Added support for broadcasting

unsupported/Eigen/CXX11/Tensor

@@ -42,6 +42,7 @@
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorExpr.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h"
+#include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.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"

unsupported/Eigen/CXX11/src/Tensor/TensorBase.h

@@ -204,6 +204,12 @@ class TensorBase<Derived, ReadOnlyAccessors>
       return TensorSelectOp<const Derived, const ThenDerived, const ElseDerived>(derived(), thenTensor.derived(), elseTensor.derived());
     }
 
+    template <typename Broadcast> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorBroadcastingOp<const Broadcast, const Derived>
+    broadcast(const Broadcast& broadcast) const {
+      return TensorBroadcastingOp<const Broadcast, const Derived>(derived(), broadcast);
+    }
+
     // Morphing operators.
     template <typename NewDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorReshapingOp<const NewDimensions, const Derived>

unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h

@@ -0,0 +1,186 @@
+// 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_BROADCASTING_H
+#define EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H
+
+namespace Eigen {
+
+/** \class TensorBroadcasting
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor broadcasting class.
+  *
+  *
+  */
+namespace internal {
+template<typename Broadcast, typename XprType>
+struct traits<TensorBroadcastingOp<Broadcast, 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 Broadcast, typename XprType>
+struct eval<TensorBroadcastingOp<Broadcast, XprType>, Eigen::Dense>
+{
+  typedef const TensorBroadcastingOp<Broadcast, XprType>& type;
+};
+
+template<typename Broadcast, typename XprType>
+struct nested<TensorBroadcastingOp<Broadcast, XprType>, 1, typename eval<TensorBroadcastingOp<Broadcast, XprType> >::type>
+{
+  typedef TensorBroadcastingOp<Broadcast, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename Broadcast, typename XprType>
+class TensorBroadcastingOp : public TensorBase<TensorBroadcastingOp<Broadcast, XprType>, WriteAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorBroadcastingOp>::Scalar Scalar;
+  typedef typename Eigen::internal::traits<TensorBroadcastingOp>::Packet Packet;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+  typedef typename Eigen::internal::nested<TensorBroadcastingOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorBroadcastingOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorBroadcastingOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBroadcastingOp(const XprType& expr, const Broadcast& broadcast)
+      : m_xpr(expr), m_broadcast(broadcast) {}
+
+    EIGEN_DEVICE_FUNC
+    const Broadcast& broadcast() const { return m_broadcast; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const Broadcast m_broadcast;
+};
+
+
+// Eval as rvalue
+template<typename Broadcast, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
+{
+  typedef TensorBroadcastingOp<Broadcast, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
+  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)
+  {
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    const Broadcast& broadcast = op.broadcast();
+    for (int i = 0; i < NumDims; ++i) {
+      eigen_assert(input_dims[i] > 0);
+      m_dimensions[i] = input_dims[i] * broadcast[i];
+    }
+
+    m_inputStrides[0] = 1;
+    m_outputStrides[0] = 1;
+    for (int i = 1; i < NumDims; ++i) {
+      m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
+      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();
+  }
+
+  // TODO: attempt to speed this up. The integer divisions and modulo are slow
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    Index inputIndex = 0;
+    for (int i = NumDims - 1; i > 0; --i) {
+      const Index idx = index / m_outputStrides[i];
+      inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
+      index -= idx * m_outputStrides[i];
+    }
+    inputIndex += (index % m_impl.dimensions()[0]);
+    return m_impl.coeff(inputIndex);
+  }
+
+  // Ignore the LoadMode and always use unaligned loads since we can't guarantee
+  // the alignment at compile time.
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    static 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());
+
+    const Index originalIndex = index;
+
+    Index inputIndex = 0;
+    for (int i = NumDims - 1; i > 0; --i) {
+      const Index idx = index / m_outputStrides[i];
+      inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
+      index -= idx * m_outputStrides[i];
+    }
+    const Index innermostLoc = index % m_impl.dimensions()[0];
+    inputIndex += innermostLoc;
+
+    // Todo: this could be extended to the second dimension if we're not
+    // broadcasting alongside the first dimension, and so on.
+    if (innermostLoc + packetSize <= m_impl.dimensions()[0]) {
+      return m_impl.template packet<Unaligned>(inputIndex);
+    } else {
+      EIGEN_ALIGN_DEFAULT CoeffReturnType values[packetSize];
+      values[0] = m_impl.coeff(inputIndex);
+      for (int i = 1; i < packetSize; ++i) {
+        values[i] = coeff(originalIndex+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> m_inputStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H

unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h

@@ -22,6 +22,7 @@ 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 XprType> class TensorReductionOp;
+template<typename Broadcast, typename XprType> class TensorBroadcastingOp;
 template<typename Dimensions, typename LeftXprType, typename RightXprType> class TensorContractionOp;
 template<typename Dimensions, typename InputXprType, typename KernelXprType> class TensorConvolutionOp;
 template<typename NewDimensions, typename XprType> class TensorReshapingOp;

unsupported/test/CMakeLists.txt

@@ -109,6 +109,7 @@ if(EIGEN_TEST_CXX11)
   ei_add_test(cxx11_tensor_intdiv "-std=c++0x")
   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_morphing "-std=c++0x")
   ei_add_test(cxx11_tensor_padding "-std=c++0x")
 #  ei_add_test(cxx11_tensor_shuffling "-std=c++0x")

unsupported/test/cxx11_tensor_broadcasting.cpp

@@ -0,0 +1,114 @@
+// 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_broadcasting()
+{
+  Tensor<float, 4> tensor(2,3,5,7);
+  tensor.setRandom();
+  array<ptrdiff_t, 4> broadcasts;
+  broadcasts[0] = 1;
+  broadcasts[1] = 1;
+  broadcasts[2] = 1;
+  broadcasts[3] = 1;
+
+  Tensor<float, 4> no_broadcast;
+  no_broadcast = tensor.broadcast(broadcasts);
+
+  VERIFY_IS_EQUAL(no_broadcast.dimension(0), 2);
+  VERIFY_IS_EQUAL(no_broadcast.dimension(1), 3);
+  VERIFY_IS_EQUAL(no_broadcast.dimension(2), 5);
+  VERIFY_IS_EQUAL(no_broadcast.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(tensor(i,j,k,l), no_broadcast(i,j,k,l));
+        }
+      }
+    }
+  }
+
+  broadcasts[0] = 2;
+  broadcasts[1] = 3;
+  broadcasts[2] = 1;
+  broadcasts[3] = 4;
+  Tensor<float, 4> broadcast;
+  broadcast = tensor.broadcast(broadcasts);
+
+  VERIFY_IS_EQUAL(broadcast.dimension(0), 4);
+  VERIFY_IS_EQUAL(broadcast.dimension(1), 9);
+  VERIFY_IS_EQUAL(broadcast.dimension(2), 5);
+  VERIFY_IS_EQUAL(broadcast.dimension(3), 28);
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 9; ++j) {
+      for (int k = 0; k < 5; ++k) {
+        for (int l = 0; l < 28; ++l) {
+          VERIFY_IS_EQUAL(tensor(i%2,j%3,k%5,l%7), broadcast(i,j,k,l));
+        }
+      }
+    }
+  }
+}
+
+
+static void test_vectorized_broadcasting()
+{
+  Tensor<float, 3> tensor(8,3,5);
+  tensor.setRandom();
+  array<ptrdiff_t, 3> broadcasts;
+  broadcasts[0] = 2;
+  broadcasts[1] = 3;
+  broadcasts[2] = 4;
+
+  Tensor<float, 3> broadcast;
+  broadcast = tensor.broadcast(broadcasts);
+
+  VERIFY_IS_EQUAL(broadcast.dimension(0), 16);
+  VERIFY_IS_EQUAL(broadcast.dimension(1), 9);
+  VERIFY_IS_EQUAL(broadcast.dimension(2), 20);
+
+  for (int i = 0; i < 16; ++i) {
+    for (int j = 0; j < 9; ++j) {
+      for (int k = 0; k < 20; ++k) {
+        VERIFY_IS_EQUAL(tensor(i%8,j%3,k%5), broadcast(i,j,k));
+      }
+    }
+  }
+
+  tensor.resize(11,3,5);
+  tensor.setRandom();
+  broadcast = tensor.broadcast(broadcasts);
+
+  VERIFY_IS_EQUAL(broadcast.dimension(0), 22);
+  VERIFY_IS_EQUAL(broadcast.dimension(1), 9);
+  VERIFY_IS_EQUAL(broadcast.dimension(2), 20);
+
+  for (int i = 0; i < 22; ++i) {
+    for (int j = 0; j < 9; ++j) {
+      for (int k = 0; k < 20; ++k) {
+        VERIFY_IS_EQUAL(tensor(i%11,j%3,k%5), broadcast(i,j,k));
+      }
+    }
+  }
+}
+
+
+void test_cxx11_tensor_broadcasting()
+{
+   CALL_SUBTEST(test_simple_broadcasting());
+   CALL_SUBTEST(test_vectorized_broadcasting());
+}