|
| 1 | +"""GPU backend coverage for issue #1941 (float16 read). |
| 2 | +
|
| 3 | +#1941 added float16 auto-promotion on read by making |
| 4 | +``tiff_dtype_to_numpy(16, SAMPLE_FORMAT_FLOAT)`` return ``float32`` and |
| 5 | +adding the on-disk ``tiff_storage_dtype`` companion. The eager numpy and |
| 6 | +dask paths are covered by ``test_float16_read_1941.py``; this module |
| 7 | +closes the GPU and dask+GPU coverage gap. |
| 8 | +
|
| 9 | +A regression that: |
| 10 | +
|
| 11 | +* dropped the ``bps_mismatch`` stripped/odd-bps fallback at |
| 12 | + ``_backends/gpu.py:357`` would route float16 stripped reads through |
| 13 | + the tiled GPU decoder and mis-decode the half-precision samples; |
| 14 | +* dropped the ``bps_first == 16 and sample_format == SAMPLE_FORMAT_FLOAT`` |
| 15 | + early-out at ``_backends/gpu.py:791`` in ``_gds_chunk_path_available`` |
| 16 | + would send tiled float16 chunked reads down the kvikIO GDS path and |
| 17 | + mis-stride the buffer; |
| 18 | +* dropped the entry at ``(16, SAMPLE_FORMAT_FLOAT) -> float32`` in |
| 19 | + ``tiff_dtype_to_numpy`` would surface as ``ValueError("Unsupported |
| 20 | + BitsPerSample=16, SampleFormat=3")`` from the GPU read paths. |
| 21 | +
|
| 22 | +Every test ships through ``read_geotiff_gpu`` directly or through |
| 23 | +``open_geotiff(..., gpu=True)`` so the dispatcher path is also wired in. |
| 24 | +``cuda-unavailable`` builds skip the suite via the project's standard |
| 25 | +``CUDA_AVAILABLE`` gate. |
| 26 | +""" |
| 27 | +from __future__ import annotations |
| 28 | + |
| 29 | +import importlib.util |
| 30 | + |
| 31 | +import numpy as np |
| 32 | +import pytest |
| 33 | +import xarray as xr |
| 34 | + |
| 35 | + |
| 36 | +def _gpu_available() -> bool: |
| 37 | + if importlib.util.find_spec("cupy") is None: |
| 38 | + return False |
| 39 | + try: |
| 40 | + import cupy |
| 41 | + |
| 42 | + return bool(cupy.cuda.is_available()) |
| 43 | + except Exception: |
| 44 | + return False |
| 45 | + |
| 46 | + |
| 47 | +_HAS_GPU = _gpu_available() |
| 48 | +pytestmark = pytest.mark.skipif( |
| 49 | + not _HAS_GPU, reason="cupy + CUDA required for GPU float16 read tests", |
| 50 | +) |
| 51 | + |
| 52 | + |
| 53 | +@pytest.fixture |
| 54 | +def float16_stripped_tif(tmp_path): |
| 55 | + """Stripped float16 GeoTIFF: triggers the bps_mismatch CPU fallback. |
| 56 | +
|
| 57 | + ``tifffile.imwrite`` without ``tile=`` produces a stripped layout, so |
| 58 | + the GPU reader hits ``bps_mismatch=True`` (file_dtype.itemsize*8 == 32 |
| 59 | + but bps == 16) and falls back to ``_read_to_array`` on CPU before |
| 60 | + copying to device. |
| 61 | + """ |
| 62 | + tifffile = pytest.importorskip("tifffile") |
| 63 | + arr = np.array( |
| 64 | + [[0.0, 1.0, 2.0, 3.0], |
| 65 | + [-1.0, -2.0, -3.0, -4.0], |
| 66 | + [0.5, 1.5, 2.5, 3.5], |
| 67 | + [100.0, 200.0, 300.0, 400.0]], |
| 68 | + dtype=np.float16, |
| 69 | + ) |
| 70 | + path = tmp_path / "f16_stripped.tif" |
| 71 | + tifffile.imwrite(str(path), arr, compression=None) |
