Eigen-Contrib  5.0.1-dev+master
 
Loading...
Searching...
No Matches
TensorChipping.h
1// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
4// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
5//
6// This Source Code Form is subject to the terms of the Mozilla
7// Public License v. 2.0. If a copy of the MPL was not distributed
8// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9// SPDX-License-Identifier: MPL-2.0
10
11#ifndef EIGEN_TENSOR_TENSOR_CHIPPING_H
12#define EIGEN_TENSOR_TENSOR_CHIPPING_H
13
14// IWYU pragma: private
15#include "./InternalHeaderCheck.h"
16
17namespace Eigen {
18
19namespace internal {
20template <DenseIndex DimId, typename XprType>
21struct traits<TensorChippingOp<DimId, XprType> > : public traits<XprType> {
22 using Scalar = typename XprType::Scalar;
23 using XprTraits = traits<XprType>;
24 using StorageKind = typename XprTraits::StorageKind;
25 using Index = typename XprTraits::Index;
26 static constexpr int NumDimensions = XprTraits::NumDimensions - 1;
27 static constexpr int Layout = XprTraits::Layout;
28 using PointerType = typename XprTraits::PointerType;
29};
30
31template <DenseIndex DimId, typename XprType>
32struct eval<TensorChippingOp<DimId, XprType>, Eigen::Dense> {
33 using type = const TensorChippingOp<DimId, XprType> EIGEN_DEVICE_REF;
34};
35
36template <DenseIndex DimId>
37struct DimensionId {
38 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DimensionId(DenseIndex dim) {
39 EIGEN_UNUSED_VARIABLE(dim);
40 eigen_assert(dim == DimId);
41 }
42 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim() const { return DimId; }
43};
44template <>
45struct DimensionId<Dynamic> {
46 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DimensionId(DenseIndex dim) : actual_dim(dim) { eigen_assert(dim >= 0); }
47 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseIndex actualDim() const { return actual_dim; }
48
49 private:
50 const DenseIndex actual_dim;
51};
52
53template <typename EvalT, typename PacketT, typename IndexType, typename = void>
54struct chipping_has_partial_packet : bool_constant<false> {};
55
56template <typename EvalT, typename PacketT, typename IndexType>
57struct chipping_has_partial_packet<
58 EvalT, PacketT, IndexType,
59 void_t<decltype(std::declval<const EvalT>().template partialPacket<PacketT>(
60 std::declval<IndexType>(), std::declval<typename unpacket_traits<PacketT>::mask_t>()))>>
61 : bool_constant<unpacket_traits<PacketT>::masked_load_available &&
62 std::is_same<PacketT, decltype(std::declval<const EvalT>().template partialPacket<PacketT>(
63 std::declval<IndexType>(),
64 std::declval<typename unpacket_traits<PacketT>::mask_t>()))>::value> {};
65
66template <typename EvalT, typename PacketT, typename IndexType,
67 bool HasPartial = chipping_has_partial_packet<EvalT, PacketT, IndexType>::value>
68struct chipping_forward_partial_packet {
69 static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketT run(const EvalT& impl, IndexType index,
70 typename unpacket_traits<PacketT>::mask_t umask) {
71 return impl.template partialPacket<PacketT>(index, umask);
72 }
73};
74
75template <typename EvalT, typename PacketT, typename IndexType>
76struct chipping_forward_partial_packet<EvalT, PacketT, IndexType, false> {
77 static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketT run(const EvalT&, IndexType,
78 typename unpacket_traits<PacketT>::mask_t) {
79 eigen_assert(false && "unreachable");
80 return pset1<PacketT>(typename unpacket_traits<PacketT>::type(0));
81 }
82};
83
84template <typename EvalT, int LoadMode, typename PacketT, typename IndexType, typename = void>
85struct chipping_has_packet_t : bool_constant<false> {};
86
87template <typename EvalT, int LoadMode, typename PacketT, typename IndexType>
88struct chipping_has_packet_t<
89 EvalT, LoadMode, PacketT, IndexType,
90 void_t<decltype(std::declval<const EvalT>().template packet<LoadMode, PacketT>(std::declval<IndexType>()))>>
