vector-playground

A tiny playground for experimenting with vectorization on Apple Silicon.

Right now it focuses on the classic SAXPY kernel (y = a * x + y) and a simple dot product, and compares:

• 🚶‍♂️ Scalar – vectorization explicitly disabled
• 🤖 Auto – let Clang’s auto-vectorizer do its thing
• 🛠 Manual NEON – hand-written NEON intrinsics with configurable loop unroll factors

The goal is to measure, not guess: how close is the compiler’s auto-vectorized code to a carefully tuned NEON kernel on an M-series Mac?

---

Highlights

• Apple Silicon–friendly C++ / CMake project.
• Benchmarks for:
  • saxpy: y[i] = a * x[i] + y[i]
  • dot: sum += x[i] * x[i] (or x[i] * y[i], depending on config)
• Three variants:
  • scalar – vectorization disabled
  • auto – auto-vectorized by Clang
  • manual – NEON intrinsics with explicit unroll factors (1, 2, 4, …)
• Benchmark output includes:
  • Total time
  • GFLOP/s
  • Checksum (for sanity-checking correctness across variants)

If you’ve ever wondered “can I beat the compiler?” on simple kernels, this repo is a place to play with that question.

---

Requirements

Tested on:

• Apple Silicon (M1 / M2 / M3…)
• C++ compiler: Clang (Apple Clang) with NEON support
• CMake 3.16+

You’ll also need a standard command-line toolchain (make, ninja, etc., depending on your CMake generator).

---

Building

From the repo root:

mkdir -p build
cd build

cmake -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_CXX_FLAGS_RELEASE="-O3 -ffast-math -march=native" \
      ..
cmake --build . -j