ARM Vector Benchmark Lab

🚧 Work in Progress - This project is actively under development. Current status: M2 (DOT Scalar) complete. See ROADMAP.md for development timeline.

A focused benchmarking suite for vector operations on Apple Silicon (M3) processors, comparing scalar, auto-vectorised, and NEON intrinsics implementations.

Quick Start

# Check system requirements
python3 setup_check.py

# Build the project
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(nproc)

# Run tests
./test_scalar
./test_dot

# Run benchmarks
./bench_scalar
./bench_dot

# Run comprehensive benchmark suite
cd ..
python3 scripts/bench.py

Project Structure

• src/: Implementation source code organised by optimisation technique
• tests/: Unit tests for each implementation
• benchmarks/: Micro-benchmarks for performance measurement
• scripts/: Python scripts for comprehensive benchmarking and visualisation
• results/: Generated benchmark results and plots

Development Status

Completed Milestones ✅

• M1: AXPY (Scalar) - Basic AXPY scalar implementation with comprehensive testing and benchmarking
• M2: DOT (Scalar) - Dot product scalar implementation with performance analysis

Planned Milestones 📋

• M3: Auto-vectorisation Analysis - Compiler optimisation exploration
• M4: NEON AXPY - Manual NEON intrinsics implementation
• M5: Results & Analysis - Comprehensive performance analysis and roofline model

See ROADMAP.md for detailed development timeline and Handoff.md for milestone documentation.

Results

M1: Scalar AXPY Performance (Apple M3)

Vector Size	Time (ms)	GFLOPS	Bandwidth (GB/s)	Time/Element (ns)
1K	0.000	41.9	251.4	0.0
4K	0.000	36.2	217.1	0.0
16K	0.001	36.8	220.6	0.1
64K	0.006	22.3	134.0	0.1
256K	0.024	21.8	131.0	0.1
1M	0.098	21.4	128.3	0.1
4M	0.505	16.6	99.6	0.1
16M	2.663	12.6	75.6	0.2

Key Insights:

• Peak performance: ~37 GFLOPS at 16K elements
• Peak bandwidth: ~250 GB/s for small vectors, ~130 GB/s sustained
• Time per element: ~0.1-0.2 ns, showing excellent cache efficiency
• Performance degrades with larger vectors due to cache hierarchy effects

M2: Scalar DOT Performance (Apple M3)

Vector Size	Time (ms)	GFLOPS	Bandwidth (GB/s)	Time/Element (ns)
16K	0.001	15.2	95.1	0.065
1M	0.067	14.9	93.2	0.067

Key Insights:

• Consistent ~15 GFLOPS performance across sizes
• Read-only memory access pattern (no writes)
• Lower bandwidth requirements compared to AXPY
• Baseline established for vectorised implementation comparison

Requirements

• macOS with Apple Silicon (M1/M2/M3)
• clang++ (Xcode command line tools)
• cmake 3.16+
• Python 3.7+ with numpy, matplotlib