forgeo-gmlib 0.6.2__cp312-cp312-musllinux_1_2_x86_64.whl

include/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorInflation.h

@@ -0,0 +1,244 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Ke Yang <yangke@gmail.com>
+//
+// 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_CXX11_TENSOR_TENSOR_INFLATION_H
+#define EIGEN_CXX11_TENSOR_TENSOR_INFLATION_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename Strides, typename XprType>
+struct traits<TensorInflationOp<Strides, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename Strides, typename XprType>
+struct eval<TensorInflationOp<Strides, XprType>, Eigen::Dense>
+{
+  typedef const TensorInflationOp<Strides, XprType>& type;
+};
+
+template<typename Strides, typename XprType>
+struct nested<TensorInflationOp<Strides, XprType>, 1, typename eval<TensorInflationOp<Strides, XprType> >::type>
+{
+  typedef TensorInflationOp<Strides, XprType> type;
+};
+
+}  // end namespace internal
+
+/**
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor inflation class.
+ */
+template <typename Strides, typename XprType>
+class TensorInflationOp : public TensorBase<TensorInflationOp<Strides, XprType>, ReadOnlyAccessors> {
+ public:
+  typedef typename Eigen::internal::traits<TensorInflationOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorInflationOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorInflationOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorInflationOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorInflationOp(const XprType& expr, const Strides& strides)
+      : m_xpr(expr), m_strides(strides) {}
+
+    EIGEN_DEVICE_FUNC
+    const Strides& strides() const { return m_strides; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Strides m_strides;
+};
+
+// Eval as rvalue
+template<typename Strides, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorInflationOp<Strides, ArgType>, Device>
+{
+  typedef TensorInflationOp<Strides, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = /*TensorEvaluator<ArgType, Device>::IsAligned*/ false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_strides(op.strides())
+  {
+    m_dimensions = m_impl.dimensions();
+    // Expand each dimension to the inflated dimension.
+    for (int i = 0; i < NumDims; ++i) {
+      m_dimensions[i] = (m_dimensions[i] - 1) * op.strides()[i] + 1;
+    }
+
+    // Remember the strides for fast division.
+    for (int i = 0; i < NumDims; ++i) {
+      m_fastStrides[i] = internal::TensorIntDivisor<Index>(m_strides[i]);
+    }
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_outputStrides[0] = 1;
+      m_inputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
+        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+      }
+    } else {  // RowMajor
+      m_outputStrides[NumDims-1] = 1;
+      m_inputStrides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
+        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
+      }
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  // Computes the input index given the output index. Returns true if the output
+  // index doesn't fall into a hole.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool getInputIndex(Index index, Index* inputIndex) const
+  {
+    eigen_assert(index < dimensions().TotalSize());
+    *inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 1; i > 0; --i) {
+        const Index idx = index / m_outputStrides[i];
+        if (idx != idx / m_fastStrides[i] * m_strides[i]) {
+          return false;
+        }
+        *inputIndex += idx / m_strides[i] * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      if (index != index / m_fastStrides[0] * m_strides[0]) {
+        return false;
+      }
+      *inputIndex += index / m_strides[0];
+      return true;
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 1; ++i) {
+        const Index idx = index / m_outputStrides[i];
+        if (idx != idx / m_fastStrides[i] * m_strides[i]) {
+          return false;
+        }
+        *inputIndex += idx / m_strides[i] * m_inputStrides[i];
+        index -= idx * m_outputStrides[i];
+      }
+      if (index != index / m_fastStrides[NumDims-1] * m_strides[NumDims-1]) {
+        return false;
+      }
+      *inputIndex += index / m_strides[NumDims - 1];
+    }
+    return true;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    Index inputIndex = 0;
+    if (getInputIndex(index, &inputIndex)) {
+     return m_impl.coeff(inputIndex);
+    } else {
+     return Scalar(0);
+    }
+  }
+
+  // TODO(yangke): optimize this function so that we can detect and produce
+  // all-zero packets
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
+
+    EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
+    EIGEN_UNROLL_LOOP
+    for (int i = 0; i < PacketSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    const double compute_cost = NumDims * (3 * TensorOpCost::DivCost<Index>() +
+                                           3 * TensorOpCost::MulCost<Index>() +
+                                           2 * TensorOpCost::AddCost<Index>());
+    const double input_size = m_impl.dimensions().TotalSize();
+    const double output_size = m_dimensions.TotalSize();
+    if (output_size == 0)
+      return TensorOpCost();
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(sizeof(CoeffReturnType) * input_size / output_size, 0,
+                        compute_cost, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh);
+  }
+#endif
+
+ protected:
+  Dimensions m_dimensions;
+  array<Index, NumDims> m_outputStrides;
+  array<Index, NumDims> m_inputStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+  const Strides m_strides;
+  array<internal::TensorIntDivisor<Index>, NumDims> m_fastStrides;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_INFLATION_H

include/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorInitializer.h

@@ -0,0 +1,82 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// 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_CXX11_TENSOR_TENSOR_INITIALIZER_H
+#define EIGEN_CXX11_TENSOR_TENSOR_INITIALIZER_H
+
+#if EIGEN_HAS_VARIADIC_TEMPLATES
+
+#include <initializer_list>
+
+namespace Eigen {
+
+namespace internal {
+
+/**
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Helper template to initialize Tensors from std::initializer_lists.
