vibeProj

Fast coordinate transforms on NVIDIA GPUs, with a NumPy fallback for CPU-only environments. vibeProj supports 24 projection families, datum shifts, and preallocated device buffers through a small pyproj-style API.

• Fused CuPy CUDA kernels keep coordinates on the GPU.
• Helmert 7/15-parameter and SVD-compressed datum corrections are built in.
• Hardware-aware transcendental strategies preserve native fallbacks.
• GeoPandas, Rasterio, Shapely, and vibeSpatial workflows are supported.

Performance

One million fp64 coordinates on an RTX 4090 and i9-13900K, measured with pyproj 3.7.2 in August 2026. GPU measurements use preallocated device inputs and outputs; results are the median of three runs with ten timed calls per run.

Parenthetical comparisons use pyproj as the baseline.

Run the same suite with:

uv run python benchmarks/bench_projections.py run --n 1000000

Install

pip install vibeproj            # NumPy
pip install vibeproj[cu12]      # CUDA 12
pip install vibeproj[cu13]      # CUDA 13

Quick start

from vibeproj import Transformer

transformer = Transformer.from_crs("EPSG:4326", "EPSG:32631")
easting, northing = transformer.transform(2.0, 49.0)

Inputs may be scalars, lists, NumPy arrays, or CuPy arrays. Geographic coordinates default to longitude/latitude order; pass always_xy=False for native CRS axis order.

For caller-owned GPU buffers:

transformer.transform_buffers(
    x_gpu,
    y_gpu,
    out_x=easting_gpu,
    out_y=northing_gpu,
)

Defaults choose a safe compute precision and transcendental implementation. Advanced callers can select precision= and transcendentals= explicitly, then inspect the decision with transformer.explain_strategy().

Documentation

• Quickstart and API
• Supported projections
• Precision and transcendental strategies
• Integration recipes for GeoPandas, Rasterio, Shapely, and vibeSpatial

Development

uv sync
uv run pytest

CUDA development environments can use uv sync --extra cu12 or uv sync --extra cu13.