Skip to content

INSTALL.md

Latest commit

 

History

History
310 lines (227 loc) · 9.88 KB

INSTALL.md

File metadata and controls

310 lines (227 loc) · 9.88 KB
Raw

GPU4S: Installation Guide

This document describes how to install all dependencies required to build GPU4S. For general information about the project, check the main README.

Global Requirements

1. GNU Compiler Collection (GCC) / CLang

In order to compile the files of the project you need to have installed GCC ≥ 7. or Clang ≥ 6 in order to support C++17.

# Install on Fedora
sudo dnf install gcc-c++ clang

# Install on Ubuntu
sudo apt update
sudo apt install build-essential clang

# Check version
g++ --version
clang --version

2. CMake

In order to build the project you need to install CMake ≥ 3.24

  # Install on Fedora
  sudo dnf install cmake
  # Install on Ubuntu
  sudo apt install cmake
  # Check version
  cmake --version

3. Git

To fetch external dependencies and CMake FetchContent modules, you need Git.

# Install on Fedora
sudo dnf install git

# Install on Ubuntu
sudo apt install git

# Check version
git --version

Specific Framework Requirements

1. NVIDIA CUDA (nvcc)

Required for compiling CUDA accelerated benchmarks (.cu targets).

Tested with version: V13.2.86

# Install on Fedora (via RPM Fusion / NVIDIA repo)
sudo dnf install xorg-x11-drv-nvidia-cuda cuda-toolkit

# Install on Ubuntu
sudo apt update
sudo apt install nvidia-cuda-toolkit

# Check version
nvcc --version

Additional lib NVIDIA cuDNN
NVIDIA CUDA Deep Neural Network library (cuDNN) is required by some of the benchmark.
Tested with V9.23.2, can be downloaded from NVIDIA website.

Installation Details

When downloading the local package installer from the NVIDIA website, select your operating system distribution:

  • Fedora: Select Linux $\rightarrow$ x86_64 $\rightarrow$ RHEL (.rpm package) $\rightarrow$ FULL.
  • Ubuntu: Select Linux $\rightarrow$ x86_64 $\rightarrow$ Ubuntu or Debian (.deb package) $\rightarrow$ FULL.

Fedora Installation

# Download the RPM repository package for cuDNN 9.23.2
wget https://developer.download.nvidia.com/compute/cudnn/9.23.2/local_installers/cudnn-local-repo-rhel10-9.23.2-1.0-1.x86_64.rpm
# Install the package to register it with the package manager
sudo rpm -i cudnn-local-repo-rhel10-9.23.2-1.0-1.x86_64.rpm
# Clean the DNF package manager 
sudo dnf clean all

# Install the cuDNN 9 library for CUDA 13
sudo dnf -y install cudnn9-cuda-13

Ubuntu Installation

# Download the local DEB package for cuDNN 9.23.2
wget https://developer.download.nvidia.com/compute/cudnn/9.23.2/local_installers/cudnn-local-repo-ubuntu2404-9.23.2_1.0-1_amd64.deb
# Install the local package to register it with APT
sudo dpkg -i cudnn-local-repo-ubuntu2404-9.23.2_1.0-1_amd64.deb
# Copy the GPG key to authenticate packages
sudo cp /var/cudnn-local-repo-ubuntu2404-9.23.2/cudnn-*-keyring.gpg /usr/share/keyrings/
# Refresh the APT package
sudo apt-get update

# Install the cuDNN 9 library for CUDA 13
sudo apt-get -y install cudnn9-cuda-13

2. AMD ROCm (hipcc)

Required for compiling AMD HIP benchmarks (.cpp HIP targets).

Tested with version: 6.4.43484-9999

# Install on Fedora
sudo dnf install rocm-hip-devel rocm-runtime

# Install on Ubuntu
sudo apt update
sudo apt install hipcc rocm-dev

# Check version
hipcc --version

3. OpenCL

Required for compiling OpenCL accelerated benchmarks (.cpp OpenCL targets).

Tested with version: OpenCL 3.0

# Install on Fedora
sudo dnf install opencl-headers ocl-icd-devel clinfo

# Install on Ubuntu
sudo apt update
sudo apt install opencl-headers ocl-icd-opencl-dev clinfo

# Check version and available devices
clinfo

Additional lib CLBlast
The tuned OpenCL BLAS library (CLBlast) is required by some benchmarks.
Tested with v1.6.3 and v1.7.0, can be downloaded from the CLBlast GitHub Releases.

