Skip to content

PERFORMANCE.md

Latest commit

 

History

History
57 lines (39 loc) · 2.28 KB

PERFORMANCE.md

File metadata and controls

57 lines (39 loc) · 2.28 KB
Raw

Performance

MathScript ships 28 Google Benchmark executables covering the numerical hot paths. There is no separate benchmark build tree: enable them in the same CMake tree as the library.

How to run

Windows (same build-msvc as tests):

.\build.ps1 -Benchmark

That configures -DMS_BUILD_BENCHMARKS=ON, builds all add_ms_bench targets, and smokes each with --benchmark_min_time=0.001s.

Linux:

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release \
  -DMS_BUILD_TESTS=ON -DMS_BUILD_BENCHMARKS=ON \
  -DMS_ENABLE_CUDA=OFF -DMS_ENABLE_AVX512=OFF
cmake --build build
bash scripts/bench_smoke.sh build

Regression vs stored medians (10% tolerance, MS_BENCH_TOLERANCE):

bash scripts/bench_regression.sh build
MS_BENCH_REGRESSION=off bash scripts/bench_regression.sh build   # smoke only
bash scripts/bench_regression.sh --write-baseline build          # refresh Linux JSON

Windows baseline refresh: .\scripts\bench_write_msvc_baseline.ps1.

Targets

bench_matmul, bench_fft, bench_linalg, bench_repl, bench_special, bench_stats, bench_rng_dispatch, bench_simd, bench_signal_linalg, bench_signal_filters, bench_ode_pde, bench_fem, bench_special_memory, bench_optim_symbolic, bench_frameworks, bench_tensorops, bench_distributed_cellai, bench_poly_domain, bench_prob, bench_optim_ml, bench_crypto, bench_graph, bench_topo, bench_image, bench_geo, bench_quantum, bench_compress, bench_finance.

Baselines

File Use
tests/performance/baselines/msvc-release.json Windows medians
tests/performance/baselines/linux-gcc13.json Linux CI medians (benchmark-linux on ubuntu-24.04, AVX-512 off)

Null median entries are skipped by the regression script.

Intentional complexity

These paths are correct-first, not performance debt:

Location Complexity Why
image::dft_magnitude O(RC·RC) DFT Visualisation helper; use ms::fft for production sizes
topo::bottleneck_distance O(n²) greedy matching Typical persistence diagrams n < 500
geo::convex_hull_3d O(n³) face enumeration Small point sets
geo::minkowski_sum_convex O(n·m) brute Fallback for polygons with fewer than 3 vertices