Home > Article > Technology peripherals > This compiler can make Python as fast as C++: up to a hundred times faster, produced by MIT
Since the rise of deep learning, Python has been one of the hottest programming languages. It dominates the fields of data science and machine learning, and even plays a starring role in scientific and mathematical computing. . Nowadays, you can find a corresponding Python package for almost any project you can imagine.
However, while the simplified syntax of a high-level language makes it easy to learn and use, it is slower than a low-level language like C or C.
Researchers at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) hope to change that with Codon, a Python-based compiler that allows users Write Python code that runs as efficiently as C or C programs, while being customizable and adaptable to different needs and environments.
The latest paper of this research, "Codon: A Compiler for High-Performance Pythonic Applications and DSLs", was published at the 32nd ACM SIGPLAN International Conference on Compiler Construction in February.
"Regular Python is compiled into so-called bytecode, which is executed in a virtual machine, which makes it much slower," said Codon, the lead author of the paper Ariya Shajii said, "With Codon, we compile natively, so you can run the final result directly on the CPU - without going through an intermediate virtual machine or interpreter."
Codon's compilation pipeline includes type checking, allowing it to run Python code more efficiently.
The Python-based compiler comes with pre-built binaries for Linux and macOS, and you can also build or generate executables from source. "With Codon, you can distribute source code like Python, or you can compile it into binaries," Shajii said. "If you want to distribute a binary, it will be the same as a language like C, such as a Linux binary or a Mac binary."To make Codon faster, research People decided to perform type checking at compile time. Type checking involves assigning a data type (such as integer, string, character, or float, etc.) to a value. For example, the number 5 can be assigned as an integer, the letter c can be assigned as a character, the word hello can be assigned as a string, and the decimal number 3.14 can be assigned as a floating point number.
"In regular Python, all types are given to the runtime," Shajii said. "Using Codon, we do type checking during compilation, which allows us to avoid all expensive type manipulation at runtime."
Saman Amarasinghe, principal researcher at MIT CSAIL, added, " If you have a dynamic language (like Python), every time you have some data, you need to keep a lot of extra metadata around it to determine the type of runtime. Codon does away with this metadata, so the code is faster Faster and smaller data size."
According to Shajii, Codon does not have any unnecessary data or type checking at runtime, so there is zero overhead. In terms of performance, "Codon is generally on par with C. We typically see 10x to 100x speed improvements compared to Python."
On the other hand, Codon The approach has its trade-offs. "We do this static type checking and don't allow the use of some of Python's dynamic features, such as dynamically changing types at runtime," Shajii said.
"There are still some Python libraries that we haven't implemented yet." Amarasinghe added, "Python has been actually tested by countless people, and Codon has not reached that level yet. It needs to run more programs to obtain More feedback, and more reinforcement. It will take some time to reach a stable level of regular Python." Codon was originally designed for genomics and bioinformatics work. The researchers tried about 10 common genomics applications written in Python and compiled them using Codon, achieving speedups of 5 to 10 times compared to the original manually optimized implementation. "Data sets in these fields have become very large today, and high-level languages like Python and R are too slow to handle the terabytes of data per sequencing set," Shajii said. "That's the gap we want to fill - by building a way to process big data without having to write C or C code to help domain experts who are not computer science or professional developers." The above chart compares the performance of Python (CPython 3), PyPy, Codon, and C on several benchmarks. The y-axis shows the speedup of the Codon implementation relative to the CPython implementation. MIT/EXALOOP/UNIVERSITY OF VICTORIA/ACM In addition to genomics, Codon can also be applied to similar applications that handle massive data sets, as well as Python-based compilation GPU programming and parallel programming supported by the processor. In fact, Codon is now being used commercially in bioinformatics, deep learning, and quantitative finance through the startup Exaloop, which Shajii founded to transform Codon from an academic project into an industry application. To enable Codon to adapt to different fields, the team developed a plug-in system. "It's like an extensible compiler," Shajii said. "You can write plug-ins for genomics or other fields, and these plug-ins can have new libraries and new compiler optimizations." In addition, companies and institutions can use Codon to prototype and develop your own applications. "One of the patterns we see is that people use Python for prototyping and testing because it's easy to use, but when it comes to something important, they have to rewrite the application or have someone else do it in C or C Rewrite and test on a larger data set," Shajii said. "With Codon, you can fully use Python and get the best of both worlds." Regarding the future of Codon, Shajii and his team are currently working on native versions of the widely used Python library. implementation, as well as library-specific optimizations to help people get better performance from these libraries. They also plan to create a popular feature: Codon's WebAssembly backend to support running code on a web browser. 
The above is the detailed content of This compiler can make Python as fast as C++: up to a hundred times faster, produced by MIT. For more information, please follow other related articles on the PHP Chinese website!