91 : bool_constant<std::is_same<PacketT, decltype(std::declval<const EvalT>().template packet<LoadMode, PacketT>(
92 std::declval<IndexType>()))>::value> {};
93
94template <int LoadMode, typename PacketT, typename ArgEvaluator,
95 bool SamePacket = std::is_same<PacketT, typename ArgEvaluator::PacketReturnType>::value,
96 bool HasPacketT = chipping_has_packet_t<ArgEvaluator, LoadMode, PacketT, typename ArgEvaluator::Index>::value,
97 bool IsRawPointer = std::is_convertible<
98 typename ArgEvaluator::EvaluatorPointerType,
99 const std::remove_const_t<std::remove_reference_t<typename ArgEvaluator::CoeffReturnType>>*>::value>
100struct chipping_packet_load {
101 using Index = typename ArgEvaluator::Index;
102 using Scalar = std::remove_const_t<std::remove_reference_t<typename ArgEvaluator::CoeffReturnType>>;
103 static constexpr int kSize = unpacket_traits<PacketT>::size;
104
105 static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketT run(const ArgEvaluator& impl, Index index) {
106 EIGEN_ALIGN_TO_BOUNDARY(unpacket_traits<PacketT>::alignment) Scalar values[kSize];
107 EIGEN_UNROLL_LOOP
108 for (int i = 0; i < kSize; ++i) {
109 values[i] = impl.coeff(index + i);
110 }
111 return pload<PacketT>(values);
112 }
113};
114
115template <int LoadMode, typename PacketT, typename ArgEvaluator, bool HasPacketT, bool IsRawPointer>
116struct chipping_packet_load<LoadMode, PacketT, ArgEvaluator, true, HasPacketT, IsRawPointer> {
117 using Index = typename ArgEvaluator::Index;
118
119 static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketT run(const ArgEvaluator& impl, Index index) {
120 return impl.template packet<LoadMode>(index);
121 }
122};
123
124template <int LoadMode, typename PacketT, typename ArgEvaluator, bool IsRawPointer>
125struct chipping_packet_load<LoadMode, PacketT, ArgEvaluator, false, true, IsRawPointer> {
126 using Index = typename ArgEvaluator::Index;
127
128 static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketT run(const ArgEvaluator& impl, Index index) {
129 return impl.template packet<LoadMode, PacketT>(index);
130 }
131};
132
133template <int LoadMode, typename PacketT, typename ArgEvaluator>
134struct chipping_packet_load<LoadMode, PacketT, ArgEvaluator, false, false, true> {
135 using Index = typename ArgEvaluator::Index;
136 using Scalar = std::remove_const_t<std::remove_reference_t<typename ArgEvaluator::CoeffReturnType>>;
137
138 static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketT run(const ArgEvaluator& impl, Index index) {
139 const Scalar* src_ptr = impl.data();
140 EIGEN_IF_CONSTEXPR (!static_cast<bool>(ArgEvaluator::RawAccess)) {
141 if (src_ptr == nullptr) {
142 return chipping_packet_load<LoadMode, PacketT, ArgEvaluator, false, false, false>::run(impl, index);
143 }
144 }
145 return ploadt_ro<PacketT, Unaligned>(src_ptr + index);
146 }
147};
148
149} // end namespace internal
150
154template <DenseIndex DimId, typename XprType>
155class TensorChippingOp : public TensorBase<TensorChippingOp<DimId, XprType> > {
156 public:
158 using Scalar = typename Eigen::internal::traits<TensorChippingOp>::Scalar;
159 using RealScalar = typename Eigen::NumTraits<Scalar>::Real;
160 using CoeffReturnType = typename XprType::CoeffReturnType;
161 using Nested = typename Eigen::internal::ref_selector<TensorChippingOp>::type;
162 using StorageKind = typename Eigen::internal::traits<TensorChippingOp>::StorageKind;
163 using Index = typename Eigen::internal::traits<TensorChippingOp>::Index;
164
165 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorChippingOp(const XprType& expr, const Index offset, const Index dim)
166 : m_xpr(expr), m_offset(offset), m_dim(dim) {
167 eigen_assert(dim < XprType::NumDimensions && dim >= 0 && "Chip_Dim_out_of_range");
168 }
169
170 EIGEN_DEVICE_FUNC const Index offset() const { return m_offset; }
171 EIGEN_DEVICE_FUNC const Index dim() const { return m_dim.actualDim(); }
172
