Version 0.7.0 – Initial Public Release
Purpose
Provides a layered Vulkan® implementation of ML APIs for environments lacking
native ML extension support.
Features
Vulkan® Loader Integration
- Dual-Layer Architecture: Two separate layers - graph (
VK_ARM_data_graph)
and tensor (VK_ARM_tensors) — providing modular ML extensions for Vulkan® support. The
graph layer depends on the tensor layer. - Vulkan® Loader Integration: Seamlessly injected by the Vulkan® Loader
without requiring application modifications
ML Extensions
- Universal Compatibility: Enables ML workloads to run on any Vulkan®
compute-capable device, regardless of native ML extensions for Vulkan® support - Fallback Implementation: Acts as a bridge solution while native driver
support for ML extensions matures across the ecosystem
Shader compilation
- SPIR-V™ Processing: Advanced shader compilation and optimization through
SPIRV-Tools, and SPIRV-Cross integration - Runtime Shader Compilation: Supports both pre-compiled and runtime shader
compilation modes for flexible deployment scenarios
Platform Support
The following platform combinations are supported:
- Linux - X86-64
- Windows® - X86-64