9.6 使用Python CFFI混合C，C++，Fortran和Python

NOTE:此示例代码可以在 https://github.com/dev-cafe/cmake-cookbook/tree/v1.0/chapter-9/recipe-06 中找到，其中有一个C++示例和一个Fortran示例。该示例在CMake 3.11版(或更高版本)中是有效的，并且已经在GNU/Linux、macOS和Windows上进行过测试。

前面的三个示例中，我们使用Cython、Boost.Python和pybind11作为连接Python和C++的工具。之前的示例中，主要连接的是C++接口。然而，可能会遇到这样的情况：将Python与Fortran或其他语言进行接口。

本示例中，我们将使用Python C的外部函数接口(CFFI，参见https://cffi.readthedocs.io)。由于C是通用语言，大多数编程语言(包括Fortran)都能够与C接口进行通信，所以Python CFFI是将Python与大量语言结合在一起的工具。Python CFFI的特性是，生成简单且非侵入性的C接口，这意味着它既不限制语言特性中的Python层，也不会对C层以下的代码有任何限制。

本示例中，将使用前面示例的银行帐户示例，通过C接口将Python CFFI应用于Python和C++。我们的目标是实现一个上下文感知的接口。接口中，我们可以实例化几个银行帐户，每个帐户都带有其内部状态。我们将通过讨论如何使用Python CFFI来连接Python和Fortran来结束本教程。

第11章第3节中，通过PyPI分发一个用CMake/CFFI构建的C/Fortran/Python项目，届时我们将重新讨论这个例子，并展示如何打包它，使它可以用pip安装。

准备工作

我们从C++实现和接口开始，把它们放在名为account/implementation的子目录中。实现文件(cpp_implementation.cpp)类似于之前的示例，但是包含有断言，因为我们将对象的状态保持在一个不透明的句柄中，所以必须确保对象在访问时已经创建：

#include "cpp_implementation.hpp"

#include <cassert>

Account::Account()
{
  balance = 0.0;
  is_initialized = true;
}
Account::~Account()
{
  assert(is_initialized);
  is_initialized = false;
}
void Account::deposit(const double amount)
{
  assert(is_initialized);
  balance += amount;
}
void Account::withdraw(const double amount)
{
  assert(is_initialized);
  balance -= amount;
}
double Account::get_balance() const
{
  assert(is_initialized);
  return balance;
}

接口文件(cpp_implementation.hpp)包含如下内容：

#pragma once

class Account
{
public:
  Account();
  ~Account();
  void deposit(const double amount);
  void withdraw(const double amount);
  double get_balance() const;

private:
  double balance;
  bool is_initialized;
};

此外，我们隔离了C-C++接口(c_cpp_interface.cpp)。这将是我们与Python CFFI连接的接口：

#include "account.h"
#include "cpp_implementation.hpp"

#define AS_TYPE(Type, Obj) reinterpret_cast<Type *>(Obj)
#define AS_CTYPE(Type, Obj) reinterpret_cast<const Type *>(Obj)

account_context_t *account_new()
{
  return AS_TYPE(account_context_t, new Account());
}
void account_free(account_context_t *context) { delete AS_TYPE(Account, context); }
void account_deposit(account_context_t *context, const double amount)
{
  return AS_TYPE(Account, context)->deposit(amount);
}
void account_withdraw(account_context_t *context, const double amount)
{
  return AS_TYPE(Account, context)->withdraw(amount);
}
double account_get_balance(const account_context_t *context)
{
  return AS_CTYPE(Account, context)->get_balance();
}

account目录下，我们声明了C接口(account.h):

#ifndef ACCOUNT_API
#include "account_export.h"
#define ACCOUNT_API ACCOUNT_EXPORT
#endif
#ifdef __cplusplus
extern "C"
{
#endif
  struct account_context;
  typedef struct account_context account_context_t;
  ACCOUNT_API
  account_context_t *account_new();
  ACCOUNT_API
  void account_free(account_context_t *context);
  ACCOUNT_API
  void account_deposit(account_context_t *context, const double amount);
  ACCOUNT_API
  void account_withdraw(account_context_t *context, const double amount);
  ACCOUNT_API
  double account_get_balance(const account_context_t *context);
#ifdef __cplusplus
}
#endif
#endif /* ACCOUNT_H_INCLUDED */

