在gfortran mac中运行openMP时出现非法指令错误

Question

I am new to openMP so bear with me. I use a macbook pro with OX Mavericks and installed a version of gfortran compatible with my OS.

I implemented a very simple dynamic programming problem to solve something that economist call the Neoclassical Growth Model.

I find no problem when I run it without openMP, however when trying to compile the program with -fopenmp option I get either

Illegal Instruction: 4 or Segmentation Fault: 11

... probably I am doing something very wrong.

I attach the main program, subroutines, modules and .sh compilation file

PROGRAM prg_dp1

    ! PRG_DP1 SOLVES THE GROWTH MODEL BY VALUE FUNCTION ITERATION OVER A DISCRETE GRID
    ! WITHOUT INTERPOLATION. EVALUATION IS DONE OVER NEXT PERIOD CAPITAL

    ! PROBLEM: PROGRAMMED AS MATRICES IT LIMITS THE SIZE OF THE PROGRAM BEFORE !SEGMENTATION FAULTS OCCUR

    USE modvar

    IMPLICIT NONE
    REAL(DP), DIMENSION(nk) :: v,v0,kp,c
    REAL(DP), DIMENSION(nk,nk) :: cm,um,vm
    REAL(DP) :: kstar,tbegin,tend
    INTEGER :: it,ik1,ik2,ind(nk)

    ! INVOCATION OF THE OUTPUT FILES WHERE THE INFORMATION IS GOING TO BE WRITTEN
    CALL CPU_TIME(tbegin)


    ! DEFINITION OF THE PARAMETERS OF THE MODEL

    p(1)=1.0001     ! Intertemporal elasticity of substitution (SIGMA)
    p(2)=0.96       ! Intertemporal discount factor (BETA)
    p(3)=0.06       ! Depreciation rate (DELTA)
    p(4)=0.36       ! Share of capital in production (ALPHA)
    p(5)=0.00       ! (Parameter not needed)

    ! COMPUTATION OF THE STEADY STATE CAPITAL STOCK

    kstar=((1.0/p(2)-(1.0-p(3)))/p(4))**(1.0/(p(4)-1.0))

    ! FIRST I ALLOCATE AND CONSTRUCT THE GRID

    slope=1.0
    gkmin=0.0001
    gkmax=5.0*kstar
  !  ALLOCATE(gk(nk),ones(nk,nk))

    ALLOCATE(gk(nk))
!   ones=1.0
    CALL sub_grid_generation(gk,nk,slope,gkmin,gkmax)

    ! DEFINITION OF THE MATRICES OF CONSUMPTION AND UTILITY

    !$OMP PARALLEL  DEFAULT(SHARED) PRIVATE(ik1,ik2)
    !$OMP DO SCHEDULE(DYNAMIC)
    DO ik1=1,nk
        DO ik2=1,nk
            cm(ik1,ik2)=gk(ik1)**p(4)+(1.0-p(3))*gk(ik1)-gk(ik2)
        END DO
    END DO
    !$OMP END DO
    !$OMP END PARALLEL

   ! cm = gk**p(4)+(1.0-p(3))*gk-gk*ones


    WHERE (cm .le. 0.0)
        um=-1.0e+6
    ELSEWHERE
        um=(cm**(1.0-p(1))-1.0)/(1.0-p(1))
    END WHERE

    ! DINAMIC PROGRAMMING STEP

    ! I first initialize the value function to zeros

    v0=0.0

    ! Main do has to be done by master-thread ... can I parallelize more?

    DO
        !$OMP PARALLEL DO PRIVATE(ik2)
        DO ik2=1,nk
            vm(:,ik2)=um(:,ik2)+p(2)*v0(ik2)
        END DO
        !$OMP END PARALLEL DO
        v=MAXVAL(vm,DIM=2)
        print *, MAXVAL(ABS(v-v0))
        IF (MAXVAL(ABS(v-v0)) .le. dp_tol) THEN
            EXIT
        ELSE
            v0=v
        END IF
    END DO

    ind=MAXLOC(v,DIM=1)

    kp=gk(ind)
    c=gk**p(4)+(1.0-p(3))*gk-kp
    open(unit=1,file='output.txt')
    DO ik1=1,nk
        write(1,'(4F10.5)') gk(ik1),v(ik1),kp(ik1),c(ik1)
    END DO
    close(1)
    DEALLOCATE(gk)

    CALL CPU_TIME(tend)

    PRINT *, tend-tbegin

END PROGRAM prg_dp1

SUBROUTINE sub_grid_generation(grid,gsize,slope,gridmin,gridmax)
    USE nrtype
INTEGER, INTENT(IN) :: gsize
    REAL(DP), INTENT(IN) :: slope,gridmin,gridmax
REAL(DP), INTENT(OUT) :: grid(gsize)
    INTEGER :: ig   
grid(1)=gridmin
    DO ig=2,gsize
    grid(ig)=gridmin+((gridmax-gridmin)/dfloat(gsize)**slope)*dfloat(ig)**slope
    END DO