+ */
+template <typename Derived, int N>
+struct Initializer {
+  typedef std::initializer_list<
+    typename Initializer<Derived, N - 1>::InitList> InitList;
+
+  static void run(TensorEvaluator<Derived, DefaultDevice>& tensor,
+                  Eigen::array<typename traits<Derived>::Index, traits<Derived>::NumDimensions>* indices,
+                  const InitList& vals) {
+    int i = 0;
+    for (const auto& v : vals) {
+      (*indices)[traits<Derived>::NumDimensions - N] = i++;
+      Initializer<Derived, N - 1>::run(tensor, indices, v);
+    }
+  }
+};
+
+template <typename Derived>
+struct Initializer<Derived, 1> {
+  typedef std::initializer_list<typename traits<Derived>::Scalar> InitList;
+
+  static void run(TensorEvaluator<Derived, DefaultDevice>& tensor,
+                  Eigen::array<typename traits<Derived>::Index, traits<Derived>::NumDimensions>* indices,
+                  const InitList& vals) {
+    int i = 0;
+    // There is likely a faster way to do that than iterating.
+    for (const auto& v : vals) {
+      (*indices)[traits<Derived>::NumDimensions - 1] = i++;
+      tensor.coeffRef(*indices) = v;
+    }
+  }
+};
+
+template <typename Derived>
+struct Initializer<Derived, 0> {
+  typedef typename traits<Derived>::Scalar InitList;
+
+  static void run(TensorEvaluator<Derived, DefaultDevice>& tensor,
+                  Eigen::array<typename traits<Derived>::Index, traits<Derived>::NumDimensions>*,
+                  const InitList& v) {
+    tensor.coeffRef(0) = v;
+  }
+};
+
+
+template <typename Derived, int N>
+void initialize_tensor(TensorEvaluator<Derived, DefaultDevice>& tensor,
+                       const typename Initializer<Derived, traits<Derived>::NumDimensions>::InitList& vals) {
+  Eigen::array<typename traits<Derived>::Index, traits<Derived>::NumDimensions> indices;
+  Initializer<Derived, traits<Derived>::NumDimensions>::run(tensor, &indices, vals);
+}
+
+}  // namespace internal
+}  // namespace Eigen
+
+#endif  // EIGEN_HAS_VARIADIC_TEMPLATES
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_INITIALIZER_H

include/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h

@@ -0,0 +1,257 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// 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_CXX11_TENSOR_TENSOR_INTDIV_H
+#define EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H
+
+
+namespace Eigen {
+
+namespace internal {
+
+  // Note: result is undefined if val == 0
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  typename internal::enable_if<sizeof(T)==4,int>::type count_leading_zeros(const T val)
+  {
+#ifdef EIGEN_GPU_COMPILE_PHASE
+    return __clz(val);
+#elif defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::clz(val);
+#elif EIGEN_COMP_MSVC
+    unsigned long index;
+    _BitScanReverse(&index, val);
+    return 31 - index;
+#else
+    EIGEN_STATIC_ASSERT(sizeof(unsigned long long) == 8, YOU_MADE_A_PROGRAMMING_MISTAKE);
+    return __builtin_clz(static_cast<uint32_t>(val));
+#endif
+  }
+
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+  typename internal::enable_if<sizeof(T)==8,int>::type count_leading_zeros(const T val)
+  {
+#ifdef EIGEN_GPU_COMPILE_PHASE
+    return __clzll(val);
+#elif defined(SYCL_DEVICE_ONLY)
+    return static_cast<int>(cl::sycl::clz(val));
+#elif EIGEN_COMP_MSVC && EIGEN_ARCH_x86_64
+    unsigned long index;
+    _BitScanReverse64(&index, val);
+    return 63 - index;
+#elif EIGEN_COMP_MSVC
+    // MSVC's _BitScanReverse64 is not available for 32bits builds.
