Failing tests with predictor = 2

[maxrjones]

There are a couple GDAL test images that I couldn't figure out how to decode properly in #12. Here's how to replicate the issue outside pytest:

Failure 1

download https://dagshub.com/maxrjones/virtual-tiff/raw/40997bc52e1c2d61014220f4a87849015a7555e0/tests/dvc/gdal_autotest/float32_LZW_predictor_2.tif

from obstore.store import LocalStore
from virtual_tiff.reader import create_manifest_store
import zarr
import rioxarray
import numpy as np

filepath = "float32_LZW_predictor_2.tif"

ms = create_manifest_store(
    filepath=filepath, group="0", store = LocalStore()
)
a = zarr.open_array(ms, mode="r", zarr_format=3, path="0")
observed = a[:]
expected = rioxarray.open_rasterio(filepath).data.squeeze()
np.testing.assert_allclose(observed, expected)

Failure 2

Download https://dagshub.com/maxrjones/virtual-tiff/raw/40997bc52e1c2d61014220f4a87849015a7555e0/tests/dvc/gdal_gcore/test_hgrid_with_subgrid.tif

from obstore.store import LocalStore
from virtual_tiff.reader import create_manifest_store
import zarr
import rioxarray
import numpy as np

filepath = "test_hgrid_with_subgrid.tif"

ms = create_manifest_store(
    filepath=filepath, group="0", store = LocalStore()
)
a = zarr.open_array(ms, mode="r", zarr_format=3, path="0")
observed = a[:]
expected = rioxarray.open_rasterio(filepath)[0]
expected = expected[filepath.replace("/", "_").lstrip("_")]
expected = expected.data.squeeze()
np.testing.assert_allclose(observed, expected)

[jsignell]

I just opened #47 which might be a duplicate of this. Predictor 2 should be a SpatialDelta as described in zarr-developers/numcodecs#583 and included in cell 3 in this notebook https://github.com/jkeifer/cog-vs-zarr/blob/main/notebooks/cog_vs_zarr.ipynb

[maxrjones]

I just opened #47 which might be a duplicate of this. Predictor 2 should be a SpatialDelta as described in zarr-developers/numcodecs#583 and included in cell 3 in this notebook jkeifer/cog-vs-zarr@main/notebooks/cog_vs_zarr.ipynb

This library includes a HorizontalDeltaCodec that is functionally equivalent to the proposed SpatialCodec, which works for many TIFFs with predictor = 2 (e.g., the uint32 version of the failing LZW + predict=2 test file). The failures are for a small subset of the predictor=2 TIFFs in GDAL's test suite, and I am not sure if there's a common cause for those issues.