C++ LAPACK

[turbozedd]

The following is a framework for using LAPACK with C++.

For more information checkout the LAPACK forums:
http://icl.cs.utk.edu/lapack-forum/

And the netlib repository
http://www.netlib.org/clapack/

There are compile/install details at the above sites for CLAPACK and BLAS.

Compile commands for test program (executable 'fmat'):

CLAPACKPATH = path to clapack
g++  Fmatrix_test.cpp -I$(CLAPACKPATH)/SRC -I$(CLAPACKPATH) -c
g++  Fmatrix_test.o -o fmat $(CLAPACKPATH)/lapack_LINUX.a $(CLAPACKPATH)/blas_LINUX.a $(CLAPACKPATH)/F2CLIBS/libf2c.a -lm

[turbozedd]

Matrix class: this is just a container-- all c++
We use a template to store the different data types allowed by LAPACK.

The matrix "data" (the mat member) is a one dimensional array, it must be this way so that it is usable to FORTRAN.
This means we have to keep track of our location in the matrix. A[i][j] is given by A[ j*m+i ] where m is the number of rows.

#ifndef _FMATRIX_H
#define _FMATRIX_H

#include<iostream>

template <class T>
class Fmatrix
{
public:
    //default
    Fmatrix(): m(0), n(0), flag(0), mat(new T[0]){ this->settype(); }
    //blank
    Fmatrix(size_t _m): m(_m), n(1), flag(0), mat(new T[_m]){ this->settype(); }
    Fmatrix(size_t _m, size_t _n): m(_m), n(_n), flag(0), mat(new T[_m*_n])
    { this->settype(); }
    //copy
    Fmatrix(const Fmatrix<T>& B) :
    m(B.numrows()), n(B.numcols()), mat(new T[B.numelms()])
    {
        for(size_t i=0; i<B.numelms(); i++){ this->mat[i] = B[i]; }
        this->type=B.gettype();
    }

    ~Fmatrix(){ if(mat) delete[] mat; }

    inline T& operator() (size_t i, size_t j)
    { return this->mat[j*(this->m)+i]; }
    inline T  operator() (size_t i, size_t j) const
    { return this->mat[j*(this->m)+i]; }

    inline T& operator[] (size_t i)       { return mat[i]; }
    inline T  operator[] (size_t i) const { return mat[i]; }

    Fmatrix operator = (const Fmatrix<T>& B)
    {
        if(&B!=this){
            delete[] mat;
            this->m=B.numrows();
            this->n=B.numcols();
            this->flag=B.error();
            mat=new T[(this->m)*(this->n)];
            for(size_t i=0; i<B.numelms(); i++){ this->mat[i] = B[i];}
        }
        return *this;
    }
    //these (+-*/) won't work for complex
    Fmatrix operator += (const Fmatrix<T>& B)
    {  for(size_t i=0; i< this->numelms(); i++){ this->mat[i]+=B[i]; }  }
    Fmatrix operator -= (const Fmatrix<T>& B)
    {  for(size_t i=0; i< this->numelms(); i++){ this->mat[i]-=B[i]; }  }
    Fmatrix operator *= (const T alpha)
    {  for(size_t i=0; i< this->numelms(); i++){ this->mat[i]*=alpha; }  }
    Fmatrix operator /= (const T beta)
    {  for(size_t i=0; i< this->numelms(); i++){ this->mat[i]/=beta; }  }

    inline T* begin(){ return mat; }

    inline const size_t numrows() const { return this->m; }
    inline const size_t numcols() const { return this->n; }
    inline const size_t numelms() const { return (this->m)*(this->n); }
    inline const bool   error()   const { return this->flag; }
    inline const char   gettype() const { return this->type; }
    
    void print(std::ostream &os=std::cout, char delim='\t')
    {
        for(size_t i=0;i<this->m;i++){for(size_t j=0;j<this->n;j++){
            os << (*this)(i,j) << delim;
        } os << std::endl; } os << std::endl;
    }
    
    void read(std::istream &is=std::cin)
    {
        for(size_t i=0;i<this->m;i++){for(size_t j=0;j<this->n;j++){
            is >> (*this)(i,j);
        } }
    }

private:
    size_t m,n;
    T *mat;
    char type;
    bool flag;

    inline const char settype()
    {
        char tstr[20], otyp[]="czu";
        strcpy(tstr,typeid(T).name());
        if(tstr[0]=='f'||tstr[0]=='d'||tstr[0]=='e') type=tstr[0];
        else if( strcmp(tstr,"7complex")==0 )        type=otyp[0];
        else if( strcmp(tstr,"13doublecomplex")==0 ) type=otyp[1];
        else                                         type=otyp[2];
    }
};

