// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
// SPDX-FileCopyrightText: The Eigen Authors
// SPDX-License-Identifier: MPL-2.0

#ifndef EIGEN_TENSOR_TENSOR_VOLUME_PATCH_H
#define EIGEN_TENSOR_TENSOR_VOLUME_PATCH_H

// IWYU pragma: private
#include "./InternalHeaderCheck.h"

namespace Eigen {

namespace internal {

template <DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
struct traits<TensorVolumePatchOp<Planes, Rows, Cols, XprType>> : traits<XprType> {
  typedef std::remove_const_t<typename XprType::Scalar> Scalar;
  typedef traits<XprType> XprTraits;
  typedef typename XprTraits::StorageKind StorageKind;
  typedef typename XprTraits::Index Index;
  static constexpr int NumDimensions = XprTraits::NumDimensions + 1;
  static constexpr int Layout = XprTraits::Layout;
  typedef typename XprTraits::PointerType PointerType;
};

template <DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
struct eval<TensorVolumePatchOp<Planes, Rows, Cols, XprType>, Eigen::Dense> {
  typedef const TensorVolumePatchOp<Planes, Rows, Cols, XprType>& type;
};

}  // end namespace internal

/**
 * \ingroup Tensor_Module
 *
 * \brief Patch extraction specialized for processing of volumetric data.
 * This assumes that the input has at least 4 dimensions ordered as follows:
 *  - channels
 *  - planes
 *  - rows
 *  - columns
 *  - (optional) additional dimensions such as time or batch size.
 * Calling the volume patch code with patch_planes, patch_rows, and patch_cols
 * is equivalent to calling the regular patch extraction code with parameters
 * d, patch_planes, patch_rows, patch_cols, and 1 for all the additional
 * dimensions.
 */
template <DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename XprType>
class TensorVolumePatchOp : public TensorBase<TensorVolumePatchOp<Planes, Rows, Cols, XprType>, ReadOnlyAccessors> {
 public:
  typedef typename Eigen::internal::traits<TensorVolumePatchOp>::Scalar Scalar;
  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
  typedef typename XprType::CoeffReturnType CoeffReturnType;
  typedef typename Eigen::internal::ref_selector<TensorVolumePatchOp>::type Nested;
  typedef typename Eigen::internal::traits<TensorVolumePatchOp>::StorageKind StorageKind;
  typedef typename Eigen::internal::traits<TensorVolumePatchOp>::Index Index;

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(
      const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
      DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides, DenseIndex in_plane_strides,
      DenseIndex in_row_strides, DenseIndex in_col_strides, DenseIndex plane_inflate_strides,
      DenseIndex row_inflate_strides, DenseIndex col_inflate_strides, PaddingType padding_type, Scalar padding_value)
      : m_xpr(expr),
        m_patch_planes(patch_planes),
        m_patch_rows(patch_rows),
        m_patch_cols(patch_cols),
        m_plane_strides(plane_strides),
        m_row_strides(row_strides),
        m_col_strides(col_strides),
        m_in_plane_strides(in_plane_strides),
        m_in_row_strides(in_row_strides),
        m_in_col_strides(in_col_strides),
        m_plane_inflate_strides(plane_inflate_strides),
        m_row_inflate_strides(row_inflate_strides),
        m_col_inflate_strides(col_inflate_strides),
        m_padding_explicit(false),
        m_padding_top_z(0),
        m_padding_bottom_z(0),
        m_padding_top(0),
        m_padding_bottom(0),
        m_padding_left(0),
        m_padding_right(0),
        m_padding_type(padding_type),
        m_padding_value(padding_value) {}

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorVolumePatchOp(
      const XprType& expr, DenseIndex patch_planes, DenseIndex patch_rows, DenseIndex patch_cols,
      DenseIndex plane_strides, DenseIndex row_strides, DenseIndex col_strides, DenseIndex in_plane_strides,
      DenseIndex in_row_strides, DenseIndex in_col_strides, DenseIndex plane_inflate_strides,
      DenseIndex row_inflate_strides, DenseIndex col_inflate_strides, DenseIndex padding_top_z,
      DenseIndex padding_bottom_z, DenseIndex padding_top, DenseIndex padding_bottom, DenseIndex padding_left,
      DenseIndex padding_right, Scalar padding_value)
      : m_xpr(expr),
        m_patch_planes(patch_planes),
        m_patch_rows(patch_rows),
        m_patch_cols(patch_cols),
        m_plane_strides(plane_strides),
        m_row_strides(row_strides),
        m_col_strides(col_strides),
        m_in_plane_strides(in_plane_strides),
        m_in_row_strides(in_row_strides),
        m_in_col_strides(in_col_strides),
        m_plane_inflate_strides(plane_inflate_strides),
        m_row_inflate_strides(row_inflate_strides),
        m_col_inflate_strides(col_inflate_strides),
        m_padding_explicit(true),
        m_padding_top_z(padding_top_z),
        m_padding_bottom_z(padding_bottom_z),
        m_padding_top(padding_top),
        m_padding_bottom(padding_bottom),
        m_padding_left(padding_left),
        m_padding_right(padding_right),
        m_padding_type(PADDING_VALID),
        m_padding_value(padding_value) {}

