Added support for patch extraction
diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index 79510fd..0dac95e 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -1,6 +1,7 @@
 // This file is part of Eigen, a lightweight C++ template library
 // for linear algebra.
 //
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
 // Copyright (C) 2013 Christian Seiler <christian@iwakd.de>
 //
 // This Source Code Form is subject to the terms of the Mozilla
@@ -27,6 +28,11 @@
 
 #include <cstddef>
 #include <cstring>
+#include <stdint.h>
+
+#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+#include <curand_kernel.h>
+#endif
 
 #include "Eigen/Core"
 
@@ -46,6 +52,7 @@
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h"
+#include "unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index cadeb3b..27c10f6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -243,6 +243,12 @@
       return TensorConcatenationOp<Axis, const Derived, const OtherDerived>(derived(), other.derived(), axis);
     }
 
+    template <typename PatchDims> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorPatchOp<const PatchDims, const Derived>
+    extract_patches(const PatchDims& patch_dims) const {
+      return TensorPatchOp<const PatchDims, const Derived>(derived(), patch_dims);
+    }
+
     // Morphing operators.
     template <typename NewDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorReshapingOp<const NewDimensions, const Derived>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 86ddd1a..67f4788 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -25,6 +25,7 @@
 template<typename Axis, typename LeftXprType, typename RightXprType> class TensorConcatenationOp;
 template<typename Dimensions, typename LeftXprType, typename RightXprType> class TensorContractionOp;
 template<typename Dimensions, typename InputXprType, typename KernelXprType> class TensorConvolutionOp;
+template<typename PatchDim, typename XprType> class TensorPatchOp;
 template<typename Broadcast, typename XprType> class TensorBroadcastingOp;
 template<std::size_t DimId, typename XprType> class TensorChippingOp;
 template<typename NewDimensions, typename XprType> class TensorReshapingOp;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
new file mode 100644
index 0000000..01f2daf
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -0,0 +1,212 @@
+// 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 {
+
+/** \class TensorPatch
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor patch class.
+  *
+  *
+  */
+namespace internal {
+template<typename PatchDim, typename XprType>
+struct traits<TensorPatchOp<PatchDim, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::packet_traits<Scalar>::type Packet;
+  typedef typename traits<XprType>::StorageKind StorageKind;
+  typedef typename traits<XprType>::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+};
+
+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
+
+
+
+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::internal::traits<TensorPatchOp>::Packet Packet;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+  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;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+  };
+
+  EIGEN_DEVICE_FUNC 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();
+    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];
+    }
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    // Find the location of the first element of the patch.
+    Index patchIndex = index / m_outputStrides[NumDims - 1];
+    // Find the offset of the element wrt the location of the first element.
+    Index patchOffset = index - patchIndex * m_outputStrides[NumDims - 1];
+
+    Index inputIndex = 0;
+    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];
+    }
+    inputIndex += (patchIndex + patchOffset);
+    return m_impl.coeff(inputIndex);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
+    EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+packetSize-1 < dimensions().TotalSize());
+
+    Index indices[2] = {index, index + packetSize - 1};
+    Index patchIndices[2] = {indices[0] / m_outputStrides[NumDims - 1],
+                             indices[1] / m_outputStrides[NumDims - 1]};
