Merged eigen/eigen into default

Eigen/src/Core/GeneralProduct.h

@@ -396,6 +396,7 @@ template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
   */
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 inline const Product<Derived, OtherDerived>
 MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
 {

Eigen/src/Core/Map.h

@@ -114,7 +114,7 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
     inline Index outerStride() const
     {
       return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
-           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? internal::traits<Map>::OuterStrideAtCompileTime
+           : internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? Index(internal::traits<Map>::OuterStrideAtCompileTime)
            : IsVectorAtCompileTime ? (this->size() * innerStride())
            : int(Flags)&RowMajorBit ? (this->cols() * innerStride())
            : (this->rows() * innerStride());

Eigen/src/Core/NoAlias.h

@@ -99,7 +99,7 @@ class NoAlias
   * \sa class NoAlias
   */
 template<typename Derived>
-NoAlias<Derived,MatrixBase> MatrixBase<Derived>::noalias()
+NoAlias<Derived,MatrixBase> EIGEN_DEVICE_FUNC MatrixBase<Derived>::noalias()
 {
   return NoAlias<Derived, Eigen::MatrixBase >(derived());
 }

Eigen/src/Core/arch/CUDA/Half.h

@@ -50,38 +50,45 @@ struct half;
 namespace half_impl {
 
 #if !defined(EIGEN_HAS_CUDA_FP16)
-
-// Make our own __half definition that is similar to CUDA's.
-struct __half {
-  EIGEN_DEVICE_FUNC __half() : x(0) {}
-  explicit EIGEN_DEVICE_FUNC __half(unsigned short raw) : x(raw) {}
+// Make our own __half_raw definition that is similar to CUDA's.
+struct __half_raw {
+  EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
+  explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
   unsigned short x;
 };
-
+#elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
+// In CUDA < 9.0, __half is the equivalent of CUDA 9's __half_raw
+typedef __half __half_raw;
 #endif
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x);
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff);
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h);
 
-struct half_base : public __half {
+struct half_base : public __half_raw {
   EIGEN_DEVICE_FUNC half_base() {}
-  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half(h) {}
-  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half(h) {}
+  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half_raw(h) {}
+  EIGEN_DEVICE_FUNC half_base(const __half_raw& h) : __half_raw(h) {}
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(*(__half_raw*)&h) {}
+#endif
 };
 
 } // namespace half_impl
 
 // Class definition.
 struct half : public half_impl::half_base {
-  #if !defined(EIGEN_HAS_CUDA_FP16)
-    typedef half_impl::__half __half;
+  #if !defined(EIGEN_HAS_CUDA_FP16) || (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000)
+    typedef half_impl::__half_raw __half_raw;
   #endif
 
   EIGEN_DEVICE_FUNC half() {}
 
-  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+  EIGEN_DEVICE_FUNC half(const __half_raw& h) : half_impl::half_base(h) {}
   EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
+  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+#endif
 
   explicit EIGEN_DEVICE_FUNC half(bool b)
       : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
@@ -269,8 +276,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
 // these in hardware. If we need more performance on older/other CPUs, they are
 // also possible to vectorize directly.
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x) {
-  __half h;
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x) {
+  __half_raw h;
   h.x = x;
   return h;
 }
@@ -280,12 +287,13 @@ union FP32 {
   float f;
 };
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
 #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
-  return __float2half(ff);
+  __half tmp_ff = __float2half(ff);
+  return *(__half_raw*)&tmp_ff;
 
 #elif defined(EIGEN_HAS_FP16_C)
-  __half h;
+  __half_raw h;
   h.x = _cvtss_sh(ff, 0);
   return h;
 
@@ -296,7 +304,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
   const FP32 f16max = { (127 + 16) << 23 };
   const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
   unsigned int sign_mask = 0x80000000u;
-  __half o;
+  __half_raw o;
   o.x = static_cast<unsigned short>(0x0u);
 
   unsigned int sign = f.u & sign_mask;
@@ -335,7 +343,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
 #endif
 }
 
-EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) {
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h) {
 #if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
   return __half2float(h);
 
@@ -512,8 +520,8 @@ struct numeric_limits<Eigen::half> {
   static const bool is_bounded = false;
   static const bool is_modulo = false;
   static const int digits = 11;
-  static const int digits10 = 2;
-  //static const int max_digits10 = ;
+  static const int digits10 = 3;      // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
+  static const int max_digits10 = 5;  // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
   static const int radix = 2;
   static const int min_exponent = -13;
   static const int min_exponent10 = -4;
@@ -612,11 +620,15 @@ struct hash<Eigen::half> {
 // Add the missing shfl_xor intrinsic
 #if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
 __device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
+  #if EIGEN_CUDACC_VER < 90000
   return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
+  #else
+  return static_cast<Eigen::half>(__shfl_xor_sync(0xFFFFFFFF, static_cast<float>(var), laneMask, width));
+  #endif
 }
 #endif
 
-// ldg() has an overload for __half, but we also need one for Eigen::half.
+// ldg() has an overload for __half_raw, but we also need one for Eigen::half.
 #if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
   return Eigen::half_impl::raw_uint16_to_half(

Eigen/src/Core/arch/CUDA/PacketMathHalf.h

@@ -100,7 +100,8 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2&
 
 template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
   half2 result;
-  result.x = a.x & 0x7FFF7FFF;
+  unsigned temp = *(reinterpret_cast<const unsigned*>(&(a)));
+  *(reinterpret_cast<unsigned*>(&(result))) = temp & 0x7FFF7FFF;
   return result;
 }
 

Eigen/src/Core/util/Macros.h

@@ -410,6 +410,16 @@
 #endif
 #endif
 
+// Does the compiler support type_trais?
+#ifndef EIGEN_HAS_TYPE_TRAITS
+#if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700)
+#define EIGEN_HAS_TYPE_TRAITS 1
+#define EIGEN_INCLUDE_TYPE_TRAITS
+#else
+#define EIGEN_HAS_TYPE_TRAITS 0
+#endif
+#endif
+
 // Does the compiler support variadic templates?
 #ifndef EIGEN_HAS_VARIADIC_TEMPLATES
 #if EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) \

Eigen/src/Core/util/Memory.h

@@ -493,7 +493,7 @@ template<typename T> struct smart_copy_helper<T,true> {
     IntPtr size = IntPtr(end)-IntPtr(start);
     if(size==0) return;
     eigen_internal_assert(start!=0 && end!=0 && target!=0);
-    memcpy(target, start, size);
+    std::memcpy(target, start, size);
   }
 };
 
@@ -696,7 +696,15 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
 /** \class aligned_allocator
 * \ingroup Core_Module
 *
-* \brief STL compatible allocator to use with with 16 byte aligned types
+* \brief STL compatible allocator to use with types requiring a non standrad alignment.
+*
+* The memory is aligned as for dynamically aligned matrix/array types such as MatrixXd.
+* By default, it will thus provide at least 16 bytes alignment and more in following cases:
+*  - 32 bytes alignment if AVX is enabled.
+*  - 64 bytes alignment if AVX512 is enabled.
+*
+* This can be controled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented
+* \link TopicPreprocessorDirectivesPerformance there \endlink.
 *
 * Example:
 * \code

Eigen/src/Core/util/XprHelper.h

@@ -34,6 +34,18 @@ inline IndexDest convert_index(const IndexSrc& idx) {
   return IndexDest(idx);
 }
 
+// true if T can be considered as an integral index (i.e., and integral type or enum)
+template<typename T> struct is_valid_index_type
+{
+  enum { value =
+#if EIGEN_HAS_TYPE_TRAITS
+   internal::is_integral<T>::value || std::is_enum<T>::value
+#else
+  // without C++11, we use is_convertible to Index instead of is_integral in order to treat enums as Index.
+  internal::is_convertible<T,Index>::value
+#endif
+  };
+};
 
 // promote_scalar_arg is an helper used in operation between an expression and a scalar, like:
 //    expression * scalar