// Running Albucore micro-benchmarks under benchmarks/, synthetic router timings, and comparing PyPI releases with uv --no-project. Use when adding benchmarks, comparing performance across versions, or documenting benchmark workflow.
Albucore image processing conventions - shapes (H,W,C), dtypes (uint8/float32), benchmark-driven backend routing (OpenCV, NumPy, LUT, NumKong), tests, and lockfile discipline. Use when implementing or modifying albucore modules, writing tests, or reviewing image-processing code.
Albucore star-exported API (__all__), routers vs albucore.functions shims, and dependents such as Albumentations. Use when changing exports, documenting API, or deciding what belongs in package __all__.
| name | albucore-benchmarks |
| description | Running Albucore micro-benchmarks under benchmarks/, synthetic router timings, and comparing PyPI releases with uv --no-project. Use when adding benchmarks, comparing performance across versions, or documenting benchmark workflow. |
benchmarks/ - Python timing scripts. Run from repo root: uv run python benchmarks/<script>.py.benchmarks/timing.py - Shared median_ms helper for scripts executed as python benchmarks/foo.py../benchmark.sh - Dataset-driven runner; expects an external benchmark package that is not always present in-tree. Prefer synthetic scripts for CI-style checks.benchmarks/benchmark_router_synthetic.py - Times public routers on synthetic uint8 and float32 arrays: HWC, plus NHWC for mean, std, and mean_std only.benchmarks/compare_router_json.py - Builds a Markdown table from two JSON outputs.Benchmark shape sweeps use channel-last Albucore conventions.
HWC images:
128x128 with 1, 3, 9 channels - small / warm-cache.256x256 with 1, 3, 9 channels - mid-size crop.512x512 with 1, 3, 9 channels - typical augmentation training crop.1024x1024 with 1, 3, 9 channels - high-res / full-image pass.DHWC volumes:
16x128x128x1, 16x128x128x3 - thin slab.32x128x128x1, 32x128x128x3 - common nnU-Net patch depth.64x128x128x3 - deeper slab.128x128x128x1 - isotropic single-channel cube.48x256x256x3 - large in-plane, multi-channel.NDHWC batch of volumes:
2x32x128x128x12x32x128x128x32x64x128x128x34x16x128x128x3Channel choices: 1 for grayscale, 3 for RGB / 3-channel, and 9 for hyperspectral paths that exceed MAX_OPENCV_WORKING_CHANNELS=4.
uv run python benchmarks/benchmark_router_synthetic.py --output-json benchmarks/results/router-main.json
uv run --no-project --with albucore==0.0.40 --with opencv-python-headless \
--with simsimd --with stringzilla --with numpy \
python benchmarks/benchmark_router_synthetic.py --output-json benchmarks/results/router-0.0.40.json
uv run python benchmarks/compare_router_json.py benchmarks/results/router-main.json \
benchmarks/results/router-0.0.40.json benchmarks/results/REPORT_router_compare.md
Use --quick for smaller shape/channel grids while iterating.
docs/numkong-performance.mddocs/research/