@smake/eigen 1.1.0 → 1.1.1

package/eigen/Eigen/src/Core/arch/GPU/Tuple.h

@@ -0,0 +1,268 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2021 The Eigen Team
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TUPLE_GPU
+#define EIGEN_TUPLE_GPU
+
+#include <type_traits>
+#include <utility>
+
+// This is a replacement of std::tuple that can be used in device code.
+
+namespace Eigen {
+namespace internal {
+namespace tuple_impl {
+
+// Internal tuple implementation.
+template <size_t N, typename... Types>
+class TupleImpl;
+
+// Generic recursive tuple.
+template <size_t N, typename T1, typename... Ts>
+class TupleImpl<N, T1, Ts...> {
+ public:
+  // Tuple may contain Eigen types.
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW
+
+  // Default constructor, enable if all types are default-constructible.
+  template <typename U1 = T1,
+            typename EnableIf = std::enable_if_t<std::is_default_constructible<U1>::value &&
+                                                 reduce_all<std::is_default_constructible<Ts>::value...>::value>>
+  constexpr EIGEN_DEVICE_FUNC TupleImpl() : head_{}, tail_{} {}
+
+  // Element constructor.
+  template <typename U1, typename... Us,
+            // Only enable if...
+            typename EnableIf = std::enable_if_t<
+                // the number of input arguments match, and ...
+                sizeof...(Us) == sizeof...(Ts) && (
+                                                      // this does not look like a copy/move constructor.
+                                                      N > 1 || std::is_convertible<U1, T1>::value)>>
+  constexpr EIGEN_DEVICE_FUNC TupleImpl(U1&& arg1, Us&&... args)
+      : head_(std::forward<U1>(arg1)), tail_(std::forward<Us>(args)...) {}
+
+  // The first stored value.
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE T1& head() { return head_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE const T1& head() const { return head_; }
+
+  // The tail values.
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE TupleImpl<N - 1, Ts...>& tail() { return tail_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE const TupleImpl<N - 1, Ts...>& tail() const { return tail_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void swap(TupleImpl& other) {
+    using numext::swap;
+    swap(head_, other.head_);
+    swap(tail_, other.tail_);
+  }
+
+  template <typename... UTypes>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TupleImpl& operator=(const TupleImpl<N, UTypes...>& other) {
+    head_ = other.head_;
+    tail_ = other.tail_;
+    return *this;
+  }
+
+  template <typename... UTypes>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TupleImpl& operator=(TupleImpl<N, UTypes...>&& other) {
+    head_ = std::move(other.head_);
+    tail_ = std::move(other.tail_);
+    return *this;
+  }
+
+ private:
+  // Allow related tuples to reference head_/tail_.
+  template <size_t M, typename... UTypes>
+  friend class TupleImpl;
+
+  T1 head_;
+  TupleImpl<N - 1, Ts...> tail_;
+};
+
+// Empty tuple specialization.
+template <>
+class TupleImpl<size_t(0)> {};
+
+template <typename TupleType>
+struct is_tuple : std::false_type {};
+
+template <typename... Types>
+struct is_tuple<TupleImpl<sizeof...(Types), Types...>> : std::true_type {};
+
+// Gets an element from a tuple.
+template <size_t Idx, typename T1, typename... Ts>
+struct tuple_get_impl {