  EIGEN_DEVICE_FUNC DenseIndex patch_planes() const { return m_patch_planes; }
  EIGEN_DEVICE_FUNC DenseIndex patch_rows() const { return m_patch_rows; }
  EIGEN_DEVICE_FUNC DenseIndex patch_cols() const { return m_patch_cols; }
  EIGEN_DEVICE_FUNC DenseIndex plane_strides() const { return m_plane_strides; }
  EIGEN_DEVICE_FUNC DenseIndex row_strides() const { return m_row_strides; }
  EIGEN_DEVICE_FUNC DenseIndex col_strides() const { return m_col_strides; }
  EIGEN_DEVICE_FUNC DenseIndex in_plane_strides() const { return m_in_plane_strides; }
  EIGEN_DEVICE_FUNC DenseIndex in_row_strides() const { return m_in_row_strides; }
  EIGEN_DEVICE_FUNC DenseIndex in_col_strides() const { return m_in_col_strides; }
  EIGEN_DEVICE_FUNC DenseIndex plane_inflate_strides() const { return m_plane_inflate_strides; }
  EIGEN_DEVICE_FUNC DenseIndex row_inflate_strides() const { return m_row_inflate_strides; }
  EIGEN_DEVICE_FUNC DenseIndex col_inflate_strides() const { return m_col_inflate_strides; }
  EIGEN_DEVICE_FUNC bool padding_explicit() const { return m_padding_explicit; }
  EIGEN_DEVICE_FUNC DenseIndex padding_top_z() const { return m_padding_top_z; }
  EIGEN_DEVICE_FUNC DenseIndex padding_bottom_z() const { return m_padding_bottom_z; }
  EIGEN_DEVICE_FUNC DenseIndex padding_top() const { return m_padding_top; }
  EIGEN_DEVICE_FUNC DenseIndex padding_bottom() const { return m_padding_bottom; }
  EIGEN_DEVICE_FUNC DenseIndex padding_left() const { return m_padding_left; }
  EIGEN_DEVICE_FUNC DenseIndex padding_right() const { return m_padding_right; }
  EIGEN_DEVICE_FUNC PaddingType padding_type() const { return m_padding_type; }
  EIGEN_DEVICE_FUNC Scalar padding_value() const { return m_padding_value; }

  EIGEN_DEVICE_FUNC const internal::remove_all_t<typename XprType::Nested>& expression() const { return m_xpr; }

 protected:
  typename XprType::Nested m_xpr;
  const DenseIndex m_patch_planes;
  const DenseIndex m_patch_rows;
  const DenseIndex m_patch_cols;
  const DenseIndex m_plane_strides;
  const DenseIndex m_row_strides;
  const DenseIndex m_col_strides;
  const DenseIndex m_in_plane_strides;
  const DenseIndex m_in_row_strides;
  const DenseIndex m_in_col_strides;
  const DenseIndex m_plane_inflate_strides;
  const DenseIndex m_row_inflate_strides;
  const DenseIndex m_col_inflate_strides;
  const bool m_padding_explicit;
  const DenseIndex m_padding_top_z;
  const DenseIndex m_padding_bottom_z;
  const DenseIndex m_padding_top;
  const DenseIndex m_padding_bottom;
  const DenseIndex m_padding_left;
  const DenseIndex m_padding_right;
  const PaddingType m_padding_type;
  const Scalar m_padding_value;
};

// Eval as rvalue
template <DenseIndex Planes, DenseIndex Rows, DenseIndex Cols, typename ArgType, typename Device>
struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, Device> {
  typedef TensorVolumePatchOp<Planes, Rows, Cols, ArgType> XprType;
  typedef typename XprType::Index Index;
  static constexpr int NumInputDims =
      internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
  static constexpr int NumDims = NumInputDims + 1;
  typedef DSizes<Index, NumDims> Dimensions;
  typedef std::remove_const_t<typename XprType::Scalar> Scalar;
  typedef typename XprType::CoeffReturnType CoeffReturnType;
  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
  static constexpr int PacketSize = PacketType<CoeffReturnType, Device>::size;
  typedef StorageMemory<CoeffReturnType, Device> Storage;
  typedef typename Storage::Type EvaluatorPointerType;

  static constexpr int Layout = TensorEvaluator<ArgType, Device>::Layout;
  enum {
    IsAligned = false,
    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
    // block() reads the argument one coefficient at a time through coeff() --
    // the contract the scalar executors already rely on for every evaluator --
    // so it requires no capability bit from the argument (same as
    // TensorReverse).
    BlockAccess = true,
    // The coeff/packet path pays ~10 divisions of index math per element; the
    // block path amortizes all of it over whole depth runs.
    PreferBlockAccess = true,
    CoordAccess = false,
    RawAccess = false
  };

