Performance Optimization and Analysis for High-Performance Computing

What you'll learn

Comprehensive Understanding of GPU vs CPU Architecture

learn the history of graphical processing unit (GPU) until the most recent products

Understand the internal structure of GPU

Understand the different types of memories and how they affect the performance

Understand the most recent technologies in the GPU internal components

Understand the basics of the CUDA programming on GPU

Start programming GPU using both CUDA on Both windows and linux

understand the most efficient ways for parallelization

Profiling and Performance Tuning

Leveraging Shared Memory

Requirements

C and C++ basics

Linux and windows basics

Computer Architecture basics

Description

This comprehensive course is designed for anyone looking to dive deep into CUDA programming and NVIDIA GPU architectures. Starting from the basics of GPU hardware, the course walks you through the evolution of NVIDIA's architectures, their key performance features, and the computational power of CUDA. With practical programming examples and step-by-step instruction, students will develop an in-depth understanding of GPU computing, CUDA programming, and performance optimization. Whether you're an experienced developer or new to parallel computing, this course provides the knowledge and skills necessary to harness the full potential of GPU programming.Here's a refined summary of what you will gain from this CUDA programming course:Comprehensive Understanding of GPU vs CPU Architecture: Students will learn the fundamental differences between GPUs and CPUs, gaining insight into how GPUs are designed for parallel processing tasks.Deep Dive into NVIDIA's GPU Architectures: The course covers the evolution of NVIDIA's GPU architectures, including Fermi, Pascal, Volta, Ampere, and Hopper, and teaches how to compare different generations based on key performance parameters.Hands-On CUDA Installation: Students will learn how to install CUDA across various operating systems, including Windows, Linux, and using WSL, while exploring the essential features that come with the CUDA toolkit.Introduction to CUDA Programming Concepts: Through practical examples, students will understand core CUDA programming principles, including thread and block management, and how to develop parallel applications like vector addition.Profiling and Performance Tuning: The course will guide students through using NVIDIA's powerful profiling tools like Nsight Compute and nvprof to measure GPU performance and optimize code by addressing issues like occupancy and latency hiding.Mastering 2D Indexing for Matrix Operations: Students will explore 2D indexing techniques for efficient matrix computations, learning to optimize memory access patterns and enhance performance.Performance Optimization Techniques: They will acquire skills to optimize GPU programs through real-world examples, including handling non-power-of-2 data sizes and fine-tuning operations for maximum efficiency.Leveraging Shared Memory: The course dives into how shared memory can boost CUDA application performance by improving data locality and minimizing global memory accesses.Understanding Warp Divergence: Students will learn about warp divergence and its impact on performance, along with strategies to minimize it and ensure smooth execution of parallel threads.Real-World Application of Profiling and Debugging: The course emphasizes practical use cases, where students will apply debugging techniques, error-checking APIs, and advanced profiling methods to fine-tune their CUDA programs for real-world applications.By the end of the course, students will be proficient in CUDA programming, profiling, and optimization, equipping them with the skills to develop high-performance GPU applications.

Overview

Section 1: Introduction to the Nvidia GPUs hardware

Lecture 1 GPU vs CPU (very important)

Lecture 2 NVidia's history (How Nvidia started dominating the GPU sector)

Lecture 3 Architectures and Generations relationship [Hopper, Ampere, GeForce and Tesla]

Lecture 4 How to know the Architecture and Generation

Lecture 5 The difference between the GPU and the GPU Chip

Lecture 6 The architectures and the corresponding chips

Lecture 7 Nvidia GPU architectures From Fermi to hopper

Lecture 8 Parameters required to compare between different Architectures

Lecture 9 Half, single and double precision operations

Lecture 10 Compute capability and utilizations of the GPUs

Lecture 11 Before reading any whitepapers !! look at this

Lecture 12 Volta+Ampere+Pascal+SIMD (Don't skip)

Section 2: Installing Cuda and other programs

Lecture 13 What features installed with the CUDA toolkit?

Lecture 14 Installing CUDA on Windows

Lecture 15 Installing WSL to use Linux on windows OS.

Lecture 16 Installing Cuda toolkits on Linux

Section 3: Introduction to CUDA programming

Lecture 17 Mapping SW from CUDA to HW + introducing CUDA.

Lecture 18 001 Hello World program (threads - Blocks)

Lecture 19 Compiling Cuda on Linux

Lecture 20 002 Hello World program ( Warp_IDs)

Lecture 21 003 : Vector addition + the Steps for any CUDA project

Lecture 22 004 : Vector addition + blocks and thread indexing + GPU performance

Lecture 23 005 levels of parallelization - Vector addition with Extra-large vectors

Section 4: Profiling

Lecture 24 Query the device properties using the Runtime APIs

Lecture 25 Nvidia-smi and its configurations (Linux User)

Lecture 26 The GPU's Occupancy and Latency hiding

Lecture 27 Allocated active blocks per SM (important)

Lecture 28 Starting with the nsight compute (first issue)

Lecture 29 All profiling tools from NVidia (Nsight systems - compute - nvprof .)

Lecture 30 Error checking APIs (look at chat GPU there is an example)

Lecture 31 Nsight Compute performance using command line analysis

Lecture 32 Graphical Nsight Compute (windows and linux)

Section 5: Performance analysis for the previous applications

Lecture 33 Performance analysis

Lecture 34 Vector addition with a size not power of 2 !!! important

Section 6: 2D Indexing

Lecture 35 Matrices addition using 2D of blocks and threads

Lecture 36 Why L1 Hit-rate is zero ?

Section 7: Shared Memory + Warp Divergence + Shuffle Operations

Lecture 37 The shared memory

Lecture 38 Warp Divergence

Section 8: Debugging tools

Lecture 39 Debugging using visual studio (important) 1

For any one interested in GPU and CUDA like engineering students, researchers and any other one