GPU JPEG-tiled 3-band read crashes with cudaErrorIllegalAddress and poisons CUDA context

[brendancol]

Bug

open_geotiff(path, gpu=True) and read_geotiff_gpu(path) raise cupy.cuda.runtime.CUDARuntimeError: cudaErrorIllegalAddress on a tiled JPEG TIFF with PhotometricInterpretation=RGB (3-band, samples=3) using JPEGTables (TIFF tag 347). The CPU path reads the same file fine. Single-band JPEG-tiled (photometric='minisblack', samples=1) also works on GPU. Only the 3-band JPEGTables splice path crashes.

After the crash the CUDA context is poisoned. Any later GPU call in the same Python process fails with the same illegal-address error, even on an unrelated file.

Reproducer

import numpy as np
import tifffile
from xrspatial.geotiff import open_geotiff

np.random.seed(0)
arr = np.random.randint(0, 255, (256, 256, 3), dtype=np.uint8)
path = '/tmp/jpeg_rgb.tif'
tifffile.imwrite(path, arr, photometric='rgb', tile=(128, 128), compression='jpeg')

cpu = open_geotiff(path)
print('CPU OK:', cpu.shape, cpu.dtype)

gpu = open_geotiff(path, gpu=True)

Observed:

CPU OK: (256, 256, 3) uint8
RuntimeWarning: read_geotiff_gpu: GPU decode failed (CUDARuntimeError:
cudaErrorIllegalAddress: an illegal memory access was encountered);
falling back to CPU.
...
File "cupy/cuda/pinned_memory.pyx", line 309, in PinnedMemoryPool.malloc
File "cupy_backends/cuda/api/runtime.pyx", line 584, in hostAlloc
File "cupy_backends/cuda/api/runtime.pyx", line 146, in check_status
cupy_backends.cuda.api.runtime.CUDARuntimeError:
cudaErrorIllegalAddress: an illegal memory access was encountered

The auto-mode wrapper at xrspatial/geotiff/__init__.py:372 catches the first failure and warns about CPU fallback, but the call with explicit gpu=True then re-raises during pinned-memory allocation. The pinned allocator failing on its first call after the JPEG decode is the signature of a corrupted CUDA context, not a fresh allocation problem.

Expected

Either decode the file, or raise a RuntimeError that the auto-mode wrapper can catch cleanly. The CUDA context must remain usable so that follow-up GPU calls (and the auto-fallback path) keep working.

Impact

A single corrupted file kills GPU support for the rest of the Python process. Worker processes in a dask cluster die. The auto-mode fallback is rendered unreliable because the warning fires but downstream GPU work still crashes.

Audit pass

Found in geotiff backend parity sweep on 2026-05-09. Reproduced cleanly on this host with tifffile==2024.9.20, cupy-cuda12x, Python 3.14.

Related: #1520 (added JPEGTables splice for tiled JPEG read on the CPU side).