Over the past years, Numba and Cython have gained a lot of attention in the data science community. They both provide a way to speed up CPU intensive tasks, but in different ways. This article describes architectural differences between them.
Numba is a just-in-time (JIT) compiler which translates Python code to native machine instructions both for CPU and GPU. The code can be compiled at import time, runtime, or ahead of time.
It's extremely easy to start using Numba, by simply putting a
import numpy as np from numba import jit a = np.arange(1, 10 ** 7) b = np.arange(-10 ** 7, -1) @jit(nopython=True) def sum_sequence(a, b): result = np.zeros_like(a) for i in range(len(a)): result[i] = a[i] - b[i] return result
>>> fast_sum_sequence = jit(int64[:](int64[:], int64[:]), nopython=True)(sum_sequence) >>> timeit.timeit('sum_sequence(a, b)', globals=globals(), number=1) Basic Python version: 4.227093786990736 >>> timeit.timeit('fast_sum_sequence(a, b)', globals=globals(), number=1) Numba version: 0.05048697197344154
As you may know, In Python, all code blocks are compiled down to bytecode:
>>> import dis >>> dis.dis(sum_sequence) 2 0 LOAD_GLOBAL 0 (np) 2 LOAD_ATTR 1 (zeros_like) 4 LOAD_FAST 0 (a) 6 CALL_FUNCTION 1 8 STORE_FAST 2 (result) 3 10 SETUP_LOOP 40 (to 52) 12 LOAD_GLOBAL 2 (range) 14 LOAD_GLOBAL 3 (len) 16 LOAD_FAST 0 (a) 18 CALL_FUNCTION 1 20 CALL_FUNCTION 1 22 GET_ITER >> 24 FOR_ITER 24 (to 50) 26 STORE_FAST 3 (i) 4 28 LOAD_FAST 0 (a) 30 LOAD_FAST 3 (i) 32 BINARY_SUBSCR 34 LOAD_FAST 1 (b) 36 LOAD_FAST 3 (i) 38 BINARY_SUBSCR 40 BINARY_SUBTRACT 42 LOAD_FAST 2 (result) 44 LOAD_FAST 3 (i) 46 STORE_SUBSCR 48 JUMP_ABSOLUTE 24 >> 50 POP_BLOCK 5 >> 52 LOAD_FAST 2 (result) 54 RETURN_VALUE
To optimize Python code, Numba takes a bytecode from a provided function and runs a set of analyzers on it. This process involves many stages, but as a result, Numba converts Python bytecode to LLVM intermediate representation (IR).
Note that LLVM IR is a low-level programming language, which is similar to assembler syntax and has nothing to do with Python.
The are two modes in Numba:
object. The former doesn't use Python runtime and produces native code without Python dependencies. The native code is statically typed and runs very fast. Whereas the object mode uses Python objects and Python C API, which often does not give significant speed improvements. In both cases, Python code is compiled using LLVM.
What is LLVM?
LLVM is a compiler, that takes a special intermediate representation (IR) of the code and compiles it down to native (machine) code. The process of compiling involves a lot of additional "passes" in which the compiler optimizes IR. LLVM toolchain is very good at optimizing IR, so not only it compiles code for Numba, but also optimizes it.
The whole system roughly looks as follows:
Advantages of Numba:
- Ease of use
- Automatic parallelization
- Support for numpy operations and objects
- GPU support
Disadvantages of Numba:
- Many layers of abstraction make it very hard to debug and optimize
- There is no way to interact with Python and its modules in
- Limited support for classes
Instead of analyzing bytecode and generating IR, Cython uses a superset of Python syntax which later translates to C code. When working with Cython, you basically writing C code with high-level Python syntax.
When working with Cython, you usually don't have to worry about wrappers and low-level API calls, because all interactions are automatically expanded to a proper C code.
Unlike Numba, all Cython code should be separated from regular Python code in special files. Cython parses and translates such files to C code and then compiles it using provided C compiler (e.g.
Python code is already valid Cython code.
def sum_sequence_cython(a, b): result = np.zeros_like(a) for i in range(len(a)): result[i] = a[i] - b[i] return result
However, typed version works a lot faster.
cimport numpy as np cpdef sum_sequence_cython(np.ndarray[np.int64_t, ndim=1] a, np.ndarray[np.int64_t, ndim=1] b): cdef int N = a.shape cdef np.ndarray[np.int64_t, ndim=1] result = np.zeros([N], dtype=np.int) for i in range(N): result[i] = a[i] - b[i] return result
>>> timeit.timeit('sum_sequence_untyped(a, b)', globals=globals(), number=1) Untyped Cython version 2.0215444170171395 >>> timeit.timeit('sum_sequence_cython(a, b)', globals=globals(), number=1) Typed Cython version 0.046073039297712967
Writing fast Cython code requires an understanding of C and Python internals. If you know C, your Cython code can run as fast as the C code.
Advantages of Cython:
- Control over Python API usage
- Easy interfacing with C/C++ libraries and C/C++ code
- Parallel execution support
- Support for Python classes, which gives object-oriented features in C
Disadvantages of Cython:
- Learning curve
- Requires expertise both in C and Python internals
- Inconvenient organization of modules