Skip to content

Optional FIM engine comparison

The optional RasFim API uses the same Rust engine source as the ras2fim-2D integration. rust/step05/SOURCE_MANIFEST.json records the matching source hashes. Build it explicitly from a repository checkout; Python does not build or download it. The Python API is installed with the fim extra.

The engine was compared with the ras2fim-2D Step05 method on one public 19-profile model. Exact finite float32 values, nodata masks and grids agreed for full maps, bounding boxes and polygons. These checks establish agreement with that catalog method under the tested conditions, not equivalence to RAS Mapper stored maps or hydraulic accuracy.

Development throughput observations from the ras2fim adapter, rather than this new RasFim wrapper, were 1.58× for one full map and 1.49–1.61× for 19-profile constrained maps. Full-domain catalog throughput remains unqualified pending controlled remeasurement. Those results used four logical CPUs on a shared Intel i9-11900KF workstation, x86-64 WSL/Linux, tmpfs working storage, Python 3.13.17, SciPy 1.16.3 and rasterio 1.4.4 (GDAL 3.10.3), with five interleaved repetitions and an excluded warm-up pair. Host load was uncontrolled. The results do not establish the performance of this wrapper or a general catalog speedup.

Python still prepares global samples and SciPy transforms and writes the GeoTIFF/floodplain products. Native profile concurrency shares one terrain cache and thread pool but does not parallelize those Python stages. Decoded cache limits and planned allocation admission are not total-process RSS caps. An AOI reduces source-pixel work while retaining a full-grid geometry lookup and, when required, the serial interpolation fallback.

See the API guide for installation, inputs, bounds, polygon CRS rules and explicit failure behavior. Native Windows builds and other untested terrain/CRS combinations require additional qualification.

Both retained engine wheels passed installed synthetic shared-terrain batch, bbox and polygon-mask tests. The Windows wheel additionally passed a local Commander public-model full/bbox/multipart-polygon-with-hole crop test. wheels/BUILD.json identifies the public build and wheels/SHA256SUMS.txt records the retained artifacts. These platform checks are distinct from the Linux 19-profile numerical qualification and from performance measurements.