+  using TupleType = TupleImpl<sizeof...(Ts) + 1, T1, Ts...>;
+  using ReturnType = typename tuple_get_impl<Idx - 1, Ts...>::ReturnType;
+
+  static constexpr EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE ReturnType& run(TupleType& tuple) {
+    return tuple_get_impl<Idx - 1, Ts...>::run(tuple.tail());
+  }
+
+  static constexpr EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE const ReturnType& run(const TupleType& tuple) {
+    return tuple_get_impl<Idx - 1, Ts...>::run(tuple.tail());
+  }
+};
+
+// Base case, getting the head element.
+template <typename T1, typename... Ts>
+struct tuple_get_impl<0, T1, Ts...> {
+  using TupleType = TupleImpl<sizeof...(Ts) + 1, T1, Ts...>;
+  using ReturnType = T1;
+
+  static constexpr EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE T1& run(TupleType& tuple) { return tuple.head(); }
+
+  static constexpr EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE const T1& run(const TupleType& tuple) { return tuple.head(); }
+};
+
+// Concatenates N Tuples.
+template <size_t NTuples, typename... Tuples>
+struct tuple_cat_impl;
+
+template <size_t NTuples, size_t N1, typename... Args1, size_t N2, typename... Args2, typename... Tuples>
+struct tuple_cat_impl<NTuples, TupleImpl<N1, Args1...>, TupleImpl<N2, Args2...>, Tuples...> {
+  using TupleType1 = TupleImpl<N1, Args1...>;
+  using TupleType2 = TupleImpl<N2, Args2...>;
+  using MergedTupleType = TupleImpl<N1 + N2, Args1..., Args2...>;
+
+  using ReturnType = typename tuple_cat_impl<NTuples - 1, MergedTupleType, Tuples...>::ReturnType;
+
+  // Uses the index sequences to extract and merge elements from tuple1 and tuple2,
+  // then recursively calls again.
+  template <typename Tuple1, size_t... I1s, typename Tuple2, size_t... I2s, typename... MoreTuples>
+  static constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ReturnType run(Tuple1&& tuple1, std::index_sequence<I1s...>,
+                                                                        Tuple2&& tuple2, std::index_sequence<I2s...>,
+                                                                        MoreTuples&&... tuples) {
+    return tuple_cat_impl<NTuples - 1, MergedTupleType, Tuples...>::run(
+        MergedTupleType(tuple_get_impl<I1s, Args1...>::run(std::forward<Tuple1>(tuple1))...,
+                        tuple_get_impl<I2s, Args2...>::run(std::forward<Tuple2>(tuple2))...),
+        std::forward<MoreTuples>(tuples)...);
+  }
+
+  // Concatenates the first two tuples.
+  template <typename Tuple1, typename Tuple2, typename... MoreTuples>
+  static constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ReturnType run(Tuple1&& tuple1, Tuple2&& tuple2,
+                                                                        MoreTuples&&... tuples) {
+    return run(std::forward<Tuple1>(tuple1), std::make_index_sequence<N1>{}, std::forward<Tuple2>(tuple2),
+               std::make_index_sequence<N2>{}, std::forward<MoreTuples>(tuples)...);
+  }
+};
+
+// Base case with a single tuple.
+template <size_t N, typename... Args>
+struct tuple_cat_impl<1, TupleImpl<N, Args...>> {
+  using ReturnType = TupleImpl<N, Args...>;
+
+  template <typename Tuple1>
+  static constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ReturnType run(Tuple1&& tuple1) {
+    return tuple1;
+  }
+};
+
+// Special case of no tuples.
+template <>
+struct tuple_cat_impl<0> {
