EVSL  1.1.0
EigenValues Slicing Library
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros
EVSL Documentation
Alt
Polynomial and Rational Filtered Lanczos and subspace iteration algorithms
For Symmetric Eigenvalue problems

Welcome to EVSL. EVSL is a C library for computing the eigenvalues of a symmetric matrix that are located in a given interval. This release includes the routines listed above and does not yet offer full parallel implementations (trivial openMP test programs are available among the test drivers). EVSL also provides tools for spectrum slicing, i.e., the technique of subdividing a given interval into p smaller subintervals and computing the eigenvalues in each subinterval independently, as well as Kernel Polynomial method (KPM) and Lanczos based density of states/spectral density estimators. EVSL implements a polynomial filtered Lanczos (thick restart, no restart) a rational filtered Lanczos (thick restart, no restart), and a polynomial filtered subspace iteration for solving standard eigenvalue problems A u = λ u and generalized eigenvalue problems A u = λ B u.

For questions/feedback send e-mail to Yousef Saad [saad@.nosp@m.umn..nosp@m.edu]


DESCRIPTION OF CONTENTS



INSTALLATION


Library: The user only needs to modify the file makefile.in [see makefile.in.example for samples of files makefile.in that are given for mac-os and for Linux].

cp makefile.in_Linux/MacOS.example makefile.in. modify makefile.in [provide C compiler and BLAS/LAPACK path] make clean; make

Test programs: In the directories under TESTS/, you will find a number of directories which contain makefiles to build sample drivers that test a few different situations.

CXSparse SuiteSparse is the default direct linear solver used in EVSL for the linear systems that arise in rational filtering methods and in generalized eigenvalue problems. CXSparse is included only to allow quick testing. However, know that it is significantly slower than other direct solvers such as those in SuiteSparse for example (see below).

NOTE: CXSparse, which is distributed with EVSL, is Copyrighted by Timothy Davis. As noted above much better performance can be achieved by other existing direct solvers. Refer to CXSparse package for its License. [http://faculty.cse.tamu.edu/davis/suitesparse.html]

SuiteSparse: As a replacement for CXSparse, bindings are provided for the package SuiteSparse. Once SuiteSparse is installed, simply swich the DIRECTSOL variable in the makefile.in, and add the path to EXTERNAL/makefile_suitesparse.in.

EVSL invokes SuiteSparse to solve linear systems with (A-SIGMA I) or (A-SIGMA B), and CHOLMOD for solving linear systems with B.

Other solvers can also be used by providing the same interface as done for SuiteSparse. Follow the examples implemented in EXTERNAL/evsl_suitesparse.c

NOTE: SuiteSparse is NOT distributed with EVSL, and it is Copyrighted by Timothy Davis. Refer to SuiteSparse package for its License. [http://faculty.cse.tamu.edu/davis/suitesparse.html]


LINKING WITH UMFPACK (SuiteSparse 4.5.3)


UMFPACK and CHOLMOD requires AMD, COLAMD, CCOLAMD and CAMD, and optionally METIS 5.1.0. Compile each of these packages to have the library file in the Lib directory. If SuiteSparse is configured with METIS, give the path to METIS (v 5.1.0) as well to make libmetis.a, in metis-5.1.0/ type

make  config; make

Please refer to SuiteSparse and METIS for installation details.


Bindings with Pardiso as a direct solver are also provided – see the directory EXTERNAL for details.


RATIONAL FILTERING


Rational filtering requires solving linear systems (where the coefficient matrix is the original matrix shifted by a complex shift). A linear solver routine must be provided.

After having computed the rational filter by

find_ratf(intv, &rat),

users can call

SetASigmaBSol(&rat, func, allf, data)

to set the solver functions and associated data for all the poles of the rational filter. func is an array of function pointers of length num of poles, i.e., rat->num. So, func[i] is the function to solve the systems with pole i, the coefficient matrix of which is A - s_i I(or, B), where s_i = rat->zk[i] is the complex shift. data is an array of (void*) of the same length, where data[i] is the data needed by func[i].

Each "func[i]" must be of the following prototype

void SolFuncC(int n, double *br, double *bz, double *xr, double *xz, void *data);

where n is the size of the system, br, bz are the right-hand side (real and imaginary parts of complex vector), xr, xz will hold the solution (complex vector) on return, and data contains all the data needed for the solver.

If all func[i] are the same, one can set func==NULL and set allf to the function

Once SetASigmaBSol is executed, rational filtering Lanczos methods should be ready to use.


MATRIX-FREE EIGEN-SOLVERS


All the eigensolvers in EVSL can be used in `matrix-free' mode. In this mode, users need only to provide their own matrix-vector product function of the following prototype:

void MVFunc(double *x, double *y, void *data);

where y = A * x and data is the pointer to the associated data to perform the matvec. The (void *) argument is to provide a uniform interface to all user-specific functions. For a particular Matvec function, one can pack all data needed by this function into a struct and pass the pointer of this struct to EVSL (after casting it to (void *)). This function needs to be passed to EVSL as well, so EVSL can call it to perform all matvecs. The user can also pass needed matrices in the standard CSR format to EVSL, in which case EVSL will use its internal MATVEC routine. This can be set by

SetAMatrix(csrMat *A)
SetBMatrix(csrMat *B)

for the matrices A and B

In TESTS/LAP, an example of matvec functions for 2D/3D Laplacian matrices is provided, where the matrix is not explicitly formed but 5pt/7pt stencil is used instead to perform the matvecs. In this example, a struct for matvec is first defined:

typedef struct _lapmv_t {  
  int  nx,  ny,  nz; 
  double  *stencil;  
} lapmv_t;

and the matvec function is implemented as:

void Lap2D3DMatvec(double  *x, double  *y,  void  *data) {  
  lapmv_t *lapmv = (lapmv_t *) data; 
  int nx = lapmv->nx; 
  int ny = lapmv->ny;
  int nz = lapmv->nz; 
  double *stencil = lapmv->stencil; 
  ...  
}

in which the pointer is first cast and all the data is unpacked. Once these are ready, they can be passed to EVSL by calling

SetAMatvec(n, &Lap2D3DMatvec, (void*) &lapmv) and
SetBMatvec(n, &Lap2D3DMatvec, (void*) &lapmv)

to set the matvec routines for A and B respectively, where the first input is the size of the "matrix", the second input is the function pointer and the third one is the data pointer. Once SetMatvecFunc is called, EVSL will use the registered matvec function to perform all matvecs with A.

Users should first create a function wrapper of the above type for an external matvec routine. Then, following the steps in the example, it will be straightforward to use it in EVSL.


GENERALIZED EIGENVALUE PROBLEM


For solving A x = λ B x, the users must also provide a solver for the B matrix by calling

SetBSol(SolFuncR func, void *data).

To tell EVSL to solve the generalized eigenvalue problem, one must call

SetGenEig()

since by default, EVSL assumes solving standard eigenvalue problem even if B is provided. Call function

SetStdEig()

for solving standard eigenvalue problem

The current version of EVSL will need solves with LT for spectrum slicing, where B=L*L' is the Cholesky factorization. Call function

SetLTSol(SolFuncR func, void *data)

to set the solver function


Initialization and Completion