10#ifndef EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H
11#define EIGEN_CXX11_TENSOR_TENSOR_CHIPPING_H
24template<DenseIndex DimId,
typename XprType>
25struct traits<TensorChippingOp<DimId, XprType> > :
public traits<XprType>
27 typedef typename XprType::Scalar Scalar;
28 typedef traits<XprType> XprTraits;
29 typedef typename XprTraits::StorageKind StorageKind;
30 typedef typename XprTraits::Index
Index;
31 typedef typename XprType::Nested Nested;
32 typedef typename remove_reference<Nested>::type _Nested;
33 static const int NumDimensions = XprTraits::NumDimensions - 1;
34 static const int Layout = XprTraits::Layout;
35 typedef typename XprTraits::PointerType PointerType;
38template<DenseIndex DimId,
typename XprType>
39struct eval<TensorChippingOp<DimId, XprType>,
Eigen::Dense>
41 typedef const TensorChippingOp<DimId, XprType> EIGEN_DEVICE_REF type;
44template<DenseIndex DimId,
typename XprType>
45struct nested<TensorChippingOp<DimId, XprType>, 1, typename eval<TensorChippingOp<DimId, XprType> >::type>
47 typedef TensorChippingOp<DimId, XprType> type;
50template <DenseIndex DimId>
53 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DimensionId(DenseIndex dim) {
54 EIGEN_UNUSED_VARIABLE(dim);
55 eigen_assert(dim == DimId);
57 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim()
const {
64 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DimensionId(DenseIndex dim) : actual_dim(dim) {
65 eigen_assert(dim >= 0);
67 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim()
const {
71 const DenseIndex actual_dim;
79template<DenseIndex DimId,
typename XprType>
80class TensorChippingOp :
public TensorBase<TensorChippingOp<DimId, XprType> >
83 typedef TensorBase<TensorChippingOp<DimId, XprType> > Base;
84 typedef typename Eigen::internal::traits<TensorChippingOp>::Scalar Scalar;
86 typedef typename XprType::CoeffReturnType CoeffReturnType;
87 typedef typename Eigen::internal::nested<TensorChippingOp>::type Nested;
88 typedef typename Eigen::internal::traits<TensorChippingOp>::StorageKind StorageKind;
89 typedef typename Eigen::internal::traits<TensorChippingOp>::Index
Index;
91 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp(
const XprType& expr,
const Index offset,
const Index dim)
92 : m_xpr(expr), m_offset(offset), m_dim(dim) {
96 const Index offset()
const {
return m_offset; }
98 const Index dim()
const {
return m_dim.actualDim(); }
101 const typename internal::remove_all<typename XprType::Nested>::type&
102 expression()
const {
return m_xpr; }
104 EIGEN_TENSOR_INHERIT_ASSIGNMENT_OPERATORS(TensorChippingOp)
107 typename XprType::Nested m_xpr;
108 const Index m_offset;
109 const internal::DimensionId<DimId> m_dim;
114template<DenseIndex DimId,
typename ArgType,
typename Device>
115struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
117 typedef TensorChippingOp<DimId, ArgType> XprType;
118 static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
119 static const int NumDims = NumInputDims-1;
120 typedef typename XprType::Index
Index;
121 typedef DSizes<Index, NumDims> Dimensions;
122 typedef typename XprType::Scalar Scalar;
123 typedef typename XprType::CoeffReturnType CoeffReturnType;
124 typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
125 static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
126 typedef StorageMemory<CoeffReturnType, Device> Storage;
127 typedef typename Storage::Type EvaluatorPointerType;
133 Layout = TensorEvaluator<ArgType, Device>::Layout,
134 PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
135 BlockAccess = TensorEvaluator<ArgType, Device>::BlockAccess,
138 IsOuterChipping = (
static_cast<int>(Layout) ==
ColMajor && DimId == NumInputDims - 1) ||
139 (
static_cast<int>(Layout) ==
RowMajor && DimId == 0),
141 IsInnerChipping = (
static_cast<int>(Layout) ==
ColMajor && DimId == 0) ||
142 (
static_cast<int>(Layout) ==
RowMajor && DimId == NumInputDims - 1),
145 PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess ||
151 typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
154 typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
155 typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
157 typedef internal::TensorBlockDescriptor<NumInputDims, Index>
159 typedef typename TensorEvaluator<const ArgType, Device>::TensorBlock
162 typedef typename internal::TensorMaterializedBlock<ScalarNoConst, NumDims,
167 EIGEN_STRONG_INLINE TensorEvaluator(
const XprType& op,
const Device& device)
168 : m_impl(op.expression(), device), m_dim(op.dim()), m_device(device)
170 EIGEN_STATIC_ASSERT((NumInputDims >= 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
171 eigen_assert(NumInputDims > m_dim.actualDim());
173 const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
174 eigen_assert(op.offset() < input_dims[m_dim.actualDim()]);
177 for (
int i = 0; i < NumInputDims; ++i) {
178 if (i != m_dim.actualDim()) {
179 m_dimensions[j] = input_dims[i];
186 if (
