Macro gemv_kernel 

macro_rules! gemv_kernel {
    (
        4,
        $is_fused:expr,
        $out_c:expr,
        $out_len:expr,
        $in_frame:expr,
        $weights:expr,
        $bias:expr,
        $out_frame:expr,
        $do_bias:expr,
        $setzero:expr,
        $load_out:expr,
        $load_bias:expr,
        $add_ps:expr,
        $load_weight:expr,
        $fmadd_ps:expr,
        $store_ps:expr
    ) => { ... };
    (
        8,
        $is_fused:expr,
        $out_c:expr,
        $out_len:expr,
        $in_frame:expr,
        $weights:expr,
        $bias:expr,
        $out_frame:expr,
        $do_bias:expr,
        $setzero:expr,
        $load_out:expr,
        $load_bias:expr,
        $add_ps:expr,
        $load_weight:expr,
        $fmadd_ps:expr,
        $store_ps:expr
    ) => { ... };
}

GEMV kernel macro — generates platform-specific FMADD accumulate loops.

Parameterized by SIMD width (4 = AVX2/256-bit, 8 = AVX-512/512-bit) and all relevant SIMD operations as inline closures. Both variants use 8 independent accumulators to maximize FMA throughput.

Safety

Caller must ensure valid pointer arithmetic and slice bounds.