NAG C Header File (Linux/Unix and Windows Implementations)

Calling NAG Fortran Library Routines
from C/C++ Language Programs Using the NAG C Header File

Shah Datardina and Ian Hounam

NAG Ltd, Oxford

The Numerical Algorithms Group Ltd, Oxford UK. 2001, 2006, 2009

1 Introduction

A great number of systems allow the C/C++ programmer to call other language routines. Indeed the ANSI standard definition of C/C++ provides a powerful argument checking facility that, given the correct definition of function prototypes, can facilitate cross language communication between C/C++ and, say, Fortran.

A header file containing the function prototypes can be included in the user's program to allow the C/C++ compiler to check argument passage. Such a header file (nagmk22.h) has been created for the current NAG Fortran Library. This document explains how to call Fortran routines from C and C++ using the NAG header file on Windows and Linux/Unix Systems. Current users of NAG header files may prefer to skip to Section 8.

2 Declaration of Fortran Routine Names in C/C++

Because case is not significant in Fortran, compilers convert names to a standard case. Most Linux and Unix compilers convert the name to lower case and append an underscore, e.g., X01AAF becomes x01aaf_.

Generally under Windows, names are converted to upper case. However with some compilers, e.g. the Intel Fortran compiler on 32-bit Windows with the option /iface:cvf an additional qualifier is attached to the subprogram name, i.e., X01AAF becomes __stdcall X01AAF. This is the standard calling convention adopted by Microsoft for the WIN32 API.

3 Argument Types

The table below lists the mapping between relevant Fortran and C argument and function types.
   DOUBLE PRECISION D       double d;
   INTEGER I                int i;
   LOGICAL L                int l;
   CHARACTER*n S            char *s; int len_s; /* See below */
   COMPLEX*16 Z             struct {double re,im;} z;

In C all scalar arguments are passed as pointers to a variable; this means that a constant must first be assigned to a variable and then the address of the variable passed.

On the other hand in C++, scalar arguments may be passed as references with the result that all scalar arguments become C++ reference arguments. This C Header File caters for C++ compilers and uses reference arguments. Hence the "address of" operator must not be used for these arguments. In user supplied functions dereferencing such arguments is also avoided. Also constants may be passed to such arguments.

Character arguments are passed as two arguments

  1. a pointer to the character string
  2. an integer (int) passed by value containing the length of the string (for a character array this is the length of each element of the array).

All Linux/Unix Fortran compilers append this argument to the end of the argument list. Under Windows, the Intel Fortran compiler places the length at the end of the argument list by default. However, with the /iface:cvf option, the length of the character array is passed immediately after the string pointer, i.e., the compiler option /iface:mixed_str_len_arg is turned on. Note also that with /iface:cvf the standard calling convention (__stdcall) is enforced.

NB: On HP-UX 64, the type of the length parameter is long.

For example to call a routine called "NAGSUB" which in Fortran looks like this

      SUBROUTINE NAGSUB(A,B,C)
      CHARACTER A
      CHARACTER*3 B
      CHARACTER*5 C(2)

With most Unix and Linux compilers the C code looks like this.

  extern void nagsub_(char* a, char *b, char c[],
                     int len_a, int len_b, int len_c);
  main()
    {
      char a = 'a';
      int len_a = 1;
      char b[] = "abc";
      int len_b = 3;
      char c[2][6] = {"abcde", "fghij"}; /* Note the value 6 to allow
                                            for the null termination character */
      int len_c = 5;

      nagsub_(&a, b, (char *)c, len_a, len_b, len_c);
    }

The C code looks like this when the Fortran code is compiled with 32-bit Windows Intel Fortran using the option /iface:cvf

  extern void __stdcall NAGSUB(char* a, int len_a,
                char *b, int len_b, char c[], int len_c);
  main()
    {
      char a = 'a';
      int len_a = 1;
      char b[] = "abc";
      int len_b = 3;
      char c[2][6] = {"abcde", "fghij"}; /* Note the value 6 to allow
                                            for the null termination character */
      int len_c = 5;

      NAGSUB(&a, len_a, b, len_b, (char *)c, len_c);
    }

Note that the declared length of char variables in the C program must allow space for the null termination in the C string.

The Fortran type COMPLEX*16 (or COMPLEX(KIND=0.0d0)) is provided in the NAG header files by the typedef "Complex", which expands to "struct {double re,im;}"

4 Subroutine and Function Types

Fortran subroutines are declared as void functions in C.

Fortran functions will map as shown in the table above for data types, except for COMPLEX*16 functions and CHARACTER functions (see below).

