我的代码是cython化的问题,更具体地说是以下(和类似的)代码片段:
cdef double [:,:] grad_d_him_d_jm
grad_d_ihm_d_jm = grad_d_im_d_jm(...)
其中grad_d_im_d_jm(...)将返回double [:,:] memoryview。 此代码将由Cython转换为以下C代码:
__pyx_t_1 = __pyx_f_24gradient_better_c_mviews_grad_d_im_d_jm(__pyx_v_i, __pyx_v_j, __pyx_v_m, __pyx_v_structure, __pyx_v_distances);
if (unlikely(!__pyx_t_1)) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 203; __pyx_clineno = __LINE__; goto __pyx_L1_error;}
__Pyx_GOTREF(__pyx_t_1);
__pyx_t_7 = __Pyx_PyObject_to_MemoryviewSlice_dsds_double(__pyx_t_1);
if (unlikely(!__pyx_t_7.memview)) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 203; __pyx_clineno = __LINE__; goto __pyx_L1_error;}
__Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
__pyx_v_grad_d_ihm_d_jm = __pyx_t_7;
__pyx_t_7.memview = NULL;
__pyx_t_7.data = NULL;
当我在循环中执行此操作时,我怀疑Python API调用会对我的代码速度产生相当大的影响。
在其他场合也会发生GOTREF / DECREF调用,以及PyFloat_asFloat:
cdef float sp
sp = scalar_product()
其中scalar_product()返回一个cdef float。此代码段已转换为
__pyx_t_1 = __pyx_f_24gradient_better_c_mviews_scalar_product(__pyx_v_i, __pyx_v_j, __pyx_v_m, __pyx_v_structure);
if (unlikely(!__pyx_t_1)) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 178; __pyx_clineno = __LINE__; goto __pyx_L1_error;}
__Pyx_GOTREF(__pyx_t_1);
__pyx_t_2 = __pyx_PyFloat_AsFloat(__pyx_t_1);
if (unlikely((__pyx_t_2 == (float)-1) && PyErr_Occurred())) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 178; __pyx_clineno = __LINE__; goto __pyx_L1_error;}
__Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
__pyx_v_sp = __pyx_t_2;
我正在运行Python 2.7.11+和Cython 0.23.4。如果您能告诉我a)这与性能无关或b)如何解决,我将非常感激。 如果我能改进这个问题,请告诉我,我很乐意这样做。
答案 0 :(得分:1)
这些似乎是Cython引用计数API的一部分,解释为Maven。
我的猜测是grad_d_im_d_jm
返回一个Python对象(例如NumPy数组),因此Cython必须在获得内存视图后递减对象引用计数器。
对于scalar_product
,我认为它是def(而不是cdef)或者是无类型的。例如以下
cdef g():
return 1.0
编译到
// ...
__Pyx_XDECREF(__pyx_r);
__Pyx_INCREF(__pyx_float_1_0);
__pyx_r = __pyx_float_1_0;
goto __pyx_L0;
但是,一旦指定了返回类型,引用计数就会消失
cdef float g():
return 1.0
变为
// ...
__pyx_r = 1.0;
goto __pyx_L0;