Cython equivalent of range-based for?

Question

How can I get cython to generate something like this C++ loop?

int get_sum(const std::vector<MyObject>& my_vect) {
  int sum=0;
  for(const auto& my_obj : my_vect) {
    sum += my_obj.value();
  }
  return sum;
}

When I try writing this in Cython using for my_obj in my_vect the code it generates creates an default-constructed MyObject, then iterates over my_vect and does a copy-assignment to that object.

How can I convince Cython to use references in the for loop instead of doing copies?

The optimizer shouldn't matter. Calling an object's operator= is semantically different from working with a reference to the object. — Thomas Johnson, Aug 21 '19 at 17:36
It matters a lot if OP is looking at a debug build. An optimized build may well perform orders of magnitude better. — Jesper Juhl, Aug 21 '19 at 17:39
It's not just a question of performance though. Suppose MyObject's operator= has side-effects (e.g., logging). That side-effect would be triggered in the generated Cython code, but not in the range-based for loop that uses references. The generated Cython code that makes a copy and the range-based for actually do different things. — Thomas Johnson, Aug 21 '19 at 17:41
Realistically you can't. Cython uses values rather than references to match Python scoping rules (so `my_obj` is accessible after the loop). If you need to generate specific C++ code then you're much better just writing the code in C++. — DavidW, Aug 22 '19 at 08:53
Can you provide a Cython example? Or, can you show where it creates a default constructed `MyObject`? I seem to get a normal behavior using typed memory views. — Arda Aytekin, Aug 23 '19 at 10:04

score 0 · Answer 1 · answered Aug 23 '19 at 10:36

This solution seems to provide the behavior you want:

# dummy.pyx
# cython: language_level = 3
cimport cython

cdef class MyObject:
  cdef int value_

  def __init__(self, value = 0):
    print("Constructor called.")
    self.value_ = value

  cpdef int value(self):
    return self.value_

@cython.boundscheck(False)
cpdef int get_sum_1(MyObject[::1] my_vect):
  cdef int sum_ = 0
  cdef Py_ssize_t len = my_vect.shape[0]

  for i in range(len):
    sum_ += (<MyObject>my_vect[i]).value()

  return sum_

cpdef int get_sum_2(MyObject[::1] my_vect):
  cdef int sum_ = 0
  # cdef MyObject obj

  for obj in my_vect:
    sum_ += obj.value()

  return sum_

get_sum_2, above, uses for obj in my_vect style (i.e., iterators). However, because the type of obj is not known in advance and iterators are used (the range of iteration is not known a priori), there is a lot of Python overhead. If you hint Cython regarding the type of obj in get_sum_2 by uncommenting the cdef MyObject obj part, you get around 30-40% speedup (see below for an application code).

You can get two orders of magnitude speedup on top of get_sum_2 by using a plain old for-loop instead of the iterators. Now, because we know how many elements the typed memory view has in advance, we can also turn off bounds checking.

The application code I have tried is as follows:

# app.py
from numpy import array, median
from timeit import repeat

import pyximport
pyximport.install()

from dummy import MyObject
from dummy import get_sum_1
from dummy import get_sum_2


my_vect = array([MyObject(i) for i in range(50000)])

get_1 = repeat("get_sum_1(my_vect)", repeat=100, number=1, globals=globals())
get_2 = repeat("get_sum_2(my_vect)", repeat=100, number=1, globals=globals())

print(f"Median of get_1: {1000*median(get_1)} ms.")
print(f"Median of get_2: {1000*median(get_2)} ms.")

Running python app.py, I receive 50,000 "Constructor called." print statements, followed by the performance measures on my laptop:

Median of get_1: 0.20261999452486634 ms.
Median of get_2: 11.251458498009015 ms.

If you run cython --annotate dummy.pyx, you should see the overheads clearly. In both of the examples, however, I do not see anything but struct MyObject * in the generated C code. This is further supported by the lack of print statements of the constructor in the application code inside the respective function calls.

My reading of the question is that OP is trying to use Cython's C++ standard library bindings. Python objects are all handled by pointer anyway so don't end up being copied unnecessary — DavidW, Aug 23 '19 at 15:32
I am sorry, then, @DavidW. But is it not then similar to, if not the same as, [this question](https://stackoverflow.com/q/21720982/4720025). — Arda Aytekin, Aug 23 '19 at 15:37
Not really. The issue is that Cython requires local variables (but not function parameters) to be default constructable. Therefore you can pass by c++ reference to a Cython function, but a loop variable as a reference doesn't work. — DavidW, Aug 23 '19 at 16:20

Cython equivalent of range-based for?

1 Answers1