Merge pull request #269 from vchamarthi/fix/absolute-nan

BUG: fix absolute() returning -NaN for +NaN on scalar path

Author: ndgrigorian

CHANGELOG.md

@@ -14,6 +14,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 * Pinned Cython in the Coverity Scan workflow so generated code stays stable between scans, and added `coverity/README.md` documenting the known Cython-boilerplate false positives and the scan review checklist [gh-266](https://github.com/IntelPython/mkl_umath/pull/266)
 
 ### Fixed
+* Fixed `absolute` (float32/float64) returning `-NaN` for `+NaN` input on the scalar fallback path (same issue as numpy/numpy@dc478c58b9, gh-31433) [gh-269](https://github.com/IntelPython/mkl_umath/pull/269)
 * Fixed an over-decref of the borrowed module dictionary reference on the module initialization error path [gh-264](https://github.com/IntelPython/mkl_umath/pull/264)
 * Fixed a leak of the module object when the NumPy C-API import fails during module initialization [gh-263](https://github.com/IntelPython/mkl_umath/pull/263)
 

mkl_umath/src/mkl_umath_loops.c.src

@@ -917,7 +917,7 @@ mkl_umath_@TYPE@_absolute(char **args, const npy_intp *dimensions, const npy_int
     else {
         UNARY_LOOP {
             const @type@ in1 = *(@type@ *)ip1;
-            const @type@ tmp = in1 > 0 ? in1 : -in1;
+            const @type@ tmp = fabs@c@(in1);
             /* add 0 to clear -0.0 */
             *((@type@ *)op1) = tmp + 0;
         }

mkl_umath/tests/test_basic.py

@@ -189,3 +189,16 @@ def test_reduce_complex(func, dtype):
     assert np.allclose(
         mkl_res, np_res
     ), f"Results for '{func}[reduce]' do not match"
+
+
+@pytest.mark.parametrize("size", [100, 8192 + 1])
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_absolute_nan_signbit(size, dtype):
+    # size 100 takes the fallback loop, 8193 takes MKL's vAbs
+    for nan in (np.nan, -np.nan):
+        a = np.full(size, nan, dtype=dtype)
+        sign = "-" if np.signbit(nan) else "+"
+        assert not np.signbit(mu.absolute(a)).any(), (
+            f"absolute({sign}nan) must be +nan for "
+            f"{dtype.__name__} at size {size}"
+        )