forgeo-gmlib 0.6.2__cp313-cp313-musllinux_1_2_x86_64.whl

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

@@ -0,0 +1,286 @@
+// 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_PATCH_H
+#define EIGEN_CXX11_TENSOR_TENSOR_PATCH_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename PatchDim, typename XprType>
+struct traits<TensorPatchOp<PatchDim, 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 + 1;
+  static const int Layout = XprTraits::Layout;
+  typedef typename XprTraits::PointerType PointerType;
+};
+
+template<typename PatchDim, typename XprType>
+struct eval<TensorPatchOp<PatchDim, XprType>, Eigen::Dense>
+{
+  typedef const TensorPatchOp<PatchDim, XprType>& type;
+};
+
+template<typename PatchDim, typename XprType>
+struct nested<TensorPatchOp<PatchDim, XprType>, 1, typename eval<TensorPatchOp<PatchDim, XprType> >::type>
+{
+  typedef TensorPatchOp<PatchDim, XprType> type;
+};
+
+}  // end namespace internal
+
+/**
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Tensor patch class.
+ */
+template <typename PatchDim, typename XprType>
+class TensorPatchOp : public TensorBase<TensorPatchOp<PatchDim, XprType>, ReadOnlyAccessors> {
+ public:
+  typedef typename Eigen::internal::traits<TensorPatchOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename Eigen::internal::nested<TensorPatchOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorPatchOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorPatchOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorPatchOp(const XprType& expr, const PatchDim& patch_dims)
+      : m_xpr(expr), m_patch_dims(patch_dims) {}
+
+    EIGEN_DEVICE_FUNC
+    const PatchDim& patch_dims() const { return m_patch_dims; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const PatchDim m_patch_dims;
+};
+
+
+// Eval as rvalue
+template<typename PatchDim, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
+{
+  typedef TensorPatchOp<PatchDim, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value + 1;
+  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 = false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+    BlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = false,
+    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)
+  {
+    Index num_patches = 1;
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    const PatchDim& patch_dims = op.patch_dims();
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = 0; i < NumDims-1; ++i) {
+        m_dimensions[i] = patch_dims[i];
+        num_patches *= (input_dims[i] - patch_dims[i] + 1);
+      }
+      m_dimensions[NumDims-1] = num_patches;
+
+      m_inputStrides[0] = 1;
+      m_patchStrides[0] = 1;
+      for (int i = 1; i < NumDims-1; ++i) {
+        m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+        m_patchStrides[i] = m_patchStrides[i-1] * (input_dims[i-1] - patch_dims[i-1] + 1);
+      }
+      m_outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
+      }
+    } else {
+      for (int i = 0; i < NumDims-1; ++i) {
+        m_dimensions[i+1] = patch_dims[i];
+        num_patches *= (input_dims[i] - patch_dims[i] + 1);
+      }
+      m_dimensions[0] = num_patches;
+
+      m_inputStrides[NumDims-2] = 1;
+      m_patchStrides[NumDims-2] = 1;
+      for (int i = NumDims-3; i >= 0; --i) {
+        m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
+        m_patchStrides[i] = m_patchStrides[i+1] * (input_dims[i+1] - patch_dims[i+1] + 1);
+      }
+      m_outputStrides[NumDims-1] = 1;
+      for (int i = NumDims-2; i >= 0; --i) {
+        m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[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();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    Index output_stride_index = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? NumDims - 1 : 0;
+    // Find the location of the first element of the patch.
+    Index patchIndex = index / m_outputStrides[output_stride_index];
+    // Find the offset of the element wrt the location of the first element.
