Show HN: Npunlock – Run custom C kernels for Intel NPUs

npunlock is a toolchain that restores a missing workflow for Intel Core Ultra NPUs by turning custom C code into runnable ACT‑SHAVE machine kernels and embedding them in NPU execution graphs. Verified on Windows x64 with Meteor Lake / NPU3720, it compiles user C into SHAVE code via Intel/Movidius MoviTools, packages a native helper DLL, and provides Python, CLI, and C APIs to insert custom kernels into graphs. Examples include FP32 and FP16 GELU implementations that run on the NPU and verify results against NumPy; the toolchain exposes libm functions (like tanhf) via the MoviTools runtime. A recent breakthrough (2026-09-23) enables a single native graph to run independent FP32-unary and FP16-binary custom branches by using an explicit ACT-group preflight to control the compiler’s branch reordering.

Installation requires Windows x64, an installed Intel NPU driver for the device, Python 3.10+, CMake and MSVC for source builds, and an extracted MoviTools MVC_DEPEND package (supplied by legacy driver payloads but not redistributed). npunlock currently supports static dense FP16 unary and two-input kernels and a verified FP32 unary path; mixed-precision and other NPU generations are experimental or unverified, with strict tensor-layout and carrier compatibility. Documentation covers kernel writing, compilation, and reverse‑engineering notes. The code is Apache‑2.0 licensed; MoviTools and Intel libraries remain proprietary. Contributions and hardware testing for Linux or newer NPUs are invited.