| 72 | + return path, arr |
| 73 | + |
| 74 | + |
| 75 | +@pytest.fixture |
| 76 | +def float16_tiled_tif(tmp_path): |
| 77 | + """Multi-tile float16 GeoTIFF: 32x32 image, 16x16 tiles (2x2 grid). |
| 78 | +
|
| 79 | + Tiled and deflate-compressed. The 2x2 tile grid exercises inter-tile |
| 80 | + reassembly in the decoder path so a regression that mis-stitched |
| 81 | + adjacent tiles would surface here. ``bps_mismatch`` short-circuits |
| 82 | + the tiled GPU decode path and routes through the CPU decoder; the |
| 83 | + GDS path is also gated off via ``_gds_chunk_path_available`` |
| 84 | + returning False for (bps=16, sf=3). |
| 85 | + """ |
| 86 | + tifffile = pytest.importorskip("tifffile") |
| 87 | + arr = np.arange(1024, dtype=np.float16).reshape(32, 32) |
| 88 | + path = tmp_path / "f16_tiled.tif" |
| 89 | + tifffile.imwrite( |
| 90 | + str(path), arr, compression="deflate", tile=(16, 16)) |
| 91 | + return path, arr |
| 92 | + |
| 93 | + |
| 94 | +@pytest.fixture |
| 95 | +def float16_tiled_uncompressed_tif(tmp_path): |
| 96 | + """Tiled uncompressed float16 GeoTIFF. |
| 97 | +
|
| 98 | + Mirrors ``float16_tiled_tif`` but with ``compression=None`` so the |
| 99 | + tile-decode path is exercised without an extra deflate codec call. |
| 100 | + Tile size 16 is the smallest tifffile allows. |
| 101 | + """ |
| 102 | + tifffile = pytest.importorskip("tifffile") |
| 103 | + arr = np.arange(256, dtype=np.float16).reshape(16, 16) |
| 104 | + path = tmp_path / "f16_tiled_none.tif" |
| 105 | + tifffile.imwrite( |
| 106 | + str(path), arr, compression=None, tile=(16, 16)) |
| 107 | + return path, arr |
| 108 | + |
| 109 | + |
| 110 | +class TestEagerGPUReadFloat16: |
| 111 | + """``read_geotiff_gpu`` returns float32 for stripped float16 input.""" |
| 112 | + |
| 113 | + def test_read_geotiff_gpu_stripped_returns_float32( |
| 114 | + self, float16_stripped_tif |
| 115 | + ): |
| 116 | + from xrspatial.geotiff import read_geotiff_gpu |
| 117 | + |
| 118 | + path, arr = float16_stripped_tif |
| 119 | + result = read_geotiff_gpu(str(path)) |
| 120 | + assert result.dtype == np.float32, ( |
| 121 | + f"GPU read of float16 must return float32, got {result.dtype}" |
| 122 | + ) |
| 123 | + np.testing.assert_array_equal( |
| 124 | + result.data.get(), arr.astype(np.float32)) |
| 125 | + |
| 126 | + def test_read_geotiff_gpu_tiled_returns_float32( |
| 127 | + self, float16_tiled_tif |
| 128 | + ): |
| 129 | + from xrspatial.geotiff import read_geotiff_gpu |
| 130 | + |
| 131 | + path, arr = float16_tiled_tif |
| 132 | + result = read_geotiff_gpu(str(path)) |
| 133 | + assert result.dtype == np.float32 |
| 134 | + np.testing.assert_array_equal( |
| 135 | + result.data.get(), arr.astype(np.float32)) |
| 136 | + |
| 137 | + def test_read_geotiff_gpu_tiled_uncompressed_returns_float32( |
| 138 | + self, float16_tiled_uncompressed_tif |
| 139 | + ): |
| 140 | + from xrspatial.geotiff import read_geotiff_gpu |
| 141 | + |
| 142 | + path, arr = float16_tiled_uncompressed_tif |