我们还描述了Python接口，将在稍后对此进行讨论(__init_ _.py):

from subprocess import check_output
from cffi import FFI
import os
import sys
from configparser import ConfigParser
from pathlib import Path

def get_lib_handle(definitions, header_file, library_file):
  ffi = FFI()
  command = ['cc', '-E'] + definitions + [header_file]
  interface = check_output(command).decode('utf-8')
  # remove possible \r characters on windows which
  # would confuse cdef
  _interface = [l.strip('\r') for l in interface.split('\n')]
  ffi.cdef('\n'.join(_interface))
  lib = ffi.dlopen(library_file)
  return lib

# this interface requires the header file and library file
# and these can be either provided by interface_file_names.cfg
# in the same path as this file
# or if this is not found then using environment variables
_this_path = Path(os.path.dirname(os.path.realpath(__file__)))
_cfg_file = _this_path / 'interface_file_names.cfg'
if _cfg_file.exists():
  config = ConfigParser()
  config.read(_cfg_file)
  header_file_name = config.get('configuration', 'header_file_name')
  _header_file = _this_path / 'include' / header_file_name
  _header_file = str(_header_file)
  library_file_name = config.get('configuration', 'library_file_name')
  _library_file = _this_path / 'lib' / library_file_name
  _library_file = str(_library_file)
else:
  _header_file = os.getenv('ACCOUNT_HEADER_FILE')
  assert _header_file is not None
  _library_file = os.getenv('ACCOUNT_LIBRARY_FILE')
  assert _library_file is not None

_lib = get_lib_handle(definitions=['-DACCOUNT_API=', '-DACCOUNT_NOINCLUDE'],
                      header_file=_header_file,
                      library_file=_library_file)
# we change names to obtain a more pythonic API
new = _lib.account_new
free = _lib.account_free
deposit = _lib.account_deposit
withdraw = _lib.account_withdraw
get_balance = _lib.account_get_balance

__all__ = [
    '__version__',
    'new',
    'free',
    'deposit',
    'withdraw',
    'get_balance',
]

我们看到，这个接口的大部分工作是通用的和可重用的，实际的接口相当薄。

项目的布局为：

.
├── account
│    ├── account.h
│    ├── CMakeLists.txt
│    ├── implementation
│    │    ├── c_cpp_interface.cpp
│    │    ├── cpp_implementation.cpp
│    │    └── cpp_implementation.hpp
│    ├── __init__.py
│    └── test.py
└── CMakeLists.txt

具体实施

现在使用CMake来组合这些文件，形成一个Python模块:

1. 主CMakeLists.txt文件包含一个头文件。此外，根据GNU标准，设置编译库的位置:

   # define minimum cmake version
   cmake_minimum_required(VERSION 3.5 FATAL_ERROR)
   
   # project name and supported language
   project(recipe-06 LANGUAGES CXX)
   
   # require C++11
   set(CMAKE_CXX_STANDARD 11)
   set(CMAKE_CXX_EXTENSIONS OFF)
   set(CMAKE_CXX_STANDARD_REQUIRED ON)
   
   # specify where to place libraries
   include(GNUInstallDirs)
   set(CMAKE_LIBRARY_OUTPUT_DIRECTORY
   ${CMAKE_BINARY_DIR}/${CMAKE_INSTALL_LIBDIR})

2. 第二步，是在account子目录下包含接口和实现的定义：

   # interface and sources
   add_subdirectory(account)

3. 主CMakeLists.txt文件以测试定义(需要Python解释器)结束：

   # turn on testing
   enable_testing()
   
   # require python
   find_package(PythonInterp REQUIRED)
   
   # define test
   add_test(
     NAME
       python_test
     COMMAND
       ${CMAKE_COMMAND} -E env ACCOUNT_MODULE_PATH=${CMAKE_CURRENT_SOURCE_DIR}
                           ACCOUNT_HEADER_FILE=${CMAKE_CURRENT_SOURCE_DIR}/account/account.h
                           ACCOUNT_LIBRARY_FILE=$<TARGET_FILE:account>
      ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/account/test.py
   )

4. account/CMakeLists.txt中定义了动态库目标：

   add_library(account
     SHARED
       plementation/c_cpp_interface.cpp
       implementation/cpp_implementation.cpp
     )
   
   target_include_directories(account
     PRIVATE
       ${CMAKE_CURRENT_SOURCE_DIR}
       ${CMAKE_CURRENT_BINARY_DIR}
     )

5. 导出一个可移植的头文件:

   include(GenerateExportHeader)
   generate_export_header(account
     BASE_NAME account
     )

6. 使用Python-C接口进行对接:

   $ mkdir -p build
   $ cd build
   $ cmake ..
   $ cmake --build .
   $ ctest
   
   Start 1: python_test
   1/1 Test #1: python_test ...................... Passed 0.14 sec
   100% tests passed, 0 tests failed out of 1

工作原理

虽然，之前的示例要求我们显式地声明Python-C接口，并将Python名称映射到C(++)符号，但Python CFFI从C头文件(示例中是account.h)推断出这种映射。我们只需要向Python CFFI层提供描述C接口的头文件和包含符号的动态库。在主CMakeLists.txt文件中使用了环境变量集来实现这一点，这些环境变量可以在__init__.py中找到:

# ...
def get_lib_handle(definitions, header_file, library_file):
  ffi = FFI()
  command = ['cc', '-E'] + definitions + [header_file]
  interface = check_output(command).decode('utf-8')

  # remove possible \r characters on windows which
  # would confuse cdef
  _interface = [l.strip('\r') for l in interface.split('\n')]

  ffi.cdef('\n'.join(_interface))
  lib = ffi.dlopen(library_file)
  return lib

# ...

_this_path = Path(os.path.dirname(os.path.realpath(__file__)))
_cfg_file = _this_path / 'interface_file_names.cfg'
if _cfg_file.exists():
    # we will discuss this section in chapter 11, recipe 3
else:
  _header_file = os.getenv('ACCOUNT_HEADER_FILE')
  assert _header_file is not None
  _library_file = os.getenv('ACCOUNT_LIBRARY_FILE')
  assert _library_file is not None
  _lib = get_lib_handle(definitions=['-DACCOUNT_API=', '-DACCOUNT_NOINCLUDE'],
  header_file=_header_file,
  library_file=_library_file)
# ...

get_lib_handle函数打开头文件(使用ffi.cdef)并解析加载库(使用ffi.dlopen)。并返回库对象。前面的文件是通用的，可以在不进行修改的情况下重用，用于与Python和C或使用Python CFFI的其他语言进行接口的其他项目。

_lib库对象可以直接导出，这里有一个额外的步骤，使Python接口在使用时，感觉更像Python:

# we change names to obtain a more pythonic API
new = _lib.account_new
free = _lib.account_free
deposit = _lib.account_deposit
withdraw = _lib.account_withdraw
get_balance = _lib.account_get_balance

__all__ = [
  '__version__',
  'new',
  'free',
  'deposit',
  'withdraw',
  'get_balance',
]

有了这个变化，可以将例子写成下面的方式:

import account
account1 = account.new()
account.deposit(account1, 100.0)

另一种选择则不那么直观:

from account import lib
account1 = lib.account_new()
lib.account_deposit(account1, 100.0)

需要注意的是，如何使用API来实例化和跟踪上下文:

account1 = account.new()
account.deposit(account1, 10.0)

account2 = account.new()
account.withdraw(account1, 5.0)
account.deposit(account2, 5.0)

为了导入account的Python模块，需要提供ACCOUNT_HEADER_FILE和ACCOUNT_LIBRARY_FILE环境变量，就像测试中那样：

add_test(
  NAME
    python_test
  COMMAND
    ${CMAKE_COMMAND} -E env ACCOUNT_MODULE_PATH=${CMAKE_CURRENT_SOURCE_DIR}
                        ACCOUNT_HEADER_FILE=${CMAKE_CURRENT_SOURCE_DIR}/account/account.h
                        ACCOUNT_LIBRARY_FILE=$<TARGET_FILE:account>
    ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/account/test.py
  )

第11章中，将讨论如何创建一个可以用pip安装的Python包，其中头文件和库文件将安装在定义良好的位置，这样就不必定义任何使用Python模块的环境变量。

讨论了Python方面的接口之后，现在看下C的接口。account.h内容为：

struct account_context;
typedef struct account_context account_context_t;

ACCOUNT_API
account_context_t *account_new();

ACCOUNT_API
void account_free(account_context_t *context);

ACCOUNT_API
void account_deposit(account_context_t *context, const double amount);

ACCOUNT_API
void account_withdraw(account_context_t *context, const double amount);

ACCOUNT_API
double account_get_balance(const account_context_t *context);

黑盒句柄account_context会保存对象的状态。ACCOUNT_API定义在account_export.h中，由account/interface/CMakeLists.txt生成：

include(GenerateExportHeader)
generate_export_header(account
  BASE_NAME account
  )

account_export.h头文件定义了接口函数的可见性，并确保这是以一种可移植的方式完成的，实现可以在cpp_implementation.cpp中找到。它包含is_initialized布尔变量，可以检查这个布尔值确保API函数按照预期的顺序调用：上下文在创建之前或释放之后都不应该被访问。

