Added support for tensor chips
diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor index b1bd2f6..5a6246a 100644 --- a/unsupported/Eigen/CXX11/Tensor +++ b/unsupported/Eigen/CXX11/Tensor
@@ -47,6 +47,7 @@ #include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionCuda.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.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" #include "unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h" #include "unsupported/Eigen/CXX11/src/Tensor/TensorShuffling.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h index d4b7846..cadeb3b 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -254,6 +254,11 @@ slice(const StartIndices& startIndices, const Sizes& sizes) const { return TensorSlicingOp<const StartIndices, const Sizes, const Derived>(derived(), startIndices, sizes); } + template <std::size_t DimId> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + const TensorChippingOp<DimId, const Derived> + chip(const Index offset) const { + return TensorChippingOp<DimId, const Derived>(derived(), offset); + } template <typename PaddingDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorPaddingOp<const PaddingDimensions, const Derived> pad(const PaddingDimensions& padding) const { @@ -327,7 +332,7 @@ template <typename NewDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorReshapingOp<const NewDimensions, Derived> - reshape(const NewDimensions& newDimensions) { + reshape(const NewDimensions& newDimensions) const { return TensorReshapingOp<const NewDimensions, Derived>(derived(), newDimensions); } template <typename StartIndices, typename Sizes> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE @@ -335,6 +340,11 @@ slice(const StartIndices& startIndices, const Sizes& sizes) const { return TensorSlicingOp<const StartIndices, const Sizes, Derived>(derived(), startIndices, sizes); } + template <std::size_t DimId> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + TensorChippingOp<DimId, Derived> + chip(const Index offset) const { + return TensorChippingOp<DimId, Derived>(derived(), offset); + } template <typename Shuffle> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorShufflingOp<const Shuffle, Derived> shuffle(const Shuffle& shuffle) const {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h new file mode 100644 index 0000000..9ecea91 --- /dev/null +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -0,0 +1,232 @@ +// 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_CHIPPING_H +#define EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H + +namespace Eigen { + +/** \class TensorKChippingReshaping + * \ingroup CXX11_Tensor_Module + * + * \brief A chip is a thin slice, corresponding to a column or a row in a 2-d tensor. + * + * + */ + +namespace internal { +template<std::size_t DimId, typename XprType> +struct traits<TensorChippingOp<DimId, 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<std::size_t DimId, typename XprType> +struct eval<TensorChippingOp<DimId, XprType>, Eigen::Dense> +{ + typedef const TensorChippingOp<DimId, XprType>& type; +}; + +template<std::size_t DimId, typename XprType> +struct nested<TensorChippingOp<DimId, XprType>, 1, typename eval<TensorChippingOp<DimId, XprType> >::type> +{ + typedef TensorChippingOp<DimId, XprType> type; +}; + +} // end namespace internal + + + +template<std::size_t DimId, typename XprType> +class TensorChippingOp : public TensorBase<TensorChippingOp<DimId, XprType> > +{ + public: + typedef typename Eigen::internal::traits<TensorChippingOp>::Scalar Scalar; + typedef typename Eigen::internal::traits<TensorChippingOp>::Packet Packet; + typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + typedef typename Eigen::internal::nested<TensorChippingOp>::type Nested; + typedef typename Eigen::internal::traits<TensorChippingOp>::StorageKind StorageKind; + typedef typename Eigen::internal::traits<TensorChippingOp>::Index Index; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp(const XprType& expr, const Index offset) + : m_xpr(expr), m_offset(offset) {} + + EIGEN_DEVICE_FUNC + const Index offset() const { return m_offset; } + + EIGEN_DEVICE_FUNC + const typename internal::remove_all<typename XprType::Nested>::type& + expression() const { return m_xpr; } + + template<typename OtherDerived> + EIGEN_DEVICE_FUNC + EIGEN_STRONG_INLINE TensorChippingOp& operator = (const OtherDerived& other) + { + typedef TensorAssignOp<TensorChippingOp, const OtherDerived> Assign; + Assign assign(*this, other); + internal::TensorExecutor<const Assign, DefaultDevice, false>::run(assign, DefaultDevice()); + return *this; + } + + protected: + typename XprType::Nested m_xpr; + const Index m_offset; +}; + + +// Eval as rvalue +template<std::size_t DimId, typename ArgType, typename