| 143 | + result = read_geotiff_gpu(str(path)) |
| 144 | + assert result.dtype == np.float32 |
| 145 | + np.testing.assert_array_equal( |
| 146 | + result.data.get(), arr.astype(np.float32)) |
| 147 | + |
| 148 | + def test_open_geotiff_gpu_dispatcher_float16(self, float16_tiled_tif): |
| 149 | + """``open_geotiff(gpu=True)`` dispatches correctly for float16.""" |
| 150 | + from xrspatial.geotiff import open_geotiff |
| 151 | + |
| 152 | + path, arr = float16_tiled_tif |
| 153 | + result = open_geotiff(str(path), gpu=True) |
| 154 | + assert result.dtype == np.float32 |
| 155 | + np.testing.assert_array_equal( |
| 156 | + result.data.get(), arr.astype(np.float32)) |
| 157 | + |
| 158 | + |
| 159 | +class TestGPUWindowedFloat16: |
| 160 | + """Windowed GPU reads honour the bps_mismatch fallback path.""" |
| 161 | + |
| 162 | + def test_read_geotiff_gpu_windowed_stripped(self, float16_stripped_tif): |
| 163 | + from xrspatial.geotiff import read_geotiff_gpu |
| 164 | + |
| 165 | + path, arr = float16_stripped_tif |
| 166 | + result = read_geotiff_gpu(str(path), window=(0, 0, 2, 2)) |
| 167 | + assert result.dtype == np.float32 |
| 168 | + assert result.shape == (2, 2) |
| 169 | + np.testing.assert_array_equal( |
| 170 | + result.data.get(), arr[:2, :2].astype(np.float32)) |
| 171 | + |
| 172 | + def test_read_geotiff_gpu_windowed_tiled(self, float16_tiled_tif): |
| 173 | + from xrspatial.geotiff import read_geotiff_gpu |
| 174 | + |
| 175 | + path, arr = float16_tiled_tif |
| 176 | + result = read_geotiff_gpu(str(path), window=(0, 0, 8, 8)) |
| 177 | + assert result.dtype == np.float32 |
| 178 | + assert result.shape == (8, 8) |
| 179 | + np.testing.assert_array_equal( |
| 180 | + result.data.get(), arr[:8, :8].astype(np.float32)) |
| 181 | + |
| 182 | + |
| 183 | +class TestDaskGPUFloat16: |
| 184 | + """``open_geotiff(chunks=, gpu=True)`` decodes float16 correctly.""" |
| 185 | + |
| 186 | + def test_dask_gpu_tiled_float16(self, float16_tiled_tif): |
| 187 | + from xrspatial.geotiff import open_geotiff |
| 188 | + |
| 189 | + path, arr = float16_tiled_tif |
| 190 | + result = open_geotiff(str(path), chunks=8, gpu=True) |
| 191 | + assert result.dtype == np.float32, ( |
| 192 | + f"dask+GPU read of float16 must return float32, got {result.dtype}" |
| 193 | + ) |
| 194 | + # Compute the dask array; under dask+cupy, .compute() yields a |
| 195 | + # cupy-backed DataArray, so the .data.get() step pulls to host. |
| 196 | + computed = result.compute() |
| 197 | + np.testing.assert_array_equal( |
| 198 | + computed.data.get(), arr.astype(np.float32)) |
| 199 | + |
| 200 | + def test_read_geotiff_gpu_chunks_kwarg_float16(self, float16_tiled_tif): |
| 201 | + """``read_geotiff_gpu(chunks=)`` also routes correctly.""" |
| 202 | + from xrspatial.geotiff import read_geotiff_gpu |
| 203 | + |
| 204 | + path, arr = float16_tiled_tif |
| 205 | + result = read_geotiff_gpu(str(path), chunks=8) |
| 206 | + assert result.dtype == np.float32 |
| 207 | + computed = result.compute() |
| 208 | + np.testing.assert_array_equal( |
