Day 01: Introduction to GPU Computing & CUDA

字数

873 字

阅读时间

6 分钟

Table of Contents

1. Overview
2. CPU vs. GPU: Conceptual Diagram
3. Core CUDA Concepts
4. Environment Setup
5. “Hello GPU” Program
6. Common Pitfalls
7. References & Further Reading
8. Next Steps

---

1. Overview

Welcome to your first day of learning CUDA! Today’s focus is on establishing a clear mental model of GPU computing and how CUDA opens up the GPU’s parallel processing power to you. By the end of this lesson, you should:

• Understand the high-level difference between CPUs and GPUs.
• Have a mental picture of CUDA’s programming approach.
• Be ready to write and run a simple “Hello GPU” kernel.

Modern GPUs thrive on parallelism, letting thousands of lightweight threads handle tasks in tandem. CUDA exposes these capabilities through an extension of C/C++ that includes special keywords and memory models for GPU programming.

---

2. CPU vs. GPU: Conceptual Diagram

Let’s start with a quick visual comparison of how CPUs and GPUs handle parallel tasks:

flowchart LR
    subgraph CPU
        A[CPU Core 1]
        B[CPU Core 2]
    end
    subgraph GPU
        C1[GPU Core 1]
        C2[GPU Core 2]
        C3[GPU Core 3]
        C4[GPU Core 4]
        Cn[... up to thousands of cores ...]
    end
    
    CPU -- Good for complex, low-latency tasks --> SerialOps[Serial or branching tasks]
    GPU -- Great for parallel, high-throughput tasks --> ParallelOps[Massive data-parallel tasks]

    style CPU fill:#dee8f7,stroke:#081a46,stroke-width:1px
    style GPU fill:#f7e8de,stroke:#8a2f0d,stroke-width:1px

• CPU Cores: Optimized for fast, complex operations on fewer threads at once.
• GPU Cores: Designed for handling thousands of smaller, simpler threads, excelling at data-parallel operations.

When you have huge arrays or matrices to process, or repetitive operations that can be performed in parallel, GPUs often deliver massive speedups.

---

3. Core CUDA Concepts

a) Execution Model

• Threads: The smallest unit of parallel work. Each thread executes the same kernel (function), but on different data (Single Instruction, Multiple Thread approach).
• Blocks: Group of threads that can share fast on-chip memory (called shared memory).
• Grid: Collection of all blocks needed to solve your problem.

b) Memory Hierarchy

• Global Memory: Large but relatively slower; accessible by all threads.
• Shared Memory: A small, fast region local to each block.
• Local/Private Memory: Registers or thread-local storage.
• Constant & Texture Memory: Specialized for read-only data or caching.

c) Kernel Functions

• Marked with __global__, these are the functions you launch on the GPU.
• Kernel launches use the triple-angle-bracket syntax: kernelName<<<blocks, threads>>>(...)

---

4. Environment Setup

1. Install CUDA Toolkit

   • Download here for your OS (Windows, Linux, macOS).
   • Ensure your NVIDIA driver is updated to the recommended version or newer.

2. Verify the Install

   • Check nvcc --version in your terminal or Command Prompt.
   • If you see a valid version (e.g., Cuda compilation tools, release 11.x), you’re set.

3. Check GPU Compatibility

   • Run nvidia-smi (Linux) or open the NVIDIA Control Panel (Windows) to see your GPU model and compute capability.
   • This helps when using advanced CUDA features or specifying compilation flags like -arch=sm_75.

---

5. “Hello GPU” Program

Let’s write a simple CUDA program that prints a message from the GPU, confirming our environment works.

#include <stdio.h>
#include <cuda_runtime.h>

// __global__ indicates this function runs on the GPU.
__global__ void helloFromGPU() {
    if (threadIdx.x == 0) {
        printf("Hello from the GPU!\n");
    }
}

int main() {
    // Launch the kernel with 1 block of 1 thread.
    helloFromGPU<<<1, 1>>>();

    // Synchronize to ensure the GPU finishes before exiting.
    cudaDeviceSynchronize();

    printf("Hello from the CPU!\n");
    return 0;
}

Compiling & Running

1. Save it as helloGPU.cu.
2. Compile with:

   nvcc helloGPU.cu -o helloGPU

3. Run the executable:

   ./helloGPU

4. Expected output:

   Hello from the GPU!
   Hello from the CPU!

---

6. Common Pitfalls

1. Missing <cuda_runtime.h>
   Leads to errors about undefined CUDA functions.
2. Using gcc Instead of nvcc
   You must use nvcc (the CUDA compiler) to recognize CUDA keywords and handle device code.
3. Driver/Toolkit Mismatch
   An outdated driver may fail to run kernels built with a newer CUDA version.
4. Forgetting cudaDeviceSynchronize()
   Without synchronization, your program may terminate before the GPU finishes.

---

7. References & Further Reading

1. CUDA C Programming Guide – Chapters 1 & 2
   Introduction and basic concepts of CUDA, including threads, blocks, and the compilation process.

2. CUDA Quick Start Guide
   Step-by-step instructions for installing the toolkit and verifying your setup.

3. GPU Gems – Chapter 1
   Provides a solid overview of GPU architecture and parallel computing fundamentals.

4. NVIDIA Developer Blog: Introduction to CUDA
   Articles that dive deeper into the “why” of GPU computing and feature beginner-friendly examples.

---

8. Next Steps

In Day 2, we’ll get into optimizing your development environment—covering debugging tools, performance measurement, and more. Keep this Day 1 code handy as a reference point for ensuring future projects compile and run correctly.

Hint: Experiment! Try changing the block and thread configuration (e.g., <<<1, 10>>>) and print out which thread is speaking to get an early feel for how parallelism works.

---

贡献者

阿啸

文件历史

最后编辑于 27 天前查看完整历史

d2467-vault backup: 2025-05-06 22:40:30