Overview

Skill Level: Intermediate 

Language: English

Workload: 2 hours total

Topic: Performance Engineering on CPUs and GPUs

Overview:  This lecture delves into performance engineering on CPUs and GPUs, focusing on optimizing computation and memory access patterns for high-performance computing applications. 

Course Description: The lecture covers simple but key architectural concepts, including pipelining, memory hierarchies, and caches. On GPUs, the emphasis is placed coalesced memory access, shared memory usage, and addressing bank conflicts. By examining real-world examples such as matrix multiplication and dot product computations, the lecture provides practical insights into maximizing computational throughput and minimizing bottlenecks. This session equips participants with the foundational knowledge to design and implement efficient algorithms for modern multicore and manycore systems.

Course Contents: 

Part 1: Introduction to CPUs

Part 2: Optimizing Memory Access Patterns for Performance 

Part 3: Utilizing the Cache and Spatial-Temporal Locality 

Part 4: Prefetching and Latency Mitigation 

Part 5: Fundamentals of GPU Architecture and Programming 

Part 6: Efficient Memory Accesses on GPUs

Part 7: Bank Conflicts and Shared Memory Performance on GPUs 

Part 8: GPU Occupancy and Parallel Matrix Multiplication 

Who Should Enroll: Anyone who thinks performance matters and want more for their codes. 

Prerequisite: Experience with C++ and CUDA

Tools, libraries, frameworks used: g++, nvcc, perf

Learning Objectives: By participating in this course, you will learn: 

simple things that can be useful to improve the performance on a CPU

simple things that can be useful to improve the performance on a GPU

About the instructor(s): Kamer Kaya is an Associate Professor at the Faculty of Engineering and Natural Sciences at Sabancı University. His research areas include high-performance computing, machine learning on sparse data, and graph algorithms.