Link a prebuilt archive

Use a prebuilt archive when its architecture, compiler and build flavor match your application. For configurable kernel profiles or float features, use source integration.

Inspect the release bundle

Download and extract the bundle from heliaRT releases. Read its MANIFEST.txt and retain the version and commit identity with your firmware dependencies. The bundle layout is:

helia-rt-.../
├── lib/libhelia-rt-cm55-gcc-release-with-logs.a
├── tensorflow/
├── third_party/
├── zephyr/
└── MANIFEST.txt

Archives use cm4 or cm55 architecture tokens, gcc, armclang or atfe compiler tokens, and debug, release-with-logs or release flavor tokens. Choose an actual file in the bundle. The cm4 artifacts target Cortex-M4 with an FPU; the Cortex-M55 float kernels require the corresponding MVE floating-point target configuration.

The bundle contains runtime and HELIA kernel objects in each archive. Do not add another heliaCORE archive for the same symbols. Header paths start at the bundle root, not an include/ subdirectory. These contracts come from the packager and prebuilt module.

Define an imported target

Add this to an existing firmware CMake project after project(). The example selects the GCC Cortex-M55 archive retaining logs:

set(HELIA_RT_BUNDLE "${CMAKE_CURRENT_SOURCE_DIR}/third_party/helia-rt")
set(HELIA_RT_ARCHIVE
    "${HELIA_RT_BUNDLE}/lib/libhelia-rt-cm55-gcc-release-with-logs.a")
if(NOT EXISTS "${HELIA_RT_ARCHIVE}")
  message(FATAL_ERROR "Selected heliaRT archive is missing")
endif()

add_library(helia_rt_prebuilt STATIC IMPORTED)
set_target_properties(helia_rt_prebuilt PROPERTIES
  IMPORTED_LOCATION "${HELIA_RT_ARCHIVE}"
  INTERFACE_INCLUDE_DIRECTORIES
    "${HELIA_RT_BUNDLE};${HELIA_RT_BUNDLE}/third_party/flatbuffers/include;${HELIA_RT_BUNDLE}/third_party/gemmlowp;${HELIA_RT_BUNDLE}/third_party/ruy;${HELIA_RT_BUNDLE}/third_party/kissfft;${HELIA_RT_BUNDLE}/third_party/ns_cmsis_nn/Include"
  INTERFACE_COMPILE_DEFINITIONS "TF_LITE_STATIC_MEMORY"
)
target_link_libraries(my_firmware PRIVATE helia_rt_prebuilt)

Replace my_firmware with your existing application target. Its toolchain file and board support must already supply startup code, linker placement, CPU/FPU flags and compatible C/C++ runtime libraries. Match the archive’s floating-point ABI and compiler ABI, including enum sizing. The generated Zephyr prebuilt integration applies -fshort-enums; consult the selected compiler’s release build flags when integrating outside Zephyr.

For Zephyr consumers, see the GCC prebuilt enum ABI caution. The bundled Zephyr template applies an enum-width flag that the GCC archive build does not explicitly request; do not copy that flag into a custom CMake integration without checking the archive ABI.

Keep the tensor layout consistent

TF_LITE_STATIC_MEMORY affects public tensor structures. Compile application code including LiteRT headers with the same definition as the archive. Use the headers delivered with that archive, rather than headers from another runtime checkout.

Build and check inference

Configure your firmware with its existing cross-toolchain file, then build:

cmake -S . -B build -DCMAKE_TOOLCHAIN_FILE=cmake/board-toolchain.cmake
cmake --build build

Use First inference for checked model loading, registration, allocation and invocation. Check known inputs and expected outputs on your target before relying on the integration.

If linking fails, first check the selected archive and its symbols, header version, ABI settings, and platform logging dependencies. If allocation fails, inspect model compatibility and diagnostics before increasing the arena. Prebuilt float support cannot be changed with application-only compiler definitions.