  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
  typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
  typedef typename internal::TensorMaterializedBlock<Scalar, NumDims, Layout, Index> TensorBlock;
  //===--------------------------------------------------------------------===//

  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
      : m_impl(op.expression(), device), m_device(device) {
    EIGEN_STATIC_ASSERT((NumDims >= 5), YOU_MADE_A_PROGRAMMING_MISTAKE);

    m_paddingValue = op.padding_value();

    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();

    // Cache a few variables.
    EIGEN_IF_CONSTEXPR (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      m_inputDepth = input_dims[0];
      m_inputPlanes = input_dims[1];
      m_inputRows = input_dims[2];
      m_inputCols = input_dims[3];
    } else {
      m_inputDepth = input_dims[NumInputDims - 1];
      m_inputPlanes = input_dims[NumInputDims - 2];
      m_inputRows = input_dims[NumInputDims - 3];
      m_inputCols = input_dims[NumInputDims - 4];
    }

    m_plane_strides = op.plane_strides();
    m_row_strides = op.row_strides();
    m_col_strides = op.col_strides();

    // Input strides and effective input/patch size
    m_in_plane_strides = op.in_plane_strides();
    m_in_row_strides = op.in_row_strides();
    m_in_col_strides = op.in_col_strides();
    m_plane_inflate_strides = op.plane_inflate_strides();
    m_row_inflate_strides = op.row_inflate_strides();
    m_col_inflate_strides = op.col_inflate_strides();

    // The "effective" spatial size after inflating data with zeros.
    m_input_planes_eff = (m_inputPlanes - 1) * m_plane_inflate_strides + 1;
    m_input_rows_eff = (m_inputRows - 1) * m_row_inflate_strides + 1;
    m_input_cols_eff = (m_inputCols - 1) * m_col_inflate_strides + 1;
    m_patch_planes_eff = op.patch_planes() + (op.patch_planes() - 1) * (m_in_plane_strides - 1);
    m_patch_rows_eff = op.patch_rows() + (op.patch_rows() - 1) * (m_in_row_strides - 1);
    m_patch_cols_eff = op.patch_cols() + (op.patch_cols() - 1) * (m_in_col_strides - 1);

    if (op.padding_explicit()) {
      m_outputPlanes =
          numext::ceil((m_input_planes_eff + op.padding_top_z() + op.padding_bottom_z() - m_patch_planes_eff + 1.f) /
                       static_cast<float>(m_plane_strides));
      m_outputRows = numext::ceil((m_input_rows_eff + op.padding_top() + op.padding_bottom() - m_patch_rows_eff + 1.f) /
                                  static_cast<float>(m_row_strides));
      m_outputCols = numext::ceil((m_input_cols_eff + op.padding_left() + op.padding_right() - m_patch_cols_eff + 1.f) /
                                  static_cast<float>(m_col_strides));
      m_planePaddingTop = op.padding_top_z();
      m_rowPaddingTop = op.padding_top();
      m_colPaddingLeft = op.padding_left();
    } else {
      // Computing padding from the type
      switch (op.padding_type()) {
        case PADDING_VALID:
          m_outputPlanes =
              numext::ceil((m_input_planes_eff - m_patch_planes_eff + 1.f) / static_cast<float>(m_plane_strides));
          m_outputRows = numext::ceil((m_input_rows_eff - m_patch_rows_eff + 1.f) / static_cast<float>(m_row_strides));
          m_outputCols = numext::ceil((m_input_cols_eff - m_patch_cols_eff + 1.f) / static_cast<float>(m_col_strides));
          m_planePaddingTop = 0;
          m_rowPaddingTop = 0;
          m_colPaddingLeft = 0;
          break;
        case PADDING_SAME: {
          m_outputPlanes = numext::ceil(m_input_planes_eff / static_cast<float>(m_plane_strides));
          m_outputRows = numext::ceil(m_input_rows_eff / static_cast<float>(m_row_strides));
          m_outputCols = numext::ceil(m_input_cols_eff / static_cast<float>(m_col_strides));
          const Index dz = (m_outputPlanes - 1) * m_plane_strides + m_patch_planes_eff - m_input_planes_eff;
          const Index dy = (m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff;
          const Index dx = (m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff;
          m_planePaddingTop = dz / 2;
          m_rowPaddingTop = dy / 2;
          m_colPaddingLeft = dx / 2;
          break;
        }
        default: {
          eigen_assert(false && "unexpected padding");
          return;
        }
      }
    }
    eigen_assert(m_outputRows > 0);
    eigen_assert(m_outputCols > 0);
    eigen_assert(m_outputPlanes > 0);