173 EIGEN_DEVICE_FUNC const internal::remove_all_t<typename XprType::Nested>& expression() const { return m_xpr; }
174
175 EIGEN_INHERIT_ASSIGNMENT_OPERATORS(TensorChippingOp)
176
177 protected:
178 typename XprType::Nested m_xpr;
179 const Index m_offset;
180 const internal::DimensionId<DimId> m_dim;
181};
182
183// Eval as rvalue
184template <DenseIndex DimId, typename ArgType, typename Device>
185struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device> {
186 using XprType = TensorChippingOp<DimId, ArgType>;
187 static constexpr int NumInputDims =
188 internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
189 static constexpr int NumDims = NumInputDims - 1;
190 using Index = typename XprType::Index;
191 using Dimensions = DSizes<Index, NumDims>;
192 using Scalar = typename XprType::Scalar;
193 using CoeffReturnType = typename XprType::CoeffReturnType;
194 using PacketReturnType = typename PacketType<CoeffReturnType, Device>::type;
195 static constexpr int PacketSize = PacketType<CoeffReturnType, Device>::size;
196 using Storage = StorageMemory<CoeffReturnType, Device>;
197 using EvaluatorPointerType = typename Storage::Type;
198 static constexpr int Layout = TensorEvaluator<ArgType, Device>::Layout;
199
200 enum {
201 // Alignment can't be guaranteed at compile time since it depends on the
202 // slice offsets.
203 IsAligned = false,
204 PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
206 // Chipping of outer-most dimension is a trivial operation, because we can
207 // read and write directly from the underlying tensor using a single offset.
208 IsOuterChipping = (Layout == ColMajor && DimId == NumInputDims - 1) || (Layout == RowMajor && DimId == 0),
209 // Chipping inner-most dimension.
210 IsInnerChipping = (Layout == ColMajor && DimId == 0) || (Layout == RowMajor && DimId == NumInputDims - 1),
211 // Prefer block access if the underlying expression prefers it, otherwise
212 // only if chipping is not trivial.
213 PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess || !IsOuterChipping,
214 CoordAccess = false, // to be implemented
215 RawAccess = false
216 };
217
218 using ScalarNoConst = std::remove_const_t<Scalar>;
219
220 //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
221 using TensorBlockDesc = internal::TensorBlockDescriptor<NumDims, Index>;
222 using TensorBlockScratch = internal::TensorBlockScratchAllocator<Device>;
223
224 using ArgTensorBlockDesc = internal::TensorBlockDescriptor<NumInputDims, Index>;
225 using ArgTensorBlock = typename TensorEvaluator<const ArgType, Device>::TensorBlock;
226
227 using TensorBlock = typename internal::TensorMaterializedBlock<ScalarNoConst, NumDims, Layout, Index>;
228 //===--------------------------------------------------------------------===//
229
230 EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
231 : m_impl(op.expression(), device), m_dim(op.dim()), m_device(device) {
232 EIGEN_STATIC_ASSERT((NumInputDims >= 1), YOU_MADE_A_PROGRAMMING_MISTAKE);
233
234 const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
235 eigen_assert(NumInputDims > m_dim.actualDim() && op.offset() < input_dims[m_dim.actualDim()]);
236
237 int j = 0;
238 for (int i = 0; i < NumInputDims; ++i) {
239 if (i != m_dim.actualDim()) {
240 m_dimensions[j] = input_dims[i];
241 ++j;
242 }
243 }
244
245 m_stride = 1;
246 m_inputStride = 1;
247 EIGEN_IF_CONSTEXPR (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
248 for (int i = 0; i < m_dim.actualDim(); ++i) {
249 m_stride *= input_dims[i];
250 m_inputStride *= input_dims[i];
251 }
252 } else {
253 for (int i = NumInputDims - 1; i > m_dim.actualDim(); --i) {
254 m_stride *= input_dims[i];
255 m_inputStride *= input_dims[i];
256 }
257 }
258 m_inputStride *= input_dims[m_dim.actualDim()];
259 m_inputOffset = m_stride * op.offset();
260
261 // Check if chipping is effectively inner or outer: product of dimensions
262 // before or after the chipped dimension is `1`, or the chipped dimension itself has size `1`.