Installation Details

Fedora Installation

# Install CLBlast development libraries directly via DNF
sudo dnf -y install clblast-devel

Ubuntu Installation

# Install CLBlast development libraries directly via APT
sudo apt-get -y install libclblast-dev

4. OpenMP

Required for multi-threaded CPU parallel execution (-fopenmp).

Tested with: OpenMP 4.5 (201511)

# Install runtime & development libraries on Fedora
sudo dnf install libgomp

# Install runtime & development libraries on Ubuntu
sudo apt install libomp-dev

# Check supported OpenMP version via compiler macro (_OPENMP)
echo | g++ -fopenmp -dM -E - | grep _OPENMP

Additional lib OpenBLAS
An optimized Basic Linear Algebra Subprograms (BLAS) library required by some benchmarks.
Tested with v0.3.x, can be downloaded from the OpenBLAS GitHub Releases.

Installation Details

Fedora Installation

# Install OpenBLAS development libraries directly via DNF
sudo dnf -y install openblas-devel

Ubuntu Installation

# Install OpenBLAS development libraries directly via APT
sudo apt-get -y install libopenblas-dev

5. Additional lib FFTW3

A C subroutine library for computing the Discrete Fourier Transform (DFT) required by some benchmarks.
Tested with v3.3.10, can be downloaded from the FFTW Official Website.

Installation Details

Fedora Installation

# Install FFTW3 development libraries directly via DNF
sudo dnf -y install fftw-devel

Ubuntu Installation

# Install FFTW3 development libraries directly via APT
sudo apt-get -y install libfftw3-dev

ANDROID specific Requirements

1. Android NDK r27d (27.3.13750724)

First, you will need to download the Android NDK in your environment to cross-compile for Android targets.

You can use other versions of the Android NDK, but the project was built and successfully tested using NDK r27d (27.3.13750724).

Recommended installation instructions:

  1. Download the Android Command Line Tools
  2. Install them and set ANDROID_HOME and sdkmanager to your bash path:
   # Put default Android Studio path 
   # or wherever you installed it
   echo 'export ANDROID_HOME=$HOME/Android/Sdk' >> ~/.bashrc  
   echo 'export PATH=$PATH:$ANDROID_HOME/cmdline-tools/latest/bin' >> ~/.bashrc
   source ~/.bashrc #reload bash configuration
  1. Download your preferred version of the NDK using the sdkmanager:
sdkmanager --install "ndk;27.3.13750724"

2. Android Debug Bridge (ADB)

In addition, if you want to push the binaries to your phone and execute them, you should install ADB (Android Debug Bridge).

# Ubuntu 
  sudo apt install android-tools-adb
# Fedora
  sudo dnf install android-tools

3. Additional Android Libs

Some benchmarks on Android will require several pre-compiled static libraries, headers, and an OpenCL stub.

In release versions, these dependencies are already pre-compiled and included under ./gpu4s_benchmark/common/android/ this version targets the following specifications:

  • Target ABIs: arm64-v8a, armeabi-v7a
  • Minimum Android API: 21
  • Built With: Android NDK r27d

If you need to target a newer architecture (like ARMv9) or running the main repo, you can use my custom toolchains to download, compile the stub, libraries and headers:

After having recompiled the libraries, you need to replace the corresponding files in the common/android/ directory with the new ones:

common/android/
├── include/
│   ├── arm64-v8a/
│   │   ├── cblas.h *
│   │   └── openblas_config.h *
│   └── armeabi-v7a/
│       ├── cblas.h *
│       └── openblas_config.h *
└── libs/
    ├── arm64-v8a/
    │   ├── libOpenCL.so *
    │   ├── libclblast.a *
    │   ├── libfftw3_omp.a *
    │   ├── libfftw3.a *
    │   └── libopenblas.a *
    └── armeabi-v7a/
        ├── libOpenCL.so *
        ├── libopenblas.a *
        ├── libfftw3_omp.a *
        ├── libfftw3.a *
        └── libopenblas.a *

OpenBLAS toolchain*  CLBlast toolchain*  fftw3 toolchain*