| 209 | + computed.data.get(), arr.astype(np.float32)) |
| 210 | + |
| 211 | + |
| 212 | +class TestGDSPathGatedOffForFloat16: |
| 213 | + """``_gds_chunk_path_available`` returns False for (bps=16, sf=3). |
| 214 | +
|
| 215 | + Direct structural test of the gating logic added in #1941 to keep the |
| 216 | + KvikIO GDS chunked path from mis-decoding half-precision tiles. A |
| 217 | + regression dropping the float16 guard would silently corrupt every |
| 218 | + chunked GPU read of a float16 source. |
| 219 | + """ |
| 220 | + |
| 221 | + def test_gds_path_gated_off_for_float16(self, float16_tiled_tif): |
| 222 | + pytest.importorskip("kvikio", exc_type=ImportError) |
| 223 | + |
| 224 | + from xrspatial.geotiff._backends.gpu import _gds_chunk_path_available |
| 225 | + from xrspatial.geotiff._header import parse_all_ifds, parse_header |
| 226 | + |
| 227 | + path, _ = float16_tiled_tif |
| 228 | + with open(str(path), "rb") as f: |
| 229 | + data = f.read() |
| 230 | + header = parse_header(data) |
| 231 | + ifds = parse_all_ifds(data, header) |
| 232 | + ifd = ifds[0] |
| 233 | + |
| 234 | + # Sanity-check fixture: tiled, bps=16, sample_format=3 (float) |
| 235 | + from xrspatial.geotiff._dtypes import SAMPLE_FORMAT_FLOAT |
| 236 | + assert ifd.is_tiled, "fixture sanity: tiled layout expected" |
| 237 | + # Mirror the production unpacking pattern at gpu.py:791 |
| 238 | + # (bps_first[0] if bps_first else 0) so an empty BitsPerSample |
| 239 | + # tuple would not raise IndexError here. |
| 240 | + bps_first = ifd.bits_per_sample |
| 241 | + if isinstance(bps_first, tuple): |
| 242 | + bps = bps_first[0] if bps_first else 0 |
| 243 | + else: |
| 244 | + bps = bps_first |
| 245 | + assert bps == 16, "fixture sanity: bps=16 expected" |
| 246 | + assert ifd.sample_format == SAMPLE_FORMAT_FLOAT |
| 247 | + |
| 248 | + result = _gds_chunk_path_available( |
| 249 | + str(path), ifd, has_sparse_tile=False, orientation=1) |
| 250 | + assert result is False, ( |
| 251 | + "_gds_chunk_path_available must return False for " |
| 252 | + "(bps=16, sf=float) so the GDS chunked path does not " |
| 253 | + "mis-decode half-precision tiles." |
| 254 | + ) |
| 255 | + |
| 256 | + def test_gds_path_allowed_for_float32_tiled(self, tmp_path): |
| 257 | + """Sanity: GDS path remains allowed for a float32 tiled file. |
| 258 | +
|
| 259 | + Pins that the float16 guard at gpu.py:791 fires only on |
| 260 | + (bps=16, sf=float), not on every tiled float file. A regression |
| 261 | + widening the guard to all floats would silently disable the |
| 262 | + GDS path on every float32 tiled COG. |
| 263 | + """ |
| 264 | + tifffile = pytest.importorskip("tifffile") |
| 265 | + pytest.importorskip("kvikio", exc_type=ImportError) |
| 266 | + |
| 267 | + arr = np.arange(256, dtype=np.float32).reshape(16, 16) |
| 268 | + path = tmp_path / "f32_tiled.tif" |
| 269 | + tifffile.imwrite( |
| 270 | + str(path), arr, compression="deflate", tile=(16, 16)) |
| 271 | + |
| 272 | + from xrspatial.geotiff._backends.gpu import _gds_chunk_path_available |
