Kernels

The registered CPU kernels share a common runtime contract: onnx-light creates one KernelBase adapter per graph node, the adapter validates concrete tensors and delegates to an immutable or per-invocation execution plan, and parallel work is submitted only through the current session’s CpuExecutor. No kernel family owns a private worker pool.

The pages below describe five implemented kernel families. Other registered families, including variadic elementwise, integer MatMul, normalization, SwiGLU, TreeEnsemble, and custom-domain kernels, are not yet covered here. Performance targets and future work remain in Next Steps.

• Unary Kernel Design
• Binary Kernel Design
• Gemm Kernel Design
• MatMul Kernel Design
• Attention Kernel Design