AMD Releases Loom v0.1 Compiler Stack for Heterogeneous Hardware
AMD has launched Loom v0.1, an open-source compiler and intermediate representation stack designed to streamline kernel development across CPUs, GPUs, and NPUs.
AMD tagged v0.1 of Loom, an open-source compiler and intermediate representation stack developed under the ROCm HIP Runtime Extended umbrella. Loom lets developers compile, link, and specialize kernels and asynchronous device programs across heterogeneous hardware, targeting AMD CPUs, GPUs, and NPUs.
Project documentation shows that Loom tracks source facts, target specifications, configuration data, correctness cases, and compiler reports through a unified authoring and execution pipeline. For hardware targets, it supports AMDGPU HSA emission for native GPU execution along with SPIR-V and Vulkan for portable workloads. Additional backends for x86 processors and WebAssembly are currently in development.
The toolkit ships via the hrx-loom SDK, bundling the compiler, runtime components, Python APIs, and command-line utilities. The SDK contains loom-fetch for installing selected model programs and assets, plus loom-serve, an HTTP-based service utility that loads generated model cards and processes inference requests over the network.
Loom connects to the Hyperloom development initiative detailed at AMD's Advancing AI event, which targets optimizations for agentic workloads on AMD computing platforms. You can inspect the v0.1 release tags on GitHub or install the package in Python environments as hrx-loom.