vibespatial-raster 0.1.4

GPU-first raster processing for vibespatial

vibespatial-raster

GPU-first raster processing for vibeSpatial. Custom NVRTC kernels and CCCL primitives for every operation — algebra, focal, zonal stats, labeling, rasterize, polygonize — with CPU fallbacks via scipy and rasterio.

[!WARNING] vibeSpatial-Raster is very early in development. Operations may be unoptimized or have multiple Host/Device transfers causing reduced performance. File an issue if you hit a problem!

Operations

Category	Operations	GPU Backend
IO	GeoTIFF, COG, JPEG2000 read/write	nvImageCodec (zero-copy decode to device)
Algebra	add, subtract, multiply, divide, apply, where, classify	CuPy element-wise
Focal	convolution, Gaussian filter, slope, aspect	NVRTC shared-memory stencil kernels
Zonal	count, sum, mean, min, max, std, median	CCCL segmented reduce
Rasterize	vector-to-raster (point-in-polygon)	NVRTC per-pixel PIP kernel
Label	connected components (4/8 connectivity)	NVRTC union-find
Morphology	erode, dilate, open, close, sieve	NVRTC tiled kernels
Polygonize	raster-to-vector (marching squares)	NVRTC classify + emit edges

Install

pip install vibespatial-raster            # core (CPU-only, scipy fallback)
pip install vibespatial-raster[io]        # + rasterio for GeoTIFF/COG
pip install vibespatial-raster[cu12]      # + CUDA 12, CuPy, nvImageCodec
pip install vibespatial-raster[cu13]      # + CUDA 13, CuPy, nvImageCodec

For development:

uv sync --group dev

Usage

from vibespatial.raster import read_raster, raster_slope, zonal_stats

# Read a GeoTIFF (GPU-native decode when available, rasterio fallback)
raster = read_raster("elevation.tif")

# Terrain analysis — NVRTC stencil kernels on GPU
slope = raster_slope(raster)

# Per-zone statistics — CCCL segmented reduce on GPU
stats = zonal_stats(zones, slope, ["mean", "max", "std"])

Connected components + polygonize

from vibespatial.raster import label_connected_components, sieve_filter, polygonize_to_gdf

labels = label_connected_components(binary_raster, connectivity=8)
labels = sieve_filter(labels, min_size=50)
gdf = polygonize_to_gdf(labels)

Zero-copy GPU pipeline

OwnedRasterArray tracks HOST/DEVICE residency. Move data to GPU once, then all operations stay on device:

from vibespatial.raster import from_numpy, raster_add

a = from_numpy(data_a, affine=affine, crs="EPSG:4326")
b = from_numpy(data_b, affine=affine, crs="EPSG:4326")
a.move_to("DEVICE")
b.move_to("DEVICE")

result = raster_add(a, b)  # no host-device round-trip

GPU-native IO

When nvImageCodec is available, read_raster() decodes GeoTIFF and JPEG2000 directly to GPU memory — no host copy, no rasterio:

raster = read_raster("large_image.tif")  # → OwnedRasterArray(DEVICE)

Architecture

• Namespace package — installs as vibespatial.raster via pkgutil.extend_path, independent release cadence from core
• Zero-copy NVRTC kernels — no cuCIM, no CuPy ndimage; all GPU ops are custom CCCL/NVRTC
• OwnedRasterArray — mirrors core's OwnedGeometryArray (HOST/DEVICE residency, diagnostic events, nodata mask)
• Dual IO — HYBRID path (rasterio on CPU) + GPU_NATIVE path (nvImageCodec direct-to-device)
• CPU fallback — scipy.ndimage, rasterio.features for every operation; tests validate GPU against CPU

Test

uv run pytest                              # all CPU tests
uv run pytest -m gpu                       # GPU kernel tests (requires CUDA)
uv run pytest tests/test_raster_algebra.py # specific module

License

Apache 2.0 — see LICENSE.