Macro simd_4acc_dot_loop 

macro_rules! simd_4acc_dot_loop {
    ($a_ptr:expr, $b_ptr:expr, $end:expr,
     $zero:expr, $load:ident, $fmadd:ident, $add:ident, $lane:expr) => { ... };
}

4-accumulator unrolled SIMD loop for dot product (ILP optimization).

Processes 4 × lane elements per iteration using 4 independent accumulators to hide FMA latency. Works across AVX2, AVX-512, and NEON by accepting ISA-specific intrinsics as parameters.

Returns (combined_accumulator, updated_a_ptr, updated_b_ptr).

Arguments

• $a_ptr, $b_ptr — Starting pointers for the two input vectors
• $end — End pointer for the main loop (aligned to 4 × lane)
• $zero — Zero-init expression (e.g., _mm256_setzero_ps())
• $load — SIMD load intrinsic (e.g., _mm256_loadu_ps)
• $fmadd — FMA intrinsic with signature fmadd(a, b, acc) → a*b + acc
• $add — SIMD add intrinsic (e.g., _mm256_add_ps)
• $lane — Number of f32 elements per SIMD register (4/8/16)

Safety

Must be invoked inside an unsafe context where the specified SIMD intrinsics are valid for the current CPU.