
NumbaSummarizer: A Python Library for Simplified Vectorization Reports Neftali Watkinson∗, Preston Taiy, Alexandru Nicolauz and Alexander Veidenbaumx Department of Computer Science University of California, Irvine Irvine, United States of America ∗[email protected], [email protected], [email protected], [email protected] Abstract—Python is a very popular programming language integration relatively easy. Because of this, Python lacks and has been adopted by many learning institutions as the first capabilities to take advantage of modern architectural features. or only programming language taught to students. While its The inclusion of multiple cores on a single chip, adoption of simple syntax and gentle learning curve makes it a very attractive option for new programmers, among its main drawbacks are Single Instruction Multiple Data (SIMD) or Vector instructions that it obviates many key concepts of programming and com- that exploit low level parallelism, and the pervasiveness of puter architecture as well as rarely optimizing code to modern Graphical Processing Units (GPU), categorically guarantee architectures. that any serious programmer will eventually interact with some There are tools that provide optimizing capability for Python level of parallel computing. This introduces a shift on how programmers. Numba is a JIT compiler for Python that among other things, optimizes Python and Numpy functions for better we learn and teach programming. As recognized by Prasad performance. Numba can exploit automatic parallelism and et al. [15] and Rague [16], instead of focusing on serial pro- vectorization. gramming, parallel programming should be introduced early In this paper we present NumbaSummarizer, a wrapper for in Computer Science curricula. Additionally, many experts Numba’s vectorization module that works as a vectorization in parallel computing agree that dependence analysis for the reporting tool. Students can use it to identify functions that are exploiting SIMD instructions. The main purpose of this tool is purpose of exploiting vectorization is a useful way to introduce to teach students about Parallel Computing, Vectorization, Data students to parallelism [20, 2, 14, 8]. Dependence, and Loop Transformations. Using Python to teach parallel computing allows for stu- We evaluated the use of NumbaSummarizer with 64 students dents to easily understand the tools and necessary coding during a course on intermediate programming. Most students skills for manual optimization of their code without the had little to no background in Computer Architecture nor Parallel Computing. When asked to expose parallelism in a set need to learn complex syntax [5, 21]. There are tools that of 8 loops, 95% percent of the students were able to transform allow programmers exploit high performance optimizations and optimize at least 6 of these loops. When testing conceptual while still keeping the practicality of Python. The work of understanding, they scored an average of 95% for parallel Marowka [11], reviews some of the most popular tools and computing concepts, and 62% for dependence analysis. workarounds to Python’s drawbacks, noting that Numba [9] is Index Terms—SIMD, Numba, Python, Vectorization a good option for exploiting optimizations. Numba is a Just- In-Time (JIT) compiler that uses LLVM [10] as a back end to I. INTRODUCTION compile entire functions. This allows for code optimizations Python is a popular language thanks to its high abstraction and use of automatic parallelism. Thanks to Numba’s simple and simple syntax. Many universities are adopting Python interface that uses decorators, it can be easily adopted in the as their main programming language to introduce students classroom. One of the major drawbacks that Numba has, is to programming and computer science. For example, the that it functions as a ”black box”. Programmers annotate the University of California, Irvine (UCI), requires that students functions they want to optimize but there’s little to no user take three quarters of Python programming. While C, C++ and friendly feedback on what’s been optimized and how they can Java are also offered, these are electives and required for some exploit more parallelism or vectorization. upper-division courses. However, there are relevant drawbacks In this paper, we present a Python library named Numba- in using a programming language that obviates many of the Summarizer that streamlines Numba’s vectorization analysis. intricacies of programming. For example, Alzahrani et al. [1] It is designed as a pedagogical tool for teaching Dependence discovered that for many basic programming tasks, Python Analysis, Loop Transformations, and Vectorization as practical users struggled more when compared to students familiarized examples of parallel computing. We used this tool along with with C++. a prepared lesson plan to introduce concepts of parallelism and Another issue is that Python is not typically a performance high performance computing to students from an intermediate oriented language. CPython, Python’s default interpreter, is programming class. locked to a single thread due to its Global Interpreter Lock, The rest of the paper is organized as follows: Related meant to keep programs being thread safe and make library Work is a summary of other tools and teaching models that are related and have inspired the work in this paper. Design refactored the loops to make them representative of the loop Overview describes NumbaSummarizer, its components and transformations that we taught. functionality. Experience outlines our teaching experience III. DESIGN OVERVIEW using the tool and the materials used along with NumbaSum- marizer. Discussion presents a brief analysis in retrospective Our contribution is two fold: of the advantages and disadvantages of using this tool, and 1) NumbaSummarizer, a wrapper that works on top of proposed improvements for future experiences. Conclusion Numba 1 module that users can import into their code gives a brief overview of why we believe this tool is useful as so they can analyze functions for vectorization and the well as propose different scenarios where it could be adopted. presence of data dependence. 2) Lecture notes and sample problems for implementing II. RELATED WORK NumbaSummarizer in a classroom. 2 Introducing concepts of high performance and distributed NumbaSummarizer eases the implementation of Numba’s vec- computing outside specialized or high level courses is not torizing functionality by encapsulating all required Numba’s new. As noted by Prasad et al. [15] and the ACM/IEEE- specific syntax into custom functions. NumbaSummarizer will CS curricula 2013 [7], it is more effective to learn about instruct Numba to request a vectorization report from LLVM. parallelism ”sprinkled” throughout the curricula. Other works Once that report is generated, NumbaSummarizer will parse have tried to put this idea into practice. The work of Garrity through it and extract information pertaining the targeted et al. [6] involves a multi-lingual platform for simple and functions. Figure 1 shows an abstract overview of how the straightforward implementation of MapReduce [4], which is a three of them interact. Numba translates Python functions powerful model for distributed computing. This is a high-level into an intermediate language (IR) that LLVM can interpret application of parallelism, and the authors report success in and pre-compile. For this reason, many details about the using it for teaching parallelism in introductory and advanced original function, such as the function name, gets lost in courses alike. Similarly, Ortiz-Ubarri and Arce-Nazario [13] the report. Also, the report contains plenty of information describe an interface for MapReduce through Python modules. relevant for the compiler but not necessarily for the Python The work of Curtsinger [3] teaches students of an Operating user. Because of this, we need to filter the information and Systems class how to use concurrency and multi threading generate a new summarized report that focuses only on the as parallelism in practice. In this case the students are expe- vectorization results for the given function, adding our own rienced with C and have previous knowledge on Computer identifiers as well. Since our annotated functions have a simple Architecture. Our work focuses on first year students and just structure with one or two loops, we show students how to assumes basic programming and algorithmic knowledge. Our map the LLVM report to them. NumbaSummarizer also has a module is introduced as part of programming optimizations correctness checker which analyzes the output of the original which include simple algorithmic analysis and finding run time loop and the one from the loop transformed by the programmer hotspots. to verify that they are the same. This behavior is accessible Our focus is on using SIMD and dependence analysis as through the following NumbaSummarizer’s functions: the main catalyst to learning parallelism. Ortiz [12] teaches the SIMD execution model in an introductory computer A. NumbaSummarizer’s functions architecture course. However this is done at the assembly • Init diagnostics() enables the LLVM flag debug-only = level, focusing on getting the student familiarized with SIMD loop-vectorize that allows for the creation of vectorization idioms. reports. As of now, this function is private and called by This paper presents the continuation of the work in Watkin- Vector wrapper() when needed. son et al. [18]. Students were given
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages7 Page
-
File Size-