something ? <BR /><BR />Thanks for your advices,<BR /><BR />Xavier<BR /> Wed, 05 Dec 2007 15:30:25 GMT xavier_lacroze 2007-12-05T15:30:25Z https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Tridiagonal-system-and-PARDISO/m-p/882539#M9724 Hi, <BR /><BR />I need to solve two tridiagonal (non symmetric though) systems at each time step (coming from a split finite difference scheme). My strategy was as follow:<BR />- define two context handles pt1 and pt2<BR />- perform symbolic factorization (phase 11) on each system<BR />- then at each time step, use a preconditioned CGS and solve the systems (phase 23)<BR /><BR />Unfortunately, I found no speed improvment wether I used preconditioned CGS or not. Also, solving this tridiagonal systems with PARDISO is like 10 times slower than solving it with LAPACK dgtsv()...<BR /><BR />Am I missing something ? <BR /><BR />Thanks for your advices,<BR /><BR />Xavier<BR /> Wed, 05 Dec 2007 15:30:25 GMT https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Tridiagonal-system-and-PARDISO/m-p/882539#M9724 xavier_lacroze 2007-12-05T15:30:25Z https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Tridiagonal-system-and-PARDISO/m-p/882540#M9725 It seems unlikely that you could improve on the dgtsv algorithm, if you require its generality. The point of CG schemes is to solve problems with in-band sparsity, which would be exclusive of those where dgtsv could be applied directly.<BR /> Wed, 05 Dec 2007 16:48:14 GMT https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Tridiagonal-system-and-PARDISO/m-p/882540#M9725 TimP 2007-12-05T16:48:14Z https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Tridiagonal-system-and-PARDISO/m-p/882541#M9726 Thank you for your reply.<BR /><BR />Xavier<BR /> Wed, 05 Dec 2007 17:04:26 GMT https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/Tridiagonal-system-and-PARDISO/m-p/882541#M9726 xavier_lacroze 2007-12-05T17:04:26Z