263 Index after_chipped_dim_product = 1;
264 for (int i = static_cast<int>(m_dim.actualDim()) + 1; i < NumInputDims; ++i) {
265 after_chipped_dim_product *= input_dims[i];
266 }
267
268 Index before_chipped_dim_product = 1;
269 for (int i = 0; i < m_dim.actualDim(); ++i) {
270 before_chipped_dim_product *= input_dims[i];
271 }
272
273 const bool is_trivial_chip_dim = input_dims[m_dim.actualDim()] == 1;
274 EIGEN_IF_CONSTEXPR (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
275 m_isEffectivelyInnerChipping = (before_chipped_dim_product == 1) && !is_trivial_chip_dim;
276 m_isEffectivelyOuterChipping = (after_chipped_dim_product == 1) || is_trivial_chip_dim;
277 } else {
278 m_isEffectivelyInnerChipping = (after_chipped_dim_product == 1) && !is_trivial_chip_dim;
279 m_isEffectivelyOuterChipping = (before_chipped_dim_product == 1) || is_trivial_chip_dim;
280 }
281 }
282
283 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
284
285 EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType) {
286 m_impl.evalSubExprsIfNeeded(nullptr);
287 return true;
288 }
289
290#ifdef EIGEN_USE_THREADS
291 template <typename EvalSubExprsCallback>
292 EIGEN_STRONG_INLINE void evalSubExprsIfNeededAsync(EvaluatorPointerType /*data*/, EvalSubExprsCallback done) {
293 m_impl.evalSubExprsIfNeededAsync(nullptr, [done](bool) { done(true); });
294 }
295#endif // EIGEN_USE_THREADS
296
297 EIGEN_STRONG_INLINE void cleanup() { m_impl.cleanup(); }
298
299 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
300 return m_impl.coeff(srcCoeff(index));
301 }
302
303 template <int LoadMode, typename PacketReturnTypeT = PacketReturnType>
304 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE PacketReturnTypeT packet(Index index) const {
305 constexpr int RequestedPacketSize = internal::unpacket_traits<PacketReturnTypeT>::size;
306 EIGEN_STATIC_ASSERT((RequestedPacketSize <= PacketSize), YOU_MADE_A_PROGRAMMING_MISTAKE);
307 eigen_assert(index >= 0 && index + RequestedPacketSize - 1 < dimensions().TotalSize());
308
309 EIGEN_IF_CONSTEXPR (IsInnerChipping) {
310 return packetInnerChipping<LoadMode, PacketReturnTypeT>(index);
311 } else EIGEN_IF_CONSTEXPR (IsOuterChipping) {
312 return packetOuterChipping<LoadMode, PacketReturnTypeT>(index);
313 } else if (m_isEffectivelyInnerChipping) {
314 // index * m_inputStride is linear in the caller's loop index; keep clang-19+ LSR
315 // from hoisting per-lane induction variables across this runtime branch.
316 Index opaqueIndex = index;
317#if EIGEN_COMP_CLANG >= 1900 && !defined(__apple_build_version__) && !defined(EIGEN_GPU_COMPILE_PHASE) && \
318 !defined(EIGEN_SYCL_DEVICE_ONLY)
319 __asm__("" : "+r"(opaqueIndex));
320#endif
321 return packetInnerChipping<LoadMode, PacketReturnTypeT>(opaqueIndex);
322 } else if (m_isEffectivelyOuterChipping) {
323 return packetOuterChipping<LoadMode, PacketReturnTypeT>(index);
324 } else {
325 const Index idx = index / m_stride;
326 const Index rem = index - idx * m_stride;
327 Index inputIndex = idx * m_inputStride + m_inputOffset + rem;
328 if (EIGEN_PREDICT_TRUE(rem + RequestedPacketSize <= m_stride)) {
329 return internal::chipping_packet_load<LoadMode, PacketReturnTypeT, TensorEvaluator<ArgType, Device>>::run(
330 m_impl, inputIndex);
331 } else {
332 EIGEN_ALIGN_TO_BOUNDARY(internal::unpacket_traits<PacketReturnTypeT>::alignment)
333 std::remove_const_t<CoeffReturnType> values[RequestedPacketSize];
