Workload Library v0.11.0

Version 0.11.0 – Initial Public Release

Purpose

The ML Workload Library for Vulkan® provides an embeddable C++17 API for
constructing, inspecting, and executing Vulkan® ML workloads. Applications can
retain ownership of their Vulkan® objects and bound resources, or use
library-owned Vulkan® contexts and resource allocations, while the library
manages the workload's execution state.

Features

Workloads and Modules

• Load VGF workloads from files or caller-owned memory, including composed
  compute and Vulkan® data-graph execution sequences.
• Construct standalone compute-shader and Vulkan® data-graph workloads.
• Use embedded SPIR-V™ modules, optionally compile GLSL or HLSL source modules,
  and supply implementations for placeholder modules before configuration.
• Inspect workload resources, executables, modules, bindings, dispatches, and
  data-graph metadata through non-owning C++ views.

Resource Binding and Execution

• Wrap caller-owned Vulkan® RAII instance, physical device, device, queue, and
  queue-family state in a non-owning context, or create a runtime-owned context
  with additional device extension and feature requirements.
• Allocate workload-compatible tensors, storage buffers, images, and backing
  memory as runtime-owned RAII objects.
• Configure reusable sessions and bind caller-owned or runtime-owned tensors,
  storage buffers, images, and their memory where required.
• Optionally provide an image's current layout when binding it so prepared
  execution can transition it to the layout required by the workload.
• Prepare immutable binding snapshots, then either run them with library-managed
  command state or record them into a caller-owned command buffer.
• Validate configuration order, workload compatibility, required bindings,
  module code, resource metadata, and dispatch dimensions.

Build and Integration

• Add package options for installed-binary and source-tree TGZ or ZIP archives,
  and for a Python wheel and source distribution containing the native install
  tree.
• Optionally build the ML Workload Library for Vulkan® as a shared library with
  ML_WORKLOAD_LIB_BUILD_SHARED for ABI and API compatibility analysis.
• Build C++ samples for inspecting and running VGF workloads, constructing a
  standalone data-graph workload, running a standalone GLSL compute workload,
  binding placeholder modules, and wrapping application-owned Vulkan® context
  objects.

Supported Platforms

The following platform combinations are supported:

• Linux - AArch64 and x86-64
• Windows® - x86-64
• Darwin - AArch64 (experimental)
• Android™ - AArch64 (experimental)