#endif /* _FMATRIX_H */

[turbozedd]

Linear Algebra Module: This hooks up with CLAPACK
Note the extern for the includes, they are written in C.
Most of the functions here call LAPACK functions, I add functionality as I need it (LAPACK is huge).
To add a new function call to LAPACK you will need to look up the function on netlib or elsewhere to figure what you need to pass.
To pass the matrix data use myMatrix.begin()

#ifndef _LINALG_H
#define _LINALG_H

#include <cstdlib>
#include <cstdio>
#include <cmath>
#include <iostream>
#include <fstream>

extern "C" {
    #include "f2c.h"
    #include "blaswrap.h"
    #include "clapack.h"
}

#include "Fmatrix.h"

//these scalar ops (+-*/) won't work for complex
template <class T>
const Fmatrix<T> operator + (const Fmatrix<T>& A, const Fmatrix<T>& B)
{  return Fmatrix<T>(A)+=B;  }
template <class T>
const Fmatrix<T> operator - (const Fmatrix<T>& A, const Fmatrix<T>& B)
{  return Fmatrix<T>(A)+=B;  }
template <class T>
const Fmatrix<T> operator * (const T alpha, const Fmatrix<T>& A)
{  return Fmatrix<T>(A)*=alpha;  }
template <class T>
const Fmatrix<T> operator / (const Fmatrix<T>& A, const T beta)
{  return Fmatrix<T>(A)/=beta;  }
template <class T>
const Fvector<T> operator * ( Fmatrix<T>& A, Fvector<T>& x)
{
    Fvector<T> y(A.numrows());
    gmv(A,x,y);
    return y;
}
template <class T>
const Fmatrix<T> operator * ( Fmatrix<T>& A, Fmatrix<T>& B)
{
    Fmatrix<T> C(A.numrows(),B.numcols());
    gmm(A,B,C);
    return C;
}

template <class T>
int gmv(Fmatrix<T>& A, Fvector<T>& x, Fvector<T>& y, char TRANSA='N',
            T alpha=1, T beta=0, integer INCX=1, integer INCY=1)
{
    integer M, N, LDA;

    M=(TRANSA=='N') ? A.numrows():A.numcols();
    N=(TRANSA=='N') ? A.numcols():A.numrows();
    LDA=M;

    //this still needs a switch on data type
    dgemv_(&TRANSA, &M, &N,
                        (doublereal*)&alpha,
                        (doublereal*)A.begin(), &LDA,
                        (doublereal*)x.begin(), &INCX,
                        (doublereal*)&beta,
                        (doublereal*)y.begin(), &INCY);


}


template <class T>
int gmm(Fmatrix<T>& A, Fmatrix<T>& B, Fmatrix<T>& C,
            char TRANSA='N', char TRANSB='N', T alpha=1, T beta=0)
{
    integer M, N, K, LDA, LDB, LDC;

    M=(TRANSA=='N') ? A.numrows():A.numcols();
    N=(TRANSB=='N') ? B.numcols():B.numrows();
    K=(TRANSA=='N') ? A.numcols():A.numrows();
    LDA=(TRANSA=='N') ? M:K;
    LDB=(TRANSB=='N') ? K:N;
    LDC=C.numrows();

    switch(A.gettype()){//cast explicitly otherwise compiler is cranky
        case 'f':
            sgemm_(&TRANSA, &TRANSB, &M, &N, &K,
                        (real*)&alpha,
                        (real*)A.begin(), &LDA,
                        (real*)B.begin(), &LDB,
                        (real*)&beta,
                        (real*)C.begin(), &LDC);
        break;
        case 'd':
            dgemm_(&TRANSA, &TRANSB, &M, &N, &K,
                        (doublereal*)&alpha,
                        (doublereal*)A.begin(), &LDA,
                        (doublereal*)B.begin(), &LDB,
                        (doublereal*)&beta,
                        (doublereal*)C.begin(), &LDC);
        break;
        case 'c':
            cgemm_(&TRANSA, &TRANSB, &M, &N, &K,
                        (complex*)&alpha,
                        (complex*)A.begin(), &LDA,
                        (complex*)B.begin(), &LDB,
                        (complex*)&beta,
                        (complex*)C.begin(), &LDC);
        break;
        case 'z':
            zgemm_(&TRANSA, &TRANSB, &M, &N, &K,
                        (doublecomplex*)&alpha,
                        (doublecomplex*)A.begin(), &LDA,
                        (doublecomplex*)B.begin(), &LDB,
                        (doublecomplex*)&beta,
                        (doublecomplex*)C.begin(), &LDC);
        break;
        default:
            std::cout << "\nERROR: Something wrong in GMM switch.\n\n";
    }
    return 0;
}

