用 cython 声明一个 numpy 数组会奇怪地产生大量开销

Question

我正在使用 Cython 重写我的一些 Python 代码。

按照in the documentation 的建议，我开始用优化的 cython 定义替换我的 python 数组。

特别是，以下应该是声明 numpy 数组的“最佳”方式：

# cython: profile=True
# cython: boundscheck=False
# cython: wraparound=False

import numpy as np
cimport numpy as np

cpdef test():

    cdef np.ndarray[np.int_t, ndim=1] seeds_idx = np.empty(10, dtype=np.int)

    pass

但是，通过cython -a my_file.pyx 分析上述代码生成的 html 文件显示如下：

+10:     cdef np.ndarray[np.int_t, ndim=1] seeds_idx = np.empty(10, dtype=np.int)
  __pyx_t_1 = __Pyx_GetModuleGlobalName(__pyx_n_s_np); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_1);
  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_empty); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_2);
  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
  __pyx_t_1 = PyDict_New(); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_1);
  __pyx_t_3 = __Pyx_GetModuleGlobalName(__pyx_n_s_np); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_3);
  __pyx_t_4 = __Pyx_PyObject_GetAttrStr(__pyx_t_3, __pyx_n_s_int); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_4);
  __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
  if (PyDict_SetItem(__pyx_t_1, __pyx_n_s_dtype, __pyx_t_4) < 0) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
  __pyx_t_4 = __Pyx_PyObject_Call(__pyx_t_2, __pyx_tuple_, __pyx_t_1); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_4);
  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
  if (!(likely(((__pyx_t_4) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_4, __pyx_ptype_5numpy_ndarray))))) __PYX_ERR(0, 10, __pyx_L1_error)
  __pyx_t_5 = ((PyArrayObject *)__pyx_t_4);
  {
    __Pyx_BufFmt_StackElem __pyx_stack[1];
    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_seeds_idx.rcbuffer->pybuffer, (PyObject*)__pyx_t_5, &__Pyx_TypeInfo_nn___pyx_t_5numpy_int_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) {
      __pyx_v_seeds_idx = ((PyArrayObject *)Py_None); __Pyx_INCREF(Py_None); __pyx_pybuffernd_seeds_idx.rcbuffer->pybuffer.buf = NULL;
      __PYX_ERR(0, 10, __pyx_L1_error)
    } else {__pyx_pybuffernd_seeds_idx.diminfo[0].strides = __pyx_pybuffernd_seeds_idx.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_seeds_idx.diminfo[0].shape = __pyx_pybuffernd_seeds_idx.rcbuffer->pybuffer.shape[0];
    }
  }
  __pyx_t_5 = 0;
  __pyx_v_seeds_idx = ((PyArrayObject *)__pyx_t_4);
  __pyx_t_4 = 0;
/* … */
  __pyx_tuple_ = PyTuple_Pack(1, __pyx_int_10); if (unlikely(!__pyx_tuple_)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_tuple_);
  __Pyx_GIVEREF(__pyx_tuple_);

这是在 Python 2.7 上使用 cython 0.24 和 numpy 1.10.4 获得的。

另一方面，非常简单的声明 seeds_idx = np.empty(10) 导致：

+10:     seeds_idx = np.empty(10)
  __pyx_t_1 = __Pyx_GetModuleGlobalName(__pyx_n_s_np); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_1);
  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_empty); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_2);
  __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
  __pyx_t_1 = __Pyx_PyObject_Call(__pyx_t_2, __pyx_tuple_, NULL); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_t_1);
  __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
  __pyx_v_seeds_idx = __pyx_t_1;
  __pyx_t_1 = 0;
/* … */
  __pyx_tuple_ = PyTuple_Pack(1, __pyx_int_10); if (unlikely(!__pyx_tuple_)) __PYX_ERR(0, 10, __pyx_L1_error)
  __Pyx_GOTREF(__pyx_tuple_);
  __Pyx_GIVEREF(__pyx_tuple_);

这里出了什么问题（如果有的话）？谢谢！

Answer 1

正如评论所述，这里没有任何问题，因此无需担心。另外，请记住，您正在检查为简单分配生成的代码，任何差异都不会影响性能。

虽然有一个小勘误，但在第二种情况下，seeds_idx = np.empty(10) 应更改为seeds_idx = np.empty(10, dtype=np.int) 以匹配第一种情况。

如果添加它，那么为存储函数调用 (np.empty) 的参数而创建的字典也会被添加：

__pyx_t_1 = PyDict_New(); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 8, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_1);

查找np.int:

__pyx_t_3 = __Pyx_GetModuleGlobalName(__pyx_n_s_np); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 10, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_3);
__pyx_t_4 = __Pyx_PyObject_GetAttrStr(__pyx_t_3, __pyx_n_s_int); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 10, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_4);
__Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;

新创建的字典中的参数设置就完成了：

if (PyDict_SetItem(__pyx_t_1, __pyx_n_s_dtype, __pyx_t_4) < 0) __PYX_ERR(0, 8, __pyx_L1_error)
__Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;

除此之外，它们之间的唯一区别如下：

if (!(likely(((__pyx_t_4) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_4, __pyx_ptype_5numpy_ndarray))))) __PYX_ERR(0, 10, __pyx_L1_error)
  __pyx_t_5 = ((PyArrayObject *)__pyx_t_4);
  {
    __Pyx_BufFmt_StackElem __pyx_stack[1];
    if (unlikely(__Pyx_GetBufferAndValidate(&__pyx_pybuffernd_seeds_idx.rcbuffer->pybuffer, (PyObject*)__pyx_t_5, &__Pyx_TypeInfo_nn___pyx_t_5numpy_int_t, PyBUF_FORMAT| PyBUF_STRIDES, 1, 0, __pyx_stack) == -1)) {
      __pyx_v_seeds_idx = ((PyArrayObject *)Py_None); __Pyx_INCREF(Py_None); __pyx_pybuffernd_seeds_idx.rcbuffer->pybuffer.buf = NULL;
      __PYX_ERR(0, 10, __pyx_L1_error)
    } else {__pyx_pybuffernd_seeds_idx.diminfo[0].strides = __pyx_pybuffernd_seeds_idx.rcbuffer->pybuffer.strides[0]; __pyx_pybuffernd_seeds_idx.diminfo[0].shape = __pyx_pybuffernd_seeds_idx.rcbuffer->pybuffer.shape[0];
    }
  }

其中，as stated in the documentation you linked 很可能是为了快速访问数据缓冲区而执行的。

到目前为止，最好的选择是使用typed memoryviews。这些是本机方式，并且很可能是在 cython 中使用数组的最简单方法。 Their performance is usually on par with numpy arrays 如果没有，您可以随时在它们之间轻松切换。