    // Dimensions for result of extraction.
    EIGEN_IF_CONSTEXPR (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      // ColMajor
      // 0: depth
      // 1: patch_planes
      // 2: patch_rows
      // 3: patch_cols
      // 4: number of patches
      // 5 and beyond: anything else (such as batch).
      m_dimensions[0] = input_dims[0];
      m_dimensions[1] = op.patch_planes();
      m_dimensions[2] = op.patch_rows();
      m_dimensions[3] = op.patch_cols();
      m_dimensions[4] = m_outputPlanes * m_outputRows * m_outputCols;
      for (int i = 5; i < NumDims; ++i) {
        m_dimensions[i] = input_dims[i - 1];
      }
    } else {
      // RowMajor
      // NumDims-1: depth
      // NumDims-2: patch_planes
      // NumDims-3: patch_rows
      // NumDims-4: patch_cols
      // NumDims-5: number of patches
      // NumDims-6 and beyond: anything else (such as batch).
      m_dimensions[NumDims - 1] = input_dims[NumInputDims - 1];
      m_dimensions[NumDims - 2] = op.patch_planes();
      m_dimensions[NumDims - 3] = op.patch_rows();
      m_dimensions[NumDims - 4] = op.patch_cols();
      m_dimensions[NumDims - 5] = m_outputPlanes * m_outputRows * m_outputCols;
      for (int i = NumDims - 6; i >= 0; --i) {
        m_dimensions[i] = input_dims[i];
      }
    }

    // Strides for the output tensor.
    EIGEN_IF_CONSTEXPR (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      m_rowStride = m_dimensions[1];
      m_colStride = m_dimensions[2] * m_rowStride;
      m_patchStride = m_colStride * m_dimensions[3] * m_dimensions[0];
      m_otherStride = m_patchStride * m_dimensions[4];
    } else {
      m_rowStride = m_dimensions[NumDims - 2];
      m_colStride = m_dimensions[NumDims - 3] * m_rowStride;
      m_patchStride = m_colStride * m_dimensions[NumDims - 4] * m_dimensions[NumDims - 1];
      m_otherStride = m_patchStride * m_dimensions[NumDims - 5];
    }

    // Strides for navigating through the input tensor.
    m_planeInputStride = m_inputDepth;
    m_rowInputStride = m_inputDepth * m_inputPlanes;
    m_colInputStride = m_inputDepth * m_inputRows * m_inputPlanes;
    m_otherInputStride = m_inputDepth * m_inputRows * m_inputCols * m_inputPlanes;

    m_outputPlanesRows = m_outputPlanes * m_outputRows;

    // Fast representations of different variables.
    m_fastOtherStride = internal::TensorIntDivisor<Index>(m_otherStride);

    m_fastPatchStride = internal::TensorIntDivisor<Index>(m_patchStride);
    m_fastColStride = internal::TensorIntDivisor<Index>(m_colStride);
    m_fastRowStride = internal::TensorIntDivisor<Index>(m_rowStride);
    m_fastInputRowStride = internal::TensorIntDivisor<Index>(m_row_inflate_strides);
    m_fastInputColStride = internal::TensorIntDivisor<Index>(m_col_inflate_strides);
    m_fastInputPlaneStride = internal::TensorIntDivisor<Index>(m_plane_inflate_strides);
    m_fastInputColsEff = internal::TensorIntDivisor<Index>(m_input_cols_eff);
    m_fastOutputPlanes = internal::TensorIntDivisor<Index>(m_outputPlanes);
    m_fastOutputPlanesRows = internal::TensorIntDivisor<Index>(m_outputPlanesRows);

    EIGEN_IF_CONSTEXPR (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
      m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[0]);
    } else {
      m_fastOutputDepth = internal::TensorIntDivisor<Index>(m_dimensions[NumDims - 1]);
    }
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }

  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
    m_impl.evalSubExprsIfNeeded(nullptr);
    return true;
  }

#ifdef EIGEN_USE_THREADS
  template <typename EvalSubExprsCallback>
  EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(EvaluatorPointerType /*data*/, EvalSubExprsCallback done) {
    m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
  }
#endif  // EIGEN_USE_THREADS

