Points to Note

For anyone who obtains the complete LAPACK package from netlib (see Chapter 1), a comprehensive installation guide is provided. We recommend installation of the complete package as the most convenient and reliable way to make LAPACK available.

People who obtain copies of a few LAPACK routines from netlib need to be aware of the following points:

1.

Double precision complex routines (names beginning Z-) use a COMPLEX*16 data type. This is an extension to the Fortran 77 standard, but is provided by many Fortran compilers on machines where double precision computation is usual. The following related extensions are also used:

• the intrinsic function DCONJG, with argument and result of type COMPLEX*16;

• the intrinsic functions DBLE and DIMAG, with COMPLEX*16 argument and DOUBLE PRECISION result, returning the real and imaginary parts, respectively;

• the intrinsic function DCMPLX, with DOUBLE PRECISION argument(s) and COMPLEX*16 result;

• COMPLEX*16 constants, formed from a pair of double precision constants in parentheses.

Some compilers provide DOUBLE COMPLEX as an alternative to COMPLEX*16, and an intrinsic function DREAL instead of DBLE to return the real part of a COMPLEX*16 argument. If the compiler does not accept the constructs used in LAPACK, the installer will have to modify the code: for example, globally change COMPLEX*16 to DOUBLE COMPLEX, or selectively change DBLE to DREAL.^6.1

2.

For optimal performance, a small set of tuning parameters must be set for each machine, or even for each configuration of a given machine (for example, different parameters may be optimal for different numbers of processors). These values, such as the block size, minimum block size, crossover point below which an unblocked routine should be used, and others, are set by calls to an inquiry function ILAENV. The default version of ILAENV provided with LAPACK uses generic values which often give satisfactory performance, but users who are particularly interested in performance may wish to modify this subprogram or substitute their own version. Further details on setting ILAENV for a particular environment are provided in section 6.2.

3.

SLAMCH/DLAMCH determines properties of the floating-point arithmetic at run-time, such as the machine epsilon, underflow threshold, overflow threshold, and related parameters. It works satisfactorily on all commercially important machines of which we are aware, but will necessarily be updated from time to time as new machines and compilers are produced.