Description
flow_direction_mfd on the numpy and cupy backends allocates an (8, rows, cols) float64 array with no memory guard. That is 64 bytes per input cell — an 8x amplifier on top of the input raster.
_cpu at xrspatial/hydro/flow_direction_mfd.py:51 does np.full((8, rows, cols), np.nan, dtype=np.float64)._run_cupy at xrspatial/hydro/flow_direction_mfd.py:285 does cupy.full((8,) + data.shape, cupy.nan, dtype='f8').
A (50000, 50000) DEM asks for ~100 GB of host or GPU memory before anything errors out.
The sister module flow_accumulation_mfd already uses _available_memory_bytes() / _check_memory(rows, cols) and _check_gpu_memory(rows, cols) (flow_accumulation_mfd.py:68-129). flow_direction_mfd is the only MFD module without it. The flow_direction_d8 and flow_direction_dinf variants allocate same-shape outputs (1x amplifiers) and do not need the guard.
Expected behavior
flow_direction_mfd() raises MemoryError with a clear message on numpy and cupy backends when the projected (8, H, W) working set exceeds available memory, matching the flow_accumulation_mfd behavior. Dask paths process per-chunk through map_overlap and inherit the per-chunk guard via _cpu.
Proposed fix
Port _available_memory_bytes / _available_gpu_memory_bytes / _check_memory / _check_gpu_memory from flow_accumulation_mfd.py (64 bytes/pixel budget, 50% threshold) and call from _run_numpy and _run_cupy before the (8, H, W) allocation runs.
Description
flow_direction_mfdon the numpy and cupy backends allocates an(8, rows, cols)float64 array with no memory guard. That is 64 bytes per input cell — an 8x amplifier on top of the input raster._cpuatxrspatial/hydro/flow_direction_mfd.py:51doesnp.full((8, rows, cols), np.nan, dtype=np.float64)._run_cupyatxrspatial/hydro/flow_direction_mfd.py:285doescupy.full((8,) + data.shape, cupy.nan, dtype='f8').A (50000, 50000) DEM asks for ~100 GB of host or GPU memory before anything errors out.
The sister module
flow_accumulation_mfdalready uses_available_memory_bytes()/_check_memory(rows, cols)and_check_gpu_memory(rows, cols)(flow_accumulation_mfd.py:68-129).flow_direction_mfdis the only MFD module without it. Theflow_direction_d8andflow_direction_dinfvariants allocate same-shape outputs (1x amplifiers) and do not need the guard.Expected behavior
flow_direction_mfd()raisesMemoryErrorwith a clear message on numpy and cupy backends when the projected(8, H, W)working set exceeds available memory, matching theflow_accumulation_mfdbehavior. Dask paths process per-chunk throughmap_overlapand inherit the per-chunk guard via_cpu.Proposed fix
Port
_available_memory_bytes/_available_gpu_memory_bytes/_check_memory/_check_gpu_memoryfromflow_accumulation_mfd.py(64 bytes/pixel budget, 50% threshold) and call from_run_numpyand_run_cupybefore the(8, H, W)allocation runs.