Intel® MKL Fortran 95 Interfaces for LAPACK Routines vs. Netlib Implementation

The following list presents general digressions of the Intel® oneAPI Math Kernel Library LAPACK95 implementation from the Netlib analog:

• The Intel® oneAPI Math Kernel Library Fortran 95 interfaces are provided for pure procedures.
• Names of interfaces do not contain the LA_ prefix.
• An optional array argument always has the target attribute.
• Functionality of the Intel® oneAPI Math Kernel Library LAPACK95 wrapper is close to the FORTRAN 77 original implementation in thegetrf, gbtrf, and potrf interfaces.
• If jobz argument value specifies presence or absence of z argument, then z is always declared as optional and jobz is restored depending on whether z is present or not.
• To avoid double error checking, processing of the info argument is limited to checking of the allocated memory and disarranging of optional arguments.
• If an argument that is present in the list of arguments completely defines another argument, the latter is always declared as optional.

You can transform an application that uses the Netlib LAPACK interfaces to ensure its work with the Intel® oneAPI Math Kernel Library interfaces providing that:

1. The application is correct, that is, unambiguous, compiler-independent, and contains no errors.
2. Each routine name denotes only one specific routine. If any routine name in the application coincides with a name of the original Netlib routine (for example, after removing the LA_ prefix) but denotes a routine different from the Netlib original routine, this name should be modified through context name replacement.

You should transform your application in the following cases:

• When using the Netlib routines that differ from the Intel® oneAPI Math Kernel Library routines only by theLA_ prefix or in the array attribute target. The only transformation required in this case is context name replacement.

• When using Netlib routines that differ from the Intel® oneAPI Math Kernel Library routines by theLA_ prefix, the target array attribute, and the names of formal arguments. In the case of positional passing of arguments, no additional transformation except context name replacement is required. In the case of the keywords passing of arguments, in addition to the context name replacement the names of mismatching keywords should also be modified.

• When using the Netlib routines that differ from the respective Intel® oneAPI Math Kernel Library routines by theLA_ prefix, the targetarray attribute, sequence of the arguments, arguments missing in Intel® oneAPI Math Kernel Library but present in Netlib and, vice versa, present in Intel® oneAPI Math Kernel Library but missing in Netlib. Remove the differences in the sequence and range of the arguments in process of all the transformations when you use the Netlib routines specified by this bullet and the preceding bullet.

• When using the getrf, gbtrf, and potrfinterfaces, that is, new functionality implemented in Intel® oneAPI Math Kernel Library but unavailable in the Netlib source. To override the differences, build the desired functionality explicitly with the Intel® oneAPI Math Kernel Library means or create a new subroutine with the new functionality, using specific MKL interfaces corresponding to LAPACK 77 routines. You can call the LAPACK 77 routines directly but using the new Intel® oneAPI Math Kernel Library interfaces is preferable. Note that if the transformed application callsgetrf, gbtrf or potrf without controlling arguments rcond and norm, just context name replacement is enough in modifying the calls into the Intel® oneAPI Math Kernel Library interfaces, as described in the first bullet above. The Netlib functionality is preserved in such cases.

• When using the Netlib auxiliary routines. In this case, call a corresponding subroutine directly, using the Intel® oneAPI Math Kernel Library LAPACK 77 interfaces.

Transform your application as follows:

1. Make sure conditions a. and b. are met.
2. Select Netlib LAPACK 95 calls. For each call, do the following:
   • Select the type of digression and do the required transformations.
   • Revise results to eliminate unneeded code or data, which may appear after several identical calls.
3. Make sure the transformations are correct and complete.