ROCm/TheRock

TheRock

TheRock (The HIP Environment and ROCm Kit) is a lightweight open source build platform for HIP and ROCm. It is designed for ROCm contributors as well as developers, researchers, and advanced users who need access to the latest ROCm capabilities without the complexity of traditional package-based installations. The project is currently in an early preview state but is under active development and welcomes contributors. Come try us out! Please see CONTRIBUTING.md for more info and the FAQ for frequently asked questions.

Features

TheRock includes:

• Nightly releases of ROCm and PyTorch
• A CMake super-project for HIP and ROCm source builds
• Support for building PyTorch and JAX with ROCm from source
• Operating system support including multiple Linux distributions and native Windows
• Tools for developing individual ROCm components
• Comprehensive CI/CD pipelines for building, testing, and releasing supported components

Installing from releases

[!IMPORTANT] See the Releases Page for instructions on how to install prebuilt ROCm, PyTorch, and JAX packages.

Project status

See the unified project HUD at https://therock-hud.amd.com/

Nightly release status

Multi-arch releases (all GPU architectures):

Job description	Status
Build ROCm artifacts/tarballs/packages
ASan instrumented build
Test ROCm artifacts
Test ROCm native Linux packages
PyTorch packages - Linux build/test
PyTorch packages - Windows build/test
PyTorch packages - full tests
JAX packages - Linux build/test

Building from source

[!WARNING] Disk Space and Build Time Requirements: Building from source requires approximately 200 GB of free disk space and can take multiple hours to compile. Builds will fail if you run out of space.

We keep the following instructions for recent, commonly used operating system versions. Most build failures are due to minor operating system differences in dependencies and project setup. Refer to the Environment Setup Guide for contributed instructions and configurations for alternatives.

[!TIP] While building from source offers the greatest flexibility, installing from releases in supported configurations is often faster and easier.

[!IMPORTANT] Frequent setup and building problems and their solutions can be found in section Common Issues.

Setup - Ubuntu (24.04)

[!TIP] dvc is used for version control of pre-compiled MIOpen kernels. dvc is not a hard requirement, but it does reduce compile time. snap install --classic dvc can be used to install on Ubuntu. Visit the DVC website for other installation methods.

# Install Ubuntu dependencies
sudo apt update
sudo apt install gfortran git ninja-build cmake g++ pkg-config xxd automake libtool python3-venv python3-dev libegl1-mesa-dev texinfo bison flex

# Clone the repository
git clone https://github.com/ROCm/TheRock.git
cd TheRock

# Install a patched patchelf from source. For details see
# https://github.com/ROCm/TheRock/blob/main/docs/environment_setup_guide.md#patchelf
sudo apt install curl make
sudo env INSTALL_PREFIX=/usr/local ./dockerfiles/install_pinned_patchelf.sh

# Install Rust 1.95 for the Mirage emulator build
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | \
  sh -s -- --default-toolchain 1.95.0
source "$HOME/.cargo/env"

# Init python virtual environment and install python dependencies
python3 -m venv .venv && source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

# Download submodules and apply patches
python3 ./build_tools/fetch_sources.py

Setup - Windows 11 (VS 2022)

[!IMPORTANT] See windows_support.md for setup instructions on Windows, in particular the section for installing tools.

If the build system is a non-English system. Make sure to switch to utf-8.

chcp 65001

# Install dependencies following the Windows support guide

# Clone the repository
git clone https://github.com/ROCm/TheRock.git
cd TheRock

# Init python virtual environment and install python dependencies
python -m venv .venv
.venv\Scripts\Activate.bat
pip install --upgrade pip
pip install -r requirements.txt

# Download submodules and apply patches
# Note that dvc is used for pulling large files
python ./build_tools/fetch_sources.py

Build configuration

The build can be customized through cmake feature flags.

Required configuration flags

• -DTHEROCK_AMDGPU_FAMILIES=

  or

• -DTHEROCK_AMDGPU_TARGETS=

[!NOTE] Not all family and targets are currently supported. See therock_amdgpu_targets.cmake file for available options.

Discovering available targets on your system

In case you don't have an existing ROCm/HIP installation from which you can run any of these tools:

You can install the rocm Python package for any architecture inside a venv and run offload-arch from there:

1. python build_tools/setup_venv.py --index-name nightly --packages rocm .tmpvenv
2. .tmpvenv/bin/offload-arch on Linux, .tmpvenv\Scripts\offload-arch on Windows
3. rm -rf .tmpvenv

Optional configuration flags

By default, the project builds everything available. The following group flags enable/disable selected subsets:

Individual features can be controlled separately (typically in combination with -DTHEROCK_ENABLE_ALL=OFF or -DTHEROCK_RESET_FEATURES=ON to force a minimal build):

hipDNN provider plugins:

[!TIP] Enabling any features will implicitly enable their minimum dependencies. Some libraries (like MIOpen) have a number of optional dependencies, which must be enabled manually if enabling/disabling individual features.

[!TIP] A report of enabled/disabled features and flags will be printed on every CMake configure.

By default, components are built from the sources fetched via the submodules. For some components, external sources can be used by setting the following couple options:

The following components accept specifying alternative source locations:

Further flags allow to build components with specific features enabled.

[!NOTE] Building components with MPI support, currently requires MPI to be pre-installed until issue #1284 is resolved.

CMake build usage

For workflows that demand frequent rebuilds, it is recommended to build it with ccache enabled to speed up the build. See instructions in the next section for Linux and Windows.

Otherwise, ROCm/HIP can be configured and build with just the following commands:

cmake -B build -GNinja . -DTHEROCK_AMDGPU_FAMILIES=gfx110X-all
cmake --build build

CCache usage on Linux

To build with the ccache compiler cache:

• You must have a recent ccache (>= 4.11 at the time of writing) that supports proper caching with the --offload-compress option used for compressing AMDGPU device code.
• export CCACHE_SLOPPINESS=include_file_ctime to support hard-linking
• Proper setup of the compiler_check directive to do safe caching in the presence of compiler bootstrapping
• Set the C/CXX compiler launcher options to cmake appropriately.

Since these options are very fiddly and prone to change over time, we recommend using the ./build_tools/setup_ccache.py script to create a .ccache directory in the repository root with hard coded configuration suitable for the project.

Example:

# Any shell used to build must eval setup_ccache.py to set environment
# variables.
eval "$(./build_tools/setup_ccache.py)"
cmake -B build -GNinja -DTHEROCK_AMDGPU_FAMILIES=gfx110X-all \
  -DCMAKE_C_COMPILER_LAUNCHER=ccache \
  -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
  .

cmake --build build

CCache usage on Windows

• You must have a recent ccache (>= 4.13.3 at the time of writing) that contains bug fixes for MSVC and supports proper caching with the --offload-compress option used for compressing AMDGPU device code.
• export CCACHE_SLOPPINESS=include_file_ctime,pch_defines,time_macros to support hard-linking and precompiled headers (amd-llvm is built with PCH).
• Proper setup of the compiler_check directive to do safe caching in the presence of compiler bootstrapping.
• Set the C/CXX compiler launcher options to cmake appropriately.

Since these options are very fiddly and prone to change over time, we recommend using the ./build_tools/setup_ccache.py script to create a .ccache directory in the repository root with hard coded configuration suitable for the project.

Example (In Command Prompt):

# Any command prompt used to build must eval setup_ccache.py to set environment
# variables.
for /f "delims=" %i in ('python build_tools/setup_ccache.py') do @%i

cmake -B build -GNinja -DTHEROCK_AMDGPU_FAMILIES=gfx110X-all \
  -DCMAKE_C_COMPILER_LAUNCHER=ccache ^
  -DCMAKE_CXX_COMPILER_LAUNCHER=ccache ^
  .

cmake --build build

Running tests

Project-wide testing can be controlled with the standard CMake -DBUILD_TESTING=ON|OFF flag. This gates both setup of build tests and compilation of installed testing artifacts.

Tests of the integrity of the build are enabled by default and can be run with ctest:

ctest --test-dir build

See TESTING.md for more details about testing.

Development manuals

• FAQ: Frequently asked questions for TheRock users.
• CONTRIBUTING.md: Documentation for the process of contributing to this project including a quick pointer to its governance.
• TESTING.md: Testing policies for TheRock itself and ROCm subprojects.
• Development Guide: Documentation on how to use TheRock as a daily driver for developing any of its contained ROCm components (i.e. vs interacting with each component build individually).
• Build System: More detailed information about TheRock's build system relevant to people looking to extend TheRock, add components, etc.
• Environment Setup Guide: Comprehensive guide for setting up a build environment, known workarounds, and other operating specific information.
• Git Chores: Procedures for managing the codebase, specifically focused on version control, upstream/downstream, etc.
• Dependencies: Further specifications on ROCm-wide standards for depending on various components.
• Dockerfiles for TheRock: Information about containers used for building, testing, and distributing ROCm using TheRock.
• Build Artifacts: Documentation about the outputs of the build system.
• Releases Page: Documentation for how to leverage our build artifacts.
• Supported GPUs: Current support status and prioritized roadmap for each AMD GPU architecture.