Added a new operation to enable more powerful tensorindexing.

1 files changed, 279 insertions, 0 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
index 75d54759b..359fd243a 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@@ -603,6 +603,285 @@ struct TensorEvaluator<TensorSlicingOp<StartIndices, Sizes, ArgType>, Device>
 };
 
 
+
+namespace internal {
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
+struct traits<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = array_size<StartIndices>::value;
+  static const int Layout = XprTraits::Layout;
+};
+
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
+struct eval<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>, Eigen::Dense>
+{
+  typedef const TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>& type;
+};
+
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
+struct nested<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType>, 1, typename eval<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> >::type>
+{
+  typedef TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> type;
+};
+
+}  // end namespace internal
+
+
+template<typename StartIndices, typename StopIndices, typename Strides, typename XprType>
+class TensorStridingSlicingOp : public TensorBase<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, XprType> >
+{
+  public:
+  typedef typename internal::traits<TensorStridingSlicingOp>::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename internal::nested<TensorStridingSlicingOp>::type Nested;
+  typedef typename internal::traits<TensorStridingSlicingOp>::StorageKind StorageKind;
+  typedef typename internal::traits<TensorStridingSlicingOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorStridingSlicingOp(
+    const XprType& expr, const StartIndices& startIndices,
+    const StopIndices& stopIndices, const Strides& strides)
+      : m_xpr(expr), m_startIndices(startIndices), m_stopIndices(stopIndices),
+        m_strides(strides) {}
+
+    EIGEN_DEVICE_FUNC
+    const StartIndices& startIndices() const { return m_startIndices; }
+    EIGEN_DEVICE_FUNC
+    const StartIndices& stopIndices() const { return m_stopIndices; }
+    EIGEN_DEVICE_FUNC
+    const StartIndices& strides() const { return m_strides; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorStridingSlicingOp& operator = (const TensorStridingSlicingOp& other)
+    {
+      typedef TensorAssignOp<TensorStridingSlicingOp, const TensorStridingSlicingOp> Assign;
+      Assign assign(*this, other);
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(
+          assign, DefaultDevice());
+      return *this;
+    }
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TensorStridingSlicingOp& operator = (const OtherDerived& other)
+    {
+      typedef TensorAssignOp<TensorStridingSlicingOp, const OtherDerived> Assign;
+      Assign assign(*this, other);
+      internal::TensorExecutor<const Assign, DefaultDevice>::run(
+          assign, DefaultDevice());
+      return *this;
+    }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const StartIndices m_startIndices;
+    const StopIndices m_stopIndices;
+    const Strides m_strides;
+};
+
+// Eval as rvalue
+template<typename StartIndices, typename StopIndices, typename Strides, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType>, Device>
+{
+  typedef TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType> XprType;
+  static const int NumDims = internal::array_size<Strides>::value;
+
+  enum {
+    // Alignment can't be guaranteed at compile time since it depends on the
+    // slice offsets and sizes.
+    IsAligned = false,
+    PacketAccess = false,
+    BlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    RawAccess = false
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_device(device), m_strides(op.strides())
+  {
+    auto clamp = [](Index value, Index min, Index max){
+      return numext::maxi(min,numext::mini(max,value));
+    };
+    // Handle degenerate intervals by gracefully clamping and allowing m_dimensions to be zero
+    DSizes<Index,NumDims> startIndicesClamped, stopIndicesClamped;
+    for (int i = 0; i < internal::array_size<Dimensions>::value; ++i) {
+      eigen_assert(m_strides[i] != 0 && "0 stride is invalid");
+      if(m_strides[i]>0){
+        startIndicesClamped[i] = clamp(op.startIndices()[i], 0, m_impl.dimensions()[i]);
+        stopIndicesClamped[i] = clamp(op.stopIndices()[i], 0, m_impl.dimensions()[i]);
+      }else{
+        /* implies m_strides[i]<0 by assert */
+        startIndicesClamped[i] = clamp(op.startIndices()[i], -1, m_impl.dimensions()[i] - 1);
+        stopIndicesClamped[i] = clamp(op.stopIndices()[i], -1, m_impl.dimensions()[i] - 1);
+      }
+      m_startIndices[i] = startIndicesClamped[i];
+    }
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+
+    // check for degenerate intervals and compute output tensor shape
+    bool degenerate = false;;
+    for(int i = 0; i < NumDims; i++){
+      Index interval = stopIndicesClamped[i] - startIndicesClamped[i];