+    Index patchOffset = index - patchIndex * m_outputStrides[output_stride_index];
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 2; i > 0; --i) {
+        const Index patchIdx = patchIndex / m_patchStrides[i];
+        patchIndex -= patchIdx * m_patchStrides[i];
+        const Index offsetIdx = patchOffset / m_outputStrides[i];
+        patchOffset -= offsetIdx * m_outputStrides[i];
+        inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i];
+      }
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 2; ++i) {
+        const Index patchIdx = patchIndex / m_patchStrides[i];
+        patchIndex -= patchIdx * m_patchStrides[i];
+        const Index offsetIdx = patchOffset / m_outputStrides[i+1];
+        patchOffset -= offsetIdx * m_outputStrides[i+1];
+        inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i];
+      }
+    }
+    inputIndex += (patchIndex + patchOffset);
+    return m_impl.coeff(inputIndex);
+  }
+
+  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());
+
+    Index output_stride_index = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? NumDims - 1 : 0;
+    Index indices[2] = {index, index + PacketSize - 1};
+    Index patchIndices[2] = {indices[0] / m_outputStrides[output_stride_index],
+                             indices[1] / m_outputStrides[output_stride_index]};
+    Index patchOffsets[2] = {indices[0] - patchIndices[0] * m_outputStrides[output_stride_index],
+                             indices[1] - patchIndices[1] * m_outputStrides[output_stride_index]};
+
+    Index inputIndices[2] = {0, 0};
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
+      for (int i = NumDims - 2; i > 0; --i) {
+        const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i],
+                                   patchIndices[1] / m_patchStrides[i]};
+        patchIndices[0] -= patchIdx[0] * m_patchStrides[i];
+        patchIndices[1] -= patchIdx[1] * m_patchStrides[i];
+
+        const Index offsetIdx[2] = {patchOffsets[0] / m_outputStrides[i],
+                                    patchOffsets[1] / m_outputStrides[i]};
+        patchOffsets[0] -= offsetIdx[0] * m_outputStrides[i];
+        patchOffsets[1] -= offsetIdx[1] * m_outputStrides[i];
+
+        inputIndices[0] += (patchIdx[0] + offsetIdx[0]) * m_inputStrides[i];
+        inputIndices[1] += (patchIdx[1] + offsetIdx[1]) * m_inputStrides[i];
+      }
+    } else {
+      EIGEN_UNROLL_LOOP
+      for (int i = 0; i < NumDims - 2; ++i) {
+        const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i],
+                                   patchIndices[1] / m_patchStrides[i]};
+        patchIndices[0] -= patchIdx[0] * m_patchStrides[i];
+        patchIndices[1] -= patchIdx[1] * m_patchStrides[i];
+
+        const Index offsetIdx[2] = {patchOffsets[0] / m_outputStrides[i+1],
+                                    patchOffsets[1] / m_outputStrides[i+1]};
+        patchOffsets[0] -= offsetIdx[0] * m_outputStrides[i+1];
+        patchOffsets[1] -= offsetIdx[1] * m_outputStrides[i+1];
+
+        inputIndices[0] += (patchIdx[0] + offsetIdx[0]) * m_inputStrides[i];
+        inputIndices[1] += (patchIdx[1] + offsetIdx[1]) * m_inputStrides[i];
+      }
+    }
+    inputIndices[0] += (patchIndices[0] + patchOffsets[0]);
+    inputIndices[1] += (patchIndices[1] + patchOffsets[1]);
+
+    if (inputIndices[1] - inputIndices[0] == PacketSize - 1) {
+      PacketReturnType rslt = m_impl.template packet<Unaligned>(inputIndices[0]);
+      return rslt;
+    }
+    else {
+      EIGEN_ALIGN_MAX CoeffReturnType values[PacketSize];
+      values[0] = m_impl.coeff(inputIndices[0]);
+      values[PacketSize-1] = m_impl.coeff(inputIndices[1]);
+      EIGEN_UNROLL_LOOP
+      for (int i = 1; i < PacketSize-1; ++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 * (TensorOpCost::DivCost<Index>() +
+                                           TensorOpCost::MulCost<Index>() +
+                                           2 * TensorOpCost::AddCost<Index>());
+    return m_impl.costPerCoeff(vectorized) +
+           TensorOpCost(0, 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-1> m_inputStrides;
+  array<Index, NumDims-1> m_patchStrides;
+
+  TensorEvaluator<ArgType, Device> m_impl;
+
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_PATCH_H

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

@@ -0,0 +1,317 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2018 Mehdi Goli <eigen@codeplay.com> Codeplay Software Ltd.
+//
+// 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_RANDOM_H
+#define EIGEN_CXX11_TENSOR_TENSOR_RANDOM_H
+
+namespace Eigen {
+namespace internal {
+
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE uint64_t get_random_seed() {
+#if defined(EIGEN_GPU_COMPILE_PHASE)
+  // We don't support 3d kernels since we currently only use 1 and
+  // 2d kernels.