template <class T>
int pca(Fmatrix<T>& A, Fmatrix<T>& U, Fmatrix<T>& S, Fmatrix<T>& PC)
{
   std::ofstream PCoutx("pcx"), PCouty("pcy"), PCoutz("pcz"), PCoutw("pcw");
        PCoutx.precision(16); PCouty.precision(16);
        PCoutz.precision(16); PCoutw.precision(16);
        Fmatrix<T> Y(A.numcols(),A.numrows());
        T s,*mean;

        mean = new T[A.numrows()];
        for(int i=0; i<A.numrows(); i++) mean[i]=0;

        //find means
        for(int i=0; i<A.numrows(); i++){ for(int j=0; j<A.numcols(); j++){
            mean[i]+=A(i,j);
        } }
        for(int i=0; i<A.numrows(); i++) mean[i]/=A.numcols();

        //subtract means from data
        for(int i=0; i<A.numrows(); i++){ for(int j=0; j<A.numcols(); j++){
            A(i,j)-=mean[i];
        } }

        //form matrix for SVD
        s=sqrt(A.numcols()-1);
        for(int i=0; i<A.numrows(); i++){ for(int j=0; j<A.numcols(); j++){
            Y(j,i)=A(i,j)/s;
        } }

        return svd(Y,U,S,PC);
}

template <class T>
int svd(Fmatrix<T>& A, Fmatrix<T>& U, Fmatrix<T>& S, Fmatrix<T>& V,
            char JOBU='A', char JOBVT='A')
{
    float  *wkf,*rwkc;
    double *wkd,*rwkz;
    complex *wkc;
    doublecomplex *wkz;
    integer M, N, mn, MN, LDA, LDU, LDVT, LWORK, INFO;

    M=A.numrows(); N=A.numcols(); LDA=M; LDU=M; LDVT=N;
    mn=min(M,N);
    MN=max(M,N);
    LWORK=2*max(3*mn+MN,5*mn);

    switch(A.gettype()){//cast explicitly otherwise compiler is cranky
        case 'f':
            wkf = new real[LWORK];
            sgesvd_(&JOBU, &JOBVT, &M, &N,
                        (real*)A.begin(), &LDA,
                        (real*)S.begin(),
                        (real*)U.begin(), &LDU,
                        (real*)V.begin(), &LDVT,
                        wkf, &LWORK, &INFO);
            delete[] wkf;
        break;
        case 'd':
            wkd = new doublereal[LWORK];
            dgesvd_(&JOBU, &JOBVT, &M, &N,
                        (doublereal*)A.begin(), &LDA,
                        (doublereal*)S.begin(),
                        (doublereal*)U.begin(), &LDU,
                        (doublereal*)V.begin(), &LDVT,
                        wkd, &LWORK, &INFO);
            delete[] wkd;
        break;
        case 'c':
            wkc = new complex[LWORK];
            rwkc = new real[5*mn];
            cgesvd_(&JOBU, &JOBVT, &M, &N,
                        (complex*)A.begin(), &LDA,
                        (real*)S.begin(),
                        (complex*)U.begin(), &LDU,
                        (complex*)V.begin(), &LDVT,
                        wkc, &LWORK, rwkc, &INFO);
            delete[] wkc; delete[] rwkc;
        break;
        case 'z':
            wkz = new doublecomplex[LWORK];
            rwkz = new doublereal[5*mn];
            zgesvd_(&JOBU, &JOBVT, &M, &N,
                        (doublecomplex*)A.begin(), &LDA,
                        (doublereal*)S.begin(),
                        (doublecomplex*)U.begin(), &LDU,
                        (doublecomplex*)V.begin(), &LDVT,
                        wkz, &LWORK, rwkz, &INFO);
            delete[] wkz; delete[] rwkz;
        break;
        default:
            std::cout << "\nERROR: Something wrong in SVD switch.\n\n";
    }

    return INFO;
}

#endif /* _LINALG_H */

[turbozedd]