334 const Index strideDelta = m_inputStride - m_stride;
335 Index lane_in_slice = rem;
336 EIGEN_UNROLL_LOOP
337 for (int i = 0; i < RequestedPacketSize; ++i) {
338 values[i] = m_impl.coeff(inputIndex);
339 ++inputIndex;
340 if (++lane_in_slice == m_stride) {
341 lane_in_slice = 0;
342 inputIndex += strideDelta;
343 }
344 }
345 return internal::pload<PacketReturnTypeT>(values);
346 }
347 }
348 }
349
350 template <
351 typename PacketReturnTypeT,
352 std::enable_if_t<PacketAccess && internal::unpacket_traits<PacketReturnTypeT>::masked_load_available, int> = 0>
353 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnTypeT
354 partialPacket(Index index, typename internal::unpacket_traits<PacketReturnTypeT>::mask_t umask) const {
355 using MaskT = typename internal::unpacket_traits<PacketReturnTypeT>::mask_t;
356 using RawCoeffType = std::remove_const_t<CoeffReturnType>;
357 using ArgEval = TensorEvaluator<ArgType, Device>;
358 constexpr int RequestedPacketSize = internal::unpacket_traits<PacketReturnTypeT>::size;
359 constexpr int kMaskBits = static_cast<int>(CHAR_BIT * sizeof(MaskT));
360 constexpr bool kHasImplPartial = internal::chipping_has_partial_packet<ArgEval, PacketReturnTypeT, Index>::value;
361
362 EIGEN_IF_CONSTEXPR (RequestedPacketSize < kMaskBits) {
363 umask &= static_cast<MaskT>((MaskT(1) << (RequestedPacketSize < kMaskBits ? RequestedPacketSize : 0)) - 1);
364 }
365 if (umask == 0) {
366 return internal::pset1<PacketReturnTypeT>(CoeffReturnType(0));
367 }
368
369 const Index first_lane = static_cast<Index>(internal::ctz(umask));
370 const Index last_lane = static_cast<Index>(kMaskBits - 1 - internal::clz(umask));
371 const Index active_start = index + first_lane;
372 const Index active_count = last_lane - first_lane + 1;
373 eigen_assert(active_start >= 0 && active_start + active_count - 1 < dimensions().TotalSize());
374
375 if (isOuterChipping()) {
376 const Index target_index = index + m_inputOffset;
377 if (kHasImplPartial && EIGEN_PREDICT_TRUE(target_index >= 0)) {
378 return internal::chipping_forward_partial_packet<ArgEval, PacketReturnTypeT, Index>::run(m_impl, target_index,
379 umask);
380 }
381 EIGEN_ALIGN_TO_BOUNDARY(internal::unpacket_traits<PacketReturnTypeT>::alignment)
382 RawCoeffType values[RequestedPacketSize];
383 EIGEN_UNROLL_LOOP
384 for (int i = 0; i < RequestedPacketSize; ++i) {
385 values[i] = RawCoeffType(0);
386 }
387 const Index srcOffset = srcCoeffOuterChipping(active_start);
388 for (Index lane = first_lane; lane <= last_lane; ++lane) {
389 if ((umask >> lane) & MaskT(1)) {
390 values[lane] = m_impl.coeff(srcOffset + (lane - first_lane));
391 }
392 }
393 return internal::pload<PacketReturnTypeT>(values);
394 }
395
396 if (isInnerChipping()) {
397 EIGEN_ALIGN_TO_BOUNDARY(internal::unpacket_traits<PacketReturnTypeT>::alignment)
398 RawCoeffType values[RequestedPacketSize];
399 EIGEN_UNROLL_LOOP
400 for (int i = 0; i < RequestedPacketSize; ++i) {
401 values[i] = RawCoeffType(0);
402 }
403 Index inputIndex = srcCoeffInnerChipping(active_start);
404 for (Index lane = first_lane; lane <= last_lane; ++lane) {
405 if ((umask >> lane) & MaskT(1)) {
406 values[lane] = m_impl.coeff(inputIndex);
407 }
408 inputIndex += m_inputStride;
409 }
410 return internal::pload<PacketReturnTypeT>(values);
411 }
412
413 const Index idx = active_start / m_stride;
414 const Index rem = active_start - idx * m_stride;
415 const Index sliceBase = idx * m_inputStride + m_inputOffset;
