cart-elc

Source code for CART-ELC
git clone git://git.laack.co/cart-elc.git
Log | Files | Refs | README | LICENSE

TensorMeta.h (8104B)


      1 // This file is part of Eigen, a lightweight C++ template library
      2 // for linear algebra.
      3 //
      4 // Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
      5 //
      6 // This Source Code Form is subject to the terms of the Mozilla
      7 // Public License v. 2.0. If a copy of the MPL was not distributed
      8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
      9 
     10 #ifndef EIGEN_CXX11_TENSOR_TENSOR_META_H
     11 #define EIGEN_CXX11_TENSOR_TENSOR_META_H
     12 
     13 namespace Eigen {
     14 
     15 template<bool cond> struct Cond {};
     16 
     17 template<typename T1, typename T2> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
     18 const T1& choose(Cond<true>, const T1& first, const T2&) {
     19   return first;
     20 }
     21 
     22 template<typename T1, typename T2> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
     23 const T2& choose(Cond<false>, const T1&, const T2& second) {
     24   return second;
     25 }
     26 
     27 
     28 template <typename T, typename X, typename Y>
     29 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
     30 T divup(const X x, const Y y) {
     31   return static_cast<T>((x + y - 1) / y);
     32 }
     33 
     34 template <typename T>
     35 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
     36 T divup(const T x, const T y) {
     37   return static_cast<T>((x + y - 1) / y);
     38 }
     39 
     40 template <size_t n> struct max_n_1 {
     41   static const size_t size = n;
     42 };
     43 template <> struct max_n_1<0> {
     44   static const size_t size = 1;
     45 };
     46 
     47 
     48 // Default packet types
     49 template <typename Scalar, typename Device>
     50 struct PacketType : internal::packet_traits<Scalar> {
     51   typedef typename internal::packet_traits<Scalar>::type type;
     52 };
     53 
     54 // For CUDA packet types when using a GpuDevice
     55 #if defined(EIGEN_USE_GPU) && defined(EIGEN_HAS_GPU_FP16)
     56 
     57 typedef ulonglong2 Packet4h2;
     58 template<>
     59 struct PacketType<half, GpuDevice> {
     60   typedef Packet4h2 type;
     61   static const int size = 8;
     62   enum {
     63     HasAdd    = 1,
     64     HasSub    = 1,
     65     HasMul    = 1,
     66     HasNegate = 1,
     67     HasAbs    = 1,
     68     HasArg    = 0,
     69     HasAbs2   = 0,
     70     HasMin    = 1,
     71     HasMax    = 1,
     72     HasConj   = 0,
     73     HasSetLinear = 0,
     74     HasBlend  = 0,
     75 
     76     HasDiv    = 1,
     77     HasSqrt   = 1,
     78     HasRsqrt  = 1,
     79     HasExp    = 1,
     80     HasExpm1  = 0,
     81     HasLog    = 1,
     82     HasLog1p  = 0,
     83     HasLog10  = 0,
     84     HasPow    = 1,
     85   };
     86 };
     87 #endif
     88 
     89 #if defined(EIGEN_USE_SYCL)
     90 
     91 namespace TensorSycl {
     92 namespace internal {
     93 
     94 template <typename Index, Index A, Index B> struct PlusOp {
     95   static constexpr Index Value = A + B;
     96 };
     97 
     98 template <typename Index, Index A, Index B> struct DivOp {
     99   static constexpr Index Value = A / B;
    100 };
    101 
    102 template <typename Index, Index start, Index end, Index step,
    103           template <class Indx, Indx...> class StepOp>
    104 struct static_for {
    105   template <typename UnaryOperator>
    106   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void loop(UnaryOperator op) {
    107     op(start);
    108     static_for<Index, StepOp<Index, start, step>::Value, end, step,
    109                StepOp>::loop(op);
    110   }
    111 };
    112 template <typename Index, Index end, Index step,
    113           template <class Indx, Indx...> class StepOp>
    114 struct static_for<Index, end, end, step, StepOp> {
    115   template <typename UnaryOperator>
    116   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void loop(UnaryOperator) {}
    117 };
    118 
    119 template <typename OutScalar, typename Device, bool Vectorizable>
    120 struct Vectorise {
    121   static const int PacketSize = 1;
    122   typedef OutScalar PacketReturnType;
    123 };
    124 
    125 template <typename OutScalar, typename Device>
    126 struct Vectorise<OutScalar, Device, true> {
    127   static const int PacketSize = Eigen::PacketType<OutScalar, Device>::size;
    128   typedef typename Eigen::PacketType<OutScalar, Device>::type PacketReturnType;
    129 };
    130 
    131 static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Index roundUp(Index x, Index y) {
    132   return ((((x) + (y)-1) / (y)) * (y));
    133 }
    134 
    135 } // namespace internal
    136 } // namespace TensorSycl
    137 
    138 template <>
    139   struct PacketType<half, SyclDevice> {
    140   typedef half type;
    141   static const int size = 1;
    142   enum {
    143     HasAdd    = 0,
    144     HasSub    = 0,
    145     HasMul    = 0,
    146     HasNegate = 0,
    147     HasAbs    = 0,
    148     HasArg    = 0,
    149     HasAbs2   = 0,
    150     HasMin    = 0,
    151     HasMax    = 0,
    152     HasConj   = 0,
    153     HasSetLinear = 0,
    154     HasBlend  = 0
    155   };
    156 };
