Question

我通过添加一些类型并编译它来转换为cython python函数。我在python和cython函数的结果之间得到了很小的数值差异。经过一些工作后，我发现差异来自于使用unsigned int而不是int来访问numpy数组。

我使用unsigned int索引来加速访问：http://docs.cython.org/src/userguide/numpy_tutorial.html#tuning-indexing-further

无论如何，我认为使用无符号的整数是无害的。

请参阅此代码：

cpdef function(np.ndarray[np.float32_t, ndim=2] response, max_loc):
    cdef unsigned int x, y   
    x, y = int(max_loc[0]), int(max_loc[1])
    x2, y2 = int(max_loc[0]), int(max_loc[1])
    print response[y,x], type(response[y,x]), response.dtype
    print response[y2,x2], type(response[y2,x2]), response.dtype   
    print 2*(response[y,x] - min(response[y,x-1], response[y,x+1]))
    print 2*(response[y2,x2] - min(response[y2,x2-1], response[y2,x2+1]))

打印：

0.959878861904 <type 'float'> float32
0.959879 <type 'numpy.float32'> float32
1.04306024313
1.04306030273

为什么会这样？!!!这是一个错误吗？

好的，这里要求的是一个SSCCE，它与我在原始函数中使用的类型和值相同

cpdef function():
    cdef unsigned int x, y  
    max_loc2 = np.asarray([ 15., 25.], dtype=float) 
    cdef np.ndarray[np.float32_t, ndim=2] response2 = np.zeros((49,49), dtype=np.float32)    
    x, y = int(max_loc2[0]), int(max_loc2[1])
    x2, y2 = int(max_loc2[0]), int(max_loc2[1])

    response2[y,x] = 0.959878861904  
    response2[y,x-1] = 0.438348740339
    response2[y,x+1] = 0.753262758255  


    print response2[y,x], type(response2[y,x]), response2.dtype
    print response2[y2,x2], type(response2[y2,x2]), response2.dtype
    print 2*(response2[y,x] - min(response2[y,x-1], response2[y,x+1]))
    print 2*(response2[y2,x2] - min(response2[y2,x2-1], response2[y2,x2+1]))

打印

0.959878861904 <type 'float'> float32
0.959879 <type 'numpy.float32'> float32
1.04306024313
1.04306030273

我使用python 2.7.3 cython 0.18和msvc9 express

Answer 1

我修改了问题中的示例，以便更简单地读取模块生成的C源代码。我只对查看创建Python float对象的逻辑感兴趣，而不是从np.float32数组中获取response个对象。

我使用pyximport编译扩展模块。它将生成的C文件保存在~/.pyxbld的子目录中（在Windows上可能为%userprofile%\.pyxbld）。

import numpy as np
import pyximport
pyximport.install(setup_args={'include_dirs': [np.get_include()]})

open('_tmp.pyx', 'w').write('''
cimport numpy as np
cpdef function(np.ndarray[np.float32_t, ndim=2] response, max_loc):
    cdef unsigned int p_one, q_one
    p_one = int(max_loc[0])
    q_one = int(max_loc[1])
    p_two = int(max_loc[0])
    q_two = int(max_loc[1])
    r_one = response[q_one, p_one]
    r_two = response[q_two, p_two]
''')

import _tmp
assert(hasattr(_tmp, 'function'))

这里是为感兴趣的部分生成的C代码（稍微重新格式化以使其更易于阅读）。事实证明，当您使用C unsigned int索引变量时，生成的代码直接从数组缓冲区中获取数据并调用PyFloat_FromDouble，这会将其强制转换为double。另一方面，当您使用Python int索引变量时，它采用通用方法。它形成一个元组并调用PyObject_GetItem。这种方式允许ndarray正确兑现np.float32 dtype。

#define __Pyx_BufPtrStrided2d(type, buf, i0, s0, i1, s1) \
    (type)((char*)buf + i0 * s0 + i1 * s1)

