10 #ifndef EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H 11 #define EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H 23 template<
typename Str
ides,
typename XprType>
24 struct traits<TensorStridingOp<Strides, XprType> > :
public traits<XprType>
26 typedef typename XprType::Scalar Scalar;
27 typedef traits<XprType> XprTraits;
28 typedef typename XprTraits::StorageKind StorageKind;
29 typedef typename XprTraits::Index Index;
30 typedef typename XprType::Nested Nested;
31 typedef typename remove_reference<Nested>::type _Nested;
32 static const int NumDimensions = XprTraits::NumDimensions;
33 static const int Layout = XprTraits::Layout;
36 template<
typename Str
ides,
typename XprType>
37 struct eval<TensorStridingOp<Strides, XprType>,
Eigen::Dense>
39 typedef const TensorStridingOp<Strides, XprType>& type;
42 template<
typename Str
ides,
typename XprType>
43 struct nested<TensorStridingOp<Strides, XprType>, 1, typename eval<TensorStridingOp<Strides, XprType> >::type>
45 typedef TensorStridingOp<Strides, XprType> type;
52 template<
typename Str
ides,
typename XprType>
53 class TensorStridingOp :
public TensorBase<TensorStridingOp<Strides, XprType> >
56 typedef typename Eigen::internal::traits<TensorStridingOp>::Scalar Scalar;
57 typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
58 typedef typename XprType::CoeffReturnType CoeffReturnType;
59 typedef typename Eigen::internal::nested<TensorStridingOp>::type Nested;
60 typedef typename Eigen::internal::traits<TensorStridingOp>::StorageKind StorageKind;
61 typedef typename Eigen::internal::traits<TensorStridingOp>::Index Index;
63 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorStridingOp(
const XprType& expr,
const Strides& dims)
64 : m_xpr(expr), m_dims(dims) {}
67 const Strides& strides()
const {
return m_dims; }
70 const typename internal::remove_all<typename XprType::Nested>::type&
71 expression()
const {
return m_xpr; }
74 EIGEN_STRONG_INLINE TensorStridingOp& operator = (
const TensorStridingOp& other)
76 typedef TensorAssignOp<TensorStridingOp, const TensorStridingOp> Assign;
77 Assign assign(*
this, other);
78 internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
82 template<
typename OtherDerived>
84 EIGEN_STRONG_INLINE TensorStridingOp& operator = (
const OtherDerived& other)
86 typedef TensorAssignOp<TensorStridingOp, const OtherDerived> Assign;
87 Assign assign(*
this, other);
88 internal::TensorExecutor<const Assign, DefaultDevice>::run(assign, DefaultDevice());
93 typename XprType::Nested m_xpr;
99 template<
typename Str
ides,
typename ArgType,
typename Device>
100 struct TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
102 typedef TensorStridingOp<Strides, ArgType> XprType;
103 typedef typename XprType::Index Index;
104 static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
105 typedef DSizes<Index, NumDims> Dimensions;
106 typedef typename XprType::Scalar Scalar;
107 typedef typename XprType::CoeffReturnType CoeffReturnType;
108 typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
109 static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
113 PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
114 Layout = TensorEvaluator<ArgType, Device>::Layout,
119 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(
const XprType& op,
const Device& device)
120 : m_impl(op.expression(), device)
122 m_dimensions = m_impl.dimensions();
123 for (
int i = 0; i < NumDims; ++i) {
124 m_dimensions[i] = ceilf(static_cast<float>(m_dimensions[i]) / op.strides()[i]);
127 const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
128 if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
129 m_outputStrides[0] = 1;
130 m_inputStrides[0] = 1;
131 for (
int i = 1; i < NumDims; ++i) {
132 m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
133 m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
134 m_inputStrides[i-1] *= op.strides()[i-1];
136 m_inputStrides[NumDims-1] *= op.strides()[NumDims-1];
138 m_outputStrides[NumDims-1] = 1;
139 m_inputStrides[NumDims-1] = 1;
140 for (
int i = NumDims - 2; i >= 0; --i) {
141 m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
142 m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
143 m_inputStrides[i+1] *= op.strides()[i+1];
145 m_inputStrides[0] *= op.strides()[0];
149 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
const Dimensions& dimensions()
const {
return m_dimensions; }
151 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
bool evalSubExprsIfNeeded(Scalar* ) {
152 m_impl.evalSubExprsIfNeeded(NULL);
155 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
void cleanup() {
159 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index)
const 161 return m_impl.coeff(srcCoeff(index));
164 template<
int LoadMode>
165 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index)
const 167 EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
168 eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
170 Index inputIndices[] = {0, 0};
171 Index indices[] = {index, index + PacketSize - 1};
172 if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
173 for (
int i = NumDims - 1; i > 0; --i) {
174 const Index idx0 = indices[0] / m_outputStrides[i];
175 const Index idx1 = indices[1] / m_outputStrides[i];
176 inputIndices[0] += idx0 * m_inputStrides[i];
177 inputIndices[1] += idx1 * m_inputStrides[i];
178 indices[0] -= idx0 * m_outputStrides[i];
179 indices[1] -= idx1 * m_outputStrides[i];
181 inputIndices[0] += indices[0] * m_inputStrides[0];
182 inputIndices[1] += indices[1] * m_inputStrides[0];
184 for (
int i = 0; i < NumDims - 1; ++i) {
185 const Index idx0 = indices[0] / m_outputStrides[i];
186 const Index idx1 = indices[1] / m_outputStrides[i];
187 inputIndices[0] += idx0 * m_inputStrides[i];
188 inputIndices[1] += idx1 * m_inputStrides[i];
189 indices[0] -= idx0 * m_outputStrides[i];
190 indices[1] -= idx1 * m_outputStrides[i];
192 inputIndices[0] += indices[0] * m_inputStrides[NumDims-1];
193 inputIndices[1] += indices[1] * m_inputStrides[NumDims-1];
195 if (inputIndices[1] - inputIndices[0] == PacketSize - 1) {
196 PacketReturnType rslt = m_impl.template packet<Unaligned>(inputIndices[0]);
200 EIGEN_ALIGN_MAX
typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
201 values[0] = m_impl.coeff(inputIndices[0]);
202 values[PacketSize-1] = m_impl.coeff(inputIndices[1]);
203 for (
int i = 1; i < PacketSize-1; ++i) {
204 values[i] = coeff(index+i);
206 PacketReturnType rslt = internal::pload<PacketReturnType>(values);
211 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(
bool vectorized)
const {
212 double compute_cost = (NumDims - 1) * (TensorOpCost::AddCost<Index>() +
213 TensorOpCost::MulCost<Index>() +
214 TensorOpCost::DivCost<Index>()) +
215 TensorOpCost::MulCost<Index>();
219 const int innerDim = (
static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? 0 : (NumDims - 1);
220 return m_impl.costPerCoeff(vectorized && m_inputStrides[innerDim] == 1) +
222 TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
225 EIGEN_DEVICE_FUNC Scalar* data()
const {
return NULL; }
228 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index)
const 230 Index inputIndex = 0;
231 if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
232 for (
int i = NumDims - 1; i > 0; --i) {
233 const Index idx = index / m_outputStrides[i];
234 inputIndex += idx * m_inputStrides[i];
235 index -= idx * m_outputStrides[i];
237 inputIndex += index * m_inputStrides[0];
239 for (
int i = 0; i < NumDims - 1; ++i) {
240 const Index idx = index / m_outputStrides[i];
241 inputIndex += idx * m_inputStrides[i];
242 index -= idx * m_outputStrides[i];
244 inputIndex += index * m_inputStrides[NumDims-1];
249 Dimensions m_dimensions;
250 array<Index, NumDims> m_outputStrides;
251 array<Index, NumDims> m_inputStrides;
252 TensorEvaluator<ArgType, Device> m_impl;
257 template<
typename Str
ides,
typename ArgType,
typename Device>
258 struct TensorEvaluator<TensorStridingOp<Strides, ArgType>, Device>
259 :
public TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
261 typedef TensorStridingOp<Strides, ArgType> XprType;
262 typedef TensorEvaluator<const XprType, Device> Base;
264 static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
269 PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
270 Layout = TensorEvaluator<ArgType, Device>::Layout,
275 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(
const XprType& op,
const Device& device)
276 : Base(op, device) { }
278 typedef typename XprType::Index Index;
279 typedef typename XprType::Scalar Scalar;
280 typedef typename XprType::CoeffReturnType CoeffReturnType;
281 typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
282 static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
284 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
286 return this->m_impl.coeffRef(this->srcCoeff(index));
289 template <
int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
290 void writePacket(Index index,
const PacketReturnType& x)
292 EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
293 eigen_assert(index+PacketSize-1 < this->dimensions().TotalSize());
295 Index inputIndices[] = {0, 0};
296 Index indices[] = {index, index + PacketSize - 1};
297 if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
298 for (
int i = NumDims - 1; i > 0; --i) {
299 const Index idx0 = indices[0] / this->m_outputStrides[i];
300 const Index idx1 = indices[1] / this->m_outputStrides[i];
301 inputIndices[0] += idx0 * this->m_inputStrides[i];
302 inputIndices[1] += idx1 * this->m_inputStrides[i];
303 indices[0] -= idx0 * this->m_outputStrides[i];
304 indices[1] -= idx1 * this->m_outputStrides[i];
306 inputIndices[0] += indices[0] * this->m_inputStrides[0];
307 inputIndices[1] += indices[1] * this->m_inputStrides[0];
309 for (
int i = 0; i < NumDims - 1; ++i) {
310 const Index idx0 = indices[0] / this->m_outputStrides[i];
311 const Index idx1 = indices[1] / this->m_outputStrides[i];
312 inputIndices[0] += idx0 * this->m_inputStrides[i];
313 inputIndices[1] += idx1 * this->m_inputStrides[i];
314 indices[0] -= idx0 * this->m_outputStrides[i];
315 indices[1] -= idx1 * this->m_outputStrides[i];
317 inputIndices[0] += indices[0] * this->m_inputStrides[NumDims-1];
318 inputIndices[1] += indices[1] * this->m_inputStrides[NumDims-1];
320 if (inputIndices[1] - inputIndices[0] == PacketSize - 1) {
321 this->m_impl.template writePacket<Unaligned>(inputIndices[0], x);
324 EIGEN_ALIGN_MAX Scalar values[PacketSize];
325 internal::pstore<Scalar, PacketReturnType>(values, x);
326 this->m_impl.coeffRef(inputIndices[0]) = values[0];
327 this->m_impl.coeffRef(inputIndices[1]) = values[PacketSize-1];
328 for (
int i = 1; i < PacketSize-1; ++i) {
329 this->coeffRef(index+i) = values[i];
338 #endif // EIGEN_CXX11_TENSOR_TENSOR_STRIDING_H Namespace containing all symbols from the Eigen library.
Definition: AdolcForward:45