  EIGEN_STRONG_INLINE void cleanup() { m_impl.cleanup(); }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
    // Patch index corresponding to the passed in index.
    const Index patchIndex = index / m_fastPatchStride;

    // Spatial offset within the patch. This has to be translated into 3D
    // coordinates within the patch.
    const Index patchOffset = (index - patchIndex * m_patchStride) / m_fastOutputDepth;

    // Batch, etc.
    const Index otherIndex = (NumDims == 5) ? 0 : index / m_fastOtherStride;
    const Index patch3DIndex = (NumDims == 5) ? patchIndex : (index - otherIndex * m_otherStride) / m_fastPatchStride;

    // Calculate column index in the input original tensor.
    const Index colIndex = patch3DIndex / m_fastOutputPlanesRows;
    const Index colOffset = patchOffset / m_fastColStride;
    const Index inputCol = colIndex * m_col_strides + colOffset * m_in_col_strides - m_colPaddingLeft;
    const Index origInputCol =
        (m_col_inflate_strides == 1) ? inputCol : ((inputCol >= 0) ? (inputCol / m_fastInputColStride) : 0);
    if (inputCol < 0 || inputCol >= m_input_cols_eff ||
        ((m_col_inflate_strides != 1) && (inputCol != origInputCol * m_col_inflate_strides))) {
      return Scalar(m_paddingValue);
    }

    // Calculate row index in the original input tensor.
    const Index rowIndex = (patch3DIndex - colIndex * m_outputPlanesRows) / m_fastOutputPlanes;
    const Index rowOffset = (patchOffset - colOffset * m_colStride) / m_fastRowStride;
    const Index inputRow = rowIndex * m_row_strides + rowOffset * m_in_row_strides - m_rowPaddingTop;
    const Index origInputRow =
        (m_row_inflate_strides == 1) ? inputRow : ((inputRow >= 0) ? (inputRow / m_fastInputRowStride) : 0);
    if (inputRow < 0 || inputRow >= m_input_rows_eff ||
        ((m_row_inflate_strides != 1) && (inputRow != origInputRow * m_row_inflate_strides))) {
      return Scalar(m_paddingValue);
    }

    // Calculate plane index in the original input tensor.
    const Index planeIndex = patch3DIndex - m_outputPlanes * (colIndex * m_outputRows + rowIndex);
    const Index planeOffset = patchOffset - colOffset * m_colStride - rowOffset * m_rowStride;
    const Index inputPlane = planeIndex * m_plane_strides + planeOffset * m_in_plane_strides - m_planePaddingTop;
    const Index origInputPlane =
        (m_plane_inflate_strides == 1) ? inputPlane : ((inputPlane >= 0) ? (inputPlane / m_fastInputPlaneStride) : 0);
    if (inputPlane < 0 || inputPlane >= m_input_planes_eff ||
        ((m_plane_inflate_strides != 1) && (inputPlane != origInputPlane * m_plane_inflate_strides))) {
      return Scalar(m_paddingValue);
    }

    constexpr int depth_index = static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
    const Index depth = index - (index / m_fastOutputDepth) * m_dimensions[depth_index];

    const Index inputIndex = depth + origInputRow * m_rowInputStride + origInputCol * m_colInputStride +
                             origInputPlane * m_planeInputStride + otherIndex * m_otherInputStride;

    return m_impl.coeff(inputIndex);
  }

  template <int LoadMode>
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const {
    eigen_assert(index + PacketSize - 1 < dimensions().TotalSize());

    if (m_in_row_strides != 1 || m_in_col_strides != 1 || m_row_inflate_strides != 1 || m_col_inflate_strides != 1 ||
        m_in_plane_strides != 1 || m_plane_inflate_strides != 1) {
      return packetWithPossibleZero(index);
    }

    const Index indices[2] = {index, index + PacketSize - 1};
    const Index patchIndex = indices[0] / m_fastPatchStride;
    if (patchIndex != indices[1] / m_fastPatchStride) {
      return packetWithPossibleZero(index);
    }
    const Index otherIndex = (NumDims == 5) ? 0 : indices[0] / m_fastOtherStride;
    eigen_assert(otherIndex == indices[1] / m_fastOtherStride);

    // Find the offset of the element wrt the location of the first element.
    Index first_entry = (indices[0] - patchIndex * m_patchStride) / m_fastOutputDepth;
    Index second_entry = PacketSize == 1 ? first_entry : (indices[1] - patchIndex * m_patchStride) / m_fastOutputDepth;

    const Index patchOffsets[2] = {first_entry, second_entry};

    const Index patch3DIndex =
        (NumDims == 5) ? patchIndex : (indices[0] - otherIndex * m_otherStride) / m_fastPatchStride;
    eigen_assert(patch3DIndex == (indices[1] - otherIndex * m_otherStride) / m_fastPatchStride);

    const Index colIndex = patch3DIndex / m_fastOutputPlanesRows;
    const Index colOffsets[2] = {patchOffsets[0] / m_fastColStride, patchOffsets[1] / m_fastColStride};

