fix: handle sparse matrix reading for big-endian byte order

src/elisa/data/base.py

@@ -21,6 +21,7 @@
     significance_lima,
 )
 from elisa.plot.misc import get_colors
+from elisa.util.misc import to_native_byteorder
 
 if TYPE_CHECKING:
     NDArray = np.ndarray
@@ -1138,8 +1139,11 @@ def __init__(
         photon_egrid = np.array(photon_egrid, dtype=np.float64, order='C')
         channel_emin = np.array(channel_emin, dtype=np.float64, order='C')
         channel_emax = np.array(channel_emax, dtype=np.float64, order='C')
-        response_matrix = coo_array(response_matrix)
-        response_matrix.eliminate_zeros()
+        if not isinstance(response_matrix, sparray):
+            # convert to native byteorder to be compatible with scipy.sparse
+            response_matrix = to_native_byteorder(response_matrix)
+            response_matrix = coo_array(response_matrix)
+            response_matrix.eliminate_zeros()
         channel = np.array(channel, dtype=str, order='C')
 
         photon_egrid_shape = np.shape(photon_egrid)

src/elisa/data/ogip.py

@@ -9,6 +9,7 @@
 from scipy.sparse import coo_array
 
 from elisa.data.base import ObservationData, ResponseData, SpectrumData
+from elisa.util.misc import to_native_byteorder
 
 if TYPE_CHECKING:
     NDArray = np.ndarray
@@ -629,6 +630,9 @@ def _read_response(self, file: str, response_id: int) -> dict:
         else:
             reduced_matrix = np.hstack(matrix)
 
+        # convert to native byteorder to be compatible with scipy.sparse
+        reduced_matrix = to_native_byteorder(reduced_matrix)
+
         sparse_matrix = coo_array(
             (reduced_matrix, (rows, cols)),
             shape=(len(response_data), channel_number),

src/elisa/util/misc.py

@@ -22,6 +22,8 @@
     from collections.abc import Sequence
     from typing import Callable, Literal, TypeVar
 
+    from numpy import ndarray as NDArray
+
     from elisa.util.typing import CompEval
 
     T = TypeVar('T')
@@ -571,3 +573,11 @@ def _wrapper_progress_bar(i, vals):
         return _wrapper_progress_bar
 
     return progress_bar_fori_loop
+
+
+def to_native_byteorder(arr: NDArray) -> NDArray:
+    """Convert an array to native byte order."""
+    if arr.dtype.byteorder != '=':
+        return arr.astype(arr.dtype.newbyteorder('='))
+    else:
+        return arr

tests/test_data.py

@@ -1,6 +1,9 @@
 import numpy as np
+import pytest
+from jax.experimental.sparse import BCSR
 
 from elisa.data.grouping import significance_gv, significance_lima
+from elisa.data.ogip import Response, ResponseData
 from elisa.models import PowerLaw
 
 
@@ -117,3 +120,36 @@ def test_data_grouping():
         data.spec_counts, data.back_counts, data.back_errors, data.back_ratio
     )
     assert np.all(sig >= scale)
+
+
+@pytest.mark.parametrize(
+    'file',
+    [
+        '../docs/notebooks/data/P011160500104_LE.rsp',
+        '../docs/notebooks/data/P011160500104_ME.rsp',
+        '../docs/notebooks/data/P011160500104_HE.rsp',
+    ],
+)
+def test_load_response(file):
+    # test Response against big-endian files
+    rsp = Response(file)
+    # test if the response matrix can be converted to a BCSR matrix in JAX
+    BCSR.from_scipy_sparse(rsp.matrix)
+    assert np.all(rsp.channel_fwhm > 0)
+
+
+def test_response():
+    # test ResponseData against different endianness
+    photon_egrid = np.linspace(1.0, 100.0, 101)
+    channel = np.arange(100)
+    channel_emin = photon_egrid[:-1]
+    channel_emax = photon_egrid[1:]
+    mat1 = np.eye(100).astype('<f4')
+    mat2 = np.eye(100).astype('>f4')
+    r1 = ResponseData(photon_egrid, channel_emin, channel_emax, mat1, channel)
+    r2 = ResponseData(photon_egrid, channel_emin, channel_emax, mat2, channel)
+    # test if the response matrix can be converted to a BCSR matrix in JAX
+    BCSR.from_scipy_sparse(r1.matrix)
+    BCSR.from_scipy_sparse(r2.matrix)
+    assert np.all(r1.matrix == r2.matrix)
+    assert np.all(r1.channel_fwhm == r2.channel_fwhm)