static_cast<int>(Layout) ==
static_cast<int>(
ColMajor)) {
187 for (
int i = 0; i < m_dim.actualDim(); ++i) {
188 m_stride *= input_dims[i];
189 m_inputStride *= input_dims[i];
192 for (
int i = NumInputDims-1; i > m_dim.actualDim(); --i) {
193 m_stride *= input_dims[i];
194 m_inputStride *= input_dims[i];
197 m_inputStride *= input_dims[m_dim.actualDim()];
198 m_inputOffset = m_stride * op.offset();
201 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
const Dimensions& dimensions()
const {
return m_dimensions; }
203 EIGEN_STRONG_INLINE
bool evalSubExprsIfNeeded(EvaluatorPointerType) {
204 m_impl.evalSubExprsIfNeeded(NULL);
208 EIGEN_STRONG_INLINE
void cleanup() {
212 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index)
const
214 return m_impl.coeff(srcCoeff(index));
217 template<
int LoadMode>
218 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index)
const
220 EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
221 eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
223 if (isInnerChipping()) {
225 eigen_assert(m_stride == 1);
226 Index inputIndex = index * m_inputStride + m_inputOffset;
227 EIGEN_ALIGN_MAX
typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
229 for (
int i = 0; i < PacketSize; ++i) {
230 values[i] = m_impl.coeff(inputIndex);
231 inputIndex += m_inputStride;
233 PacketReturnType rslt = internal::pload<PacketReturnType>(values);
235 }
else if (isOuterChipping()) {
237 eigen_assert(m_stride > index);
238 return m_impl.template packet<LoadMode>(index + m_inputOffset);
240 const Index idx = index / m_stride;
241 const Index rem = index - idx * m_stride;
242 if (rem + PacketSize <= m_stride) {
243 Index inputIndex = idx * m_inputStride + m_inputOffset + rem;
244 return m_impl.template packet<LoadMode>(inputIndex);
247 EIGEN_ALIGN_MAX
typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
249 for (
int i = 0; i < PacketSize; ++i) {
250 values[i] = coeff(index);
253 PacketReturnType rslt = internal::pload<PacketReturnType>(values);
259 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
260 costPerCoeff(
bool vectorized)
const {
262 if ((
static_cast<int>(Layout) ==
static_cast<int>(
ColMajor) &&
263 m_dim.actualDim() == 0) ||
264 (
static_cast<int>(Layout) ==
static_cast<int>(
RowMajor) &&
265 m_dim.actualDim() == NumInputDims - 1)) {
266 cost += TensorOpCost::MulCost<Index>() + TensorOpCost::AddCost<Index>();
267 }
else if ((
static_cast<int>(Layout) ==
static_cast<int>(
ColMajor) &&
268 m_dim.actualDim() == NumInputDims - 1) ||
269 (
static_cast<int>(Layout) ==
static_cast<int>(
RowMajor) &&
270 m_dim.actualDim() == 0)) {
271 cost += TensorOpCost::AddCost<Index>();
273 cost += 3 * TensorOpCost::MulCost<Index>() + TensorOpCost::DivCost<Index>() +
274 3 * TensorOpCost::AddCost<Index>();
277 return m_impl.costPerCoeff(vectorized) +
278 TensorOpCost(0, 0, cost, vectorized, PacketSize);
281 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
282 internal::TensorBlockResourceRequirements getResourceRequirements()
const {
283 const size_t target_size = m_device.lastLevelCacheSize();
284 return internal::TensorBlockResourceRequirements::merge(
285 internal::TensorBlockResourceRequirements::skewed<Scalar>(target_size),
286 m_impl.getResourceRequirements());
289 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
290 block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
291 bool root_of_expr_ast =
false)
const {
292 const Index chip_dim = m_dim.actualDim();
294 DSizes<Index, NumInputDims> input_block_dims;
295 for (
int i = 0; i < NumInputDims; ++i) {
297 = i < chip_dim ? desc.dimension(i)
298 : i > chip_dim ? desc.dimension(i - 1)
302 ArgTensorBlockDesc arg_desc(srcCoeff(desc.offset()), input_block_dims);
305 if (desc.HasDestinationBuffer()) {
306 DSizes<Index, NumInputDims> arg_destination_strides;
307 for (
int i = 0; i < NumInputDims; ++i) {
308 arg_destination_strides[i]
309 = i < chip_dim ? desc.destination().strides()[i]
310 : i > chip_dim ? desc.destination().strides()[i - 1]
314 arg_desc.template AddDestinationBuffer<Layout>(
315 desc.destination().template data<ScalarNoConst>(),
316 arg_destination_strides);
319 ArgTensorBlock arg_block = m_impl.block(arg_desc, scratch, root_of_expr_ast);
320 if (!arg_desc.HasDestinationBuffer()) desc.DropDestinationBuffer();
322 if (arg_block.data() != NULL) {
324 return TensorBlock(arg_block.kind(), arg_block.data(),
331 const typename TensorBlock::Storage block_storage =
332 TensorBlock::prepareStorage(desc, scratch);
334 typedef internal::TensorBlockAssignment<
335 ScalarNoConst, NumInputDims,
typename ArgTensorBlock::XprType,
Index>
336 TensorBlockAssignment;
338 TensorBlockAssignment::Run(
339 TensorBlockAssignment::target(
340 arg_desc.dimensions(),
341 internal::strides<Layout>(arg_desc.dimensions()),
342 block_storage.data()),
345 return block_storage.AsTensorMaterializedBlock();
349 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
typename Storage::Type data()