Procedure arguments, i.e. function or subroutine names, are passed by address in the normal C manner.

5 COMPLEX*16 functions

The code examples in this section use the Unix conventions. On 32-bit Windows with the Intel Fortran Compiler using the /iface:cvf option, the __stdcall attribute should be used.

5.1 Most Unix and Windows Compilers

With newer compilers such as NAG Fortran, the HP-UX and Sun SPARC Solaris (64-bit implementation), COMPLEX*16 FUNCTIONs return a structure identical to the "Complex" typedef in the NAG C Header File, e.g.
      COMPLEX*16 FUNCTION F(X)
      COMPLEX*16 X

has the following prototype in C:

Complex (*f)(Complex *)
and a user callable function, e.g S01EAF, looks like this
extern Complex s01eaf_(const Complex *z,int *ifail);

5.2 Older Unix (e.g. 32-bit Sun SPARC Solaris)

Some older compilers such as the Sun SPARC Solaris 32-bit Fortran compiler use the convention that Fortran COMPLEX*16 functions are actually implemented as C void functions with an extra first argument which is a pointer to the result e.g.

The prototype for the user supplied function F with these compilers is the following.

      extern void  f(Complex *, Complex *);

and a user callable function, e.g S01EAF, looks like this

extern void  s01eaf_(Complex *return_value, const Complex *z,
                              int *ifail);

5.3 IBM AIX

With the native Fortran compiler on IBM AIX, the C programmer wishing to access Fortran COMPLEX*16 FUNCTIONs has real problems. The conventions used by the native Fortran compiler on this system does not allow the C programmer to access the return value of COMPLEX*16 FUNCTIONs directly because the COMPLEX*16 return value is stored in registers that do not map to any C data type.

One solution is to use the NAG Fortran Compiler implementation of the NAG Fortran Library, if available, please see above.

As it is impossible to access the return value of a COMPLEX*16 FUNCTION from C, another solution is to write a Fortran "jacket" routine to convert the COMPLEX*16 FUNCTION to a SUBROUTINE with an extra initial argument containing a pointer to the return value. The "jacket" routine then calls S01EAF.

For example to call S01EAF, write a jacket routine called S01EAFJ which would have the following prototype.

extern void s01eafj_(Complex *ret_val, CONST Complex *z,
                     int *ifail);
The jacket routine is then called from the C program rather than S01EAF.
      SUBROUTINE S01EAFJ(RET_VAL, Z, IFAIL)
      COMPLEX*16 RET_VAL, Z, S01EAF
      INTEGER IFAIL
      RET_VAL = S01EAF(Z, IFAIL)
      END
This routine can be compiled and linked with the C files and NAG Fortran Library in the normal way.

Similar jacket routines can be written for user supplied functions.

The supplied C Header file does not contain prototypes for these jacket functions. You may need to modify these prototypes appropriately.

6 Character functions

A character function is implemented as a void function with two extra arguments added to the start of the argument list, the first is a pointer to the returned string and the second its length.

e.g.

      CHARACTER*10 FUNCTION F(I)
is called thus:
     extern void f_(char *,int,int *);
      char c[10];
      int clen = 10,i;
      f_(&c,clen,&i)

The HP-UX 64-bit compiler uses long instead of int for the hidden length argument.

7 Multi-Dimension Arrays

As Fortran stores multi-dimension arrays in column major order whereas C/C++ store in row major order, either

  1. the C/C++ routine must store and manipulate the transpose of the problem matrix, or
  2. the C/C++ routine must transpose the matrix before and after calling the Fortran routine.
There is no syntax in the C language to specify an arbitrary array dimension and it is not possible to know in advance the dimensions of the arrays in your program. The best that the C Header Files can do is to declare the multi-dimensional arrays as one dimension arrays with a comment stating the actual number of dimensions, e.g. for 2 dimensions:
int array[] /* 2 dimension */

and for 3 dimensions:

double array[] /* 3 dimension */

The code examples in this section are for the Intel Fortran Compiler on 32-bit Windows with the /iface:cvf option. On other systems the __stdcall attribute should be omitted.

The prototype for a hypothetical NAG Fortran routine with a 2 dimensional DOUBLE PRECISION array argument would look like this:

extern void __stdcall NAGSUB(double[] /* 2 dimension */);

A simple program to call this routine might look like this:

main ()
{
  double p[2][2];
  NAGSUB((double *)p);
}

Note that we need to cast the 2 dimensional C array actual argument to (double *).