    // Calculate col indices in the original input tensor.
    const Index inputCols[2] = {colIndex * m_col_strides + colOffsets[0] - m_colPaddingLeft,
                                colIndex * m_col_strides + colOffsets[1] - m_colPaddingLeft};
    if (inputCols[1] < 0 || inputCols[0] >= m_inputCols) {
      return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
    }

    if (inputCols[0] != inputCols[1]) {
      return packetWithPossibleZero(index);
    }

    const Index rowIndex = (patch3DIndex - colIndex * m_outputPlanesRows) / m_fastOutputPlanes;
    const Index rowOffsets[2] = {(patchOffsets[0] - colOffsets[0] * m_colStride) / m_fastRowStride,
                                 (patchOffsets[1] - colOffsets[1] * m_colStride) / m_fastRowStride};
    eigen_assert(rowOffsets[0] <= rowOffsets[1]);
    // Calculate row indices in the original input tensor.
    const Index inputRows[2] = {rowIndex * m_row_strides + rowOffsets[0] - m_rowPaddingTop,
                                rowIndex * m_row_strides + rowOffsets[1] - m_rowPaddingTop};

    if (inputRows[1] < 0 || inputRows[0] >= m_inputRows) {
      return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
    }

    if (inputRows[0] != inputRows[1]) {
      return packetWithPossibleZero(index);
    }

    const Index planeIndex = patch3DIndex - m_outputPlanes * (colIndex * m_outputRows + rowIndex);
    const Index planeOffsets[2] = {patchOffsets[0] - colOffsets[0] * m_colStride - rowOffsets[0] * m_rowStride,
                                   patchOffsets[1] - colOffsets[1] * m_colStride - rowOffsets[1] * m_rowStride};
    eigen_assert(planeOffsets[0] <= planeOffsets[1]);
    const Index inputPlanes[2] = {planeIndex * m_plane_strides + planeOffsets[0] - m_planePaddingTop,
                                  planeIndex * m_plane_strides + planeOffsets[1] - m_planePaddingTop};

    if (inputPlanes[1] < 0 || inputPlanes[0] >= m_inputPlanes) {
      return internal::pset1<PacketReturnType>(Scalar(m_paddingValue));
    }

    if (inputPlanes[0] >= 0 && inputPlanes[1] < m_inputPlanes) {
      // no padding
      constexpr int depth_index = static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 0 : NumDims - 1;
      const Index depth = index - (index / m_fastOutputDepth) * m_dimensions[depth_index];
      const Index inputIndex = depth + inputRows[0] * m_rowInputStride + inputCols[0] * m_colInputStride +
                               m_planeInputStride * inputPlanes[0] + otherIndex * m_otherInputStride;
      return m_impl.template packet<Unaligned>(inputIndex);
    }

    return packetWithPossibleZero(index);
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE internal::TensorBlockResourceRequirements getResourceRequirements() const {
    const size_t target_size = m_device.firstLevelCacheSize();
    // In-bounds output coefficients read the argument once and every output
    // coefficient is stored once (padding runs make this a slight
    // over-estimate). Pass the full cost explicitly rather than adding to
    // skewed()'s default load+store seed, which would double-count the
    // baseline byte traffic and halve the tile size.
    const TensorOpCost cost_per_coeff = m_impl.costPerCoeff(/*vectorized=*/false) + TensorOpCost(0, sizeof(Scalar), 0);
    return internal::TensorBlockResourceRequirements::withShapeAndSize<Scalar>(
        internal::TensorBlockShapeType::kSkewedInnerDims, target_size, cost_per_coeff);
  }

  // Materializes the block by iterating patch/col/row/plane coordinates and
  // either copying the (always input-contiguous) depth run or filling it with
  // the padding value. All per-coordinate index math and bounds checks are
  // amortized over a whole depth run instead of paid per coefficient.
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
                                                          bool /*root_of_expr_ast*/ = false) const {
    constexpr bool is_col_major = static_cast<int>(Layout) == static_cast<int>(ColMajor);

    if (desc.size() == 0) {
      return TensorBlock(internal::TensorBlockKind::kView, nullptr, desc.dimensions());
    }

    typename TensorBlock::Storage block_storage = TensorBlock::prepareStorage(desc, scratch);
    Scalar* block_buffer = block_storage.data();