| 273 | + from xrspatial.geotiff._header import parse_all_ifds, parse_header |
| 274 | + |
| 275 | + with open(str(path), "rb") as f: |
| 276 | + data = f.read() |
| 277 | + header = parse_header(data) |
| 278 | + ifds = parse_all_ifds(data, header) |
| 279 | + |
| 280 | + result = _gds_chunk_path_available( |
| 281 | + str(path), ifds[0], has_sparse_tile=False, orientation=1) |
| 282 | + assert result is True, ( |
| 283 | + "_gds_chunk_path_available must remain True for " |
| 284 | + "(bps=32, sf=float) tiled files so the kvikio GDS chunk " |
| 285 | + "path still applies." |
| 286 | + ) |
| 287 | + |
| 288 | + |
| 289 | +class TestBackendParityFloat16: |
| 290 | + """All four backends agree pixel-exact on float16 input.""" |
| 291 | + |
| 292 | + def test_eager_numpy_equals_gpu(self, float16_tiled_tif): |
| 293 | + from xrspatial.geotiff import open_geotiff |
| 294 | + |
| 295 | + path, _ = float16_tiled_tif |
| 296 | + cpu = open_geotiff(str(path)) |
| 297 | + gpu = open_geotiff(str(path), gpu=True) |
| 298 | + |
| 299 | + assert cpu.dtype == gpu.dtype == np.float32 |
| 300 | + np.testing.assert_array_equal(np.asarray(cpu), gpu.data.get()) |
| 301 | + |
| 302 | + def test_eager_numpy_equals_dask_gpu(self, float16_tiled_tif): |
| 303 | + from xrspatial.geotiff import open_geotiff |
| 304 | + |
| 305 | + path, _ = float16_tiled_tif |
| 306 | + cpu = open_geotiff(str(path)) |
| 307 | + dask_gpu = open_geotiff(str(path), chunks=8, gpu=True).compute() |
| 308 | + |
| 309 | + assert cpu.dtype == dask_gpu.dtype == np.float32 |
| 310 | + np.testing.assert_array_equal( |
| 311 | + np.asarray(cpu), dask_gpu.data.get()) |
| 312 | + |
| 313 | + def test_dask_numpy_equals_dask_gpu(self, float16_tiled_tif): |
| 314 | + from xrspatial.geotiff import open_geotiff, read_geotiff_dask |
| 315 | + |
| 316 | + path, _ = float16_tiled_tif |
| 317 | + dask_cpu = read_geotiff_dask(str(path), chunks=8).compute() |
| 318 | + dask_gpu = open_geotiff(str(path), chunks=8, gpu=True).compute() |
| 319 | + |
| 320 | + np.testing.assert_array_equal( |
| 321 | + np.asarray(dask_cpu), dask_gpu.data.get()) |
| 322 | + |
| 323 | + |
| 324 | +class TestPredictor3Float16GPU: |
| 325 | + """Predictor=3 + float16 on disk also decodes correctly on GPU.""" |
| 326 | + |
| 327 | + def test_predictor3_float16_gpu_round_trip(self, tmp_path): |
| 328 | + tifffile = pytest.importorskip("tifffile") |
| 329 | + pytest.importorskip("imagecodecs") # required for predictor=3 |
| 330 | + |
| 331 | + from xrspatial.geotiff import read_geotiff_gpu |
| 332 | + |
| 333 | + arr = np.linspace(-1.0, 1.0, 16).astype(np.float16).reshape(4, 4) |
| 334 | + path = tmp_path / "pred3_f16.tif" |
| 335 | + tifffile.imwrite( |
| 336 | + str(path), arr, predictor=3, compression="deflate") |
| 337 | + |
| 338 | + result = read_geotiff_gpu(str(path)) |
| 339 | + assert result.dtype == np.float32 |
| 340 | + np.testing.assert_array_equal( |
| 341 | + result.data.get(), arr.astype(np.float32)) |
0 commit comments