const {
350 typename Storage::Type result = constCast(m_impl.data());
351 if (isOuterChipping() && result) {
352 return result + m_inputOffset;
359 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
void bind(cl::sycl::handler &cgh)
const {
365 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
Index srcCoeff(Index index)
const
368 if (isInnerChipping()) {
370 eigen_assert(m_stride == 1);
371 inputIndex = index * m_inputStride + m_inputOffset;
372 }
else if (isOuterChipping()) {
375 eigen_assert(m_stride > index);
376 inputIndex = index + m_inputOffset;
378 const Index idx = index / m_stride;
379 inputIndex = idx * m_inputStride + m_inputOffset;
380 index -= idx * m_stride;
386 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
bool isInnerChipping()
const {
387 return IsInnerChipping ||
388 (
static_cast<int>(Layout) ==
ColMajor && m_dim.actualDim() == 0) ||
389 (
static_cast<int>(Layout) ==
RowMajor && m_dim.actualDim() == NumInputDims - 1);
392 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
bool isOuterChipping()
const {
393 return IsOuterChipping ||
394 (
static_cast<int>(Layout) ==
ColMajor && m_dim.actualDim() == NumInputDims-1) ||
395 (
static_cast<int>(Layout) ==
RowMajor && m_dim.actualDim() == 0);
398 Dimensions m_dimensions;
402 TensorEvaluator<ArgType, Device> m_impl;
403 const internal::DimensionId<DimId> m_dim;
404 const Device EIGEN_DEVICE_REF m_device;
409template<DenseIndex DimId,
typename ArgType,
typename Device>
410struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
411 :
public TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
413 typedef TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device> Base;
414 typedef TensorChippingOp<DimId, ArgType> XprType;
415 static const int NumInputDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
416 static const int NumDims = NumInputDims-1;
417 typedef typename XprType::Index
Index;
418 typedef DSizes<Index, NumDims> Dimensions;
419 typedef typename XprType::Scalar Scalar;
420 typedef typename XprType::CoeffReturnType CoeffReturnType;
421 typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
422 static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
426 PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
427 BlockAccess = TensorEvaluator<ArgType, Device>::RawAccess,
428 Layout = TensorEvaluator<ArgType, Device>::Layout,
433 typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
436 EIGEN_STRONG_INLINE TensorEvaluator(
const XprType& op,
const Device& device)
440 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index)
442 return this->m_impl.coeffRef(this->srcCoeff(index));
445 template <
int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
446 void writePacket(Index index,
const PacketReturnType& x)
448 EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
450 if (this->isInnerChipping()) {
452 eigen_assert(this->m_stride == 1);
453 EIGEN_ALIGN_MAX
typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
454 internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
455 Index inputIndex = index * this->m_inputStride + this->m_inputOffset;
457 for (
int i = 0; i < PacketSize; ++i) {
458 this->m_impl.coeffRef(inputIndex) = values[i];
459 inputIndex += this->m_inputStride;
461 }
else if (this->isOuterChipping()) {
463 eigen_assert(this->m_stride > index);
464 this->m_impl.template writePacket<StoreMode>(index + this->m_inputOffset, x);
466 const Index idx = index / this->m_stride;
467 const Index rem = index - idx * this->m_stride;
468 if (rem + PacketSize <= this->m_stride) {
469 const Index inputIndex = idx * this->m_inputStride + this->m_inputOffset + rem;
470 this->m_impl.template writePacket<StoreMode>(inputIndex, x);
473 EIGEN_ALIGN_MAX
typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
474 internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
476 for (
int i = 0; i < PacketSize; ++i) {
477 this->coeffRef(index) = values[i];
484 template <
typename TensorBlock>
485 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
void writeBlock(
486 const TensorBlockDesc& desc,
const TensorBlock& block) {
487 assert(this->m_impl.data() != NULL);
489 const Index chip_dim = this->m_dim.actualDim();
491 DSizes<Index, NumInputDims> input_block_dims;
492 for (
int i = 0; i < NumInputDims; ++i) {
493 input_block_dims[i] = i < chip_dim ? desc.dimension(i)
494 : i > chip_dim ? desc.dimension(i - 1)
498 typedef TensorReshapingOp<const DSizes<Index, NumInputDims>,
499 const typename TensorBlock::XprType>
502 typedef internal::TensorBlockAssignment<Scalar, NumInputDims,
503 TensorBlockExpr,
Index>
506 TensorBlockAssign::Run(
507 TensorBlockAssign::target(
509 internal::strides<Layout>(this->m_impl.dimensions()),
510 this->m_impl.data(), this->srcCoeff(desc.offset())),
511 block.expr().reshape(input_block_dims));
Namespace containing all symbols from the Eigen library.
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index