The example prototype below shows how to call a hypothetical NAG routine that takes a single subroutine argument. This user supplied subroutine takes a 2 dimension DOUBLE PRECISION array and an integer which specifies the leading dimension of the Fortran array.

extern void __stdcall NAGSUB(void (*f) (double[] /* 2 dimension */,
                             int *));

The C code for the user supplied function is listed below. The 2 dimension array is passed as a pointer to double and the code must carry out array indexing using the dimension information passed from Fortran. In this case, the macro P uses the leading dimension of the Fortran array, which is the trailing dimension of the C array, to index into the array p. The array p is only referenced through this macro.

void __stdcall fun(double p[], int *tdp)
{
#define P(I,J) p[(I)*(*tdp) + (J)]
  P(0,0) = 0.0;
  P(1,1) = 1.0;
}

The main function looks like this:

main ()
{
  void __stdcall fun(double p[], int *tdp);
  NAGSUB(fun);
}

Example 2 below shows a complete program that illustrates these concepts.

8 The nagmk22.h header file

In the past we have provided a number of header files for our Windows and Linux/Unix C/C++ users. This time round we are providing just the one header file which we believe will cater to the needs of most of our users. If you find that this header file does not meet your needs please contact us at support@nag.co.uk.

C/C++ users on Windows 64, Linux and Solaris 64 can use the header file as is.

Users on 32-bit Windows using the NAG Fortran library compiled with Intel Fortran using the /iface:cvf option will need to provide the option /DUSE_STDCALL to their C/C++ compiler in order to use the prototypes in which character lengths follow the character arguments and to activate the __stdcall attribute.

Users on systems where a complex function returns a complex structure again need do nothing. For compilers such as Sun SPARC Solaris 32-bit, the compiler option -DRETURN_COMPLEX_PARAM will have to be specified.

Users with the 64-bit HP-UX compiler need to edit the header file to change the typedef of Charlen to long.

The C Header files contain an 'extern "C"' declaration for C++ compilers, i.e.

#ifdef __cplusplus
extern "C" {
#endif
(with a matching "}" at the end of the file). As described in section 3, scalar arguments are passed by reference, so it follows that C++ users must not supply the address of scalar variables.

9 Linking your program

It is always necessary to link your program with the Fortran run time library routines. This can most easily be achieved on most Unix systems by linking using the f77 command, not cc or ld. For example,
cc -c myprog.c
f77 -o myprog myprog.o -lnag
On some systems this will not work and in some circumstances it is necessary to link using ld or cc, in which case it is necessary to specify the Fortran run time libraries in the load command. The libraries may be documented in the f77 man page or, failing that, the commands generated by the compiler driver may be viewed using the -v, -dryrun or -# options to display the ld command line.

10 Examples

The examples below are translations of NAG Fortran Library example programs into C. For these examples, where the Fortran code reads data, this data is coded directly into the initialisation of the C variables to make self-contained programs with no need for data files. Further examples, some of which have separate data files, are included in the accompanying materials. C++ version of the examples are also provided. Please refer to the NAG Fortran Library documentation for details of all these example programs.

10.1 Example 1

This example shows how to call C05AJF using a procedure parameter, the function "f".

C05AJF Example Program (most systems)

C05AJF Example Program (Win32 with CVF interface and STDCALL)

10.2 Example 2

This example illustrates the use of two dimensional arrays in user supplied functions in the NAG Fortran Library routine D03PCF.

D03PCF Example Program (most systems)

D03PCF Example Program (Win32 with CVF interface and STDCALL)

10.3 Example 3

This example illustrates the specification of two dimensional arrays in column major order in the C translation of the E04NFF example program. Results are printed by the NAG routine.

E04NFF Example Program (most systems)

E04NFF Example Program (Win32 with CVF interface and STDCALL)

10.4 Example 4

The Fortran Library routine F01CTF is called with the matrices in the natural C row major order. The arguments TRANSA and TRANSB are used to specify that the matrices should be transposed. The result matrix is then transposed back to row major order using F01CRF. This example also illustrates character string arguments.

F01CTF Example Program (most systems)

F01CTF Example Program (Win32 with CVF interface and STDCALL)

10.5 Example 5

DGBTRS is called with one matrix specified in column major order. Another matrix is loaded with its data using a macro to define the Fortran matrix ordering. Results are printed using X04CAF. This example also illustrates character string arguments.

DGBTRS/F07BEF Example Program (most systems)

DGBTRS/F07BEF Example Program (Win32 with CVF interface and STDCALL)