+  using ReturnType = TupleImpl<0>;
+  static constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ReturnType run() { return ReturnType{}; }
+};
+
+// For use in make_tuple, unwraps a reference_wrapper.
+template <typename T>
+struct unwrap_reference_wrapper {
+  using type = T;
+};
+
+template <typename T>
+struct unwrap_reference_wrapper<std::reference_wrapper<T>> {
+  using type = T&;
+};
+
+// For use in make_tuple, decays a type and unwraps a reference_wrapper.
+template <typename T>
+struct unwrap_decay {
+  using type = typename unwrap_reference_wrapper<typename std::decay<T>::type>::type;
+};
+
+/**
+ * Utility for determining a tuple's size.
+ */
+template <typename Tuple>
+struct tuple_size;
+
+template <typename... Types>
+struct tuple_size<TupleImpl<sizeof...(Types), Types...>> : std::integral_constant<size_t, sizeof...(Types)> {};
+
+/**
+ * Gets an element of a tuple.
+ * \tparam Idx index of the element.
+ * \tparam Types ... tuple element types.
+ * \param tuple the tuple.
+ * \return a reference to the desired element.
+ */
+template <size_t Idx, typename... Types>
+constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename tuple_get_impl<Idx, Types...>::ReturnType& get(
+    const TupleImpl<sizeof...(Types), Types...>& tuple) {
+  return tuple_get_impl<Idx, Types...>::run(tuple);
+}
+
+template <size_t Idx, typename... Types>
+constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename tuple_get_impl<Idx, Types...>::ReturnType& get(
+    TupleImpl<sizeof...(Types), Types...>& tuple) {
+  return tuple_get_impl<Idx, Types...>::run(tuple);
+}
+
+/**
+ * Concatenate multiple tuples.
+ * \param tuples ... list of tuples.
+ * \return concatenated tuple.
+ */
+template <typename... Tuples, typename EnableIf = std::enable_if_t<
+                                  internal::reduce_all<is_tuple<typename std::decay<Tuples>::type>::value...>::value>>
+constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    typename tuple_cat_impl<sizeof...(Tuples), typename std::decay<Tuples>::type...>::ReturnType
+    tuple_cat(Tuples&&... tuples) {
+  return tuple_cat_impl<sizeof...(Tuples), typename std::decay<Tuples>::type...>::run(std::forward<Tuples>(tuples)...);
+}
+
+/**
+ * Tie arguments together into a tuple.
+ */
+template <typename... Args, typename ReturnType = TupleImpl<sizeof...(Args), Args&...>>
+constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ReturnType tie(Args&... args) noexcept {
+  return ReturnType{args...};
+}
+
+/**
+ * Create a tuple of l-values with the supplied arguments.
+ */
+template <typename... Args, typename ReturnType = TupleImpl<sizeof...(Args), typename unwrap_decay<Args>::type...>>
+constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ReturnType make_tuple(Args&&... args) {
+  return ReturnType{std::forward<Args>(args)...};
+}
+
+/**
+ * Forward a set of arguments as a tuple.
+ */
+template <typename... Args>
+constexpr EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TupleImpl<sizeof...(Args), Args...> forward_as_tuple(Args&&... args) {
+  return TupleImpl<sizeof...(Args), Args...>(std::forward<Args>(args)...);
+}
+
+/**
+ * Alternative to std::tuple that can be used on device.
+ */
+template <typename... Types>
+using tuple = TupleImpl<sizeof...(Types), Types...>;
+
+}  // namespace tuple_impl
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_TUPLE_GPU

