Investigate C++26 SIMD for use in FERS

[davidbits]

[davidbits]

Platform / compiler compatibility (blocking)

• Linux x86-64 (GCC ≥15, libstdc++) → ✅ usable (primary target; AVX2/AVX-512)
• Linux ARM64 (GCC) → ⚠️ compiles, limited validation (NEON backend immature)
• Clang/libc++ (all platforms) → ❌ no <simd>
• MSVC / Windows → ❌ no <simd>
• macOS (Apple Clang) → ❌ no <simd>

→ SIMD must be optional + auto-detected; scalar path remains default-safe.

---

Goal
Adopt C++26 <simd> to accelerate DSP hot loops in libfers without changing ABI or numerical behavior.

Scope

• Target: FIR/convolution, phase noise, thermal noise add, ADC quantization, reductions
• Exclude: channel model, I/O, public API, RealType changes
• No intrinsics / third-party SIMD libs

Approach

• Gate via FERS_ENABLE_SIMD + feature probe (FERS_HAS_SIMD)
• Keep scalar fallback with strict parity (bitwise / ULP-bounded)
• Use SoA (split real/imag) for complex ops
• Width-agnostic (std::simd<T>::size())

Phases

1. Infra + build flags
2. Elementwise kernels (adds, reductions, quantization)
3. Transcendentals (sin/cos, noise)
4. FIR + convolution (largest perf gain)
5. Optional CW refactor

Testing

• SIMD vs scalar parity (bitwise or ≤1–2 ULP)
• HDF5 regression tolerance checks
• Benchmarks (target ≥2× AVX2, ≥3× AVX-512 on FIR)

Risks

• Partial compiler support → gated + fallback
• Numerical drift → enforced via tests
• Alignment/ABI → isolated helpers, no public changes

Acceptance

• Builds with/without SIMD
• Parity tests pass
• Measurable FIR speedup
• CI + docs updated