[BE] Nested namespace in ATen/native headers (pytorch#103753)

Use nested namespace and `enum class` in `ATen/native` headers.
In particular, it helps avoid polluting global namespace with `MAX`,`MIN` enum values.

Pull Request resolved: pytorch#103753
Approved by: https://github.com/atalman, https://github.com/Skylion007

Author: malfet

aten/src/ATen/native/Activation.h

@@ -14,10 +14,10 @@ struct TensorIteratorBase;
 class TensorBase;
 }
 
-namespace at { namespace native {
+namespace at::native {
 
 // These constants control the approximation behavior of gelu function.
-enum GeluType {
+enum class GeluType {
   None,             // Baseline Gelu
   Tanh,             // Tahn Gelu Approximation
   END
@@ -33,6 +33,14 @@ static GeluType get_gelutype_enum(const c10::string_view approximate) {
   }
 }
 
+static std::string gelutype_to_string(const GeluType type) {
+  switch(type) {
+    case GeluType::None: return "none";
+    case GeluType::Tanh: return "tanh";
+    default: TORCH_CHECK(false, "unknown GELU type: ", static_cast<int>(type));
+  }
+}
+
 using structured_activation_fn = void (*)(TensorIteratorBase&);
 using structured_activation_backward_fn = void (*)(TensorIteratorBase&);
 
@@ -87,6 +95,4 @@ DECLARE_DISPATCH(activation_backward_fn, mish_backward_stub);
 DECLARE_DISPATCH(activation_fn, prelu_stub);
 DECLARE_DISPATCH(activation_backward_fn, prelu_backward_stub);
 
-} // namespace native
-
-} // namespace at
+} // namespace at::native

aten/src/ATen/native/AdaptivePooling.h

@@ -6,9 +6,7 @@
 #include <c10/util/irange.h>
 #include <cmath>
 
-namespace at {
-
-namespace native {
+namespace at::native {
 
 using adaptive_avg_pooling_fn = void(*)(Tensor& output, const Tensor& input, IntArrayRef output_size);
 using adaptive_avg_pooling_backward_fn = void(*)(Tensor& grad_input, const Tensor& grad_output);
@@ -38,4 +36,4 @@ static inline void adaptive_pool_empty_output_check(const Tensor& gradOutput_, c
   }
 }
 
-}} // namespace at::native
+} // namespace at::native

aten/src/ATen/native/BatchLinearAlgebra.h

@@ -15,7 +15,7 @@ enum class TransposeType;
 
 }
 
-namespace at { namespace native {
+namespace at::native {
 
 enum class LapackLstsqDriverType : int64_t { Gels, Gelsd, Gelsy, Gelss};
 
@@ -317,4 +317,4 @@ using ldl_solve_fn = void (*)(
     bool /*upper*/,
     bool /*hermitian*/);
 DECLARE_DISPATCH(ldl_solve_fn, ldl_solve_stub);
-}} // namespace at::native
+} // namespace at::native

aten/src/ATen/native/BinaryOps.h

@@ -21,7 +21,7 @@ struct TensorIterator;
 struct TensorIteratorBase;
 }
 
-namespace at { namespace native {
+namespace at::native {
 
 inline void alpha_check(const ScalarType dtype, const Scalar& alpha) {
   TORCH_CHECK(! alpha.isBoolean() || dtype == ScalarType::Bool,
@@ -180,4 +180,4 @@ DECLARE_DISPATCH(structured_binary_fn, shifted_chebyshev_polynomial_u_stub);
 DECLARE_DISPATCH(structured_binary_fn, shifted_chebyshev_polynomial_v_stub);
 DECLARE_DISPATCH(structured_binary_fn, shifted_chebyshev_polynomial_w_stub);
 
-}} // namespace at::native
+} // namespace at::native

aten/src/ATen/native/BucketizationUtils.h

@@ -10,8 +10,7 @@
 #include <ATen/ops/result_type.h>
 #endif
 
-namespace at {
-namespace native {
+namespace at::native {
 
 // original values given by raw_*. If an original value is not contiguous, will make a contiguous copy to
 // the corresponding trimmed_* value. Additionally, if the dtypes of the boundary and input tensor do not
@@ -171,4 +170,4 @@ inline void searchsorted_pre_check(
   }
 }
 
-}}
+} // namespace at::native

aten/src/ATen/native/CPUBlas.h

@@ -7,9 +7,7 @@
 #include <c10/core/ScalarType.h>
 #include <c10/core/Scalar.h>
 
-namespace at {
-namespace native {
-namespace cpublas {
+namespace at::native::cpublas {
 
 namespace internal {
 void normalize_last_dims(
@@ -161,4 +159,4 @@ void copy(int64_t n, const float *x, int64_t incx, float *y, int64_t incy);
 void copy(int64_t n, const c10::complex<double> *x, int64_t incx, c10::complex<double> *y, int64_t incy);
 void copy(int64_t n, const c10::complex<float> *x, int64_t incx, c10::complex<float> *y, int64_t incy);
 
-}}}  // namespace at::native::cpublas
+}  // namespace at::native::cpublas

aten/src/ATen/native/CPUFallback.h

@@ -7,7 +7,7 @@
 #include <c10/util/Metaprogramming.h>
 #include <torch/library.h>
 
-namespace at { namespace native {
+namespace at::native {
 
 // This function implements a boxed fallback to CPU.
 // External backends can add their own custom logging on top if it to customize their own CPU fallbacks.
@@ -42,5 +42,4 @@ using call_fallback_fn_symint = _call_fallback_fn<fallback_fn, Op, true, typenam
 template<c10::KernelFunction::BoxedKernelFunction* fallback_fn, class Op>
 using call_fallback_fn = _call_fallback_fn<fallback_fn, Op, false, typename Op::schema>;
 
-} // namespace native
-} // namespace at
+} // namespace at::native

aten/src/ATen/native/CanUse32BitIndexMath.h

@@ -6,8 +6,8 @@ namespace at {
 class TensorBase;
 }
 
-namespace at { namespace native {
+namespace at::native {
 
 TORCH_API bool canUse32BitIndexMath(const at::TensorBase &t, int64_t max_elem=std::numeric_limits<int32_t>::max());
 
-}}
+}

aten/src/ATen/native/ComplexHelper.h

@@ -15,7 +15,7 @@
 // WARNING: this header contains non-inline functions and should be only
 // included from ONE cpp file
 
-namespace at { namespace native {
+namespace at::native {
 
 // View tensor with new dtype, storage offset, sizes and strides
 inline Tensor view_tensor(
@@ -94,4 +94,4 @@ Tensor view_as_complex(const Tensor& self) {
   return view_tensor(self, complex_type, new_storage_offset, new_sizes, new_strides);
 }
 
-}} // namespace at::native
+} // namespace at::native

aten/src/ATen/native/CompositeRandomAccessor.h

@@ -2,7 +2,7 @@
 
 #include <ATen/native/CompositeRandomAccessorCommon.h>
 
-namespace at { namespace native {
+namespace at::native {
 
 struct TupleInfoCPU {
   template <typename ...Types>
@@ -31,4 +31,4 @@ auto get(references_holder<Values, References> rh) -> decltype(std::get<N>(rh.da
   return std::get<N>(rh.data());
 }
 
-}} // namespace at::native
+} // namespace at::native