  /* "_tmp.pyx":9
 *     p_two = int(max_loc[0])
 *     q_two = int(max_loc[1])
 *     r_one = response[q_one, p_one]             # <<<<<<<<<<<<<<
 *     r_two = response[q_two, p_two]
 */
  __pyx_t_3 = __pyx_v_q_one;
  __pyx_t_4 = __pyx_v_p_one;
  __pyx_t_5 = -1;

  if (unlikely(__pyx_t_3 >= (size_t)__pyx_bshape_0_response))
    __pyx_t_5 = 0;
  if (unlikely(__pyx_t_4 >= (size_t)__pyx_bshape_1_response))
    __pyx_t_5 = 1;

  if (unlikely(__pyx_t_5 != -1)) {
    __Pyx_RaiseBufferIndexError(__pyx_t_5);
    {
      __pyx_filename = __pyx_f[0];
      __pyx_lineno = 9;
      __pyx_clineno = __LINE__;
      goto __pyx_L1_error;
    }
  }

  __pyx_t_1 = PyFloat_FromDouble((
    *__Pyx_BufPtrStrided2d(
      __pyx_t_5numpy_float32_t *,
      __pyx_bstruct_response.buf,
      __pyx_t_3, __pyx_bstride_0_response,
      __pyx_t_4, __pyx_bstride_1_response)));

  if (unlikely(!__pyx_t_1)) {
    __pyx_filename = __pyx_f[0];
    __pyx_lineno = 9;
    __pyx_clineno = __LINE__;
    goto __pyx_L1_error;
  }

  __Pyx_GOTREF(__pyx_t_1);
  __pyx_v_r_one = __pyx_t_1;
  __pyx_t_1 = 0;

  /* "_tmp.pyx":10
 *     q_two = int(max_loc[1])
 *     r_one = response[q_one, p_one]
 *     r_two = response[q_two, p_two]             # <<<<<<<<<<<<<<
 */
  __pyx_t_1 = PyTuple_New(2);

  if (unlikely(!__pyx_t_1)) {
    __pyx_filename = __pyx_f[0];
    __pyx_lineno = 10;
    __pyx_clineno = __LINE__;
    goto __pyx_L1_error;
  }

  __Pyx_GOTREF(((PyObject *)__pyx_t_1));
  __Pyx_INCREF(__pyx_v_q_two);
  PyTuple_SET_ITEM(__pyx_t_1, 0, __pyx_v_q_two);
  __Pyx_GIVEREF(__pyx_v_q_two);
  __Pyx_INCREF(__pyx_v_p_two);
  PyTuple_SET_ITEM(__pyx_t_1, 1, __pyx_v_p_two);
  __Pyx_GIVEREF(__pyx_v_p_two);

  __pyx_t_2 = PyObject_GetItem(
    ((PyObject *)__pyx_v_response),
    ((PyObject *)__pyx_t_1));

  if (!__pyx_t_2) {
    __pyx_filename = __pyx_f[0];
    __pyx_lineno = 10;
    __pyx_clineno = __LINE__;
    goto __pyx_L1_error;
  }

  __Pyx_GOTREF(__pyx_t_2);
  __Pyx_DECREF(((PyObject *)__pyx_t_1));
  __pyx_t_1 = 0;
  __pyx_v_r_two = __pyx_t_2;
  __pyx_t_2 = 0;

Answer 2

在我的机器上玩这个，我看不出有什么区别。我正在使用带有cython magic的ipython笔记本：

In [1]:

%load_ext cythonmagic

In [12]:

%%cython

import numpy as np
cimport numpy as np

cpdef function(np.ndarray[np.float32_t, ndim=2] response, max_loc):
    cdef unsigned int x, y   
    x, y = int(max_loc[0]), int(max_loc[1])
    x2, y2 = int(max_loc[0]), int(max_loc[1])
    #return 2*(response[y,x] - min(response[y,x-1], response[y,x+1])), 2*(response[y2,x2] - min(response[y2,x2-1], response[y2,x2+1]))
    print response[y,x], type(response[y,x]), response.dtype
    print response[y2,x2], type(response[y2,x2]), response.dtype   
    print 2*(response[y,x] - min(response[y,x-1], response[y,x+1]))
    print 2*(response[y2,x2] - min(response[y2,x2-1], response[y2,x2+1])) 

In [13]:

a = np.random.normal(size=(10,10)).astype(np.float32)
m = [3,2]
function(a,m)

0.586090564728 <type 'float'> float32
0.586091 <type 'numpy.float32'> float32
4.39655685425
4.39655685425

第一对结果，区别仅在于print语句的输出精度。您使用的是什么版本的Cython？索引极不可能影响答案，因为它只是访问numpy数组的数据属性存储的固定长度的内存。

Cython：numpy数组的unsigned int索引给出了不同的结果

2 个答案: