ROCm/rocm-systems

ROCm Systems – a unified super‑repo for AMD’s GPU‑compute libraries

What it is

• A monorepo that gathers many individual ROCm (Radeon Open Compute) system components – things like hip, rocprofiler, rocm‑smi, rocshmem, etc. – under one umbrella. The goal is to make building, testing and contributing to these low‑level GPU libraries easier and more consistent.

Why it matters for AI/ML

• ROCm provides the driver‑level and runtime support that frameworks such as PyTorch, TensorFlow and JAX need to run on AMD GPUs. By consolidating the various ROCm libraries, this repo streamlines the development of the stack that underpins GPU‑accelerated training and inference on AMD hardware.

Key components (each lives in projects/ and is a fully‑fledged library)

Component	Purpose
hip	AMD’s CUDA‑compatible language/runtime for writing GPU kernels.
rocprofiler‑sdk / rocprofiler‑systems	Profiling and performance‑analysis tools for ROCm applications.
amdsmi / rocm‑smi‑lib	System‑management interface exposing GPU health, temperature, power, etc.
rocshmem	Shared‑memory API for multi‑GPU communication.
roctracer	Tracing library for capturing GPU activity.
rocdecode / rocjpeg	Media‑processing libraries (image/video decode) that run on the GPU.
rccl (not listed in the table but part of ROCm)	Collective communication library for multi‑GPU training.
hipfile, hipother, hip‑tests	Helper utilities and test suites for the HIP ecosystem.
rocdbgapi	Debugging API for ROCm kernels.
rocprofiler‑compute, rocprofiler‑register	Additional profiling back‑ends and registration helpers.

CI / health

• The README shows a detailed status table with badges for each component’s continuous‑integration pipelines (build, CodeQL, formatting, sanitizers, etc.). Most components are marked Completed, meaning they have already been migrated into the super‑repo and are considered the source of truth.

How to get started

1. Clone the repo and follow the instructions in CONTRIBUTING.md (sparse‑checkout, dev environment setup, PR workflow).
2. Each component can be built individually using the ROCm build system TheRock; the top‑level CI also runs multi‑component tests.
3. For developers interested in a specific library (e.g., hip or rocprofiler-sdk), navigate to projects/<component>/ and consult that sub‑project’s README for build flags and usage examples.

License

• The repository does not impose a single license at the root. Every sub‑project retains its original license (see the LICENSE* files inside each projects/ or shared/ directory). Users should check the license of the specific component they intend to use.

Who should care

• AMD GPU developers building or extending low‑level compute libraries.
• Framework maintainers (PyTorch, TensorFlow, etc.) who need a stable ROCm stack for AMD hardware.
• Performance engineers looking for profiling, tracing, or system‑monitoring tools for AMD GPUs.
• Contributors who want to work across multiple ROCm libraries without juggling separate repositories.

Bottom line ROCm Systems is a genuine, production‑grade codebase that consolidates the essential building blocks of AMD’s GPU compute ecosystem, directly supporting AI/ML workloads on AMD hardware.