ML Workload Library for Vulkan®

mlworkloadlib provides a public C++ API for constructing executable ML workloads for Vulkan® and running them through a common resource-binding and execution model.

Public API

The installed public headers are included directly as needed:

#include <mlworkloadlib/workload.hpp>
#include <mlworkloadlib/context.hpp>
#include <mlworkloadlib/session.hpp>

The API exposes mlsdk::workloadlib::Workload, mlsdk::workloadlib::Context, mlsdk::workloadlib::Session, mlsdk::workloadlib::BindingSet, and mlsdk::workloadlib::PreparedExecution. Supported workloads are VGF-backed workloads created with Workload::fromVGF(...) and standalone compute shader workloads created with Workload::fromComputeShader(...), and standalone Vulkan® data graph workloads created with Workload::fromDataGraph(...); callers provide application-owned Vulkan® objects through Context::wrap(...), or let Context::create() create a runtime-owned Vulkan® context.

Building the ML Workload Library for Vulkan® from source

The build system must have:

• C/C++ 17 compiler: GCC or Clang on Linux, Clang on Darwin, or MSVC on Windows®.
• CMake 3.25 or later.
• Ninja 1.8.2 or later.
• Python 3.10 or later. Required python libraries for building are listed in tooling-requirements.txt.
• Doxygen 1.9.1 or later. (When building documentation)

To create an archive containing the build artifacts, pass the --package-type option with an archive type such as zip or tgz. Use --package-dir to choose where the archive is written; by default, packages are written to the build directory.

For more command line options, see the help output:

python3 scripts/build.py --help

Build-tree usage

Enable the library when configuring the standalone project:

cmake -S . -B build

Link one of the build-tree targets:

target_link_libraries(my_target PRIVATE mlworkloadlib)

or, if you prefer a namespaced target in the same build:

target_link_libraries(my_target PRIVATE MLWorkloadLibraryForVulkan::mlworkloadlib)

Installed package usage

When the ML Workload Library for Vulkan® is installed, the package exports mlworkloadlib through the MLWorkloadLibraryForVulkan package. Downstream consumers should use MLWorkloadLibraryForVulkan::mlworkloadlib.

find_package(MLWorkloadLibraryForVulkan CONFIG REQUIRED)
target_link_libraries(my_target PRIVATE MLWorkloadLibraryForVulkan::mlworkloadlib)

The package target expects the VGF and VulkanHeaders packages to be discoverable at configure time.

Optional source module support

SPIR-V™ modules are supported by the base library. GLSL and HLSL source compilation are optional and can be enabled when building:

python3 scripts/build.py --enable-glsl-support
python3 scripts/build.py --enable-hlsl-support

Use --glslang-path or --dxc-path when the dependency is not in the default location. For direct CMake configuration, use ML_WORKLOAD_LIB_ENABLE_GLSL_SUPPORT with GLSLANG_PATH, or ML_WORKLOAD_LIB_ENABLE_HLSL_SUPPORT with DXC_PATH.

Applications can query the source backends available in the linked library with mlsdk::workloadlib::supports(...).

Documentation

Install Doxygen and the Python documentation tool dependencies and build the HTML documentation:

python3 -m pip install -r tooling-requirements.txt
python3 scripts/build.py --doc

The generated documentation is available at build/docs/out/index.html.

Consumption model

• mlworkloadlib is install-tree consumable today via find_package(MLWorkloadLibraryForVulkan).
• mlworkloadlib depends on external packages (VGF, VulkanHeaders), and those dependencies are resolved via the standalone package config.

License

The ML Workload Library for Vulkan® is distributed under the software licenses in the LICENSES directory.

Trademark notice

Arm® is a registered trademark of Arm Limited (or its subsidiaries) in the US and/or elsewhere.

Khronos®, Vulkan®, and SPIR-V™ are trademarks of the Khronos® Group.