diff options
Diffstat (limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h')
-rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h | 77 |
1 files changed, 11 insertions, 66 deletions
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h index 11e510414..72594a05c 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h @@ -32,7 +32,6 @@ struct traits<TensorImagePatchOp<Rows, Cols, XprType> > : public traits<XprType> { typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar; typedef traits<XprType> XprTraits; - typedef typename packet_traits<Scalar>::type Packet; typedef typename XprTraits::StorageKind StorageKind; typedef typename XprTraits::Index Index; typedef typename XprType::Nested Nested; @@ -60,10 +59,8 @@ class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprT { public: typedef typename Eigen::internal::traits<TensorImagePatchOp>::Scalar Scalar; - typedef typename Eigen::internal::traits<TensorImagePatchOp>::Packet Packet; typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; typedef typename Eigen::internal::nested<TensorImagePatchOp>::type Nested; typedef typename Eigen::internal::traits<TensorImagePatchOp>::StorageKind StorageKind; typedef typename Eigen::internal::traits<TensorImagePatchOp>::Index Index; @@ -167,7 +164,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device> IsAligned = false, PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess, Layout = TensorEvaluator<ArgType, Device>::Layout, - CoordAccess = NumDims == 5, + CoordAccess = false, + RawAccess = false }; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) @@ -228,8 +226,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device> 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)); // Calculate the padding - m_rowPaddingTop = ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2; - m_colPaddingLeft = ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2; + m_rowPaddingTop = numext::maxi<Index>(0, ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2); + m_colPaddingLeft = numext::maxi<Index>(0, ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2); break; case PADDING_SAME: m_outputRows = numext::ceil(m_input_rows_eff / static_cast<float>(m_row_strides)); @@ -296,8 +294,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device> m_fastOtherStride = internal::TensorIntDivisor<Index>(m_otherStride); m_fastPatchStride = internal::TensorIntDivisor<Index>(m_patchStride); m_fastColStride = internal::TensorIntDivisor<Index>(m_colStride); - m_fastInputRowStride = internal::TensorIntDivisor<Index>(m_row_inflate_strides); - m_fastInputColStride = internal::TensorIntDivisor<Index>(m_col_inflate_strides); + m_fastInflateRowStride = internal::TensorIntDivisor<Index>(m_row_inflate_strides); + m_fastInflateColStride = internal::TensorIntDivisor<Index>(m_col_inflate_strides); m_fastInputColsEff = internal::TensorIntDivisor<Index>(m_input_cols_eff); // Number of patches in the width dimension. @@ -310,7 +308,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device> } typedef typename XprType::CoeffReturnType CoeffReturnType; - typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType; EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; } @@ -338,7 +336,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device> const Index colIndex = patch2DIndex / m_fastOutputRows; 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); + const Index origInputCol = (m_col_inflate_strides == 1) ? inputCol : ((inputCol >= 0) ? (inputCol / m_fastInflateColStride) : 0); if (inputCol < 0 || inputCol >= m_input_cols_eff || ((m_col_inflate_strides != 1) && (inputCol != origInputCol * m_col_inflate_strides))) { return Scalar(m_paddingValue); @@ -348,7 +346,7 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device> const Index rowIndex = patch2DIndex - colIndex * m_outputRows; const Index rowOffset = patchOffset - colOffset * m_colStride; 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); + const Index origInputRow = (m_row_inflate_strides == 1) ? inputRow : ((inputRow >= 0) ? (inputRow / m_fastInflateRowStride) : 0); if (inputRow < 0 || inputRow >= m_input_rows_eff || ((m_row_inflate_strides != 1) && (inputRow != origInputRow * m_row_inflate_strides))) { return Scalar(m_paddingValue); @@ -436,59 +434,6 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device> Index rowInflateStride() const { return m_row_inflate_strides; } Index colInflateStride() const { return m_col_inflate_strides; } - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array<Index, NumDims>& coords) const - { - // Location of the first element of the patch. - // ColMajor - // 0: d, 1: patch_rows, 2: patch_cols, 3: number of patches, 4: number of batches - // RowMajor - // 0: number of batches, 1: number of patches, 2: patch_cols , 3: patch_rows, 4: d - const Index patch2DIndex = coords[static_cast<int>(Layout) == static_cast<int>(ColMajor) ? 3 : 1]; - - array<Index, NumDims-1> inputCoords; - Index input_col_idx = patch2DIndex / m_fastInputColsEff; - Index inputCol = input_col_idx + coords[1] * m_in_row_strides - m_rowPaddingTop; - Index inputRow = patch2DIndex - input_col_idx * m_input_cols_eff + coords[2] * m_in_col_strides - m_colPaddingLeft; - const Index origInputCol = (m_col_inflate_strides == 1) ? inputCol : ((inputCol >= 0) ? (inputCol / m_fastInputColStride) : 0); - const Index origInputRow = (m_row_inflate_strides == 1) ? inputRow : ((inputRow >= 0) ? (inputRow / m_fastInputRowStride) : 0); - if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) { - inputCoords[0] = coords[0]; // depth - inputCoords[1] = origInputCol; - inputCoords[2] = origInputRow; - inputCoords[3] = coords[4]; // batch - } else { - inputCoords[3] = coords[4]; // depth - inputCoords[2] = origInputCol; - inputCoords[1] = origInputRow; - inputCoords[0] = coords[0]; // batch - } - // If the computed coordinates are outside the original image perimeter, return 0. - if (inputCol < 0 || inputCol >= m_input_cols_eff || inputRow < 0 || inputRow >= m_input_rows_eff || - ((m_col_inflate_strides != 1) && (inputCol != origInputCol * m_col_inflate_strides)) || - ((m_row_inflate_strides != 1) && (inputRow != origInputRow * m_row_inflate_strides))) { - return Scalar(m_paddingValue); - } - if (TensorEvaluator<ArgType, Device>::CoordAccess) { - return m_impl.coeff(inputCoords); - } else { - Index inputIndex; - if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) { - inputIndex = - inputCoords[3] * m_patchInputStride + - inputCoords[2] * m_colInputStride + - inputCoords[1] * m_rowInputStride + - inputCoords[0]; - } else { - inputIndex = - inputCoords[1] * m_patchInputStride + - inputCoords[2] * m_colInputStride + - inputCoords[3] * m_rowInputStride + - inputCoords[4]; - } - return m_impl.coeff(inputIndex); - } - } - protected: EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const { @@ -522,8 +467,8 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device> internal::TensorIntDivisor<Index> m_fastOtherStride; internal::TensorIntDivisor<Index> m_fastPatchStride; internal::TensorIntDivisor<Index> m_fastColStride; - internal::TensorIntDivisor<Index> m_fastInputRowStride; - internal::TensorIntDivisor<Index> m_fastInputColStride; + internal::TensorIntDivisor<Index> m_fastInflateRowStride; + internal::TensorIntDivisor<Index> m_fastInflateColStride; internal::TensorIntDivisor<Index> m_fastInputColsEff; Index m_rowInputStride; |