416 if (EIGEN_PREDICT_TRUE(rem + active_count <= m_stride)) {
417 const Index target_index = sliceBase + rem - first_lane;
418 if (kHasImplPartial && EIGEN_PREDICT_TRUE(target_index >= 0)) {
419 return internal::chipping_forward_partial_packet<ArgEval, PacketReturnTypeT, Index>::run(m_impl, target_index,
420 umask);
421 }
422 }
423
424 EIGEN_ALIGN_TO_BOUNDARY(internal::unpacket_traits<PacketReturnTypeT>::alignment)
425 RawCoeffType values[RequestedPacketSize];
426 EIGEN_UNROLL_LOOP
427 for (int i = 0; i < RequestedPacketSize; ++i) {
428 values[i] = RawCoeffType(0);
429 }
430 const Index strideDelta = m_inputStride - m_stride;
431 Index inputIndex = sliceBase + rem;
432 Index lane_in_slice = rem;
433 for (Index lane = first_lane; lane <= last_lane; ++lane) {
434 if ((umask >> lane) & MaskT(1)) {
435 values[lane] = m_impl.coeff(inputIndex);
436 }
437 ++inputIndex;
438 if (++lane_in_slice == m_stride) {
439 lane_in_slice = 0;
440 inputIndex += strideDelta;
441 }
442 }
443 return internal::pload<PacketReturnTypeT>(values);
444 }
445
446 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
447 double cost = 0;
448 if (isOuterChipping()) {
449 cost += TensorOpCost::AddCost<Index>();
450 } else if (isInnerChipping()) {
451 cost += TensorOpCost::MulCost<Index>() + TensorOpCost::AddCost<Index>();
452 } else {
453 cost += 3 * TensorOpCost::MulCost<Index>() + TensorOpCost::DivCost<Index>() + 3 * TensorOpCost::AddCost<Index>();
454 }
455
456 return m_impl.costPerCoeff(vectorized) + TensorOpCost(0, 0, cost, vectorized, PacketSize);
457 }
458
459 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE internal::TensorBlockResourceRequirements getResourceRequirements() const {
460 const size_t target_size = m_device.lastLevelCacheSize();
461 return internal::TensorBlockResourceRequirements::merge(
462 internal::TensorBlockResourceRequirements::skewed<Scalar>(target_size), m_impl.getResourceRequirements());
463 }
464
465 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock block(TensorBlockDesc& desc, TensorBlockScratch& scratch,
466 bool root_of_expr_ast = false) const {
467 const Index chip_dim = m_dim.actualDim();
468
469 DSizes<Index, NumInputDims> input_block_dims;
470 input_block_dims[chip_dim] = 1;
471 for (int i = 0; i < NumDims; ++i) {
472 input_block_dims[i < chip_dim ? i : i + 1] = desc.dimension(i);
473 }
474
475 ArgTensorBlockDesc arg_desc(srcCoeff(desc.offset()), input_block_dims);
476
477 // Try to reuse destination buffer for materializing argument block.
478 if (desc.HasDestinationBuffer()) {
479 DSizes<Index, NumInputDims> arg_destination_strides;
480 arg_destination_strides[chip_dim] = 0; // The size-one chipped dimension does not use its stride.
481 for (int i = 0; i < NumDims; ++i) {
482 arg_destination_strides[i < chip_dim ? i : i + 1] = desc.destination().strides()[i];
483 }
484
485 arg_desc.template AddDestinationBuffer<Layout>(desc.destination().template data<ScalarNoConst>(),
486 arg_destination_strides);
487 }
488
489 ArgTensorBlock arg_block = m_impl.block(arg_desc, scratch, root_of_expr_ast);
490 if (!arg_desc.HasDestinationBuffer()) desc.DropDestinationBuffer();
491
492 if (arg_block.data() != nullptr) {
493 // Forward argument block buffer if possible.
494 return TensorBlock(arg_block.kind(), arg_block.data(), desc.dimensions());
495
496 } else {
497 // Assign argument block expression to a buffer.
498
499 // Prepare storage for the materialized chipping result.
