Workaround arm64 gcc error in std::copysign (pytorch#51900)

Summary:
Move definition of copysign template and specialization for
bfloat16/half types before first use of copysign in that file

Add comment explaining why this is necessary

Fixes pytorch#51889

Pull Request resolved: pytorch#51900

Reviewed By: walterddr

Differential Revision: D26321741

Pulled By: malfet

fbshipit-source-id: 888858b11d9708fa140fe9c0570cc5a24599205b

aten/src/ATen/native/cpu/BinaryOpsKernel.cpp

@@ -19,6 +19,27 @@ namespace {
 
 using namespace vec256;
 
+// Note: Explicit implementation of copysign for Half and BFloat16
+// is needed to workaround g++-7/8 crash on aarch64, but also makes
+// copysign faster for the half-precision types
+template<typename T>
+T copysign(T a, T b) {
+  return std::copysign(a, b);
+}
+
+// Implement copysign for half precision floats using bit ops
+// Sign is the most significant bit for both half and bfloat16 types
+template<>
+c10::Half copysign(c10::Half a, c10::Half b) {
+  return c10::Half((a.x&0x7fff) | (b.x&0x8000), c10::Half::from_bits());
+}
+
+template<>
+c10::BFloat16 copysign(c10::BFloat16 a, c10::BFloat16 b) {
+   return c10::BFloat16((a.x&0x7fff) | (b.x&0x8000), c10::BFloat16::from_bits());
+}
+
+
 // Note: Undefined behavior when performing addition is intentionally
 // ignored.
 void add_kernel(TensorIteratorBase& iter, Scalar alpha_scalar) {
@@ -180,7 +201,7 @@ void div_floor_kernel(TensorIterator& iter) {
                 floordiv += scalar_t(1.0);
               }
             } else {
-              floordiv = std::copysign(scalar_t(0), a / b);
+              floordiv = copysign(scalar_t(0), a / b);
             }
             return floordiv;
           });
@@ -889,23 +910,6 @@ void heaviside_kernel(TensorIterator& iter) {
   });
 }
 
-template<typename T>
-T copysign(T a, T b) {
-  return std::copysign(a, b);
-}
-
-// Implement copysign for half precision floats using bit ops
-// Sign is the most significant bit for both half and bfloat16 types
-template<>
-c10::Half copysign(c10::Half a, c10::Half b) {
-  return c10::Half((a.x&0x7fff) | (b.x&0x8000), c10::Half::from_bits());
-}
-
-template<>
-c10::BFloat16 copysign(c10::BFloat16 a, c10::BFloat16 b) {
-   return c10::BFloat16((a.x&0x7fff) | (b.x&0x8000), c10::BFloat16::from_bits());
-}
-
 void copysign_kernel(TensorIterator& iter) {
   AT_DISPATCH_FLOATING_TYPES_AND2(kBFloat16, kHalf, iter.common_dtype(), "copysign_cpu", [&]() {
     cpu_kernel(iter, [](scalar_t a, scalar_t b) -> scalar_t {