Relax sin/cos ULP test for float32 on non-FMA

Allow up to 3 ULP error for float32 sin/cos when native
FMA is not available.

Author: seiko2plus

numpy/_core/tests/test_umath.py

@@ -34,6 +34,7 @@
 from numpy.testing._private.utils import (
     LONG_DOUBLE_IS_IBM_DOUBLE_DOUBLE,
     _glibc_older_than,
+    _ufunc_has_fma,
 )
 
 UFUNCS = [obj for obj in np._core.umath.__dict__.values()
@@ -2202,19 +2203,30 @@ def test_sincos_float32(self):
             # this is known to be problematic on old glibc, so skip it there
             x_f32[index] = np.float32(10E+10 * np.random.rand(M))
         x_f64 = np.float64(x_f32)
-        assert_array_max_ulp(np.sin(x_f32), np.float32(np.sin(x_f64)), maxulp=2)
-        assert_array_max_ulp(np.cos(x_f32), np.float32(np.cos(x_f64)), maxulp=2)
+        # numpy-simd-rounding gives ~3 ulp when fma is not available and kLowAccuracy
+        # been set in the precision mode
+        maxulp_f32 = 2 if _ufunc_has_fma(np.sin, np.float32) else 3
+        assert_array_max_ulp(
+            np.sin(x_f32),
+            np.float32(np.sin(x_f64)),
+            maxulp=maxulp_f32
+        )
+        assert_array_max_ulp(
+            np.cos(x_f32),
+            np.float32(np.cos(x_f64)),
+            maxulp=maxulp_f32
+        )
         # test aliasing(issue #17761)
         tx_f32 = x_f32.copy()
         assert_array_max_ulp(
             np.sin(x_f32, out=x_f32),
             np.float32(np.sin(x_f64)),
-            maxulp=2,
+            maxulp=maxulp_f32,
         )
         assert_array_max_ulp(
             np.cos(tx_f32, out=tx_f32),
             np.float32(np.cos(x_f64)),
-            maxulp=2,
+            maxulp=maxulp_f32,
         )
 
     def test_strided_float32(self):

numpy/testing/_private/utils.py

@@ -2733,6 +2733,16 @@ def check_free_memory(free_bytes):
     return msg if mem_free < free_bytes else None
 
 
+# avx2 implies fma
+_fma_pattern = re.compile("fma|avx2|avx512|vsx|vx|asimd|vpfv4|rvv|lsx", re.IGNORECASE)
+
+def _ufunc_has_fma(ufunc, dtype) -> bool:
+    """Check if a ufunc has SIMD kernel with native FMA support for a given dtype."""
+    ufunc_name = ufunc.__name__
+    res = np.lib.introspect.opt_func_info(ufunc_name, signature=dtype.__name__)
+    dres = res[ufunc_name]
+    return all(bool(_fma_pattern.search(v["current"])) for v in dres.values())
+
 def _parse_size(size_str):
     """Convert memory size strings ('12 GB' etc.) to float"""
     suffixes = {'': 1, 'b': 1,