numpy · howjmay · Aug 28, 2021
diff --git a/benchmarks/benchmarks/bench_ufunc.py b/benchmarks/benchmarks/bench_ufunc.py
@@ -112,6 +112,9 @@ def time_double_add(self):
 
     def time_double_add_temp(self):
         1. + self.d + 1.
+
+    def time_copysign(self):
+        np.copysign(self.d, self.d)
 
 
 class CustomScalar(Benchmark):
@@ -166,7 +169,6 @@ def time_add_scalar_conv(self):
     def time_add_scalar_conv_complex(self):
         (self.y + self.z)
 
-
 class ArgPack:
     __slots__ = ['args', 'kwargs']
     def __init__(self, *args, **kwargs):

diff --git a/numpy/core/src/_simd/_simd.dispatch.c.src b/numpy/core/src/_simd/_simd.dispatch.c.src
@@ -395,7 +395,7 @@ SIMD_IMPL_INTRIN_2(@intrin@_@sfx@, v@sfx@, v@sfx@, v@sfx@)
 
 #if @fp_only@
 /**begin repeat1
- * #intrin = maxp, minp#
+ * #intrin = maxp, minp, copysign#
  */
 SIMD_IMPL_INTRIN_2(@intrin@_@sfx@, v@sfx@, v@sfx@, v@sfx@)
 /**end repeat1**/
@@ -629,7 +629,7 @@ SIMD_INTRIN_DEF(@intrin@_@sfx@)
 
 #if @fp_only@
 /**begin repeat1
- * #intrin = maxp, minp#
+ * #intrin = maxp, minp, copysign#
  */
 SIMD_INTRIN_DEF(@intrin@_@sfx@)
 /**end repeat1**/

diff --git a/numpy/core/src/common/simd/avx2/math.h b/numpy/core/src/common/simd/avx2/math.h
@@ -105,4 +105,14 @@ NPY_FINLINE npyv_s64 npyv_min_s64(npyv_s64 a, npyv_s64 b)
     return _mm256_blendv_epi8(a, b, _mm256_cmpgt_epi64(a, b));
 }
 
+// copysign
+NPY_FINLINE npyv_f32 npyv_copysign_f32(npyv_f32 a, npyv_f32 b)
+{
+    return _mm256_or_ps(a, _mm256_and_ps(b, _mm256_set1_ps(-0.0)));
+}
+NPY_FINLINE npyv_f64 npyv_copysign_f64(npyv_f64 a, npyv_f64 b)
+{
+    return _mm256_or_pd(a, _mm256_and_pd(b, _mm256_set1_pd(-0.0)));
+}
+
 #endif // _NPY_SIMD_AVX2_MATH_H
diff --git a/numpy/core/src/common/simd/avx512/operators.h b/numpy/core/src/common/simd/avx512/operators.h
@@ -321,4 +321,14 @@ NPY_FINLINE npyv_b32 npyv_notnan_f32(npyv_f32 a)
 NPY_FINLINE npyv_b64 npyv_notnan_f64(npyv_f64 a)
 { return _mm512_cmp_pd_mask(a, a, _CMP_ORD_Q); }
 
+// copysign
+NPY_FINLINE npyv_f32 npyv_copysign_f32(npyv_f32 a, npyv_f32 b)
+{
+    return npyv_or_f32(a, npyv_and_f32(b, _mm512_set1_ps(-0.0)));
+}
+NPY_FINLINE npyv_f64 npyv_copysign_f64(npyv_f64 a, npyv_f64 b)
+{
+    return npyv_or_f64(a, npyv_and_f64(b, _mm512_set1_pd(-0.0)));
+}
+
 #endif // _NPY_SIMD_AVX512_OPERATORS_H
diff --git a/numpy/core/src/common/simd/neon/math.h b/numpy/core/src/common/simd/neon/math.h
@@ -153,4 +153,16 @@ NPY_FINLINE npyv_s64 npyv_min_s64(npyv_s64 a, npyv_s64 b)
     return vbslq_s64(npyv_cmplt_s64(a, b), a, b);
 }
 
+// copysign
+NPY_FINLINE npyv_f32 npyv_copysign_f32(npyv_f32 a, npyv_f32 b)
+{
+    return vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a), vandq_u32(vreinterpretq_u32_f32(b), vdupq_n_u32(0x80000000))));
+}
+#if NPY_SIMD_F64
+    NPY_FINLINE npyv_f64 npyv_copysign_f64(npyv_f64 a, npyv_f64 b)
+    {
+        return vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a), vandq_u64(vreinterpretq_u64_f64(b), vdupq_n_u64(0x8000000000000000))));
+    }
+#endif
+
 #endif // _NPY_SIMD_NEON_MATH_H
diff --git a/numpy/core/src/common/simd/sse/math.h b/numpy/core/src/common/simd/sse/math.h
@@ -143,4 +143,14 @@ NPY_FINLINE npyv_s64 npyv_min_s64(npyv_s64 a, npyv_s64 b)
     return npyv_select_s64(npyv_cmplt_s64(a, b), a, b);
 }
 
+// copysign
+NPY_FINLINE npyv_f32 npyv_copysign_f32(npyv_f32 a, npyv_f32 b)
+{
+    return _mm_or_ps(a, _mm_and_ps(b, _mm_set1_ps(-0.0)));
+}
+NPY_FINLINE npyv_f64 npyv_copysign_f64(npyv_f64 a, npyv_f64 b)
+{
+    return _mm_or_pd(a, _mm_and_pd(b, _mm_set1_pd(-0.0)));
+}
+
 #endif // _NPY_SIMD_SSE_MATH_H
diff --git a/numpy/core/src/common/simd/vsx/math.h b/numpy/core/src/common/simd/vsx/math.h
@@ -69,4 +69,14 @@ NPY_FINLINE npyv_f64 npyv_square_f64(npyv_f64 a)
 #define npyv_min_u64 vec_min
 #define npyv_min_s64 vec_min
 
+// copysign
+NPY_FINLINE npyv_f32 npyv_copysign_f32(npyv_f32 a, npyv_f32 b)
+{
+    return npyv_or_f32(a, npyv_and_f32(b, npyv_setall_f32(-0.0)));
+}
+NPY_FINLINE npyv_f64 npyv_copysign_f64(npyv_f64 a, npyv_f64 b)
+{
+    return npyv_or_f64(a, npyv_and_f64(b, npyv_setall_f64(-0.0)));
+}
+
 #endif // _NPY_SIMD_VSX_MATH_H
diff --git a/numpy/core/tests/test_simd.py b/numpy/core/tests/test_simd.py
@@ -415,6 +415,16 @@ def test_special_cases(self):
         nnan = self.notnan(self.setall(self._nan()))
         assert nnan == [0]*self.nlanes
 
+    def test_copysign(self):
+        data_a = self._data()
+        data_b = self._data(reverse=True)
+        vdata_a = self.load(data_a)
+        vdata_b = self.load(data_b)
+
+        data_copysign = [math.copysign(a, b) for a, b in zip(data_a, data_b)]
+        vcopysign = self.copysign(vdata_a, vdata_b)
+        assert vcopysign == data_copysign
+
 class _SIMD_ALL(_Test_Utility):
     """
     To test all vector types at once