+  gpu_assert(threadIdx.z == 0);
+  return blockIdx.x * blockDim.x + threadIdx.x 
+         + gridDim.x * blockDim.x * (blockIdx.y * blockDim.y + threadIdx.y);
+#else
+  // Rely on Eigen's random implementation.
+  return random<uint64_t>();
+#endif
+}
+
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE unsigned PCG_XSH_RS_generator(uint64_t* state, uint64_t stream) {
+  // TODO: Unify with the implementation in the non blocking thread pool.
+  uint64_t current = *state;
+  // Update the internal state
+  *state = current * 6364136223846793005ULL + (stream << 1 | 1);
+  // Generate the random output (using the PCG-XSH-RS scheme)
+  return static_cast<unsigned>((current ^ (current >> 22)) >> (22 + (current >> 61)));
+}
+
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE uint64_t PCG_XSH_RS_state(uint64_t seed) {
+  seed = seed ? seed : get_random_seed();
+  return seed * 6364136223846793005ULL + 0xda3e39cb94b95bdbULL;
+}
+
+template <typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+T RandomToTypeUniform(uint64_t* state, uint64_t stream) {
+  unsigned rnd = PCG_XSH_RS_generator(state, stream);
+  return static_cast<T>(rnd);
+}
+
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+Eigen::half RandomToTypeUniform<Eigen::half>(uint64_t* state, uint64_t stream) {
+  // Generate 10 random bits for the mantissa, merge with exponent.
+  unsigned rnd = PCG_XSH_RS_generator(state, stream);
+  const uint16_t half_bits = static_cast<uint16_t>(rnd & 0x3ffu) | (static_cast<uint16_t>(15) << 10);
+  Eigen::half result = Eigen::numext::bit_cast<Eigen::half>(half_bits);
+  // Return the final result
+  return result - Eigen::half(1.0f);
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+Eigen::bfloat16 RandomToTypeUniform<Eigen::bfloat16>(uint64_t* state, uint64_t stream) {
+
+  // Generate 7 random bits for the mantissa, merge with exponent.
+  unsigned rnd = PCG_XSH_RS_generator(state, stream);
+  const uint16_t half_bits = static_cast<uint16_t>(rnd & 0x7fu) | (static_cast<uint16_t>(127) << 7);
+  Eigen::bfloat16 result = Eigen::numext::bit_cast<Eigen::bfloat16>(half_bits);
+  // Return the final result
+  return result - Eigen::bfloat16(1.0f);
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float RandomToTypeUniform<float>(uint64_t* state, uint64_t stream) {
+  typedef union {
+    uint32_t raw;
+    float fp;
+  } internal;
+  internal result;
+  // Generate 23 random bits for the mantissa mantissa
+  const unsigned rnd = PCG_XSH_RS_generator(state, stream);
+  result.raw = rnd & 0x7fffffu;
+  // Set the exponent
+  result.raw |= (static_cast<uint32_t>(127) << 23);
+  // Return the final result
+  return result.fp - 1.0f;
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double RandomToTypeUniform<double>(uint64_t* state, uint64_t stream) {
+  typedef union {
+    uint64_t raw;
+    double dp;
+  } internal;
+  internal result;
+  result.raw = 0;
+  // Generate 52 random bits for the mantissa
+  // First generate the upper 20 bits
+  unsigned rnd1 = PCG_XSH_RS_generator(state, stream) & 0xfffffu;
+  // The generate the lower 32 bits
+  unsigned rnd2 = PCG_XSH_RS_generator(state, stream);
+  result.raw = (static_cast<uint64_t>(rnd1) << 32) | rnd2;
+  // Set the exponent
+  result.raw |= (static_cast<uint64_t>(1023) << 52);
+  // Return the final result
+  return result.dp - 1.0;
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+std::complex<float> RandomToTypeUniform<std::complex<float> >(uint64_t* state, uint64_t stream) {
+  return std::complex<float>(RandomToTypeUniform<float>(state, stream),
+                             RandomToTypeUniform<float>(state, stream));
+}
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+std::complex<double> RandomToTypeUniform<std::complex<double> >(uint64_t* state, uint64_t stream) {
+  return std::complex<double>(RandomToTypeUniform<double>(state, stream),
+                              RandomToTypeUniform<double>(state, stream));
+}
+
+template <typename T> class UniformRandomGenerator {
+ public:
+  static const bool PacketAccess = true;
+
+  // Uses the given "seed" if non-zero, otherwise uses a random seed.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE UniformRandomGenerator(
+      uint64_t seed = 0) {
+    m_state = PCG_XSH_RS_state(seed);
+    #ifdef EIGEN_USE_SYCL
+    // In SYCL it is not possible to build PCG_XSH_RS_state in one step.