+      if(interval == 0 || ((interval<0) != (m_strides[i]<0))){
+        m_dimensions[i] = 0;
+        degenerate = true;
+      }else{
+        m_dimensions[i] = interval / m_strides[i]
+                          + (interval % m_strides[i] != 0 ? 1 : 0);
+        eigen_assert(m_dimensions[i] >= 0);
+      }
+    }
+    Strides output_dims = m_dimensions;
+
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      m_inputStrides[0] = m_strides[0];
+      m_offsets[0] = startIndicesClamped[0];
+      Index previousDimProduct = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        previousDimProduct *= input_dims[i-1];
+        m_inputStrides[i] = previousDimProduct * m_strides[i];
+        m_offsets[i] = startIndicesClamped[i] * previousDimProduct;
+      }
+
+      // Don't initialize m_fastOutputStrides[0] since it won't ever be accessed.
+      m_outputStrides[0] = 1;
+      for (int i = 1; i < NumDims; ++i) {
+        m_outputStrides[i] = m_outputStrides[i-1] * output_dims[i-1];
+        // NOTE: if tensor is degenerate, we send 1 to prevent TensorIntDivisor constructor crash
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(degenerate ? 1 : m_outputStrides[i]);
+      }
+    } else {
+      m_inputStrides[NumDims-1] = m_strides[NumDims-1];
+      m_offsets[NumDims-1] = startIndicesClamped[NumDims-1];
+      Index previousDimProduct = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        previousDimProduct *= input_dims[i+1];
+        m_inputStrides[i] = previousDimProduct * m_strides[i];
+        m_offsets[i] = startIndicesClamped[i] * previousDimProduct;
+      }
+
+      m_outputStrides[NumDims-1] = 1;
+      for (int i = NumDims - 2; i >= 0; --i) {
+        m_outputStrides[i] = m_outputStrides[i+1] * output_dims[i+1];
+        // NOTE: if tensor is degenerate, we send 1 to prevent TensorIntDivisor constructor crash
+        m_fastOutputStrides[i] = internal::TensorIntDivisor<Index>(degenerate ? 1 : m_outputStrides[i]);
+      }
+    }
+    m_block_total_size_max = numext::maxi(static_cast<std::size_t>(1),
+                                          device.lastLevelCacheSize() /
+                                          sizeof(Scalar));
+  }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::remove_const<Scalar>::type ScalarNonConst;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef Strides Dimensions;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType*) {
+    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));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, NumDims);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const {
+    return nullptr;
+  }
+
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const
+  {
+    Index inputIndex = 0;
+    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      for (int i = NumDims - 1; i >= 0; --i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        inputIndex += idx * m_inputStrides[i] + m_offsets[i];
+        index -= idx * m_outputStrides[i];
+      }
+    } else {
+      for (int i = 0; i < NumDims; ++i) {
+        const Index idx = index / m_fastOutputStrides[i];
+        inputIndex += idx * m_inputStrides[i] + m_offsets[i];
+        index -= idx * m_outputStrides[i];
+      }
+    }
+    return inputIndex;
+  }
+
+  array<Index, NumDims> m_outputStrides;
+  array<internal::TensorIntDivisor<Index>, NumDims> m_fastOutputStrides;
+  array<Index, NumDims> m_inputStrides;
+  TensorEvaluator<ArgType, Device> m_impl;
+  const Device& m_device;
+  DSizes<Index, NumDims> m_startIndices; // clamped startIndices
+  DSizes<Index, NumDims> m_dimensions;
+  DSizes<Index, NumDims> m_offsets; // offset in a flattened shape
+  const Strides m_strides;
+  std::size_t m_block_total_size_max;
+};
+
+// Eval as lvalue
+template<typename StartIndices, typename StopIndices, typename Strides, typename ArgType, typename Device>
+struct TensorEvaluator<TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType>, Device>
+  : public TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType>, Device>
+{
+  typedef TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType>, Device> Base;
+  typedef TensorStridingSlicingOp<StartIndices, StopIndices, Strides, ArgType> XprType;
+  static const int NumDims = internal::array_size<Strides>::value;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = false,
+    BlockAccess = false,
+    Layout = TensorEvaluator<ArgType, Device>::Layout,
+    CoordAccess = TensorEvaluator<ArgType, Device>::CoordAccess,
+    RawAccess = false
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+    : Base(op, device)
+    { }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::remove_const<Scalar>::type ScalarNonConst;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  typedef Strides Dimensions;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
+  {
+    return this->m_impl.coeffRef(this->srcCoeff(index));
+  }
+};
+
+
 } // end namespace Eigen
 
 #endif // EIGEN_CXX11_TENSOR_TENSOR_MORPHING_H