+    unsigned int lo = (unsigned int)(val&0xffffffff);
+    unsigned int hi = (unsigned int)((val>>32)&0xffffffff);
+    int n;
+    if(hi==0)
+      n = 32 + count_leading_zeros<unsigned int>(lo);
+    else
+      n = count_leading_zeros<unsigned int>(hi);
+    return n;
+#else
+    EIGEN_STATIC_ASSERT(sizeof(unsigned long long) == 8, YOU_MADE_A_PROGRAMMING_MISTAKE);
+    return __builtin_clzll(static_cast<uint64_t>(val));
+#endif
+  }
+
+  template <typename T>
+  struct UnsignedTraits {
+    typedef typename conditional<sizeof(T) == 8, uint64_t, uint32_t>::type type;
+  };
+
+  template <typename T>
+  struct DividerTraits {
+    typedef typename UnsignedTraits<T>::type type;
+    static const int N = sizeof(T) * 8;
+  };
+
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t muluh(const uint32_t a, const T b) {
+#if defined(EIGEN_GPU_COMPILE_PHASE)
+    return __umulhi(a, b);
+#elif defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::mul_hi(a, static_cast<uint32_t>(b));
+#else
+    return (static_cast<uint64_t>(a) * b) >> 32;
+#endif
+  }
+
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
+#if defined(EIGEN_GPU_COMPILE_PHASE)
+    return __umul64hi(a, b);
+#elif defined(SYCL_DEVICE_ONLY)
+    return cl::sycl::mul_hi(a, static_cast<uint64_t>(b));
+#elif EIGEN_HAS_BUILTIN_INT128
+    __uint128_t v = static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b);
+    return static_cast<uint64_t>(v >> 64);
+#else
+    return (TensorUInt128<static_val<0>, uint64_t>(a) * TensorUInt128<static_val<0>, uint64_t>(b)).upper();
+#endif
+  }
+
+  template <int N, typename T>
+  struct DividerHelper {
+    static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t computeMultiplier(const int log_div, const T divider) {
+      EIGEN_STATIC_ASSERT(N == 32, YOU_MADE_A_PROGRAMMING_MISTAKE);
+      return static_cast<uint32_t>((static_cast<uint64_t>(1) << (N+log_div)) / divider - (static_cast<uint64_t>(1) << N) + 1);
+    }
+  };
+
+  template <typename T>
+  struct DividerHelper<64, T> {
+    static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
+#if EIGEN_HAS_BUILTIN_INT128 && !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY)
+      return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1);
+#else
+      const uint64_t shift = 1ULL << log_div;
+      TensorUInt128<uint64_t, uint64_t> result = TensorUInt128<uint64_t, static_val<0> >(shift, 0) / TensorUInt128<static_val<0>, uint64_t>(divider)
+                                               - TensorUInt128<static_val<1>, static_val<0> >(1, 0)
+                                               + TensorUInt128<static_val<0>, static_val<1> >(1);
+      return static_cast<uint64_t>(result);
+#endif
+    }
+  };
+
+/** \internal
+ *
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Fast integer division by a constant.
+ *
+ * See the paper from Granlund and Montgomery for explanation.
+ *   (at https://doi.org/10.1145/773473.178249)
+ *
+ * \sa Tensor
+ */
+template <typename T, bool div_gt_one = false>
+struct TensorIntDivisor {
+ public:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() {
+    multiplier = 0;
+    shift1 = 0;
+    shift2 = 0;
+  }
+
+  // Must have 0 < divider < 2^31. This is relaxed to
+  // 0 < divider < 2^63 when using 64-bit indices on platforms that support
+  // the __uint128_t type.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor(const T divider) {
+    const int N = DividerTraits<T>::N;
+    eigen_assert(static_cast<typename UnsignedTraits<T>::type>(divider) < NumTraits<UnsignedType>::highest()/2);
+    eigen_assert(divider > 0);
+
+    // fast ln2
+    const int leading_zeros = count_leading_zeros(static_cast<UnsignedType>(divider));
+    int log_div = N - leading_zeros;
+    // if divider is a power of two then log_div is 1 more than it should be.