    // Output coordinates of the block's corner.
    const DSizes<Index, NumDims> output_strides = internal::strides<Layout>(m_dimensions);
    array<Index, NumDims> coords;
    Index remaining = desc.offset();
    EIGEN_IF_CONSTEXPR (is_col_major) {
      for (int i = NumDims - 1; i > 0; --i) {
        coords[i] = remaining / output_strides[i];
        remaining -= coords[i] * output_strides[i];
      }
      coords[0] = remaining;
    } else {
      for (int i = 0; i < NumDims - 1; ++i) {
        coords[i] = remaining / output_strides[i];
        remaining -= coords[i] * output_strides[i];
      }
      coords[NumDims - 1] = remaining;
    }

    // Output dimensions: depth, patch plane/row/col offset, 3d patch index.
    const int dd = is_col_major ? 0 : NumDims - 1;
    const int nd = is_col_major ? 1 : NumDims - 2;
    const int rd = is_col_major ? 2 : NumDims - 3;
    const int cd = is_col_major ? 3 : NumDims - 4;
    const int pd = is_col_major ? 4 : NumDims - 5;

    const Index depth_start = coords[dd];
    const Index depth_size = desc.dimension(dd);
    const Index plane_start = coords[nd];
    const Index plane_size = desc.dimension(nd);
    const Index row_start = coords[rd];
    const Index row_size = desc.dimension(rd);
    const Index col_start = coords[cd];
    const Index col_size = desc.dimension(cd);
    const Index patch_start = coords[pd];
    const Index patch_size = desc.dimension(pd);

    // Odometer over the remaining (batch etc.) dimensions, tracking the input
    // offset they contribute.
    array<Index, NumDims> other_sizes;
    array<Index, NumDims> other_src_stride;
    array<Index, NumDims> other_count;
    int num_other = 0;
    Index src_other = 0;
    {
      Index in_stride = m_otherInputStride;
      for (int k = 5; k < NumDims; ++k) {
        const int d = is_col_major ? k : NumDims - 1 - k;
        other_sizes[num_other] = desc.dimension(d);
        other_src_stride[num_other] = in_stride;
        other_count[num_other] = 0;
        src_other += coords[d] * in_stride;
        in_stride *= m_dimensions[d];
        ++num_other;
      }
    }

    typedef internal::StridedLinearBufferCopy<Scalar, Index> LinCopy;

    // The loop nest below visits the block in exactly its memory order (the
    // storage returned by prepareStorage() is dense with the block's own
    // layout-order strides), so the destination is one running cursor.
    Index dst = 0;
    for (;;) {
      for (Index p = 0; p < patch_size; ++p) {
        const Index patch3DIndex = patch_start + p;
        const Index colIndex = patch3DIndex / m_fastOutputPlanesRows;
        const Index rowIndex = (patch3DIndex - colIndex * m_outputPlanesRows) / m_fastOutputPlanes;
        const Index planeIndex = patch3DIndex - m_outputPlanes * (colIndex * m_outputRows + rowIndex);

        for (Index c = 0; c < col_size; ++c) {
          const Index colOffset = col_start + c;
          const Index inputCol = colIndex * m_col_strides + colOffset * m_in_col_strides - m_colPaddingLeft;
          Index origInputCol = inputCol;
          bool col_valid = inputCol >= 0 && inputCol < m_input_cols_eff;
          if (col_valid && m_col_inflate_strides != 1) {
            origInputCol = inputCol / m_fastInputColStride;
            col_valid = (inputCol == origInputCol * m_col_inflate_strides);
          }

          for (Index r = 0; r < row_size; ++r) {
            const Index rowOffset = row_start + r;
            bool row_valid = col_valid;
            Index origInputRow = 0;
            if (row_valid) {
              const Index inputRow = rowIndex * m_row_strides + rowOffset * m_in_row_strides - m_rowPaddingTop;
              row_valid = inputRow >= 0 && inputRow < m_input_rows_eff;
              if (row_valid) {
                origInputRow = inputRow;
                if (m_row_inflate_strides != 1) {
                  origInputRow = inputRow / m_fastInputRowStride;
                  row_valid = (inputRow == origInputRow * m_row_inflate_strides);
                }
              }
            }

            for (Index n = 0; n < plane_size; ++n) {
              const Index planeOffset = plane_start + n;
              bool valid = row_valid;
              Index origInputPlane = 0;
              if (valid) {
                const Index inputPlane =
                    planeIndex * m_plane_strides + planeOffset * m_in_plane_strides - m_planePaddingTop;
                valid = inputPlane >= 0 && inputPlane < m_input_planes_eff;
                if (valid) {
                  origInputPlane = inputPlane;
                  if (m_plane_inflate_strides != 1) {
                    origInputPlane = inputPlane / m_fastInputPlaneStride;
                    valid = (inputPlane == origInputPlane * m_plane_inflate_strides);
                  }
                }
              }

