Hi,
I want to use Extended Eigensolver RCI routines, but documentation is not enought detailed.
In particular, I don't understand what I need to do, when ijob = 10. Where do I need to put the (ZeB-A) result?
And what I need to put in Aq and As arrays ("Specifically, they contain the matrices for the reduced eigenvalue problem." - what does it mean?).
I want to change only method for solving linear system, but don't want to change something else (matrix format, factorization the Green's function and etc.). And all others I want to use predefined.
May be smb can give me example for RCI Interface, where used standart procedures from MKL?
Thanks in advance.
連結已複製
Dear Nikita,
Let us assume we want to solve a real generalized eigenvalue problem.
In the case of ijob=10 we need to compute a double complex precision matrix Ze*B-A and the representation of the output matrix Ze*B-A depends on the solver you are going to use. For example, if we going to use the MKL PARDISO, we have to compute Ze*B-A in the compressed sparse row format. If you are using a sparse direct solver, we need to factorize Ze*B-A with the help of this direct solver. In the case of a preconditioned iterative solver, you might compute a preconditioner for the linear system (Ze*A-B) X= Y. Of course, you can compute a preconditioner or performs the matrix factorization later when FEAST ask you to solve the system.
Here is a pseudocode for a real symmetric problem:
ijob=-1 ! initialization
do while (ijob/=0)
call dfeast_srci(ijob, N, Ze, work, workc, Aq, Sq, fpm, epsout,loop, Emin, Emax, M0, E, X, mode,res,info)
select case(ijob)
case(10) !! Form (ZeB-A) and Factorize the complex matrix (ZeB-A)
!! Let (ssa,sisa,sjsa) be the CSR representation of the input matrix A,
!! (ssb,sisb,sjsb) be the CSR representation of the input matrix B
!! and let (saz,isaz,jsaz) be the CSR representation of (Ze*A-B)
call <compute the resolvent matrix > (N, ssa,sisa,sjsa,Ze,ssb,sisb,sjsb,saz,isaz,jsaz) !! get saz
!! factorize the matrix Ze*B-A or
!! compute a preconditioner for (Ze*A-B) if a preconditioned iterative solver is used
PHASE=22
call pardiso(….,saz,isaz,jsaz,….)
case(11) !! Solve the complex linear system (ZeB-A)workc(1:N,1:M0) = workc(1:N,1:M0) result in workc
PHASE=33
call pardiso(…, saz,isaz,jsaz, … ,workc(:,1:M0), …)
case(30) !! Perform multiplication A*X(1:N,i:j) result in work(1:N,i:j)
!! where i=fpm(24) and j=fpm(24)+fpm(25)−1
call <sparse matrix-matrix multiply> (…, ssa, sjsa, sisa, … , X(1,fpm(24)), .., work(1,fpm(24)))
case(40) !! Perform multiplication B*X(1:N,i:j) result in work(1:N,i:j)
!! where i=fpm(24) and j=fpm(24)+fpm(25)−1
call <sparse matrix-matrix multiply> (…, ssb, sjsb, sisb, … , X(1,fpm(24)), .., work(1,fpm(24)))
end select
end do
Hope it helps
All the best
Sergey
When I call dfeast_srci first time, it work normally. Then nothing to change and nothing to do on (ijob=10). And then I call dfeast_srci second time and my programm crash with message:
Extended Eigensolvers: PROBLEM with input parameters
==>INFO code =: 1816549376
What does it mean? Why does the first call not generate an error?
After first calling of dfeast_srci the value of variable mode changing to negative, why can it be? And it's the reason why on the second call of dfeast_srcithe program is crashing with error
Extended Eigensolvers: PROBLEM with input parameters
==>INFO code =: 1816549376
But I think real reason somewhere else. Why can it be? Nothing happened after first call of this function...
Hi Nikita,
I'd assume that it is not enough memory for one or more input arrays. I'd recommend to check the type and size of each input array. If it doesn't help, please provide a reproducer.
All the best
Sergey
I try to calculate for problem with very small matrix, only 8 rows. And it's correctly work with predefined eigen solver. That cann't be memory problems with this size, I think...
I check all massives, but I didn't find an error. Here is my function code in attachment.
And it returns this in console:
---> phase -1 <---
Extended Eigensolvers: double precision driver
Extended Eigensolvers: List of input parameters fpm(1:64)-- if different from default
Extended Eigensolvers: fpm(1)=1
Search interval [1.000000000000000e+004;2.000000000000000e+006]
Extended Eigensolvers: Size subspace 8
#Loop | #Eig | Trace | Error-Trace | Max-Residual---> phase 10 <---
Extended Eigensolvers: PROBLEM with input parameters
==>INFO code =: 1816549376
I checked all arrays and found nothing wrong with it's values or size.
There is my function code in attachment. It returns the next:
---> phase -1 <---
Extended Eigensolvers: double precision driver
Extended Eigensolvers: List of input parameters fpm(1:64)-- if different from default
Extended Eigensolvers: fpm(1)=1
Search interval [1.000000000000000e+004;2.000000000000000e+006]
Extended Eigensolvers: Size subspace 8
#Loop | #Eig | Trace | Error-Trace | Max-Residual---> phase 10 <---
Extended Eigensolvers: PROBLEM with input parameters
==>INFO code =: 1816549376
And another question is the next:
On phase (ijob=11) I need to solve complex linear system and want to use iterative solver for it (that the reason why I don't use predefined eigen solver, I have very big matrices), but I can't find iterative complex RSI solver in MKL.
Maybe I can bad looking? Maybe there is exist iterative predefined eigen solver or iterative complex RSI linear solver?
Dear Nikita,
I investigated the issue and the problem appears because of a bug in RCI Extended Eigensolver interfaces from the MKL LP64 binaries. Everything should work well if you link with MKL ILP64 binaries like this
icc -I$MKL_ROOT/__release_lnx/mkl/include -DMKL_INT="long long" code.cpp $MKL_ROOT/__release_lnx/mkl/lib/intel64/libmkl_intel_ilp64.a -Wl,--start-group $MKL_ROOT/__release_lnx/mkl/lib/intel64/lib/libmkl_intel_thread.a $MKL_ROOT/__release_lnx/mkl/lib/intel64/libmkl_core.a -Wl,--end-group $MKL_ROOT/__release_lnx/compiler/lib/intel64/libiomp5.so -lpthread
Besides ILP64 RCI interfaces work a bit faster since the computational kernels are ILP64 and additional wrappers are not needed.
Unfortunately the Intel MKL doesn’t contain any iterative solver for complex linear systems. Feel free to submit a feature request. But you can try the MKL Out-of-Core PARDISO with MKL Extended Eigensolver in the case of large sparse matrices.
Thanks for catching this issue.
All the best
Sergey
Sergey Kuznetsov (Intel) wrote:Does you plan to fix this bug in MKL LP64 ? I don't want to use ILP because it causes many problems in my code in other places... =(
I investigated the issue and the problem appears because of a bug in RCI Extended Eigensolver interfaces from the MKL LP64 binaries. Everything should work well if you link with MKL ILP64 binaries like this
Sergey Kuznetsov (Intel) wrote:Where I need to submit future request? I think this problem was actual 2 years ago: http://software.intel.com/en-us/forums/topic/284068 - but nothing changed to this moment?
Unfortunately the Intel MKL doesn’t contain any iterative solver for complex linear systems. Feel free to submit a feature request