package/eigen/Eigen/src/Core/arch/GPU/TypeCasting.h

@@ -10,71 +10,68 @@
 #ifndef EIGEN_TYPE_CASTING_GPU_H
 #define EIGEN_TYPE_CASTING_GPU_H
 
+// IWYU pragma: private
+#include "../../InternalHeaderCheck.h"
+
 namespace Eigen {
 
 namespace internal {
 
 #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
-  (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
-
+    (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
 
 template <>
 struct type_casting_traits<Eigen::half, float> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 1,
-    TgtCoeffRatio = 2
-  };
+  enum { VectorizedCast = 1, SrcCoeffRatio = 1, TgtCoeffRatio = 2 };
 };
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) {
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) {
   float2 r1 = __half22float2(a);
   float2 r2 = __half22float2(b);
   return make_float4(r1.x, r1.y, r2.x, r2.y);
 }
 
-
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4h2 pcast<float4, Packet4h2>(const float4& a, const float4& b) {
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet4h2 pcast<float4, Packet4h2>(const float4& a, const float4& b) {
   Packet4h2 r;
-  half2* r_alias=reinterpret_cast<half2*>(&r);
-  r_alias[0]=__floats2half2_rn(a.x,a.y);
-  r_alias[1]=__floats2half2_rn(a.z,a.w);
-  r_alias[2]=__floats2half2_rn(b.x,b.y);
-  r_alias[3]=__floats2half2_rn(b.z,b.w);
+  half2* r_alias = reinterpret_cast<half2*>(&r);
+  r_alias[0] = __floats2half2_rn(a.x, a.y);
+  r_alias[1] = __floats2half2_rn(a.z, a.w);
+  r_alias[2] = __floats2half2_rn(b.x, b.y);
+  r_alias[3] = __floats2half2_rn(b.z, b.w);
   return r;
 }
 
 template <>
 struct type_casting_traits<float, Eigen::half> {
-  enum {
-    VectorizedCast = 1,
-    SrcCoeffRatio = 2,
-    TgtCoeffRatio = 1
-  };
+  enum { VectorizedCast = 1, SrcCoeffRatio = 2, TgtCoeffRatio = 1 };
 };
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<Packet4h2, float4>(const Packet4h2& a) {
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<Packet4h2, float4>(const Packet4h2& a) {
   // Simply discard the second half of the input
   float4 r;
-  const half2* a_alias=reinterpret_cast<const half2*>(&a);
+  const half2* a_alias = reinterpret_cast<const half2*>(&a);
   float2 r1 = __half22float2(a_alias[0]);
   float2 r2 = __half22float2(a_alias[1]);
-  r.x=static_cast<float>(r1.x);
-  r.y=static_cast<float>(r1.y);
-  r.z=static_cast<float>(r2.x);
-  r.w=static_cast<float>(r2.y);
+  r.x = static_cast<float>(r1.x);
+  r.y = static_cast<float>(r1.y);
+  r.z = static_cast<float>(r2.x);
+  r.w = static_cast<float>(r2.y);
   return r;
 }
 
-template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) {
+template <>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) {
   // Simply discard the second half of the input
   return __floats2half2_rn(a.x, a.y);
 }
 
 #endif
 
-} // end namespace internal
+}  // end namespace internal
 
-} // end namespace Eigen
+}  // end namespace Eigen
 
-#endif // EIGEN_TYPE_CASTING_GPU_H
+#endif  // EIGEN_TYPE_CASTING_GPU_H

package/eigen/Eigen/src/Core/arch/HIP/hcc/math_constants.h

@@ -1,5 +1,5 @@
 /*
- * math_constants.h - 
+ * math_constants.h -
  *  HIP equivalent of the CUDA header of the same name
  */
 
@@ -8,16 +8,16 @@
 
 /* single precision constants */
 
-#define HIPRT_INF_F        __int_as_float(0x7f800000)
-#define HIPRT_NAN_F        __int_as_float(0x7fffffff)
+#define HIPRT_INF_F __int_as_float(0x7f800000)
+#define HIPRT_NAN_F __int_as_float(0x7fffffff)
 #define HIPRT_MIN_DENORM_F __int_as_float(0x00000001)
 #define HIPRT_MAX_NORMAL_F __int_as_float(0x7f7fffff)
-#define HIPRT_NEG_ZERO_F   __int_as_float(0x80000000)
-#define HIPRT_ZERO_F       0.0f
-#define HIPRT_ONE_F        1.0f
+#define HIPRT_NEG_ZERO_F __int_as_float(0x80000000)
+#define HIPRT_ZERO_F 0.0f
+#define HIPRT_ONE_F 1.0f
 
 /* double precision constants */
-#define HIPRT_INF          __hiloint2double(0x7ff00000, 0x00000000)
-#define HIPRT_NAN          __hiloint2double(0xfff80000, 0x00000000)
+#define HIPRT_INF __hiloint2double(0x7ff00000, 0x00000000)
+#define HIPRT_NAN __hiloint2double(0xfff80000, 0x00000000)
 
 #endif