Device> +struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device> +{ + typedef TensorChippingOp<DimId, ArgType> XprType; + static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value; + static const int NumDims = NumInputDims-1; + typedef typename XprType::Index Index; + typedef DSizes<Index, NumDims> Dimensions; + + enum { + // Alignment can't be guaranteed at compile time since it depends on the + // slice offsets. + IsAligned = false, + PacketAccess = false, // not yet implemented + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : m_impl(op.expression(), device), m_device(device) + { + // We could also support the case where NumInputDims==1 if needed. + EIGEN_STATIC_ASSERT(NumInputDims >= 2, YOU_MADE_A_PROGRAMMING_MISTAKE); + EIGEN_STATIC_ASSERT(NumInputDims > DimId, YOU_MADE_A_PROGRAMMING_MISTAKE); + + const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions(); + int j = 0; + for (int i = 0; i < NumInputDims; ++i) { + if (i != DimId) { + m_dimensions[j] = input_dims[i]; + ++j; + } + } + + m_stride = 1; + m_inputStride = 1; + for (int i = 0; i < DimId; ++i) { + m_stride *= input_dims[i]; + m_inputStride *= input_dims[i]; + } + m_inputStride *= input_dims[DimId]; + m_inputOffset = m_stride * op.offset(); + } + + typedef typename XprType::Scalar Scalar; + 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 + { + return m_impl.coeff(srcCoeff(index)); + } + + /* to be done + template<int LoadMode> + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const + { + + }*/ + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { return NULL; } + + protected: + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const + { + Index inputIndex; + if (DimId == 0) { + // m_stride is equal to 1, so let's avoid the integer division. + eigen_assert(m_stride == 1); + inputIndex = index * m_inputStride + m_inputOffset; + } else if (DimId == NumInputDims-1) { + // m_stride is aways greater than index, so let's avoid the integer division. + eigen_assert(m_stride > index); + inputIndex = index + m_inputOffset; + } else { + const Index idx = index / m_stride; + inputIndex = idx * m_inputStride + m_inputOffset; + index -= idx * m_stride; + inputIndex += index; + } + return inputIndex; + } + + Dimensions m_dimensions; + Index m_stride; + Index m_inputOffset; + Index m_inputStride; + TensorEvaluator<ArgType, Device> m_impl; + const Device& m_device; +}; + + +// Eval as lvalue +template<std::size_t DimId, typename ArgType, typename Device> +struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device> + : public TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device> +{ + typedef TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device> Base; + typedef TensorChippingOp<DimId, ArgType> XprType; + static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value; + static const int NumDims = NumInputDims-1; + typedef typename XprType::Index Index; + typedef DSizes<Index, NumDims> Dimensions; + + enum { + IsAligned = false, + PacketAccess = false, + }; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) + : Base(op, device) + { } + + typedef typename XprType::Scalar Scalar; + typedef typename XprType::CoeffReturnType CoeffReturnType; + typedef typename XprType::PacketReturnType PacketReturnType; + + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) + { + return this->m_impl.coeffRef(this->srcCoeff(index)); + } + + /* to be done + template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE + void writePacket(Index index, const PacketReturnType& x) + { + } */ +}; + + +} // end namespace Eigen + +#endif // EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h index bc67586..86ddd1a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -21,11 +21,12 @@ template<typename UnaryOp, typename XprType> class TensorCwiseUnaryOp; template<typename BinaryOp, typename LeftXprType, typename RightXprType> class TensorCwiseBinaryOp; template<typename IfXprType, typename ThenXprType, typename ElseXprType> class TensorSelectOp; -template<typename Broadcast, typename XprType> class TensorBroadcastingOp; template<typename Op, typename Dims, typename XprType> class TensorReductionOp; 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 Broadcast, typename XprType> class TensorBroadcastingOp; +template<std::size_t DimId, typename XprType> class TensorChippingOp; template<typename NewDimensions, typename XprType> class TensorReshapingOp; template<typename StartIndices, typename Sizes, typename XprType> class TensorSlicingOp; template<typename PaddingDimensions, typename XprType> class TensorPaddingOp;
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt index ac2ccaf..48435eb 100644 --- a/unsupported/test/CMakeLists.txt +++ b/unsupported/test/CMakeLists.txt
