10 #ifndef EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H 11 #define EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H 23 template<
typename Broadcast,
typename XprType>
24 struct traits<TensorBroadcastingOp<Broadcast, 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 Broadcast,
typename XprType>
37 struct eval<TensorBroadcastingOp<Broadcast, XprType>,
Eigen::Dense>
39 typedef const TensorBroadcastingOp<Broadcast, XprType>& type;
42 template<
typename Broadcast,
typename XprType>
43 struct nested<TensorBroadcastingOp<Broadcast, XprType>, 1, typename eval<TensorBroadcastingOp<Broadcast, XprType> >::type>
45 typedef TensorBroadcastingOp<Broadcast, XprType> type;
48 template <
typename Dims>
49 struct is_input_scalar {
50 static const bool value =
false;
53 struct is_input_scalar<Sizes<> > {
54 static const bool value =
true;
56 #ifndef EIGEN_EMULATE_CXX11_META_H 57 template <
typename std::size_t... Indices>
58 struct is_input_scalar<Sizes<Indices...> > {
59 static const bool value = (Sizes<Indices...>::total_size == 1);
67 template<
typename Broadcast,
typename XprType>
68 class TensorBroadcastingOp :
public TensorBase<TensorBroadcastingOp<Broadcast, XprType>, ReadOnlyAccessors>
71 typedef typename Eigen::internal::traits<TensorBroadcastingOp>::Scalar Scalar;
72 typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
73 typedef typename XprType::CoeffReturnType CoeffReturnType;
74 typedef typename Eigen::internal::nested<TensorBroadcastingOp>::type Nested;
75 typedef typename Eigen::internal::traits<TensorBroadcastingOp>::StorageKind StorageKind;
76 typedef typename Eigen::internal::traits<TensorBroadcastingOp>::Index Index;
78 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBroadcastingOp(
const XprType& expr,
const Broadcast& broadcast)
79 : m_xpr(expr), m_broadcast(broadcast) {}
82 const Broadcast& broadcast()
const {
return m_broadcast; }
85 const typename internal::remove_all<typename XprType::Nested>::type&
86 expression()
const {
return m_xpr; }
89 typename XprType::Nested m_xpr;
90 const Broadcast m_broadcast;
95 template<
typename Broadcast,
typename ArgType,
typename Device>
96 struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
98 typedef TensorBroadcastingOp<Broadcast, ArgType> XprType;
99 typedef typename XprType::Index Index;
100 static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
101 typedef DSizes<Index, NumDims> Dimensions;
102 typedef typename XprType::Scalar Scalar;
103 typedef typename TensorEvaluator<ArgType, Device>::Dimensions InputDimensions;
104 typedef typename XprType::CoeffReturnType CoeffReturnType;
105 typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
106 static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
110 PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
111 Layout = TensorEvaluator<ArgType, Device>::Layout,
115 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(
const XprType& op,
const Device& device)
116 : m_broadcast(op.broadcast()),m_impl(op.expression(), device)
121 EIGEN_STATIC_ASSERT((NumDims > 0), YOU_MADE_A_PROGRAMMING_MISTAKE);
122 const InputDimensions& input_dims = m_impl.dimensions();
123 const Broadcast& broadcast = op.broadcast();
124 for (
int i = 0; i < NumDims; ++i) {
125 eigen_assert(input_dims[i] > 0);
126 m_dimensions[i] = input_dims[i] * broadcast[i];
129 if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
130 m_inputStrides[0] = 1;
131 m_outputStrides[0] = 1;
132 for (
int i = 1; i < NumDims; ++i) {
133 m_inputStrides[i] = m_inputStrides[i-1] * input_dims[i-1];
134 m_outputStrides[i] = m_outputStrides[i-1] * m_dimensions[i-1];
137 m_inputStrides[NumDims-1] = 1;
138 m_outputStrides[NumDims-1] = 1;
139 for (
int i = NumDims-2; i >= 0; --i) {
140 m_inputStrides[i] = m_inputStrides[i+1] * input_dims[i+1];
141 m_outputStrides[i] = m_outputStrides[i+1] * m_dimensions[i+1];
146 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
const Dimensions& dimensions()
const {
return m_dimensions; }
148 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
bool evalSubExprsIfNeeded(Scalar* ) {
149 m_impl.evalSubExprsIfNeeded(NULL);
153 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
void cleanup() {
157 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE CoeffReturnType coeff(Index index)
const 159 if (internal::is_input_scalar<
typename internal::remove_all<InputDimensions>::type>::value) {
160 return m_impl.coeff(0);
163 if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
164 return coeffColMajor(index);
166 return coeffRowMajor(index);
171 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffColMajor(Index index)
const 173 Index inputIndex = 0;
174 for (
int i = NumDims - 1; i > 0; --i) {
175 const Index idx = index / m_outputStrides[i];
176 if (internal::index_statically_eq<Broadcast>(i, 1)) {
177 eigen_assert(idx < m_impl.dimensions()[i]);
178 inputIndex += idx * m_inputStrides[i];
180 if (internal::index_statically_eq<InputDimensions>(i, 1)) {
181 eigen_assert(idx % m_impl.dimensions()[i] == 0);
183 inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
186 index -= idx * m_outputStrides[i];
188 if (internal::index_statically_eq<Broadcast>(0, 1)) {
189 eigen_assert(index < m_impl.dimensions()[0]);
192 if (internal::index_statically_eq<InputDimensions>(0, 1)) {
193 eigen_assert(index % m_impl.dimensions()[0] == 0);
195 inputIndex += (index % m_impl.dimensions()[0]);
198 return m_impl.coeff(inputIndex);
201 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeffRowMajor(Index index)
const 203 Index inputIndex = 0;
204 for (
int i = 0; i < NumDims - 1; ++i) {
205 const Index idx = index / m_outputStrides[i];
206 if (internal::index_statically_eq<Broadcast>(i, 1)) {
207 eigen_assert(idx < m_impl.dimensions()[i]);
208 inputIndex += idx * m_inputStrides[i];
210 if (internal::index_statically_eq<InputDimensions>(i, 1)) {
211 eigen_assert(idx % m_impl.dimensions()[i] == 0);
213 inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
216 index -= idx * m_outputStrides[i];
218 if (internal::index_statically_eq<Broadcast>(NumDims-1, 1)) {
219 eigen_assert(index < m_impl.dimensions()[NumDims-1]);
222 if (internal::index_statically_eq<InputDimensions>(NumDims-1, 1)) {
223 eigen_assert(index % m_impl.dimensions()[NumDims-1] == 0);
225 inputIndex += (index % m_impl.dimensions()[NumDims-1]);
228 return m_impl.coeff(inputIndex);
231 template<
int LoadMode>
232 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketReturnType packet(Index index)
const 234 if (internal::is_input_scalar<
typename internal::remove_all<InputDimensions>::type>::value) {
235 return internal::pset1<PacketReturnType>(m_impl.coeff(0));
238 if (static_cast<int>(Layout) ==
static_cast<int>(ColMajor)) {
239 return packetColMajor<LoadMode>(index);
241 return packetRowMajor<LoadMode>(index);
247 template<
int LoadMode>
248 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetColMajor(Index index)
const 250 EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
251 eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
253 const Index originalIndex = index;
255 Index inputIndex = 0;
256 for (
int i = NumDims - 1; i > 0; --i) {
257 const Index idx = index / m_outputStrides[i];
258 if (internal::index_statically_eq<Broadcast>(i, 1)) {
259 eigen_assert(idx < m_impl.dimensions()[i]);
260 inputIndex += idx * m_inputStrides[i];
262 if (internal::index_statically_eq<InputDimensions>(i, 1)) {
263 eigen_assert(idx % m_impl.dimensions()[i] == 0);
265 inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
268 index -= idx * m_outputStrides[i];
271 if (internal::index_statically_eq<Broadcast>(0, 1)) {
272 eigen_assert(index < m_impl.dimensions()[0]);
273 innermostLoc = index;
275 if (internal::index_statically_eq<InputDimensions>(0, 1)) {
276 eigen_assert(index % m_impl.dimensions()[0] == 0);
279 innermostLoc = index % m_impl.dimensions()[0];
282 inputIndex += innermostLoc;
286 if (innermostLoc + PacketSize <= m_impl.dimensions()[0]) {
287 return m_impl.template packet<Unaligned>(inputIndex);
289 EIGEN_ALIGN_MAX
typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
290 values[0] = m_impl.coeff(inputIndex);
291 for (
int i = 1; i < PacketSize; ++i) {
292 values[i] = coeffColMajor(originalIndex+i);
294 PacketReturnType rslt = internal::pload<PacketReturnType>(values);
299 template<
int LoadMode>
300 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetRowMajor(Index index)
const 302 EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
303 eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
305 const Index originalIndex = index;
307 Index inputIndex = 0;
308 for (
int i = 0; i < NumDims - 1; ++i) {
309 const Index idx = index / m_outputStrides[i];
310 if (internal::index_statically_eq<Broadcast>(i, 1)) {
311 eigen_assert(idx < m_impl.dimensions()[i]);
312 inputIndex += idx * m_inputStrides[i];
314 if (internal::index_statically_eq<InputDimensions>(i, 1)) {
315 eigen_assert(idx % m_impl.dimensions()[i] == 0);
317 inputIndex += (idx % m_impl.dimensions()[i]) * m_inputStrides[i];
320 index -= idx * m_outputStrides[i];
323 if (internal::index_statically_eq<Broadcast>(NumDims-1, 1)) {
324 eigen_assert(index < m_impl.dimensions()[NumDims-1]);
325 innermostLoc = index;
327 if (internal::index_statically_eq<InputDimensions>(NumDims-1, 1)) {
328 eigen_assert(index % m_impl.dimensions()[NumDims-1] == 0);
331 innermostLoc = index % m_impl.dimensions()[NumDims-1];
334 inputIndex += innermostLoc;
338 if (innermostLoc + PacketSize <= m_impl.dimensions()[NumDims-1]) {
339 return m_impl.template packet<Unaligned>(inputIndex);
341 EIGEN_ALIGN_MAX
typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
342 values[0] = m_impl.coeff(inputIndex);
343 for (
int i = 1; i < PacketSize; ++i) {
344 values[i] = coeffRowMajor(originalIndex+i);
346 PacketReturnType rslt = internal::pload<PacketReturnType>(values);
351 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
352 costPerCoeff(
bool vectorized)
const {
353 double compute_cost = TensorOpCost::AddCost<Index>();
355 for (
int i = NumDims - 1; i > 0; --i) {
356 compute_cost += TensorOpCost::DivCost<Index>();
357 if (internal::index_statically_eq<Broadcast>(i, 1)) {
359 TensorOpCost::MulCost<Index>() + TensorOpCost::AddCost<Index>();
361 if (!internal::index_statically_eq<InputDimensions>(i, 1)) {
362 compute_cost += TensorOpCost::MulCost<Index>() +
363 TensorOpCost::ModCost<Index>() +
364 TensorOpCost::AddCost<Index>();
368 TensorOpCost::MulCost<Index>() + TensorOpCost::AddCost<Index>();
371 return m_impl.costPerCoeff(vectorized) +
372 TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
375 EIGEN_DEVICE_FUNC Scalar* data()
const {
return NULL; }
377 const TensorEvaluator<ArgType, Device>& impl()
const {
return m_impl; }
379 Broadcast functor()
const {
return m_broadcast; }
382 const Broadcast m_broadcast;
383 Dimensions m_dimensions;
384 array<Index, NumDims> m_outputStrides;
385 array<Index, NumDims> m_inputStrides;
386 TensorEvaluator<ArgType, Device> m_impl;
392 #endif // EIGEN_CXX11_TENSOR_TENSOR_BROADCASTING_H Namespace containing all symbols from the Eigen library.
Definition: AdolcForward:45