小贝子编程

Julia回调函数中的Fortran数组



我正在尝试为twpvpc(带有remmesh的ODE BVP) Fortran-77求解器编写包装器。求解器需要签名为

的输入函数
subroutine fsub(ncomp, x, u, f, rpar, ipar)

,

  • ncomp为整数(向量的长度)，
  • x (in)为浮点数，
  • u (in)为长度为ncomp的矢量，
  • f (out)是结果的位置，长度为ncomp的矢量
  • rpar和ipar是浮点型和整型外部参数数组;Julia的关闭是更可取的方式，但显然有困难(见这里的博客文章)。但暂时可以忽略它们。

在Julia中，我通常会使用签名

来写fsub
function fsub_julia(x :: Float64, y :: Vector{Float64}, dy :: Vector{Float64})
        dy[1] = ...
        dy[2] = ...
        ...
end

ncomp似乎不是必要的，因为可以通过length或size获得长度(然而，Julia可以检测到从Fortran传递的数组的大小吗?对于测试代码，我明确地知道ncomp，所以现在它不是问题)。

所以，为了符合twpvpc格式，我写了一个包装器: 
function fsub_par(n :: Int64, x :: Float64, y :: Vector{Float64}, dy :: Vector{Float64}, rpar :: Vector{Float64}, ipar :: Vector{Float64})
        fsub_julia(x, y, dy)
end

现在，要将这个函数传递给Fortran例程，需要使用cfunction进行转换以声明类型。问题是怎么做到的?

如果我把它写成:

cf_fsub = cfunction(fsub_par, Void, (Ref{Int64}, Ref{Float64}, Ref{Float64}, Ref{Float64}, Ref{Float64}, Ref{Int64}))
当从Fortran调用

时，我得到错误:

ERROR: LoadError: MethodError: no method matching (::TWPBVP.#fsub_par#1{TWPBVP_Test.#f})(::Int64, ::Float64, ::Float64, ::Float64, ::Float64, ::Int64)
Closest candidates are:
  fsub_par(::Int64, ::Float64, !Matched::Array{Float64,1}, !Matched::Array{Float64,1}, !Matched::Array{Float64,1}, !Matched::Array{Float64,1})

所以，不知何故，签名与数组参数不匹配…

如果我用Ref{Array{Float64,1}}替换Ref{Float64}作为数组参数(它看起来有点奇怪，虽然…):

cf_fsub = cfunction(fsub_par, Void, (Ref{Int64}, Ref{Float64}, Ref{Array{Float64,1}}, Ref{Array{Float64,1}}, Ref{Array{Float64,1}}, Ref{Array{Int64,1}}))

在Fortran代码中调用fsub_par (cf_fsub)时，我得到一个分段错误(这大约位于，因为错误没有给出确切的位置)。

将Ref{Float54}替换为Ptr{Float64}也不会做任何事情。

我在Fortran代码中发现了一件有趣的事情，那就是fsub是如何被调用的: 
call fsub (ncomp, xx(1), u(1,1), fval(1,1),rpar,ipar)

其中u和fval声明为:

dimension xx(nmsh), u(nudim,nmsh), fval(ncomp,nmsh)

所以，我猜，它使用了Fortran通过引用传递所有参数的事实，对u(1,1)的引用应该是指向矩阵第一列的指针(据我所知，Fortran作为Julia，以列优先顺序存储矩阵)。

出路是什么?我是否需要更改fsub_julia的签名以接受指针并手动将它们转换为数组(这就是ODEInterface。Jl在较低级别工作)?