500 const typename TensorBlock::Storage block_storage = TensorBlock::prepareStorage(desc, scratch);
501
502 using TensorBlockAssignment =
503 internal::TensorBlockAssignment<ScalarNoConst, NumInputDims, typename ArgTensorBlock::XprType, Index>;
504
505 TensorBlockAssignment::Run(
506 TensorBlockAssignment::target(arg_desc.dimensions(), internal::strides<Layout>(arg_desc.dimensions()),
507 block_storage.data()),
508 arg_block.expr());
509
510 return block_storage.AsTensorMaterializedBlock();
511 }
512 }
513
514 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Storage::Type data() const {
515 typename Storage::Type result = constCast(m_impl.data());
516 if (isOuterChipping() && result) {
517 return result + m_inputOffset;
518 } else {
519 return nullptr;
520 }
521 }
522
523 protected:
524 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const {
525 EIGEN_IF_CONSTEXPR (IsInnerChipping) {
526 return srcCoeffInnerChipping(index);
527 } else EIGEN_IF_CONSTEXPR (IsOuterChipping) {
528 return srcCoeffOuterChipping(index);
529 } else if (m_isEffectivelyInnerChipping) {
530 return srcCoeffInnerChipping(index);
531 } else if (m_isEffectivelyOuterChipping) {
532 return srcCoeffOuterChipping(index);
533 } else {
534 const Index idx = index / m_stride;
535 Index inputIndex = idx * m_inputStride + m_inputOffset;
536 index -= idx * m_stride;
537 inputIndex += index;
538 return inputIndex;
539 }
540 }
541
542 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeffInnerChipping(Index index) const {
543 // m_stride is equal to 1, so let's avoid the integer division.
544 eigen_assert(m_stride == 1);
545 return index * m_inputStride + m_inputOffset;
546 }
547
548 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeffOuterChipping(Index index) const {
549 // m_stride > index (or m_stride == m_inputStride when chipped dim has size 1).
550 eigen_assert(m_stride > index || m_stride == m_inputStride);
551 return index + m_inputOffset;
552 }
553
554 template <int LoadMode, typename PacketReturnTypeT = PacketReturnType>
555 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnTypeT packetInnerChipping(Index index) const {
556 constexpr int RequestedPacketSize = internal::unpacket_traits<PacketReturnTypeT>::size;
557 EIGEN_ALIGN_TO_BOUNDARY(internal::unpacket_traits<PacketReturnTypeT>::alignment)
558 std::remove_const_t<CoeffReturnType> values[RequestedPacketSize];
559 Index inputIndex = srcCoeffInnerChipping(index);
560 EIGEN_UNROLL_LOOP
561 for (int i = 0; i < RequestedPacketSize; ++i) {
562 values[i] = m_impl.coeff(inputIndex);
563 inputIndex += m_inputStride;
564 }
565 return internal::pload<PacketReturnTypeT>(values);
566 }
567
568 template <int LoadMode, typename PacketReturnTypeT = PacketReturnType>
569 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnTypeT packetOuterChipping(Index index) const {
570 return internal::chipping_packet_load<LoadMode, PacketReturnTypeT, TensorEvaluator<ArgType, Device>>::run(
571 m_impl, srcCoeffOuterChipping(index));
572 }
573
574 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool isInnerChipping() const {
575 return IsInnerChipping || m_isEffectivelyInnerChipping;
576 }
577
578 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool isOuterChipping() const {
579 return IsOuterChipping || m_isEffectivelyOuterChipping;
580 }
581
582 Dimensions m_dimensions;
583 Index m_stride;
584 Index m_inputOffset;
585 Index m_inputStride;
586 TensorEvaluator<ArgType, Device> m_impl;
587 const internal::DimensionId<DimId> m_dim;
588 const Device EIGEN_DEVICE_REF m_device;
589
590 // If product of all dimensions after or before the chipped dimension is `1`,
591 // or the chipped dimension itself has size `1`, it is effectively the same
592 // as chipping innermost or outermost dimension.