+    if ((static_cast<typename UnsignedTraits<T>::type>(1) << (log_div-1)) == static_cast<typename UnsignedTraits<T>::type>(divider))
+      log_div--;
+
+    multiplier = DividerHelper<N, T>::computeMultiplier(log_div, divider);
+    shift1 = log_div > 1 ? 1 : log_div;
+    shift2 = log_div > 1 ? log_div-1 : 0;
+  }
+
+  // Must have 0 <= numerator. On platforms that don't support the __uint128_t
+  // type numerator should also be less than 2^32-1.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T divide(const T numerator) const {
+    eigen_assert(static_cast<typename UnsignedTraits<T>::type>(numerator) < NumTraits<UnsignedType>::highest()/2);
+    //eigen_assert(numerator >= 0); // this is implicitly asserted by the line above
+
+    UnsignedType t1 = muluh(multiplier, numerator);
+    UnsignedType t = (static_cast<UnsignedType>(numerator) - t1) >> shift1;
+    return (t1 + t) >> shift2;
+  }
+
+ private:
+  typedef typename DividerTraits<T>::type UnsignedType;
+  UnsignedType multiplier;
+  int32_t shift1;
+  int32_t shift2;
+};
+
+
+// Optimized version for signed 32 bit integers.
+// Derived from Hacker's Delight.
+// Only works for divisors strictly greater than one
+template <>
+class TensorIntDivisor<int32_t, true> {
+ public:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorIntDivisor() {
+    magic = 0;
+    shift = 0;
+  }
+  // Must have 2 <= divider
+  EIGEN_DEVICE_FUNC TensorIntDivisor(int32_t divider)  {
+    eigen_assert(divider >= 2);
+    calcMagic(divider);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE int divide(const int32_t n) const {
+#ifdef EIGEN_GPU_COMPILE_PHASE
+    return (__umulhi(magic, n) >> shift);
+#elif defined(SYCL_DEVICE_ONLY)
+    return (cl::sycl::mul_hi(magic, static_cast<uint32_t>(n)) >> shift);
+#else
+    uint64_t v = static_cast<uint64_t>(magic) * static_cast<uint64_t>(n);
+    return (static_cast<uint32_t>(v >> 32) >> shift);
+#endif
+  }
+
+private:
+  // Compute the magic numbers. See Hacker's Delight section 10 for an in
+  // depth explanation.
+  EIGEN_DEVICE_FUNC void calcMagic(int32_t d) {
+   const unsigned two31 = 0x80000000;     // 2**31.
+   unsigned ad = d;
+   unsigned t = two31 + (ad >> 31);
+   unsigned anc = t - 1 - t%ad;     // Absolute value of nc.
+   int p = 31;                      // Init. p.
+   unsigned q1 = two31/anc;         // Init. q1 = 2**p/|nc|.
+   unsigned r1 = two31 - q1*anc;    // Init. r1 = rem(2**p, |nc|).
+   unsigned q2 = two31/ad;          // Init. q2 = 2**p/|d|.
+   unsigned r2 = two31 - q2*ad;     // Init. r2 = rem(2**p, |d|).
+   unsigned delta = 0;
+   do {
+      p = p + 1;
+      q1 = 2*q1;           // Update q1 = 2**p/|nc|.
+      r1 = 2*r1;           // Update r1 = rem(2**p, |nc|).
+      if (r1 >= anc) {     // (Must be an unsigned
+         q1 = q1 + 1;      // comparison here).
+         r1 = r1 - anc;}
+      q2 = 2*q2;           // Update q2 = 2**p/|d|.
+      r2 = 2*r2;           // Update r2 = rem(2**p, |d|).
+      if (r2 >= ad) {      // (Must be an unsigned
+         q2 = q2 + 1;      // comparison here).
+         r2 = r2 - ad;}
+      delta = ad - r2;
+   } while (q1 < delta || (q1 == delta && r1 == 0));
+
+   magic = (unsigned)(q2 + 1);
+   shift = p - 32;
+  }
+
+  uint32_t magic;
+  int32_t shift;
+};
+
+
+template <typename T, bool div_gt_one>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator / (const T& numerator, const TensorIntDivisor<T, div_gt_one>& divisor) {
+  return divisor.divide(numerator);
+}
+
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_INTDIV_H