+    Index patchOffsets[2] = {indices[0] - patchIndices[0] * m_outputStrides[NumDims - 1],
+                             indices[1] - patchIndices[1] * m_outputStrides[NumDims - 1]};
+
+    Index inputIndices[2] = {0, 0};
+    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];
+    }
+    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_DEFAULT CoeffReturnType values[packetSize];
+      values[0] = m_impl.coeff(inputIndices[0]);
+      values[packetSize-1] = m_impl.coeff(inputIndices[1]);
+      for (int i = 1; i < packetSize-1; ++i) {
+        values[i] = coeff(index+i);
+      }
+      PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+      return rslt;
+    }
+  }
+
+  Scalar* data() const { return NULL; }
+
+ 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
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 99593b5..d6c4359 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -119,6 +119,7 @@
   ei_add_test(cxx11_tensor_concatenation "-std=c++0x")
   ei_add_test(cxx11_tensor_morphing "-std=c++0x")
   ei_add_test(cxx11_tensor_padding "-std=c++0x")
+  ei_add_test(cxx11_tensor_patch "-std=c++0x")
   ei_add_test(cxx11_tensor_reduction "-std=c++0x")
   ei_add_test(cxx11_tensor_shuffling "-std=c++0x")
   ei_add_test(cxx11_tensor_striding "-std=c++0x")
diff --git a/unsupported/test/cxx11_tensor_patch.cpp b/unsupported/test/cxx11_tensor_patch.cpp
new file mode 100644
index 0000000..e2ba5bf
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_patch.cpp
@@ -0,0 +1,103 @@
+// 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/.
+
+#include "main.h"
+
+#include <Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+static void test_simple_patch()
+{
+  Tensor<float, 4> tensor(2,3,5,7);
+  tensor.setRandom();
+  array<ptrdiff_t, 4> patch_dims;
+  patch_dims[0] = 1;
+  patch_dims[1] = 1;
+  patch_dims[2] = 1;
+  patch_dims[3] = 1;
+
+  Tensor<float, 5> no_patch;
+  no_patch = tensor.extract_patches(patch_dims);
+
+  VERIFY_IS_EQUAL(no_patch.dimension(0), 1);
+  VERIFY_IS_EQUAL(no_patch.dimension(1), 1);
+  VERIFY_IS_EQUAL(no_patch.dimension(2), 1);
+  VERIFY_IS_EQUAL(no_patch.dimension(3), 1);
+  VERIFY_IS_EQUAL(no_patch.dimension(4), tensor.size());
+
+  for (int i = 0; i < tensor.size(); ++i) {
+    VERIFY_IS_EQUAL(tensor.data()[i], no_patch.data()[i]);
+  }
+
+  patch_dims[0] = 1;
+  patch_dims[1] = 2;
+  patch_dims[2] = 2;
+  patch_dims[3] = 1;
+  Tensor<float, 5> twod_patch;
+  twod_patch = tensor.extract_patches(patch_dims);
+
+  VERIFY_IS_EQUAL(twod_patch.dimension(0), 1);
+  VERIFY_IS_EQUAL(twod_patch.dimension(1), 2);
+  VERIFY_IS_EQUAL(twod_patch.dimension(2), 2);
+  VERIFY_IS_EQUAL(twod_patch.dimension(3), 1);
+  VERIFY_IS_EQUAL(twod_patch.dimension(4), 2*2*4*7);
+
+  for (int i = 0; i < 2; ++i) {
+    for (int j = 0; j < 2; ++j) {
+      for (int k = 0; k < 4; ++k) {
+        for (int l = 0; l < 7; ++l) {
+          int patch_loc = i + 2 * (j + 2 * (k + 4 * l));
+          for (int x = 0; x < 2; ++x) {
+            for (int y = 0; y < 2; ++y) {
+              VERIFY_IS_EQUAL(tensor(i,j+x,k+y,l), twod_patch(0,x,y,0,patch_loc));
+            }
+          }
+        }
+      }
+    }
+  }
+
+  patch_dims[0] = 1;
+  patch_dims[1] = 2;
+  patch_dims[2] = 3;
+  patch_dims[3] = 5;
+  Tensor<float, 5> threed_patch;
+  threed_patch = tensor.extract_patches(patch_dims);
+
+  VERIFY_IS_EQUAL(threed_patch.dimension(0), 1);
+  VERIFY_IS_EQUAL(threed_patch.dimension(1), 2);
+  VERIFY_IS_EQUAL(threed_patch.dimension(2), 3);
+  VERIFY_IS_EQUAL(threed_patch.dimension(3), 5);
+  VERIFY_IS_EQUAL(threed_patch.dimension(4), 2*2*3*3);
+
+  for (int i = 0; i < 2; ++i) {
+    for (int j = 0; j < 2; ++j) {
+      for (int k = 0; k < 3; ++k) {
+        for (int l = 0; l < 3; ++l) {
+          int patch_loc = i + 2 * (j + 2 * (k + 3 * l));
+          for (int x = 0; x < 2; ++x) {
+            for (int y = 0; y < 3; ++y) {
+              for (int z = 0; z < 5; ++z) {
+                VERIFY_IS_EQUAL(tensor(i,j+x,k+y,l+z), threed_patch(0,x,y,z,patch_loc));
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+
+void test_cxx11_tensor_patch()
+{
+   CALL_SUBTEST(test_simple_patch());
+   //   CALL_SUBTEST(test_expr_shuffling());
+}