Merge pull request #190 from asmeurer/sign-fix

Add a wrapper for sign for NumPy-likes

array_api_compat/common/_aliases.py

@@ -12,7 +12,7 @@
 from typing import NamedTuple
 import inspect
 
-from ._helpers import array_namespace, _check_device, device, is_torch_array
+from ._helpers import array_namespace, _check_device, device, is_torch_array, is_cupy_namespace
 
 # These functions are modified from the NumPy versions.
 
@@ -530,11 +530,26 @@ def unstack(x: ndarray, /, xp, *, axis: int = 0) -> Tuple[ndarray, ...]:
         raise ValueError("Input array must be at least 1-d.")
     return tuple(xp.moveaxis(x, axis, 0))
 
+# numpy 1.26 does not use the standard definition for sign on complex numbers
+
+def sign(x: ndarray, /, xp, **kwargs) -> ndarray:
+    if isdtype(x.dtype, 'complex floating', xp=xp):
+        out = (x/xp.abs(x, **kwargs))[...]
+        # sign(0) = 0 but the above formula would give nan
+        out[x == 0+0j] = 0+0j
+    else:
+        out = xp.sign(x, **kwargs)
+    # CuPy sign() does not propagate nans. See
+    # https://github.com/data-apis/array-api-compat/issues/136
+    if is_cupy_namespace(xp) and isdtype(x.dtype, 'real floating', xp=xp):
+        out[xp.isnan(x)] = xp.nan
+    return out[()]
+
 __all__ = ['arange', 'empty', 'empty_like', 'eye', 'full', 'full_like',
            'linspace', 'ones', 'ones_like', 'zeros', 'zeros_like',
            'UniqueAllResult', 'UniqueCountsResult', 'UniqueInverseResult',
            'unique_all', 'unique_counts', 'unique_inverse', 'unique_values',
            'astype', 'std', 'var', 'cumulative_sum', 'clip', 'permute_dims',
            'reshape', 'argsort', 'sort', 'nonzero', 'ceil', 'floor', 'trunc',
            'matmul', 'matrix_transpose', 'tensordot', 'vecdot', 'isdtype',
-           'unstack']
+           'unstack', 'sign']

array_api_compat/cupy/_aliases.py

@@ -62,6 +62,7 @@
 matmul = get_xp(cp)(_aliases.matmul)
 matrix_transpose = get_xp(cp)(_aliases.matrix_transpose)
 tensordot = get_xp(cp)(_aliases.tensordot)
+sign = get_xp(cp)(_aliases.sign)
 
 _copy_default = object()
 
@@ -109,13 +110,6 @@ def asarray(
 
         return cp.array(obj, dtype=dtype, **kwargs)
 
-def sign(x: ndarray, /) -> ndarray:
-    # CuPy sign() does not propagate nans. See
-    # https://github.com/data-apis/array-api-compat/issues/136
-    out = cp.sign(x)
-    out[cp.isnan(x)] = cp.nan
-    return out
-
 # These functions are completely new here. If the library already has them
 # (i.e., numpy 2.0), use the library version instead of our wrapper.
 if hasattr(cp, 'vecdot'):

array_api_compat/dask/array/_aliases.py

@@ -104,7 +104,7 @@ def _dask_arange(
 trunc = get_xp(np)(_aliases.trunc)
 matmul = get_xp(np)(_aliases.matmul)
 tensordot = get_xp(np)(_aliases.tensordot)
-
+sign = get_xp(np)(_aliases.sign)
 
 # asarray also adds the copy keyword, which is not present in numpy 1.0.
 def asarray(

array_api_compat/numpy/_aliases.py

@@ -62,6 +62,7 @@
 matmul = get_xp(np)(_aliases.matmul)
 matrix_transpose = get_xp(np)(_aliases.matrix_transpose)
 tensordot = get_xp(np)(_aliases.tensordot)
+sign = get_xp(np)(_aliases.sign)
 
 def _supports_buffer_protocol(obj):
     try:

torch-xfails.txt

@@ -56,6 +56,10 @@ array_api_tests/test_operators_and_elementwise_functions.py::test_pow[__pow__(x1
 array_api_tests/test_operators_and_elementwise_functions.py::test_remainder[__imod__(x1, x2)]
 array_api_tests/test_operators_and_elementwise_functions.py::test_subtract[__sub__(x1, x2)]
 
+# inverse trig functions are too inaccurate on CPU
+array_api_tests/test_operators_and_elementwise_functions.py::test_acos
+array_api_tests/test_operators_and_elementwise_functions.py::test_atan
+array_api_tests/test_operators_and_elementwise_functions.py::test_asin
 
 # overflow near float max
 array_api_tests/test_operators_and_elementwise_functions.py::test_log1p