Merge pull request #2627 from djhoese/bugfix-abi-16bit-chunks

Fix ABI readers using wrong dtype for resolution-based chunks

satpy/readers/abi_base.py

@@ -94,10 +94,10 @@ def _chunk_bytes_for_resolution(self) -> int:
         # this is true for all CSPP Geo GRB output (226 for all sectors) and full disk from other sources
         # 250 has been seen for AWS/CLASS CONUS, Mesoscale 1, and Mesoscale 2 files
         # we align this with 4 on-disk chunks at 500m, so it will be 2 on-disk chunks for 1km, and 1 for 2km
-        high_res_elems_disk_aligned = np.round(max(num_high_res_elems_per_dim / (4 * 226), 1)) * (4 * 226)
+        high_res_elems_disk_aligned = round(max(num_high_res_elems_per_dim / (4 * 226), 1)) * (4 * 226)
         low_res_factor = int(self.filetype_info.get("resolution", 2000) // 500)
         res_elems_per_dim = int(high_res_elems_disk_aligned / low_res_factor)
-        return (res_elems_per_dim ** 2) * 4
+        return (res_elems_per_dim ** 2) * 2  # 16-bit integers on disk
 
     @staticmethod
     def _rename_dims(nc):

satpy/tests/reader_tests/test_abi_l1b.py

@@ -136,18 +136,14 @@ def generate_l1b_filename(chan_name: str) -> str:
 
 @pytest.fixture()
 def c01_refl(tmp_path) -> xr.DataArray:
-    # 4 bytes for 32-bit floats
-    # 4 on-disk chunks for 500 meter data
-    # 226 on-disk chunk size
-    # Square (**2) for 2D size
-    with dask.config.set({"array.chunk-size": ((226 * 4) ** 2) * 4}):
+    with _apply_dask_chunk_size():
         reader = _create_reader_for_data(tmp_path, "C01", None, 1000)
         return reader.load(["C01"])["C01"]
 
 
 @pytest.fixture()
 def c01_rad(tmp_path) -> xr.DataArray:
-    with dask.config.set({"array.chunk-size": ((226 * 4) ** 2) * 4}):
+    with _apply_dask_chunk_size():
         reader = _create_reader_for_data(tmp_path, "C01", None, 1000)
         return reader.load([DataQuery(name="C01", calibration="radiance")])["C01"]
 
@@ -169,14 +165,14 @@ def c01_rad_h5netcdf(tmp_path) -> xr.DataArray:
             "valid_range": (0, 4095),
         },
     )
-    with dask.config.set({"array.chunk-size": ((226 * 4) ** 2) * 4}):
+    with _apply_dask_chunk_size():
         reader = _create_reader_for_data(tmp_path, "C01", rad, 1000)
         return reader.load([DataQuery(name="C01", calibration="radiance")])["C01"]
 
 
 @pytest.fixture()
 def c01_counts(tmp_path) -> xr.DataArray:
-    with dask.config.set({"array.chunk-size": ((226 * 4) ** 2) * 4}):
+    with _apply_dask_chunk_size():
         reader = _create_reader_for_data(tmp_path, "C01", None, 1000)
         return reader.load([DataQuery(name="C01", calibration="counts")])["C01"]
 
@@ -187,7 +183,7 @@ def _load_data_array(
         clip_negative_radiances: bool = False,
     ):
         rad = _fake_c07_data()
-        with dask.config.set({"array.chunk-size": ((226 * 4) ** 2) * 4}):
+        with _apply_dask_chunk_size():
             reader = _create_reader_for_data(
                 tmp_path,
                 "C07",
@@ -241,14 +237,22 @@ def _create_reader_for_data(
     return load_readers([str(data_path)], "abi_l1b", reader_kwargs=reader_kwargs)["abi_l1b"]
 
 
+def _apply_dask_chunk_size():
+    # 226 on-disk chunk size
+    # 8 on-disk chunks for 500 meter data
+    # Square (**2) for 2D size
+    # 4 bytes for 32-bit floats
+    return dask.config.set({"array.chunk-size": ((226 * 8) ** 2) * 4})
+
+
 def _get_and_check_array(data_arr: xr.DataArray, exp_dtype: npt.DTypeLike) -> npt.NDArray:
     data_np = data_arr.data.compute()
     assert isinstance(data_arr, xr.DataArray)
     assert isinstance(data_arr.data, da.Array)
     assert isinstance(data_np, np.ndarray)
     res = 1000 if RAD_SHAPE[1000][0] == data_np.shape[0] else 2000
-    assert data_arr.chunks[0][0] == 226 * (4 / (res / 500))
-    assert data_arr.chunks[1][0] == 226 * (4 / (res / 500))
+    assert data_arr.chunks[0][0] == 226 * (8 / (res / 500))
+    assert data_arr.chunks[1][0] == 226 * (8 / (res / 500))
 
     assert data_np.dtype == data_arr.dtype
     assert data_np.dtype == exp_dtype