    157 template <typename Scalar>
    158 struct PacketType<Scalar, SyclDevice> : internal::default_packet_traits {
    159   typedef Scalar type;
    160   typedef Scalar half;
    161   enum {
    162     Vectorizable = 0,
    163     size = 1,
    164     AlignedOnScalar = 0,
    165     HasHalfPacket = 0
    166   };
    167   enum {
    168     HasAdd    = 0,
    169     HasSub    = 0,
    170     HasMul    = 0,
    171     HasNegate = 0,
    172     HasAbs    = 0,
    173     HasAbs2   = 0,
    174     HasMin    = 0,
    175     HasMax    = 0,
    176     HasConj   = 0,
    177     HasSetLinear = 0
    178   };
    179 
    180 };
    181 
    182 template <typename Scalar>
    183 struct PacketType<Scalar, const SyclDevice> : PacketType<Scalar, SyclDevice>{};
    184 
    185 #ifndef EIGEN_DONT_VECTORIZE_SYCL
    186 #define PACKET_TYPE(CVQual, Type, val, lengths, DEV)\
    187 template<> struct PacketType<CVQual Type, DEV> : internal::sycl_packet_traits<val, lengths> \
    188 {\
    189   typedef typename internal::packet_traits<Type>::type type;\
    190   typedef typename internal::packet_traits<Type>::half half;\
    191 };
    192 
    193 
    194 PACKET_TYPE(const, float, 1, 4, SyclDevice)
    195 PACKET_TYPE(, float, 1, 4, SyclDevice)
    196 PACKET_TYPE(const, float, 1, 4, const SyclDevice)
    197 PACKET_TYPE(, float, 1, 4, const SyclDevice)
    198 
    199 PACKET_TYPE(const, double, 0, 2, SyclDevice)
    200 PACKET_TYPE(, double, 0, 2, SyclDevice)
    201 PACKET_TYPE(const, double, 0, 2, const SyclDevice)
    202 PACKET_TYPE(, double, 0, 2, const SyclDevice)
    203 #undef PACKET_TYPE
    204 
    205 template<> struct PacketType<half, const SyclDevice>: PacketType<half, SyclDevice>{};
    206 template<> struct PacketType<const half, const SyclDevice>: PacketType<half, SyclDevice>{};
    207 #endif
    208 #endif
    209 
    210 // Tuple mimics std::pair but works on e.g. nvcc.
    211 template <typename U, typename V> struct Tuple {
    212  public:
    213   U first;
    214   V second;
    215 
    216   typedef U first_type;
    217   typedef V second_type;
    218 
    219   EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    220   Tuple() : first(), second() {}
    221 
    222   EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    223   Tuple(const U& f, const V& s) : first(f), second(s) {}
    224 
    225   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    226   void swap(Tuple& rhs) {
    227     using numext::swap;
    228     swap(first, rhs.first);
    229     swap(second, rhs.second);
    230   }
    231 };
    232 
    233 template <typename U, typename V>
    234 EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    235 bool operator==(const Tuple<U, V>& x, const Tuple<U, V>& y) {
    236   return (x.first == y.first && x.second == y.second);
    237 }
    238 
    239 template <typename U, typename V>
    240 EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    241 bool operator!=(const Tuple<U, V>& x, const Tuple<U, V>& y) {
    242   return !(x == y);
    243 }
    244 
    245 
    246 // Can't use std::pairs on cuda devices
    247 template <typename Idx> struct IndexPair {
    248   EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexPair() : first(0), second(0) {}
    249   EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE IndexPair(Idx f, Idx s) : first(f), second(s) {}
    250 
    251   EIGEN_DEVICE_FUNC void set(IndexPair<Idx> val) {
    252     first = val.first;
    253     second = val.second;
    254   }
    255 
    256   Idx first;
    257   Idx second;
    258 };
    259 
    260 
    261 #ifdef EIGEN_HAS_SFINAE
    262 namespace internal {
    263 
    264   template<typename IndexType, typename Index, Index... Is>
    265   EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    266   array<Index, sizeof...(Is)> customIndices2Array(IndexType& idx, numeric_list<Index, Is...>) {
    267     return { idx[Is]... };
    268   }
    269   template<typename IndexType, typename Index>
    270   EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    271   array<Index, 0> customIndices2Array(IndexType&, numeric_list<Index>) {
    272     return array<Index, 0>();
    273   }
    274 
    275   /** Make an array (for index/dimensions) out of a custom index */
    276   template<typename Index, std::size_t NumIndices, typename IndexType>
    277   EIGEN_CONSTEXPR EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
    278   array<Index, NumIndices> customIndices2Array(IndexType& idx) {
    279     return customIndices2Array(idx, typename gen_numeric_list<Index, NumIndices>::type{});
    280   }
    281 
    282 
    283   template <typename B, typename D>
    284   struct is_base_of
    285   {
    286 
    287     typedef char (&yes)[1];
    288     typedef char (&no)[2];
    289 
    290     template <typename BB, typename DD>
    291     struct Host
    292     {
    293       operator BB*() const;
    294       operator DD*();
    295     };
    296 
    297     template<typename T>
    298     static yes check(D*, T);
    299     static no check(B*, int);
    300 
    301     static const bool value = sizeof(check(Host<B,D>(), int())) == sizeof(yes);
    302   };
    303 
    304 }
    305 #endif
    306 
    307 
    308 
    309 }  // namespace Eigen
    310 
    311 #endif  // EIGEN_CXX11_TENSOR_TENSOR_META_H