onnx_light.tools.kernel_inventory#

Enumerates every registered native kernel path and classifies its coverage.

This module implements Step D of Kernel parallelization and tuning sequence: it builds one inventory row per (domain, op_type, device, element_type, implementation) path covering every operator registered in the built-in onnx_kernels::BuiltinKernelFunctions dispatch table (onnx_light/onnx_extensions/kernels/kernel_dispatch_table.cc).

Each row is classified into exactly one coverage state:

serial

No ParallelFor call site was found in the kernel’s implementation file and no tuning schema is registered for it.

parallel_fixed_policy

The kernel calls ParallelFor but has no registered tuning schema, so its grain size and participant limits are compiled constants.

tunable

The kernel registers a KernelTuningSchema (a portable default that may be overridden by a persisted profile).

calibratable

A tunable path that additionally registers a KernelCalibrationFunction.

Building the inventory only reads static kernel source files and the in-process tuning registry (via onnx_light.kernel_tuning.kernel_tuning_parameters()); it never mutates the kernel tuning cache.

onnx_light.tools.kernel_inventory.build_kernel_inventory(*, library: str = 'onnx_light', dispatch_table_path: Path | str | None = None, kernel_source_root: Path | str | None = None, tuning_report: dict[str, Any] | None = None) → list[dict[str, Any]]#

Builds the full kernel-path inventory (Step D).

Parameters:

  • library – Tuning library identifier passed to onnx_light.kernel_tuning.kernel_tuning_parameters() when tuning_report is not supplied.

  • dispatch_table_path – Optional override of the dispatch-table source path.

  • kernel_source_root – Optional override of the kernel implementation source root.

  • tuning_report – Optional pre-fetched kernel_tuning_parameters() result, mainly for tests that must not depend on the compiled extension.

Returns:

A list of inventory rows, sorted by (domain, op_type, element_type, implementation). Every registered kernel path appears exactly once.

onnx_light.tools.kernel_inventory.dispatch_table_entries(dispatch_table_path: Path | str | None = None) → list[tuple[str, str, str]]#

Parses the built-in kernel dispatch table into (domain, op_type, class_name) rows.

Parameters:

dispatch_table_path – Optional override of the kernel_dispatch_table.cc path, mainly for tests.

Returns:

One entry per registered (domain, op_type) identifier, in file order.

onnx_light.tools.kernel_inventory.validate_inventory(rows: list[dict[str, Any]]) → None#

Raises ValueError when a kernel path is missing or duplicated.

Parameters:

rows – The result of build_kernel_inventory().

Raises:

ValueError – When two rows share the same (domain, op_type, device, element_type, implementation) identity, or when a row is missing an explicit coverage_state.