Preserve NumPy sample values across non-native byte orders

soundfile.py

@@ -934,7 +934,8 @@ def read(self, frames: int = -1, dtype: dtype_str = 'float64',
             array instead of creating a new array. In this case, the
             arguments *dtype* and *always_2d* are silently ignored! If
             *frames* is not given, it is obtained from the length of
-            *out*.
+            *out*. The array must be C-contiguous and have native byte
+            order.
 
         Examples
         --------
@@ -1086,7 +1087,8 @@ def write(self, data: AudioData) -> None:
         import numpy as np
 
         # no copy is made if data has already the correct memory layout:
-        data = np.ascontiguousarray(data)
+        data = np.asarray(data)
+        data = np.ascontiguousarray(data, dtype=data.dtype.newbyteorder('='))
         written = self._array_io('write', data, len(data))
         assert written == len(data)
         self._update_frames(written)
@@ -1421,6 +1423,8 @@ def _array_io(self, action, array, frames):
             raise ValueError("Data must be C-contiguous")
         ctype = self._check_dtype(array.dtype.name)
         assert array.dtype.itemsize == _ffi.sizeof(ctype)
+        if not array.dtype.isnative:
+            raise ValueError("Data must have native byte order")
         cdata = _ffi.cast(ctype + '*', array.__array_interface__['data'][0])
         return self._cdata_io(action, cdata, ctype, frames)
 

tests/test_byteorder.py

@@ -0,0 +1,77 @@
+import io
+import sys
+
+import numpy as np
+import pytest
+from numpy.testing import assert_array_equal
+
+import soundfile as sf
+
+
+@pytest.mark.parametrize(
+    'dtype,subtype',
+    [
+        ('int16', 'PCM_16'),
+        ('int32', 'PCM_32'),
+        ('float32', 'FLOAT'),
+        ('float64', 'DOUBLE'),
+    ],
+)
+@pytest.mark.parametrize('channels', [1, 2])
+@pytest.mark.parametrize('strided', [False, True])
+def test_write_non_native_endian_preserves_samples(dtype, subtype, channels, strided):
+    order = '>' if sys.byteorder == 'little' else '<'
+    values = (
+        np.array([1, 256, -2345, 42], dtype=dtype)
+        if dtype.startswith('int')
+        else np.array([0.25, 0.5, -0.75, 0.125], dtype=dtype)
+    )
+    if channels == 2:
+        values = np.column_stack((values, -values))
+    swapped = values.astype(np.dtype(dtype).newbyteorder(order))
+    if strided:
+        backing = np.empty((len(values) * 2,) + values.shape[1:], dtype=swapped.dtype)
+        backing[::2] = swapped
+        swapped = backing[::2]
+    snapshot = swapped.tobytes()
+    output = io.BytesIO()
+    sf.write(output, swapped, 8000, format='WAV', subtype=subtype)
+    output.seek(0)
+    decoded, _ = sf.read(output, dtype=dtype)
+    assert_array_equal(decoded, values)
+    assert swapped.tobytes() == snapshot
+
+
+@pytest.mark.parametrize(
+    'dtype,subtype',
+    [
+        ('int16', 'PCM_16'),
+        ('int32', 'PCM_32'),
+        ('float32', 'FLOAT'),
+        ('float64', 'DOUBLE'),
+    ],
+)
+@pytest.mark.parametrize('channels', [1, 2])
+@pytest.mark.parametrize('fill', [None, 17])
+def test_read_non_native_endian_out_rejected_without_mutation(
+    dtype, subtype, channels, fill
+):
+    order = '>' if sys.byteorder == 'little' else '<'
+    values = (
+        np.array([1, 256, -2345, 42], dtype=dtype)
+        if dtype.startswith('int')
+        else np.array([0.25, 0.5, -0.75, 0.125], dtype=dtype)
+    )
+    if channels == 2:
+        values = np.column_stack((values, -values))
+    output = io.BytesIO()
+    sf.write(output, values, 8000, format='WAV', subtype=subtype)
+    output.seek(0)
+    shape = (6,) if channels == 1 else (6, 2)
+    out = np.full(shape, 23, dtype=np.dtype(dtype).newbyteorder(order))
+    snapshot = out.tobytes()
+    with sf.SoundFile(output) as file:
+        with pytest.raises(ValueError, match='native byte order'):
+            file.read(out=out, fill_value=fill)
+        assert file.tell() == 0
+    assert out.tobytes() == snapshot