vibeProj

GPU-accelerated coordinate projection library. Extracted from RAPIDS cuProj, re-engineered as a pure Python + CuPy package, and expanded from 1 to 24 projections — each with a fused NVRTC kernel that runs the full transform pipeline in a single GPU kernel launch.

Performance

On an RTX 4090 vs i9-13900k, 1M coordinates: (Note: datacenter GPUs will see far higher speedups due to better double precision performance)

Projection	GPU	vs CPU
Transverse Mercator / UTM	0.49 ms	281x
Lambert Conformal Conic	0.54 ms	96x
Albers Equal Area	0.27 ms	143x
Web Mercator	0.15 ms	126x
Equal Earth	0.43 ms	145x
Plate Carrée	0.04 ms	311x
Oblique Mercator (Hotine)	0.76 ms	115x
Krovak	2.08 ms	173x

All 24 projections run in under 3 ms at 1M coordinates. See full benchmark in the repo.

Supported Projections

Projection	Internal Name	EPSG Examples
Transverse Mercator / UTM	tmerc	32601–32760, 27700
Web Mercator	webmerc	3857
Mercator (ellipsoidal)	merc	3395
Lambert Conformal Conic	lcc	2154
Albers Equal Area	aea	5070
Polar Stereographic	stere	3031, 3413
Lambert Azimuthal Equal Area	laea	3035
Oblique Stereographic	sterea	28992
Plate Carrée	eqc	4087
Sinusoidal	sinu	—
Equal Earth	eqearth	8857
Cylindrical Equal Area	cea	6933
Orthographic	ortho	—
Gnomonic	gnom	—
Mollweide	moll	—
Robinson	robin	—
Winkel Tripel	wintri	—
Natural Earth	natearth	—
Azimuthal Equidistant	aeqd	—
Geostationary Satellite	geos	—
Oblique Mercator (Hotine)	omerc	3375
Krovak	krovak	5514
Eckert IV	eck4	—
Eckert VI	eck6	—

Install

pip install vibeproj            # CPU-only (NumPy fallback)
pip install vibeproj[cu12]      # CUDA 12
pip install vibeproj[cu13]      # CUDA 13

For development:

uv sync                         # CPU-only
uv sync --extra cu12            # CUDA 12
uv sync --extra cu13            # CUDA 13

Usage

from vibeproj import Transformer

# Default: always_xy=True — (lon, lat) order, matches shapely/geopandas
t = Transformer.from_crs("EPSG:4326", "EPSG:32631")
x, y = t.transform(2.0, 49.0)           # (lon, lat) in, (easting, northing) out

# always_xy=False: native CRS axis order (matches pyproj default)
t = Transformer.from_crs("EPSG:4326", "EPSG:32631", always_xy=False)
x, y = t.transform(49.0, 2.0)           # (lat, lon) in, (easting, northing) out

Cross-datum transforms (Helmert)

# Cross-datum: Helmert 7/15-parameter shift applied automatically
t = Transformer.from_crs("EPSG:4326", "EPSG:27700")  # WGS84 → OSGB36
x, y = t.transform(-0.1278, 51.5074)

# With ellipsoidal height — z is transformed through the ECEF intermediate
x, y, z = t.transform(-0.1278, 51.5074, z=45.0)

# Same-datum: z passes through unchanged, zero overhead
t = Transformer.from_crs("EPSG:4326", "EPSG:32631")
x, y, z = t.transform(2.0, 49.0, z=45.0)  # z == 45.0

For datum pairs with baked SVD corrections (e.g. NAD27 to NAD83), vibeProj achieves sub-5cm accuracy without external grid files. You can also generate your own SVD corrections for any datum pair that pyproj supports using the included fitting tool (tools/fit_datum_corrections.py). For other grid-only datum pairs without a baked or custom correction, vibeProj warns and proceeds without a datum shift.

Integration with CPU libraries

vibeProj works with popular geospatial Python libraries. GPU acceleration is automatic when CuPy is installed; otherwise it falls back to NumPy transparently.

• GeoPandas — bulk GeoDataFrame reprojection
• Rasterio — GPU-accelerated raster coordinate grids
• Shapely — geometry transforms via shapely.transform()

vibeSpatial Integration (zero-copy GPU)

# Pre-allocated output, no intermediate allocations, stays on GPU
t = Transformer.from_crs(src_crs, dst_crs, always_xy=True)
new_x = cp.empty_like(buf.x)
new_y = cp.empty_like(buf.y)
t.transform_buffers(buf.x, buf.y, out_x=new_x, out_y=new_y)

# 3D: z is transformed through Helmert when crossing datums
new_z = cp.empty_like(buf.z)
t.transform_buffers(buf.x, buf.y, buf.z, out_x=new_x, out_y=new_y, out_z=new_z)

transform_buffers() accepts pre-allocated CuPy output arrays, writes results directly into them, and returns the same objects. No host round-trip, no intermediate allocation. Designed for vibeSpatial's OwnedGeometryArray coordinate buffers.

Architecture

• Pure Python + CuPy — no compiled extensions, no CMake
• Fused NVRTC kernels — each projection's full pipeline (axis swap, deg/rad, central meridian, projection math, scale/offset) runs in a single CUDA kernel launch via CuPy RawKernel
• NumPy fallback — all projections work on CPU when CuPy is unavailable
• Helmert datum shifts — 7/15-parameter (time-dependent) datum transformation with 3D ellipsoidal height support, runs on its own GPU kernel
• SVD datum corrections — baked SVD-compressed grid corrections for sub-5cm accuracy on supported datum pairs (e.g. NAD27 to NAD83), no external grid files needed
• pyproj for CRS metadata — EPSG codes resolved via pyproj, transform math is ours
• fp64 I/O — input/output arrays always double precision (ADR-0002 compliant)
• Auto GPU detection — queries SingleToDoublePrecisionPerfRatio to classify consumer vs datacenter GPU

Test

uv run pytest                    # all tests
uv run pytest tests/test_fused_kernels.py  # GPU kernel tests (requires CuPy)