v7.27.0

7.27.0 (2026-06-22)

Features

• add arm_where_s16 kernel + unit test (634aa2d)
• add tensor shaping kernels (tile, broadcast_to, scatter_nd, mirror_pad, select_v2, where, reverse_sequence, dynamic_update_slice) (8335274)
• add tensor shaping kernels (tile, broadcast_to, scatter_nd, mirror_pad, select_v2, where, reverse_sequence, dynamic_update_slice) (8335274)
• add tensor shaping kernels (tile, broadcast_to, scatter_nd, mirror_pad, select_v2, where, reverse_sequence, dynamic_update_slice) (dd04ea4)
• Update docs with MVE vs DSP vs Scalar C kernel-benchmarks.md (8246c60)
• Update docs with MVE vs DSP vs Scalar C kernel-benchmarks.md (f33a188)

Bug Fixes

• add new operator groups to CMake build options (3a3c1d7)
• avoid single-rounding requantize overflow (#197) (d6735fd)
• broadcast_to: decouple stride computation from broadcast mask (0b14668)
• Correct formatting in arm_tile_s8.c (6266e9e)
• Correct kernel-benchmarks.md cycle counts (9e75bea)
• pass CI contracts (formatting, PDSC, SSoT, Zephyr/NSX wiring) (1cd2050)
• Update benchmark with LP mode results (8c1eebc)
• Update conv benchmark paragraph and add extra prj.conf setting to Zephyr doc (c2fa2c5)
• zephyr: align Kconfig with renamed-knob contract (e2b1a61)

Refactoring

• nsx: make ns-cmsis-nn SoC compatibility wildcard (#200) (3dc5bf2)
• use arm_memcpy_s8/s16 instead of raw memcpy (ea005a6)