quad precision -- natlib -- mkl

[diedro]

Dear All,

I would like to compile my source code with double precision. However, I have to useGETRI andGETRF.

TimP suggested me to use "netlib functions".

I usually compile with:

[bash] ifort -r8 -i8 *.f90 -L$MKLPATH/lib/em64t  -lmkl_blas95_ilp64 -lmkl_lapack95_ilp64 -lmkl_intel_ilp64 -lmkl_sequential -lmkl_core -lpthread -warn all -check all -debug all[/bash]

Is it possible to do the same withnetlib in intel fortran?

Thank to every one

[mecej4]

Your request is somewhat inconsistent. The title talks about quad precision, but in the body you mention wanting double precision, and the use of the -r8 option suggests that you wish to promote source code written for single-precision to double precision. I do not know why you use the -i8 option.

You do not have to use Netlib source code. Here is an example of using GETRF.

[fortran]program invert
use mkl95_lapack,only : getrf,getri
implicit none
integer, parameter :: N = 7
integer :: i,j
double precision, dimension(N,N) :: A = &
  (/ 1.0658d0,   -0.6313d0, 0.0000d0, 0.0000d0, 0.0000d0, 0.0000d0, 0.0000d0, &
   -0.6313d0, 4.2632d0,  -1.8940d0, 0.0000d0, 0.0000d0, 0.0000d0, 0.0000d0, &
  0.0000d0,  -1.8940d0, 8.5265d0,  -3.1567d0, 0.0000d0, 0.0000d0, 0.0000d0, &
  0.0000d0, 0.0000d0,  -3.1567d0,  23.2603d0,  -3.6892d0, 0.0000d0, 0.0000d0, &
  0.0000d0, 0.0000d0, 0.0000d0,  -3.6892d0,  37.5680d0, -4.8192d0, 0.0000d0, &
  0.0000d0, 0.0000d0, 0.0000d0, 0.0000d0,  -4.8192d0,  47.5467d0,  -5.9492d0, &
  0.0000d0, 0.0000d0, 0.0000d0, 0.0000d0, 0.0000d0,  -5.9492d0,  57.5255d0 /)
integer, dimension(N) :: piv
integer :: info

call getrf( a, piv, info )
call getri( a, piv, info )
write(*,10)((a(i,j),j=1,N),i=1,N)
stop

10 format(1x,7F12.6)
end program invert
[/fortran]

[diedro]

hi,

my fault, really sorry.

I would like to compile as:

[bash]ifort -r16 *.f90 -L$MKLPATH/lib/em64t  -lmkl_blas95_ilp64 -lmkl_lapack95_ilp64 -lmkl_intel_ilp64 -lmkl_sequential -lmkl_core -lpthread -warn all -check all -debug all
[/bash]

so quad precision.

Now concerning -i8

Consider the following program:

[bash]PROGRAM TestPrecisione
USE lapack95
INTEGER    ::i
REAL       ::r
INTEGER                 :: info
INTEGER, DIMENSION(3)   :: IPIV
REAL,    DIMENSION(3,3) :: INV
i=3
r=3
write(*,*) INT
CALL GETRI(INV,IPIV,info)
CALL GETRF(INV,IPIV,info)
ENDPROGRAM[/bash]

If I compile with

[bash]ifort -r8 *.f90 -L$MKLPATH/lib/em64t  -lmkl_blas95_ilp64 -lmkl_lapack95_ilp64 -lmkl_intel_ilp64 -lmkl_sequential -lmkl_core -lpthread -warn all -check all -debug all[/bash]

I get the errors:

[bash]There is no matching specific subroutine for this generic subroutine call.   [GETRI]
CALL GETRI(INV,IPIV,info)
-----^
error #6285: There is no matching specific subroutine for this generic subroutine call.   [GETRF]
CALL GETRF(INV,IPIV,info)
[/bash]

it seems to work only with -i8

[pbkenned1]

It only works with -i8 because you're linking against the 64-bit integer MKL libraries (the *ilp64 libraries), but you've declared standard 32-bit integers. You should either declare your integers as 64-bit (INTEGER(8) :: iFoo)), or link against the 32-bit MKL libraries (the *lp64 libraries). I chose the latter option

$ ifort -V

Intel Fortran Intel 64 Compiler XE for applications running on Intel 64, Version 12.1.1.256 Build 20111011

$ ifort U101663.f90 -L$MKLROOT/lib/intel64 -I$F95ROOT/include/intel64/lp64 -lmkl_lapack95_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -o U101663.x

$ ./U101663.x

3

$

P.S. -- I'm not sure it's safe to use -i8 with source declared 32-bit integers && link against the *ilp64 MKL libraries. I'll try to confirm.

Patrick Kennedy
Intel Developer Support

[TimP]

Source built with -i8, linking to MKL ilp64, is a model on which many customers rely. It requires use of the Intel64 compiler throughout and the 64-bit OS.

[diedro]

hi,

thank you.

I would like to compile with only -r8

I tryed with:

[bash]ifort -r8  *.f90 -L$MKLPATH -I$MKLROOT -lmkl_blas95_lp64 -lmkl_lapack95_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -warn all -check all -debug all -o prova.exe[/bash]

[bash]
[/bash]

However I get the same error

Is the link wrong?

Thank a lot

[TimP]

You have omitted the required -lmkl_intel_lp64.

[diedro]

Hi,

my fault:

I went to bash file and I change the path:

[bash]PATH="/opt/intel/composerxe-2011.0.084/mkl/bin/intel64$PATH"
source /opt/intel/composerxe-2011.0.084/mkl/bin/mklvars.sh intel64 mod lp64
export PATH[/bash]

Now it seems to work

Regarding the initial question. How can have quad precision (I need to invert a Square matrix) ?

Thanks a lot

[TimP]

MKL doesn't offer any quad precision libraries. You're free to compile the subroutines of your choice from open source code.

[mecej4]

> You're free to compile the subroutines of your choice from open source code.

True, but there is no guarantee that any Lapack algorithm when run with auto-promotion to quad-precision will deliver results with more precision than double-precision.

Performing iterative refinement using quad-precision and starting with double-precision results may suffice in some cases.