+    // Therefor, we need two step to initializate the m_state.
+    // IN SYCL, the constructor of the functor is s called on the CPU
+    // and we get the clock seed here from the CPU. However, This seed is
+    //the same for all the thread. As unlike CUDA, the thread.ID, BlockID, etc is not a global function.
+    // and only  available on the Operator() function (which is called on the GPU).
+    // Thus for CUDA (((CLOCK  + global_thread_id)* 6364136223846793005ULL) + 0xda3e39cb94b95bdbULL) is passed to each thread
+    // but for SYCL ((CLOCK * 6364136223846793005ULL) + 0xda3e39cb94b95bdbULL) is passed to each thread and each thread adds
+    // the  (global_thread_id* 6364136223846793005ULL) for itself only once, in order to complete the construction
+    // similar to CUDA Therefore, the thread Id injection is not available at this stage.
+    //However when the operator() is called the thread ID will be avilable. So inside the opeator,
+    // we add the thrreadID, BlockId,... (which is equivalent of i)
+    //to the seed and construct the unique m_state per thead similar to cuda.
+    m_exec_once =false;
+   #endif
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE UniformRandomGenerator(
+      const UniformRandomGenerator& other) {
+    m_state = other.m_state;
+    #ifdef EIGEN_USE_SYCL
+     m_exec_once =other.m_exec_once;
+    #endif
+  }
+
+  template<typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T operator()(Index i) const {
+    #ifdef EIGEN_USE_SYCL
+      if(!m_exec_once) {
+      // This is the second stage of adding thread Id to the CPU clock seed and build unique seed per thread
+      // The (i * 6364136223846793005ULL) is the remaining part of the PCG_XSH_RS_state on the GPU side
+       m_state += (i * 6364136223846793005ULL);
+       m_exec_once =true;
+      }
+    #endif
+    T result = RandomToTypeUniform<T>(&m_state, i);
+    return result;
+  }
+
+  template<typename Packet, typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Packet packetOp(Index i) const {
+    const int packetSize = internal::unpacket_traits<Packet>::size;
+    EIGEN_ALIGN_MAX T values[packetSize];
+      #ifdef EIGEN_USE_SYCL
+      if(!m_exec_once) {
+      // This is the second stage of adding thread Id to the CPU clock seed and build unique seed per thread
+       m_state += (i * 6364136223846793005ULL);
+       m_exec_once =true;
+      }
+    #endif
+    EIGEN_UNROLL_LOOP
+    for (int j = 0; j < packetSize; ++j) {
+      values[j] = RandomToTypeUniform<T>(&m_state, i);
+    }
+    return internal::pload<Packet>(values);
+  }
+
+ private:
+  mutable uint64_t m_state;
+  #ifdef EIGEN_USE_SYCL
+  mutable bool m_exec_once;
+  #endif
+};
+
+template <typename Scalar>
+struct functor_traits<UniformRandomGenerator<Scalar> > {
+  enum {
+    // Rough estimate for floating point, multiplied by ceil(sizeof(T) / sizeof(float)).
+    Cost = 12 * NumTraits<Scalar>::AddCost *
+           ((sizeof(Scalar) + sizeof(float) - 1) / sizeof(float)),
+    PacketAccess = UniformRandomGenerator<Scalar>::PacketAccess
+  };
+};
+
+
+
+template <typename T> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+T RandomToTypeNormal(uint64_t* state, uint64_t stream) {
+  // Use the ratio of uniform method to generate numbers following a normal
+  // distribution. See for example Numerical Recipes chapter 7.3.9 for the
+  // details.