END SUBROUTINE sub_grid_generation

MODULE nrtype
    INTEGER, PARAMETER :: I4B = SELECTED_INT_KIND(9)
    INTEGER, PARAMETER :: I2B = SELECTED_INT_KIND(4)
    INTEGER, PARAMETER :: I1B = SELECTED_INT_KIND(2)
    INTEGER, PARAMETER :: SP = KIND(1.0)
    INTEGER, PARAMETER :: DP = KIND(1.0D0)
    INTEGER, PARAMETER :: SPC = KIND((1.0,1.0))
    INTEGER, PARAMETER :: DPC = KIND((1.0D0,1.0D0))
    INTEGER, PARAMETER :: LGT = KIND(.true.)
    REAL(SP), PARAMETER :: PI=3.141592653589793238462643383279502884197_sp
    REAL(SP), PARAMETER :: PIO2=1.57079632679489661923132169163975144209858_sp
    REAL(SP), PARAMETER :: TWOPI=6.283185307179586476925286766559005768394_sp
    REAL(SP), PARAMETER :: SQRT2=1.41421356237309504880168872420969807856967_sp
    REAL(SP), PARAMETER :: EULER=0.5772156649015328606065120900824024310422_sp
    REAL(DP), PARAMETER :: PI_D=3.141592653589793238462643383279502884197_dp
    REAL(DP), PARAMETER :: PIO2_D=1.57079632679489661923132169163975144209858_dp
    REAL(DP), PARAMETER :: TWOPI_D=6.283185307179586476925286766559005768394_dp
    REAL(DP), PARAMETER :: gr=(5.0**0.5-1.0)/2.0
    TYPE sprs2_sp
        INTEGER(I4B) :: n,len
        REAL(SP), DIMENSION(:), POINTER :: val
        INTEGER(I4B), DIMENSION(:), POINTER :: irow
        INTEGER(I4B), DIMENSION(:), POINTER :: jcol
    END TYPE sprs2_sp
    TYPE sprs2_dp
        INTEGER(I4B) :: n,len
        REAL(DP), DIMENSION(:), POINTER :: val
        INTEGER(I4B), DIMENSION(:), POINTER :: irow
        INTEGER(I4B), DIMENSION(:), POINTER :: jcol
    END TYPE sprs2_dp
END MODULE nrtype

MODULE modvar
    USE nrtype
    IMPLICIT NONE
    REAL(DP), PARAMETER :: r_tol=1e-8
    REAL(DP), PARAMETER :: p_tol=1e-6

    REAL(DP), PARAMETER :: dp_tol=1e-6
    REAL(DP), PARAMETER :: c_tol=0.001
    REAL(DP), PARAMETER :: adj=0.5

    INTEGER, PARAMETER :: r_m=10000

    ! PARAMETER THAT DEFINE THE DIMENSION OF THE PROBLEM

    INTEGER, PARAMETER :: nk=5000
    INTEGER, PARAMETER :: nz=21
    INTEGER, PARAMETER :: np=20000
    INTEGER, PARAMETER :: nt=5000

    INTEGER, PARAMETER :: maxit=10000

    INTEGER, PARAMETER :: dist_maxit=5000

    ! POLICY PARAMETER, COMMON ENDOGENOUS VARIABLES AND OTHER ALLOCATABLE ARRAYS

    REAL(DP), PARAMETER :: nw=0.0
    REAL(DP), PARAMETER :: ft=0.33
    REAL(DP) :: p(5),gkmin,gkmax,slope
    REAL(DP), ALLOCATABLE :: gk(:),gz(:),m(:,:),mss(:),ones(:,:)
END MODULE modvar

and the .sh file I use to compile

export OMP_NUM_THREADS=8
gfortran -O2 -fopenmp -c nrtype.f90 modvar.f90 sub_grid_generation.f90 prg_dp1.f90

gfortran -O2 -fopenmp -fcheck=bounds -o myprog nrtype.o modvar.o sub_grid_generation.o prg_dp1.o

---

I know this is tedious but I would appreciate some help

Thank you

Answer 1

The other options is to make the global arrays cm,um,vm , and possibly also the smaller other ones, allocatable. This will become handy when you change the problem size, read it from somewhere and maintain the executable.

REAL(DP), DIMENSION(:,:),allocatable :: cm,um,vm

allocate(cm(nk,nk),um(nk,nk),vm(nk,nk))

Answer 2

It is a stack space issue. I tried running it with ifort and even without openmp I get illegal instruction and I had to specify -heap-arrays in order to get it to run properly. Once I added openmp the illegal instruction error came back. The WHERE statement seems to be the problem code. In both the openmp and non-openmp runs that is the part that causes it to fail

OS X stack space is rather limited and you are creating large arrays. Using -heap-arrays helps, but once you use openmp that is no longer a possibility and ulimit is maxed out as ~64 MB.

I found adding this to your compilation works:

-Wl,-stack_size,0x40000000,-stack_addr,0xf0000000

Which increases the stack size to 1GB. This could probably be fine tuned, but I tried using 256 MB and it was still not enough.