Fast Python?

https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/python3-gcc.html

Where’s the problem?

def add_two_numbers(x, y):
    return x + y


z = add_two_numbers(123, 456)
print(z)

Howto solve it?

• AOT compiler
  • Cython
• JIT compiler
  • Numba

Cython

• Superset of Python
• Compiled language
  • in fact it is transpiller to C
  • .pyx -> .c -> .so -> launch.py
• Explicit data types are optional
• nogil
• Calling native functions

Compile to C

cdef add_two_numbers(x, y):
    return x + y


z = add_two_numbers(123, 456)
print(z)

Explicit parameter types

cdef add_two_numbers(int x, int y):
    return x + y


z = add_two_numbers(123, 456)
print(z)

Explicit return type

cdef int add_two_numbers(int x, int y):
    return x + y


z = add_two_numbers(123, 456)
print(z)

Disable GIL-related locks

cdef int add_two_numbers(int x, int y) nogil:
    return x + y


z = add_two_numbers(123, 456)
print(z)

Use C standard function

from libc.stdio cimport printf


cdef int add_two_numbers(int x, int y) nogil:
    printf("%i\n", x)
    return x + y


z = add_two_numbers(123, 456)
print(z)

Numba

• JIT for Python

Decorator

from numba import jit

@jit
def funkce1():
    pass

Simpler and faster print

• Only numbers and strings
• no file or sep argument

Force JIT

@jit(nopython=True)

Some comparisons

• ANSI C: ANSI C variant (no Python)
• Cython #1: basic variant
• Cython #2: type hints
• Cython #3: optimizations + nogil
• Numba #1: original file
• Numba #2: @jit annotation
• Numba #3: native print function
• Numba #4: native print function + @jit(nopython=True)