+  T u, v, q;
+  do {
+    u = RandomToTypeUniform<T>(state, stream);
+    v = T(1.7156) * (RandomToTypeUniform<T>(state, stream) - T(0.5));
+    const T x = u - T(0.449871);
+    const T y = numext::abs(v) + T(0.386595);
+    q = x*x + y * (T(0.196)*y - T(0.25472)*x);
+  } while (q > T(0.27597) &&
+           (q > T(0.27846) || v*v > T(-4) * numext::log(u) * u*u));
+
+  return v/u;
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+std::complex<float> RandomToTypeNormal<std::complex<float> >(uint64_t* state, uint64_t stream) {
+  return std::complex<float>(RandomToTypeNormal<float>(state, stream),
+                             RandomToTypeNormal<float>(state, stream));
+}
+template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+std::complex<double> RandomToTypeNormal<std::complex<double> >(uint64_t* state, uint64_t stream) {
+  return std::complex<double>(RandomToTypeNormal<double>(state, stream),
+                              RandomToTypeNormal<double>(state, stream));
+}
+
+
+template <typename T> class NormalRandomGenerator {
+ public:
+  static const bool PacketAccess = true;
+
+  // Uses the given "seed" if non-zero, otherwise uses a random seed.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE NormalRandomGenerator(uint64_t seed = 0) {
+    m_state = PCG_XSH_RS_state(seed);
+    #ifdef EIGEN_USE_SYCL
+    // In SYCL it is not possible to build PCG_XSH_RS_state in one step.
+    // Therefor, we need two steps to initializate the m_state.
+    // IN SYCL, the constructor of the functor is s called on the CPU
+    // and we get the clock seed here from the CPU. However, This seed is
+    //the same for all the thread. As unlike CUDA, the thread.ID, BlockID, etc is not a global function.
+    // and only  available on the Operator() function (which is called on the GPU).
+    // Therefore, the thread Id injection is not available at this stage. However when the operator()
+    //is called the thread ID will be avilable. So inside the opeator,
+    // we add the thrreadID, BlockId,... (which is equivalent of i)
+    //to the seed and construct the unique m_state per thead similar to cuda.
+    m_exec_once =false;
+   #endif
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE NormalRandomGenerator(
+      const NormalRandomGenerator& other) {
+    m_state = other.m_state;
+#ifdef EIGEN_USE_SYCL
+    m_exec_once=other.m_exec_once;
+#endif
+  }
+
+ template<typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T operator()(Index i) const {
+    #ifdef EIGEN_USE_SYCL
+    if(!m_exec_once) {
+      // This is the second stage of adding thread Id to the CPU clock seed and build unique seed per thread
+      m_state += (i * 6364136223846793005ULL);
+      m_exec_once =true;
+    }
+    #endif
+    T result = RandomToTypeNormal<T>(&m_state, i);
+    return result;
+  }
+
+  template<typename Packet, typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Packet packetOp(Index i) const {
+    const int packetSize = internal::unpacket_traits<Packet>::size;
+    EIGEN_ALIGN_MAX T values[packetSize];
+    #ifdef EIGEN_USE_SYCL
+    if(!m_exec_once) {
+      // This is the second stage of adding thread Id to the CPU clock seed and build unique seed per thread
+      m_state += (i * 6364136223846793005ULL);
+      m_exec_once =true;
+    }
+    #endif
+    EIGEN_UNROLL_LOOP
+    for (int j = 0; j < packetSize; ++j) {
+      values[j] = RandomToTypeNormal<T>(&m_state, i);
+    }
+    return internal::pload<Packet>(values);
+  }
+
+ private:
+  mutable uint64_t m_state;
+   #ifdef EIGEN_USE_SYCL
+  mutable bool m_exec_once;
+  #endif
+};
+
+
+template <typename Scalar>
+struct functor_traits<NormalRandomGenerator<Scalar> > {
+  enum {
+    // On average, we need to generate about 3 random numbers
+    // 15 mul, 8 add, 1.5 logs
+    Cost = 3 * functor_traits<UniformRandomGenerator<Scalar> >::Cost +
+           15 * NumTraits<Scalar>::AddCost + 8 * NumTraits<Scalar>::AddCost +
+           3 * functor_traits<scalar_log_op<Scalar> >::Cost / 2,
+    PacketAccess = NormalRandomGenerator<Scalar>::PacketAccess
+  };
+};
+
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_RANDOM_H