README.md

Study_CUDA_Programming (based on C++ ver11)

• All materials in this repository are based on lectures and code from Inflean's CUDA course.
• The link of the course is as follows:
  • https://www.inflearn.com/roadmaps/654#community
• All lecture materials and codes follow the instructor's license and are only for educational purposes for the course attended.
• Commercial use is prohibited !!

Must prepare as follows:

• Nvidia GPU
• OS: Ubuntu 20.04 (for me), Windows 10 over., Mac
• Install CUDA (for me, v12.1)
• Pure python Env (Not conda Env)
  • If you want to set python3 for main python module, please set.

sudo apt update
sudo apt install python3.8
sudo update-alternatives --install /usr/bin/python python /usr/bin/python3.8 10
sudo update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.8 10 # If don't use python3, no need it
sudo update-alternatives --config python3 # If don't use python3, no need it

Install glfw3 packages

sudo apt-get install libglfw3-dev libglfw3

Install cmake 3.30

sudo apt purge cmake
sudo apt install wget build-essential

wget https://github.com/Kitware/CMake/releases/download/v3.30.0/cmake-3.30.0.tar.gz
tar -xvzf cmake-3.30.0.tar.gz
cd cmake-3.30.0
./bootstrap --prefix=/usr/local
make
sudo make install
cmake --version

• If you don't find cmake version, please edit as follows:

vi ~./bashrc

PATH=/usr/local/bin:$PATH:$HOME/bin

source ~./bashrc

Download CUDA Samples (for me, v12.1)

wget https://github.com/NVIDIA/cuda-samples/archive/refs/tags/v12.1.tar.gz
tar -zxvf v12.1.tar.gz

wget https://github.com/NVIDIA/cuda-samples/archive/refs/tags/v12.1.zip
unzip v12.1.zip
make
sudo make install

CUDA for Ubuntu

• $ ubuntu-drivers devices
• $ sudo apt install nvidia-driver-xx
  • reboot !
• $ nvidia-smi (only for checking your NVIDIA driver)
  • visit CUDA-zone to get the CUDA toolkit
• $ sudo apt get install build-essential (to get GCC compilers)
• $ nvcc -V (now you should get the NVIDIA CUDA Compiler messages)

CUDA Tutorial

• in each section, build the project as shown below and run the generated file.

mkdir build
cd build
cmake ..
make
./generated_execution_file

This tutorial is structured as follows:

1. part1_cuda_kernel: Start CUDA programming | Certificate

• print hello cuda (on Ubuntu)
• memory copy
• add vector by using cpu or CUDA
• error check

2. part2_vector_addition: Study CUDA kernel launch | Certificate

• elapsed time
• CUDA kernel launch
• 1d vector addition
• Giga vector addition
• AXPY and FMA
  • single precision
  • linear interpolation
• thread and GPU

3. part3_memory_structure: Memory Structure | Certificate

• 메모리 계층 구조
• CUDA 전용의 2D 메모리 할당 함수, pitched point 사용법
• 3D 행렬 사용 및 pitched point 사용법
• CUDA 메모리 계층 구조
• 인접 원소끼리 차이 구하기: shared memory 활용

4. part4_matrix_multiply: Matrix Multiply | Certificate

• matrix copy
• Matrix Transpose 전치 행렬
• Matrix Multiplication
• GEMM: general matrix-to-matrix multiplication
• 메모리에 따른 CUDA 변수 스피드 측정
• 정밀도와 속도개선

5. part5_atommic_operation: Atomic Operation | Certificate

• Control Flow
  • if 문 과 for loop 문 어떻게 최적화 할것인지?
  • shared 메모리를 사용하는 경우라면, half-by-half를 사용하는 even-odd 보다 조금더 빠르다.!!
• race conditions 문제의 해결방법으로 Atomic Operation 사용
• atomic operation 사용하여 histogram 구하기
• Reduction Problem 솔루션
• GEMV operation

6. part6_search_sort: Search & Sort | Certificate

• Linear Search 선형 탐색
• Search All 모든 위치 모두 찾기
  • CUDA에서 stride 사용하는 것이 제일 빠르다.
• Binary Search 이진 탐색
  • CUDA 사용해서, binary search는 효과적이지 못하다.
  • 그냥 CPU 사용하세요!. 특히 STL 짱짱 빠름.
• CUDA 에서 Sort 하는 방법.. 본격적으로 얘기해 보자!!
  • 블럭 단위 parallel sorting
    • CUDA even-odd sort: 엄청 빨라 짐
  • global 메모리 활용 parallel sort 할때는,
    • CUDA (even-odd) 에서 도차도 상당히 느리다.
• Bitonic Sort 바이토닉 소트
  • 병렬 처리를 위한, 소팅 방법이라고 보면 됨
• Counting Merge Sort 카운팅 방식 머지 소트 (병합 정렬)
  • 병렬 처리에 가장 적합한 Large Scale Parallel Counting Merge Sort 방법

Additional Comments

• All description in the materials have been modified by myself, Hyunkoo Kim.
• (c) 2024. hyunkookim.me@gmail.com. All rights reserved.