[veblush]

MVE VLD20.8/VLD21.8 (vld2q_s8) require 4-byte alignment and raise an alignment UsageFault otherwise, even when CCR.UNALIGN_TRP == 0. In arm_nn_vec_mat_mult_t_s4, arm_nn_mat_mult_nt_t_s4, and arm_convolve_even_s4, row pointers after row 0 are only byte-aligned when the row stride is not a multiple of 4.

Add an arm_nn_vld2q_s8 helper that falls back to byte gather loads when the pointer is not 4-byte aligned, with the alignment check hoisted outside each inner loop.

[mansnils]

Thanks for this. Can you make datasets reproducible through existing generators and TFL/TFLM reference kernels following unit test conventions (Tests/UnitTest/README.md:147)?

[veblush]

Thanks for this. Can you make datasets reproducible through existing generators and TFL/TFLM reference kernels following unit test conventions (Tests/UnitTest/README.md:147)?

Done. Rebased on main and switched the s4 conv and FC tests over to datasets generated from test_plan.json.