更新

遵循在ODEInterface中如何完成。结合在void* -thunk参数中传递Julia函数的想法，我想到了这个:

immutable TWPBVPCProblem
    fsub :: Function            # RHS function
    dfsub :: Function           # Jacobian of RHS
    gsub :: Function            # BC function
    dgsub :: Function           # gradients of BC function
end
function unsafe_fsub(rn :: Ref{Int64}, rx :: Ref{Float64}, py :: Ptr{Float64}, pdy :: Ptr{Float64}, rpar :: Ptr{Float64}, ipar :: Ptr{Int64}) :: Void
    x = rx[]
    n = rn[]
    y = unsafe_wrap(Array, py, n)
    dy = unsafe_wrap(Array, pdy, n)
    problem = unsafe_pointer_to_objref(rpar) :: TWPBVPCProblem
    problem.fsub(x, y, dy)
    return nothing
end
const fsub_ptr = cfunction(unsafe_fsub, Void, (Ref{Int64}, Ref{Float64}, Ptr{Float64}, Ptr{Float64}, Ptr{Float64}, Ptr{Int64}))

当我调用求解器时(它相当长):

function twpbvpc(nlbc :: Int64,
        aleft :: Float64, aright :: Float64,
        fixpnt :: Nullable{Vector{Float64}},
        ltol :: Vector{Int64}, tol :: Vector{Float64},
        linear :: Bool, givmsh :: Bool, giveu :: Bool, nmsh :: Ref{Int64},
        xx :: Vector{Float64}, u :: Array{Float64, 2}, nmax :: Ref{Int64},
        wrk :: Vector{Float64}, iwrk :: Vector{Int64},
        fsub :: Function, dfsub :: Function,
        gsub :: Function, dgsub :: Function,
        ckappa1 :: Ref{Float64}, gamma1 :: Ref{Float64},
        ckappa :: Ref{Float64},
        # rpar :: Vector{Float64},
        # ipar :: Vector{Int64},
        iflbvp :: Ref{Int64})
    # Keep problem functions in rpar
    # HACK!
    rpar = TWPBVPCProblem(fsub, dfsub, gsub, dgsub)
    # Fake external parameters
    # Can't have it 0-length as it would be Any[0] and not Float64[0]
    # local rpar :: Vector{Float64} = [0.0]
    local ipar :: Vector{Int64} = [0]
    # No need to pass these parameters
    # u is a matrix for the solution only!
    ncomp, nucol = size(u)
    # Get the maximum of xx
    nxxdim = length(xx)
    # max for mesh points must be the same as the number of column points of u
    assert(nucol == nxxdim)
    # Sizes of work arrays
    lwrkfl = length(wrk)
    lwrkin = length(iwrk)
    # Number of fixed mesh points
    if isnull(fixpnt)
        nfxpnt = 0
        fixpnt_v = [0.0]
    else
        fixpnt_v = get(fixpnt)
        nfxpnt = length(fixpnt_v)
    end
    # Size of tolerance vector ≤ ncomp
    ntol = length(ltol)
    ccall((:twpbvpc_, libtwpbvpc), Void,
        (Ref{Int64}, Ref{Int64},                    # ncomp, nlbc,
        Ref{Float64}, Ref{Float64},                 # aleft, aright
        Ref{Int64}, Ptr{Float64},                   # nfxpnt, fixpnt
        Ref{Int64}, Ptr{Int64}, Ptr{Float64},       # ntol, ltol, tol
        Ref{Int64}, Ref{Int64}, Ref{Int64},         # linear, givmsh, giveu
        Ref{Int64}, Ref{Int64},                     # nmsh, nxxdim
        Ptr{Float64}, Ref{Int64},                   # xx, nudim
        Ptr{Float64}, Ref{Int64},                   # u, nmax
        Ref{Int64}, Ptr{Float64},                   # lwrkfl, wrk
        Ref{Int64}, Ptr{Int64},                     # lwrkin, iwrk
        Ptr{Void}, Ptr{Void}, Ptr{Void}, Ptr{Void}, # fsub, dfsub, gsub, dgsub
        Ref{Float64}, Ref{Float64},                 # ckappa1, gamma1
        Ref{Float64}, Any, Ptr{Int64},              # ckappa, rpar, ipar
        Ref{Int64}),                                # iflbvp
        ncomp, nlbc, aleft, aright,
        nfxpnt, fixpnt_v, ntol, ltol, tol,
        linear, givmsh, giveu, nmsh,
        nxxdim, xx, nucol, u, nmax,                 # nudim = nucol
        lwrkfl, wrk, lwrkin, iwrk,
        fsub_ptr, dfsub_ptr, gsub_ptr, dgsub_ptr,
        ckappa1,gamma1,ckappa,pointer_from_objref(rpar),ipar,iflbvp)
end