              if (valid) {
                const Index src = depth_start + origInputPlane * m_planeInputStride + origInputRow * m_rowInputStride +
                                  origInputCol * m_colInputStride + src_other;
                for (Index d = 0; d < depth_size; ++d) {
                  block_buffer[dst + d] = m_impl.coeff(src + d);
                }
              } else {
                LinCopy::template Run<LinCopy::Kind::FillLinear>(typename LinCopy::Dst(dst, 1, block_buffer),
                                                                 typename LinCopy::Src(0, 0, &m_paddingValue),
                                                                 depth_size);
              }
              dst += depth_size;
            }
          }
        }
      }

      int k = 0;
      for (; k < num_other; ++k) {
        if (++other_count[k] < other_sizes[k]) {
          src_other += other_src_stride[k];
          break;
        }
        other_count[k] = 0;
        src_other -= other_src_stride[k] * (other_sizes[k] - 1);
      }
      if (k == num_other) break;
    }
    eigen_assert(dst == desc.size());

    return block_storage.AsTensorMaterializedBlock();
  }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
    const double compute_cost =
        10 * TensorOpCost::DivCost<Index>() + 21 * TensorOpCost::MulCost<Index>() + 8 * TensorOpCost::AddCost<Index>();
    return TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
  }

  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return nullptr; }

  const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }

  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index planePaddingTop() const { return m_planePaddingTop; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowPaddingTop() const { return m_rowPaddingTop; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colPaddingLeft() const { return m_colPaddingLeft; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputPlanes() const { return m_outputPlanes; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputRows() const { return m_outputRows; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index outputCols() const { return m_outputCols; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userPlaneStride() const { return m_plane_strides; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userRowStride() const { return m_row_strides; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userColStride() const { return m_col_strides; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInPlaneStride() const { return m_in_plane_strides; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInRowStride() const { return m_in_row_strides; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index userInColStride() const { return m_in_col_strides; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index planeInflateStride() const { return m_plane_inflate_strides; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowInflateStride() const { return m_row_inflate_strides; }
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colInflateStride() const { return m_col_inflate_strides; }

 protected:
  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const {
    EIGEN_ALIGN_TO_BOUNDARY(internal::unpacket_traits<PacketReturnType>::alignment)
    std::remove_const_t<CoeffReturnType> 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;
  }

  Dimensions m_dimensions;

  // Parameters passed to the constructor.
  Index m_plane_strides;
  Index m_row_strides;
  Index m_col_strides;

  Index m_outputPlanes;
  Index m_outputRows;
  Index m_outputCols;

  Index m_planePaddingTop;
  Index m_rowPaddingTop;
  Index m_colPaddingLeft;

  Index m_in_plane_strides;
  Index m_in_row_strides;
  Index m_in_col_strides;

  Index m_plane_inflate_strides;
  Index m_row_inflate_strides;
  Index m_col_inflate_strides;

  // Cached input size.
  Index m_inputDepth;
  Index m_inputPlanes;
  Index m_inputRows;
  Index m_inputCols;

  // Other cached variables.
  Index m_outputPlanesRows;

  // Effective input/patch post-inflation size.
  Index m_input_planes_eff;
  Index m_input_rows_eff;
  Index m_input_cols_eff;
  Index m_patch_planes_eff;
  Index m_patch_rows_eff;
  Index m_patch_cols_eff;

  // Strides for the output tensor.
  Index m_otherStride;
  Index m_patchStride;
  Index m_rowStride;
  Index m_colStride;

  // Strides for the input tensor.
  Index m_planeInputStride;
  Index m_rowInputStride;
  Index m_colInputStride;
  Index m_otherInputStride;

  internal::TensorIntDivisor<Index> m_fastOtherStride;
  internal::TensorIntDivisor<Index> m_fastPatchStride;
  internal::TensorIntDivisor<Index> m_fastColStride;
  internal::TensorIntDivisor<Index> m_fastRowStride;
  internal::TensorIntDivisor<Index> m_fastInputPlaneStride;
  internal::TensorIntDivisor<Index> m_fastInputRowStride;
  internal::TensorIntDivisor<Index> m_fastInputColStride;
  internal::TensorIntDivisor<Index> m_fastInputColsEff;
  internal::TensorIntDivisor<Index> m_fastOutputPlanesRows;
  internal::TensorIntDivisor<Index> m_fastOutputPlanes;
  internal::TensorIntDivisor<Index> m_fastOutputDepth;

  Scalar m_paddingValue;

  TensorEvaluator<ArgType, Device> m_impl;
  const Device EIGEN_DEVICE_REF m_device;
};

}  // end namespace Eigen

#endif  // EIGEN_TENSOR_TENSOR_VOLUME_PATCH_H