Linux Kernel & OS

AMD ROCm HRX Loom 0.2 expands C++ authoring, unifies GPU/NPU/Vulkan targeting

Key Takeaway

Loom 0.2 replaces the legacy ROCr driver with a unified libamdf-backed HAL supporting RDNA, CDNA, and AIE2P through a single compiler targeting x86, AMDGPU, SPIR-V, and WebAssembly.

Summary

AMD has released HRX Loom 0.2, the second version of its compiler, runtime, and serving infrastructure designed for AI agent workflows on ROCm. The update significantly broadens C++ program support and introduces runtime-managed objects alongside stateful finite services. A major architectural shift replaces the former ROCr-backed driver and its legacy-amdgpu selector with a single libamdf-backed AMD HAL, unifying execution paths for CPU, GPU, and NPU resources.

The release removes separate runtime and memory ownership paths that previously complicated deployment. Native AMD GPU launch and replay capabilities now support both RDNA and CDNA architectures with shared memory contracts, explicit compute-resource selection, cooperative execution, virtual mappings, and reusable command buffers. The shared target compiler extends across x86, AIE2P, AMDGPU, SPIR-V, and WebAssembly backends. Performance improvements address JIT compilation overhead, graph construction and updates, pipeline planning, and long-context serving latency.

Compiler architecture and hardware abstraction changes

  • Unified HAL replacement: The legacy ROCr-backed driver and legacy-amdgpu selector are removed entirely; all Loom tools, examples, serving infrastructure, sanitizers, and hardware qualification now use one libamdf-backed AMD HAL accessed via --device=amdgpu.
  • Native GPU execution path: The new native path supports RDNA and CDNA execution with shared CPU/GPU/NPU memory contracts, explicit compute-resource selections, cooperative execution, virtual mappings, and reusable command buffers.
  • Expanded target compiler coverage: The shared target compiler now spans x86, AIE2P (AI Engine 2P), AMDGPU, SPIR-V, and WebAssembly from a single codebase.
  • Packaged execution backends: Loom 0.2 adds packaged device execution for AMD GPU/NPU and Vulkan, enabling broader deployment scenarios beyond raw ROCm installations.
  • Performance and scalability fixes: The release removes fixed and superlinear costs from JIT compilation, graph construction and updates, GPU launch and replay, pipeline planning, and long-context serving operations.

Why it matters

For developers building AI agent infrastructure on AMD hardware, Loom 0.2's unified HAL reduces the fragmentation that previously forced choices between ROCr legacy paths and newer stacks. The inclusion of WebAssembly and SPIR-V targets alongside native AMDGPU support suggests AMD is positioning Loom as a portable compilation layer competitive with MLIR and IREE, which could attract independent compiler and runtime developers who prioritize hardware flexibility over CUDA lock-in. The explicit removal of ROCr-backed paths indicates AMD is consolidating its ROCm tooling around a cleaner abstraction, though teams with existing ROCr-dependent deployments will need to migrate.

Original reference: www.phoronix.com →