Fortran的twpbvpc看起来是这样的(很明显是开头):

  subroutine twpbvpc(ncomp, nlbc, aleft, aright,
 *       nfxpnt, fixpnt, ntol, ltol, tol,
 *       linear, givmsh, giveu, nmsh,
 *       nxxdim, xx, nudim, u, nmax,
 *       lwrkfl, wrk, lwrkin, iwrk,
 *       fsub, dfsub, gsub, dgsub,
 *       ckappa1,gamma1,ckappa,rpar,ipar,iflbvp)
  implicit double precision (a-h,o-z)
  dimension rpar(*),ipar(*)
  dimension fixpnt(*), ltol(*), tol(*)
  dimension xx(*), u(nudim,*)
  dimension wrk(lwrkfl), iwrk(lwrkin)
  logical linear, givmsh, giveu
  external fsub, dfsub, gsub, dgsub
  logical pdebug, use_c, comp_c
  common/algprs/ nminit, pdebug, iprint, idum, uval0, use_c, comp_c
  ...

用build.jl编译Fortran代码:

cd(joinpath(Pkg.dir("TWPBVP"), "deps"))
pic = @windows ? "" : "-fPIC"
run(`gfortran -m$WORD_SIZE -fdefault-real-8 -fdefault-integer-8 -ffixed-form $pic -shared -O3 -o libtwpbvpc.so twpbvpc.f`)

所以，我将rpar传递为Any(应该相当于Ptr{Void}): Fortran期望一个浮点数数组，但这并不重要。

现在，当我尝试运行一个简单的程序(在Pkg.test("TWPBVP")上)时，我得到了分段错误:

signal (11): Segmentation fault
while loading /home/alexey/.julia/v0.5/TWPBVP/test/runtests.jl, in expression starting on line 58
unknown function (ip: 0xffffffffffffffff)
Allocations: 1400208 (Pool: 1399373; Big: 835); GC: 0

由于代码变得非常长，这里是github上完整代码的链接:https://github.com/mobius-eng/TWPBVP.jl

我是否需要更改fsub_julia的签名以接受指针并手动将它们转换为数组(这就是ODEInterface如何实现的)?Jl在较低级别工作)?

是的，ODEInterface。Jl模型看起来是一个很好的模型。

您需要找出的第一件事是fortran INTEGER类型的大小(Int32或Int64)。对于下面的代码，我将借用ODEInterface。jl并使用FInt(它可以是类型参数，也可以是typealias)

结果回退应该看起来像:

# SUBROUTINE FSUB(NCOMP,X,Z,F,RPAR,IPAR)
# IMPLICIT NONE
# INTEGER NCOMP, IPAR
# DOUBLE PRECISION F, Z, RPAR, X
# DIMENSION Z(*),F(*)
# DIMENSION RPAR(*), IPAR(*)
function unsafe_fsub(ncomp::Ref{FInt}, x::Ref{Float64}, z::Ptr{Float64}, 
        f::Ptr{Float64}, rpar::Ptr{Float64}, ipar::Ptr{FInt})::Void
    xx = x[]
    zz = unsafe_wrap(Array, z, ncomp[])
    ff = unsafe_wrap(Array, f, ncomp[])
    fsub!(xx, zz, ff) # function which updates array ff
    return nothing
end
const fsub_ptr = cfunction(unsafe_fsub, Void,
    (Ref{FInt},Ref{Float64},Ptr{Float64},Ptr{Float64},Ptr{Float64},Ptr{FInt}))

相关内容

最新更新


热门标签：

javascript python java c# php android html jquery c++ css ios sql mysql arrays asp.net json python-3.x ruby-on-rails .net sql-server django objective-c excel regex ruby linux ajax iphone xml vba spring asp.net-mvc database wordpress string postgresql wpf windows xcode bash git oracle list vb.net multithreading eclipse algorithm macos powershell visual-studio image forms numpy scala function api selenium