Fmatrix_test.cpp : showing some functionality

#include <cstdlib>
#include <cstdio>
#include <cmath>
#include <iostream>

#include "Flinalg.h"

using namespace std;

int main(int argc, char** argv)
{
    cout.precision(2);
    size_t M,N;

    cout << "input numrows: ";
    cin >> M;
    cout << "input numcols: ";
    cin >> N;
    Fmatrix<double> A(M,N); //instantiate MxN matrix
    Fmatrix<double> v(N),u(M); //instantiate Mx1 matrix i.e. vector
    cout << "A is " <<  A.numrows() << 'x' << A.numcols() << ", numelms: " << A.numelms() << endl;
    cout << "v is " <<  v.numrows() << 'x' << v.numcols() << ", numelms: " << v.numelms() << endl;
    cout << "u is " <<  u.numrows() << 'x' << u.numcols() << ", numelms: " << u.numelms() << endl << endl;

    cout << "Input data for matrix A by rows (" << A.numelms() << " numbers):" << endl;
    A.read();
    cout << endl << "A =" << endl;
    A.print();
    
    //Matrix-vector multiply by operator overload
    cout << endl << "Setting v to ones vector..." << endl;
    for(int i=0;i<v.numelms();i++)
        v[i]=1;
    cout << "A*v =" << endl;
    u=A*v;
    u.print();
    
    //generic matrix-matrix multiply
    Fmatrix<double> C(A.numrows(),A.numrows());
    gmm(A,A,C,'N','T');
    cout << "A*A' =" << endl; 
    C.print();
    
    cout << "computing svd..." << endl;
    Fmatrix<double> U(M,M),S(min(M,N)),V(N,N); //min is defined in clapack.h
    cout << "INFO: " << svd(A,U,S,V) << endl; //do svd, output error code
    cout << "singular values:" << endl;
    for (size_t i=0; i < min(M,N); i++)
    { cout << "S[ " << i+1 << " ]\t= " << S[i] << endl; } cout << endl;
    cout << "singular vectors:" << endl;     //rows of V so print transpose
    for(int i=0; i<V.numrows(); i++){ cout << "v" << i+1 <<'\t'; } cout << endl;
    for(int i=0; i<V.numrows(); i++){ for(size_t j=0;j<V.numcols();j++){
        cout << V(j,i) << '\t';
    } cout << endl; }

    return 0;
}

[turbozedd]

Sample run (input is bold):

input numrows: 6
input numcols: 5
A is 6x5, numelms: 30
v is 5x1, numelms: 5
u is 6x1, numelms: 6

Input data for matrix A by rows (30 numbers):
2 5 2 6 4
3 3.4 2 6 7
2 2 2 2 0
0 0 0 1.234 0
3 23 46 93 .001
1 2 3 4 5

A =
2       5       2       6       4
3       3.4     2       6       7
2       2       2       2       0
0       0       0       1.2     0
3       23      46      93      0.001
1       2       3       4       5


Setting v to ones vector...
A*v =
19
21
8
1.2
1.7e+02
15



A*A' =
85      91      30      7.4     7.7e+02 62
91      1.1e+02 29      7.4     7.4e+02 75
30      29      16      2.5     3.3e+02 20
7.4     7.4     2.5     1.5     1.1e+02 4.9
7.7e+02 7.4e+02 3.3e+02 1.1e+02 1.1e+04 5.6e+02
62      75      20      4.9     5.6e+02 55

computing svd...
INFO: 0
singular values:
S[ 1 ]  = 1.1e+02
S[ 2 ]  = 11
S[ 3 ]  = 2.9
S[ 4 ]  = 1.9
S[ 5 ]  = 1.2

singular vectors:
v1      v2      v3      v4      v5
-0.032  -0.33   0.48    -0.35   -0.74
-0.22   -0.34   0.72    0.16    0.54
-0.43   0.088   -0.12   -0.84   0.29
-0.87   0.062   -0.13   0.39    -0.26
-0.0092 -0.88   -0.47   0.0089  0.079

 

[areyoupp]

hi . Can lapack solve equation ? for example: 4x+2=54-7x+3 etc.

[turbozedd]

It will solve systems of equations. So yes, but the amount of work needed to set up a single-variable single-equation system is equivalent to solving it.

e.g.:
To ask it to solve

4x+2=54-7x+3

You would tell it 11x=55

Which is a waste of time (formulating your input for lapack).

But something like:

12x +14y - 8z = 2
14x - 5y +  z = 4
  x +  y +  z = 4

Is worhtwhile.
You can also solve (sort of) under-determined systems (systems with more variables than equations).

You want a computer algebra system (CAS). Maple is pretty good for symbol manipulation (but big $$$). Octave/Matlab will do it too. Octave and Matlab actually use LAPACK

[areyoupp]

than you for your answer.this case too hard for me now :D. i must learn many something