I am trying to reduce a particularly computational heavy piece of Python code through compiling as C with Cython. Surprisingly I have virtually no success in speeding it up. With the original Python module the code runs in about 45sec and with the Cython compiled module it runs in about 45secs.
Compiling with annotate=True, I am left with a sea of yellow(ish) lines suggesting there is still a huge amount of Python interaction.
I have responded to this by switching off both boundscheck (False) and cdivision (True). This had no effect. The following snippet of code is an extract from this module and it insists on interacting with Python. Why is this? Nothing here need to interact with any pre-existing Python module, it is all very simple arithmatic?
cpdef float __distance_between(point1, point2) except? -2:
return ((point1[0] - point2[0]) ** 2 + (point1[1] - point2[1]) ** 2) ** (1/2)
point1 and point2 are Python lists each containing 2 doubles eg: [611811.997, -871083.372] The resulting C-code that is generated by Cython for the return line in the snippet is as follows:
+05: cpdef float __distance_between(point1, point2) except? -2:
+06: return ((point1[0] - point2[0]) ** 2 + (point1[1] - point2[1]) ** 2)**(1/2)
__pyx_t_1 = __Pyx_GetItemInt(__pyx_v_point1, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_1);
__pyx_t_2 = __Pyx_GetItemInt(__pyx_v_point2, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 0); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_2);
__pyx_t_3 = PyNumber_Subtract(__pyx_t_1, __pyx_t_2); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_3);
__Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
__Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
__pyx_t_2 = PyNumber_Power(__pyx_t_3, __pyx_int_2, Py_None); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_2);
__Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
__pyx_t_3 = __Pyx_GetItemInt(__pyx_v_point1, 1, long, 1, __Pyx_PyInt_From_long, 0, 0, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_3);
__pyx_t_1 = __Pyx_GetItemInt(__pyx_v_point2, 1, long, 1, __Pyx_PyInt_From_long, 0, 0, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_1);
__pyx_t_4 = PyNumber_Subtract(__pyx_t_3, __pyx_t_1); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_4);
__Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
__Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
__pyx_t_1 = PyNumber_Power(__pyx_t_4, __pyx_int_2, Py_None); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_1);
__Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
__pyx_t_4 = PyNumber_Add(__pyx_t_2, __pyx_t_1); if (unlikely(!__pyx_t_4)) __PYX_ERR(0, 6, __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;
__pyx_t_1 = __Pyx_PyInt_From_long((1 / 2)); if (unlikely(!__pyx_t_1)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_1);
__pyx_t_2 = PyNumber_Power(__pyx_t_4, __pyx_t_1, Py_None); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_GOTREF(__pyx_t_2);
__Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
__Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
__pyx_t_5 = __pyx_PyFloat_AsFloat(__pyx_t_2); if (unlikely((__pyx_t_5 == (float)-1) && PyErr_Occurred())) __PYX_ERR(0, 6, __pyx_L1_error)
__Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
__pyx_r = __pyx_t_5;
goto __pyx_L0;
How can I better optimise the code or the compilation to make this snippet independent of Python?
There are many other lines in this module that are all rendered pretty yellow, but starting with this question may provide me with a better understanding of how to tackle the rest.