Aim: Write a small test subroutine in Fortran using basic functionality and some calls to the International Mathematics and Statistics Library (IMSL) – compile a dll and call the subroutine from R taking some memory allocation problems into account.
Prerequisites: A working installation of Microsoft Visual Studio and Intel Fortran Compiler v.10 w. IMSL and R 2.6 This link may also be useful.
Save the following code as test.f90 in C:\Fortrancode
!—begin code—
subroutine test(M,N,S1,V,NU,S2,D,INVX)
!DEC$ ATTRIBUTES DLLEXPORT :: test !Routines necessary for BSKS, !used in calculation of !modified Bessel Function of second kind use BSKS_INT use UMACH_INT
!We use linear_operators for elegant !basic linear algebra syntax use linear_operators
implicit none
!input: !V vector with values used as input in Bessel fct. !M number of numbers in vector V !XDUM matrix with dimensions (M,N) !D dimension of dummy square matrix MX !NU fractional order used in calculation of !value of Bessel fct. modulo 1. Must be positive. integer, intent(in) :: M,N,D double precision, intent(in) :: NU, V(M) double precision :: INVX((D*D)), MX(D,D), XDUM(M,N)
!output: !S1 sum of elements in XDUM !S2 values of Bessel fct. !INVX is both an input and an output vector, !converted to a square matrix of dim. D !the inverse is calculated and returned as INVX double precision, intent(out) :: S1,S2(M)
!some dummy variables integer :: i,j,k double precision :: XX, BS(N)
!Set all M*N elements in XDUM to 0.5 XDUM=0.5
!Sum all elements in XDUM and return the result to S1 S1=sum(XDUM)
!Calculation of Bessel fct. values !corresponding to order NU+N and the !M different arguments found in V. !The M results are returned in S2. do i = 1, M XX=V(i) call BSKS (NU, XX, N, BS) S2(i)=BS(N) end do
!Transform the INVX vector into a D*D array, store result in MX !Note use of forall – which is a fast alternative to a !ordinary for loop. forall (i =1:D, j=1:D) MX(i,j)=INVX(i+(j-1)*D) end forall
!Calculate inverse of MX MX = .i. MX
!and return result in INVX forall (i =1:D, j=1:D) INVX(i+(j-1)*D)=MX(i,j) end forall
end subroutine test !—end code—
Run Microsoft Visual Studio as administrator Choose File -> New Project Select Intel(R) Fortran -> Library and click on Dynamic-link library Then give the project a name (test) and click OK In the right hand side frame in the Visual Studio main window right-click on Source Files Choose Add -> Existing Item… Browse and select the test.f90 file and click OK
Project-> Properties Select Fortran->Optimization Set Optimization to Maximize Speed Select Fortran->Data Set SEQUENCE Types Obey Alignment Rules to Yes (/align:sequence) Select Fortran->External Procedures Set Name Case Interpretation to Lower Case (/names:lowercase) and Append Underscore to External Names to Yes (\assume:underscore) Go to Fortran-> Command Line Write /heap-arrays in Additional Options: Click OK
Setting the /heap-arrays option circumvent problems when calling the subroutine from R. (Dummy arrays are allocated using the heap and not the stack.)
Since we use routines from the IMSL library (and maybe IntelMKL) we need a few more adjustments (using 32 bit versions): Tools->Options…->Intel Fortran->General Add C:\Program Files\VNI\imsl\fnl600\IA32\include\STATIC to Includes: (click … add the path to the list and click OK) and
Add C:\Program Files\VNI\imsl\fnl600\IA32\lib to Libraries: and Executables: In Project->Properties->Linker->Input in the “Additional Dependencies” line, add the following IMSL libraries and Intel MKL library according to your requirement (I highlighted the ones I use): Static link for 32 bit application: imsl.lib imslsuperlu.lib imslhpc_l.lib imsls_err.lib imslmpistub.lib mkl_c.lib libguide.lib Dynamic link for 32 bit application: imslmkl_dll.lib mkl_c_dll.lib libguide40.lib Static link Intel 64 bit application: imsl.lib imslsuperlu.lib imslscalar.lib mkl_em64t.lib libguide.lib imsls_err.lib imslmpistub.lib Dynamic link Intel 64 bit application: imslmkl_dll.lib mkl_dll.lib libguide40.lib In Linker ->Command write /force:multiple in Additional Options:
In the main window change the option above the file editor frame from Debug to Release
Choose Build -> Build test (and wait) The file test.dll should now be available in C:\Users\hellmund\Documents\Visual Studio 2005\Projects\test\test\Release (if your Windows username is hellmund)
In Intel Fortran Compiler version 10 it is not possible to change the default selection of a multi-threaded runtime library to a single-threaded. We thus have to provide som dlls in order to use our test.dll with dyn.load in R. Copy the files libifcoremd.dll and libmmd.dll from C:\Program Files\Intel\Compiler\Fortran\10.0.025\IA32\Lib to C:\Fortrancode
I have used the following R-code to call the subroutine. When writing R-code, I prefer the editor Tinn-R
#begin code
#Load the dll file dyn.load(“C:/Fortrancode/test.dll”); #Is the dll file loaded? is.loaded(“test”); #Set some parameters M=100000; V=sqrt(1:M); D=3; INVX=matrix(c(1,2,3,4,1,6,7,8,1),ncol=3,byrow=F); N=5; S1=0.0; NU=0.5; S2=rep(0,M); #Call the routine F=.Fortran(“test”,as.integer(M),as.integer(N),as.double(S1),as.double(V),as.double(NU),as.double(S2),as.integer(D),as.double(INVX))