cachemat

cachemat is a small C++ playground for exploring cache-aware matrix multiplication:

• Naive O(M·N·K) triple-loop matmul
• Cache-blocked / tiled matmul
• A simple autotuner that searches over block sizes and reports the fastest configuration

The goal isn’t to beat BLAS, but to make it easy to see how tiling and cache behavior affect performance, and to have a concrete codebase you can link from blog posts or talks.

---

Features

• Naive vs blocked matmul for C = A × B (row-major, single precision)
• Command-line benchmark driver: bench_matmul
• Autotuning mode that sweeps block sizes and reports the best configuration
• Prints:
  • Matrix dimensions (M, N, K)
  • Chosen mode (naive, blocked, auto)
  • Time in seconds
  • Derived GFLOP/s (assuming 2·M·N·K flops)
  • A checksum for correctness sanity checks

---

Project Layout

(Adjust this if your folders differ slightly.)

cachemat/
├── include/
│   └── cachemat/         # Public headers (matmul APIs, utilities)
├── src/                  # Naive / blocked / autotune implementations
├── CMakeLists.txt
└── benchmarks/           # bench_matmul target (depending on CMake setup)

Building

mkdir -p build
cd build

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

This should produce a bench_matmul binary under build/ (or build/benchmarks/, depending on your CMake config).

Flags:

-O3 – aggressive optimization

-march=native – tune codegen for your local CPU (Apple Silicon / x86_64)

-ffast-math – allow FP reassociation and FMA (great for benchmarking, not always for strict numerics)

Usage

./bench_matmul \
  --m <M> \
  --n <N> \
  --k <K> \
  --mode <naive|blocked|auto> \
  [--block-m BM --block-n BN --block-k BK] \
  [--iters ITERS]