更多信息

设计Python-C接口时，必须仔细考虑在哪一端分配数组：数组可以在Python端分配并传递给C(++)实现，也可以在返回指针的C(++)实现上分配。后一种方法适用于缓冲区大小事先未知的情况。但返回到分配给C(++)端的数组指针可能会有问题，因为这可能导致Python垃圾收集导致内存泄漏，而Python垃圾收集不会“查看”分配给它的数组。我们建议设计C API，使数组可以在外部分配并传递给C实现。然后，可以在__init__.py中分配这些数组，如下例所示：

from cffi import FFI
import numpy as np

_ffi = FFI()

def return_array(context, array_len):
  # create numpy array
  array_np = np.zeros(array_len, dtype=np.float64)

  # cast a pointer to its data
  array_p = _ffi.cast("double *", array_np.ctypes.data)

  # pass the pointer
  _lib.mylib_myfunction(context, array_len, array_p)

  # return the array as a list
  return array_np.tolist()

return_array函数返回一个Python列表。因为在Python端完成了所有的分配工作，所以不必担心内存泄漏，可以将清理工作留给垃圾收集。

对于Fortran示例，读者可以参考以下Git库:https://github.com/dev-cafe/cmake-cookbook/tree/v1.0/chapter09/recipe06/Fortran-example 。与C++实现的主要区别在于，account库是由Fortran 90源文件编译而成的，我们在account/CMakeLists.txt中使用了Fortran 90源文件:

add_library(account
  SHARED
      implementation/fortran_implementation.f90
  )

上下文保存在用户定义的类型中:

type :: account
  private
  real(c_double) :: balance
  logical :: is_initialized = .false.
end type

Fortran实现可以使用iso_c_binding模块解析account.h中定义的符号和方法:

module account_implementation

  use, intrinsic :: iso_c_binding, only: c_double, c_ptr

  implicit none

  private

  public account_new
  public account_free
  public account_deposit
  public account_withdraw
  public account_get_balance

  type :: account
    private
    real(c_double) :: balance
    logical :: is_initialized = .false.
  end type

contains

  type(c_ptr) function account_new() bind (c)
    use, intrinsic :: iso_c_binding, only: c_loc
    type(account), pointer :: f_context
    type(c_ptr) :: context

    allocate(f_context)
    context = c_loc(f_context)
    account_new = context
    f_context%balance = 0.0d0
    f_context%is_initialized = .true.
  end function

  subroutine account_free(context) bind (c)
    use, intrinsic :: iso_c_binding, only: c_f_pointer
    type(c_ptr), value :: context
    type(account), pointer :: f_context

    call c_f_pointer(context, f_context)
    call check_valid_context(f_context)
    f_context%balance = 0.0d0
    f_context%is_initialized = .false.
    deallocate(f_context)
  end subroutine

  subroutine check_valid_context(f_context)
    type(account), pointer, intent(in) :: f_context
    if (.not. associated(f_context)) then
        print *, 'ERROR: context is not associated'
        stop 1
    end if
    if (.not. f_context%is_initialized) then
        print *, 'ERROR: context is not initialized'
        stop 1
    end if
  end subroutine

  subroutine account_withdraw(context, amount) bind (c)
  use, intrinsic :: iso_c_binding, only: c_f_pointer
  type(c_ptr), value :: context
  real(c_double), value :: amount
  type(account), pointer :: f_context

  call c_f_pointer(context, f_context)
  call check_valid_context(f_context)
  f_context%balance = f_context%balance - amount
  end subroutine

  subroutine account_deposit(context, amount) bind (c)
    use, intrinsic :: iso_c_binding, only: c_f_pointer
    type(c_ptr), value :: context
    real(c_double), value :: amount
    type(account), pointer :: f_context

    call c_f_pointer(context, f_context)
    call check_valid_context(f_context)
    f_context%balance = f_context%balance + amount
  end subroutine

  real(c_double) function account_get_balance(context) bind (c)
    use, intrinsic :: iso_c_binding, only: c_f_pointer
    type(c_ptr), value, intent(in) :: context
    type(account), pointer :: f_context

    call c_f_pointer(context, f_context)
    call check_valid_context(f_context)
    account_get_balance = f_context%balance
  end function
end module

这个示例和解决方案的灵感来自Armin Ronacher的帖子“Beautiful Native Libraries”： http://lucumr.pocoo.org/2013/8/18/beautiful-native-libraries/