你将学到什么
Mapping algorithms to architectures
Extracting parallelism at multiple levels
Amortizing data movement over computation
Understanding performance data
Managing complexity through layering of software
课程概况
Is my code fast? Can it be faster? Scientific computing, machine learning, and data science are about solving problems that are compute intensive. Choosing the right algorithm, extracting parallelism at various levels, and amortizing the cost of data movement are vital to achieving scalable speedup and high performance.
In this course, the simple but important example of matrix-matrix multiplication is used to illustrate fundamental techniques for attaining high-performance on modern CPUs. A carefully designed and scaffolded sequence of exercises leads the learner from a naive implementation to one that effectively utilizes instruction level parallelism and culminates in a high-performance multithreaded implementation. Along the way, it is discovered that careful attention to data movement is key to efficient computing.
Prerequisites for this course are a basic understanding of matrix computations (roughly equivalent toWeeks 1-5 of Linear Algebra: Foundations to Frontiers on edX) and an exposure to programming. Hands-on exercises start with skeletal code in the C programming language that is progressively modified, so that extensive experience with C is not required. Access to a relatively recent x86 processor such as Intel Haswell or AMD Ryzen (or newer) running Linux is required.
MATLAB Online licenses will be made available to the participants free of charge for the duration of the course.
Join us to satisfy your need for speed!
课程大纲
0 Getting Started
1 Loops and More Loops
2 Start Your Engines
3 Pushing the Limits
4 Multithreaded Parallelism
预备知识
Exposure to programming and Linux. Basic understanding of matrix-matrix multiplication.
