Using Lapack with 128 bit precision

Question

I am trying to use Lapack for a 128 bit precision calculation of a matrix singular value decomposition (SVD) and I found out that there is some black compiler magic to accomplish this. The Intel Fortran compiler (ifort) supports the option -r16 which instructs the compiler to take all variables declared as DOUBLE PRECISION to be 128 bit reals. So I compiled Lapack and BLAS using:

ifort -O3 -r16 -c isamax.f -o isamax.o
ifort -O3 -r16 -c sasum.f -o sasum.o
...

To incorporate this in my program (which is C++) I can use the Intel C++ compiler (icc) with the option -Qoption,cpp,--extended_float_type which creates a data type _Quad that is a 128 bit floating point variable. My SVD example looks like this:

#include "stdio.h"
#include "iostream"
#include "vector"

using namespace std;
typedef _Quad scalar;

//FORTRAN BINDING
extern "C" void dgesvd_(char *JOBU, char *JOBVT, int *M, int *N,
    scalar *A, int *LDA,
    scalar *S,
    scalar *U, int *LDU,
    scalar *VT, int *LDVT,
    scalar *WORK, int *LWORK, int *INFO);

int main() {
    cout << "Size of scalar: " << sizeof(scalar) << endl;
    int N=2;
    vector< scalar > A(N*N);
    vector< scalar > S(N);
    vector< scalar > U(N*N);
    vector< scalar > VT(N*N);

    // dummy input matrix
    A[0] = 1.q;
    A[1] = 2.q;
    A[2] = 2.q;
    A[3] = 3.q;
    cout << "Input matrix: " << endl;
    for(int i = 0; i < N; i++) {
        for(int j = 0;j < N; j++) 
            cout << double(A[i*N+j]) << "\t";
        cout << endl;
    }
    cout << endl;

    char JOBU='A';
    char JOBVT='A';
    int LWORK=-1;
    scalar test;
    int INFO;

    // allocate memory
    dgesvd_(&JOBU, &JOBVT, &N, &N,
        &A[0], &N,
        &S[0],
        &U[0], &N,
        &VT[0], &N,
        &test, &LWORK, &INFO);
    LWORK=test;
    int size=int(test);
    cout<<"Needed workspace size: "<<int(test)<<endl<<endl;
    vector< scalar > WORK(size);

    // run...
    dgesvd_(&JOBU, &JOBVT, &N, &N,
        &A[0], &N,
        &S[0],
        &U[0], &N,
        &VT[0], &N,
        &WORK[0], &LWORK, &INFO);
    // output as doubles
    cout << "Singular values: " << endl;
    for(int i = 0;i < N; i++)
        cout << double(S[i]) << endl;
    cout << endl;
    cout << "U: " << endl;
    for(int i = 0;i < N; i++) {
    for(int j = 0;j < N; j++)
        cout << double(U[N*i+j]) << "\t";
    cout << endl;
    }
    cout << "VT: " << endl;
    for(int i = 0;i < N; i++) {
    for(int j = 0;j < N; j++)
        cout << double(VT[N*i+j]) << "\t";
    cout << endl;
    }
    return 0;
}

compiled with

icc test.cpp -g -Qoption,cpp,--extended_float_type -lifcore ../lapack-3.4.0/liblapack.a ../BLAS/blas_LINUX.a

Everything works fine this far. But the output is:

Size of scalar: 16
Input matrix: 
1       2
2       3

Needed workspace size: 134

Singular values: 
inf
inf

U: 
-0.525731       -0.850651
-0.850651       0.525731
VT: 
-0.525731       0.850651
-0.850651       -0.525731

I checked that U and VT are correct, but the singular values are obviously not. Has anyone got an idea why this happens or how one could circumvent it?
Thanks for your help.

Answer 1

When using external libraries with extended precision, also check whether they use old-style d1mach.f, r1mach.f, i1mach.f to get information on machine arithmetic. There may be some values to tweak here.

It can't be a problem with Lapack, which uses dlamch.f (here http://www.netlib.org/lapack/util/dlamch.f ), which uses Fortran 90 intrinsics to get these machine constants.

But it may become problematic for example with BLAS or SLATEC, if you use them.