geotiff: _apply_nodata_mask_gpu allocates instead of mutating in-place

[brendancol]

Summary

_apply_nodata_mask_gpu in xrspatial/geotiff/_backends/_gpu_helpers.py uses cupy.where(mask, nan, arr_gpu) to replace nodata sentinels with NaN. cupy.where allocates a fresh output buffer the same shape as the input array. Writing the NaN into the existing buffer with cupy.putmask (or boolean assignment) avoids the chunk-sized allocation and the corresponding device-to-device copy.

Locations

• _apply_nodata_mask_gpu (line ~45 of _gpu_helpers.py):
  • Float path: arr_gpu = cupy.where(arr_gpu == sentinel, nan, arr_gpu)
  • Integer path: arr_gpu = arr_gpu.astype(cupy.float64) followed by arr_gpu = cupy.where(mask, nan, arr_gpu)

The integer path already allocates a new float64 buffer via astype; the trailing cupy.where then allocates a second buffer of the same size. Both can collapse into a single allocation followed by an in-place write.

Call sites

All call sites pass a freshly decoded GPU buffer that is not shared with caller-visible state:

• _backends/gpu.py:387 - arr_gpu returned from read_to_array's GPU stripped path
• _backends/gpu.py:684 - arr_gpu returned from the GPU tile decoder
• _backends/gpu.py:1072 - per-chunk arr from _decode_window_gpu_direct

In-place mutation is safe at each.

Proposed fix

Use cupy.putmask(arr_gpu, mask, nan) (or arr_gpu[mask] = nan) so the existing buffer is reused. Saves one chunk-sized device buffer per call on GPU-bound workloads with large chunks.

Severity

LOW. Allocator pressure reduction rather than a correctness issue. Filed during the 2026-05-15 deep sweep on geotiff despite earlier classification as marginal.