@@ -115,6 +115,7 @@ ei_add_test(cxx11_tensor_lvalue "-std=c++0x") ei_add_test(cxx11_tensor_map "-std=c++0x") ei_add_test(cxx11_tensor_broadcasting "-std=c++0x") + ei_add_test(cxx11_tensor_chipping "-std=c++0x") 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")
diff --git a/unsupported/test/cxx11_tensor_chipping.cpp b/unsupported/test/cxx11_tensor_chipping.cpp new file mode 100644 index 0000000..8c8a0ce --- /dev/null +++ b/unsupported/test/cxx11_tensor_chipping.cpp
@@ -0,0 +1,244 @@ +// 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_chip() +{ + Tensor<float, 5> tensor(2,3,5,7,11); + tensor.setRandom(); + + Tensor<float, 4> chip1; + chip1 = tensor.chip<0>(1); + VERIFY_IS_EQUAL(chip1.dimension(0), 3); + VERIFY_IS_EQUAL(chip1.dimension(1), 5); + VERIFY_IS_EQUAL(chip1.dimension(2), 7); + VERIFY_IS_EQUAL(chip1.dimension(3), 11); + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip1(i,j,k,l), tensor(1,i,j,k,l)); + } + } + } + } + + Tensor<float, 4> chip2 = tensor.chip<1>(1); + VERIFY_IS_EQUAL(chip2.dimension(0), 2); + VERIFY_IS_EQUAL(chip2.dimension(1), 5); + VERIFY_IS_EQUAL(chip2.dimension(2), 7); + VERIFY_IS_EQUAL(chip2.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip2(i,j,k,l), tensor(i,1,j,k,l)); + } + } + } + } + + Tensor<float, 4> chip3 = tensor.chip<2>(2); + VERIFY_IS_EQUAL(chip3.dimension(0), 2); + VERIFY_IS_EQUAL(chip3.dimension(1), 3); + VERIFY_IS_EQUAL(chip3.dimension(2), 7); + VERIFY_IS_EQUAL(chip3.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + VERIFY_IS_EQUAL(chip3(i,j,k,l), tensor(i,j,2,k,l)); + } + } + } + } + + Tensor<float, 4> chip4(tensor.chip<3>(5)); + VERIFY_IS_EQUAL(chip4.dimension(0), 2); + VERIFY_IS_EQUAL(chip4.dimension(1), 3); + VERIFY_IS_EQUAL(chip4.dimension(2), 5); + VERIFY_IS_EQUAL(chip4.dimension(3), 11); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(chip4(i,j,k,l), tensor(i,j,k,5,l)); + } + } + } + } + + Tensor<float, 4> chip5(tensor.chip<4>(7)); + VERIFY_IS_EQUAL(chip5.dimension(0), 2); + VERIFY_IS_EQUAL(chip5.dimension(1), 3); + VERIFY_IS_EQUAL(chip5.dimension(2), 5); + VERIFY_IS_EQUAL(chip5.dimension(3), 7); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + VERIFY_IS_EQUAL(chip5(i,j,k,l), tensor(i,j,k,l,7)); + } + } + } + } +} + + +static void test_chip_in_expr() { + Tensor<float, 5> input1(2,3,5,7,11); + input1.setRandom(); + Tensor<float, 4> input2(3,5,7,11); + input2.setRandom(); + + Tensor<float, 4> result = input1.chip<0>(0) + input2; + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 5; ++j) { + for (int k = 0; k < 7; ++k) { + for (int l = 0; l < 11; ++l) { + float expected = input1(0,i,j,k,l) + input2(i,j,k,l); + VERIFY_IS_EQUAL(result(i,j,k,l), expected); + } + } + } + } + + Tensor<float, 3> input3(3,7,11); + input3.setRandom(); + Tensor<float, 3> result2 = input1.chip<0>(0).chip<1>(2) + input3; + for (int i = 0; i < 3; ++i) { + for (int j = 0; j < 7; ++j) { + for (int k = 0; k < 11; ++k) { + float expected = input1(0,i,2,j,k) + input3(i,j,k); + VERIFY_IS_EQUAL(result2(i,j,k), expected); + } + } + } +} + + +static void test_chip_as_lvalue() +{ + Tensor<float, 5> input1(2,3,5,7,11); + input1.setRandom(); + + Tensor<float, 4> input2(3,5,7,11); + input2.setRandom(); + Tensor<float, 5> tensor = input1; + tensor.chip<0>(1) = input2; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (i != 1) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input2(j,k,l,m)); + } + } + } + } + } + } + + Tensor<float, 4> input3(2,5,7,11); + input3.setRandom(); + tensor = input1; + tensor.chip<1>(1) = input3; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (j != 1) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input3(i,k,l,m)); + } + } + } + } + } + } + + Tensor<float, 4> input4(2,3,7,11); + input4.setRandom(); + tensor = input1; + tensor.chip<2>(3) = input4; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (k != 3) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input4(i,j,l,m)); + } + } + } + } + } + } + + Tensor<float, 4> input5(2,3,5,11); + input5.setRandom(); + tensor = input1; + tensor.chip<3>(4) = input5; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (l != 4) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input5(i,j,k,m)); + } + } + } + } + } + } + + Tensor<float, 4> input6(2,3,5,7); + input6.setRandom(); + tensor = input1; + tensor.chip<4>(5) = input6; + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + for (int k = 0; k < 5; ++k) { + for (int l = 0; l < 7; ++l) { + for (int m = 0; m < 11; ++m) { + if (m != 5) { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input1(i,j,k,l,m)); + } else { + VERIFY_IS_EQUAL(tensor(i,j,k,l,m), input6(i,j,k,l)); + } + } + } + } + } + } +} + + +void test_cxx11_tensor_chipping() +{ + CALL_SUBTEST(test_simple_chip()); + CALL_SUBTEST(test_chip_in_expr()); + CALL_SUBTEST(test_chip_as_lvalue()); +}