High-performance C++ neural network for MNIST digit recognition with SIMD kernels, OpenMP, and reproducible benchmarks.

• SIMD-accelerated matrix ops (AVX2/AVX-512/NEON) with aligned storage.
• OpenMP-aware hot paths for dot, transpose, and axpy.
• P2 PGM parser with in-memory + on-disk cache for repeat runs.
• CLI training + evaluation pipeline with configurable epochs.
• Catch2 tests wired to CTest.
• Google Benchmark suite with published results + charts.
• Doxygen docs target and clang-format config.
• CI on Linux/macOS/Windows via GitHub Actions.

python3 tools/run.py

This downloads MNIST, builds the project, and runs a training pass. Use python3 tools/run.py --help for flags.

Run files:

• docs/benchmarks/runs/bench-20251226-154121-baseline.json
• docs/benchmarks/runs/bench-20251226-154121-native.json
• docs/benchmarks/runs/bench-20251226-154121-openmp-native.json

Configs:

• baseline: OpenMP off, native off
• native: OpenMP off, native on
• openmp+native: OpenMP on, native on

Environment:

• Apple M2, macOS 15.5
• Apple clang 17.0.0
• Release (-O3, OpenMP on/off, -march=native on/off)

Matrix ops (ns/op, lower is better):

Training/inference throughput (img/s, higher is better):

OpenMP overhead shows up on smaller sizes; the line charts illustrate where parallelism starts to pay off.

See docs/benchmarks/benchmarks.md for methodology and scripts.

python3 tools/run_benchmarks.py --openmp --native

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build
ctest --test-dir build

macOS quickstart:

./tools/bootstrap_macos.sh

./build/fast_mnist_cli data 5000 10 TrainingSetList.txt TestingSetList.txt

clang-format -i src/*.cpp include/fast_mnist/*.h apps/*.cpp

cmake -S . -B build -DFAST_MNIST_ENABLE_DOXYGEN=ON
cmake --build build --target docs

python3 tools/prepare_mnist.py --output data --list-dir .

The script auto-installs tqdm for progress bars; pass --no-auto-install to skip that step.

MIT -- see LICENSE.