593 bool m_isEffectivelyInnerChipping;
594 bool m_isEffectivelyOuterChipping;
595};
596
597// Eval as lvalue
598template <DenseIndex DimId, typename ArgType, typename Device>
599struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
600 : public TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device> {
601 using Base = TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>;
602 using XprType = TensorChippingOp<DimId, ArgType>;
603 static constexpr int NumInputDims =
604 internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value;
605 static constexpr int NumDims = NumInputDims - 1;
606 using Index = typename XprType::Index;
607 using Dimensions = DSizes<Index, NumDims>;
608 using Scalar = typename XprType::Scalar;
609 using CoeffReturnType = typename XprType::CoeffReturnType;
610 using PacketReturnType = typename PacketType<CoeffReturnType, Device>::type;
611 static constexpr int PacketSize = PacketType<CoeffReturnType, Device>::size;
612
613 enum {
614 IsAligned = false,
615 PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
616 BlockAccess = TensorEvaluator<ArgType, Device>::RawAccess,
617 Layout = TensorEvaluator<ArgType, Device>::Layout,
618 RawAccess = false
619 };
620
621 //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
622 using TensorBlockDesc = internal::TensorBlockDescriptor<NumDims, Index>;
623 //===--------------------------------------------------------------------===//
624
625 EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device) : Base(op, device) {}
626
627 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType& coeffRef(Index index) const {
628 return this->m_impl.coeffRef(this->srcCoeff(index));
629 }
630
631 template <int StoreMode>
632 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writePacket(Index index, const PacketReturnType& x) const {
633 EIGEN_IF_CONSTEXPR (Base::IsInnerChipping) {
634 writePacketInnerChipping<StoreMode>(index, x);
635 } else EIGEN_IF_CONSTEXPR (Base::IsOuterChipping) {
636 writePacketOuterChipping<StoreMode>(index, x);
637 } else if (this->m_isEffectivelyInnerChipping) {
638 writePacketInnerChipping<StoreMode>(index, x);
639 } else if (this->m_isEffectivelyOuterChipping) {
640 writePacketOuterChipping<StoreMode>(index, x);
641 } else {
642 const Index idx = index / this->m_stride;
643 const Index rem = index - idx * this->m_stride;
644 Index inputIndex = idx * this->m_inputStride + this->m_inputOffset + rem;
645 if (EIGEN_PREDICT_TRUE(rem + PacketSize <= this->m_stride)) {
646 this->m_impl.template writePacket<StoreMode>(inputIndex, x);
647 } else {
648 EIGEN_ALIGN_TO_BOUNDARY(internal::unpacket_traits<PacketReturnType>::alignment)
649 std::remove_const_t<CoeffReturnType> values[PacketSize];
650 internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
651 const Index strideDelta = this->m_inputStride - this->m_stride;
652 Index lane_in_slice = rem;
653 EIGEN_UNROLL_LOOP
654 for (int i = 0; i < PacketSize; ++i) {
655 this->m_impl.coeffRef(inputIndex) = values[i];
656 ++inputIndex;
657 if (++lane_in_slice == this->m_stride) {
658 lane_in_slice = 0;
659 inputIndex += strideDelta;
660 }
661 }
662 }
663 }
664 }
665
666 template <typename TensorBlock>
667 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writeBlock(const TensorBlockDesc& desc, const TensorBlock& block) {
668 eigen_assert(this->m_impl.data() != nullptr);
669
670 const Index chip_dim = this->m_dim.actualDim();
671
672 DSizes<Index, NumInputDims> input_block_dims;
673 input_block_dims[chip_dim] = 1;
674 for (int i = 0; i < NumDims; ++i) {
675 input_block_dims[i < chip_dim ? i : i + 1] = desc.dimension(i);
676 }
677
678 using TensorBlockExpr = TensorReshapingOp<const DSizes<Index, NumInputDims>, const typename TensorBlock::XprType>;
679
680 using TensorBlockAssign = internal::TensorBlockAssignment<Scalar, NumInputDims, TensorBlockExpr, Index>;
681
682 TensorBlockAssign::Run(
683 TensorBlockAssign::target(input_block_dims, internal::strides<Layout>(this->m_impl.dimensions()),
684 this->m_impl.data(), this->srcCoeff(desc.offset())),
685 block.expr().reshape(input_block_dims));
686 }
687
688 private:
689 template <int StoreMode>
690 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writePacketInnerChipping(Index index, const PacketReturnType& x) const {
691 EIGEN_ALIGN_TO_BOUNDARY(internal::unpacket_traits<PacketReturnType>::alignment)
692 std::remove_const_t<CoeffReturnType> values[PacketSize];
693 internal::pstore<CoeffReturnType, PacketReturnType>(values, x);
694 Index inputIndex = this->srcCoeffInnerChipping(index);
695 EIGEN_UNROLL_LOOP
696 for (int i = 0; i < PacketSize; ++i) {
697 this->m_impl.coeffRef(inputIndex) = values[i];
698 inputIndex += this->m_inputStride;
699 }
700 }
701
702 template <int StoreMode>
703 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void writePacketOuterChipping(Index index, const PacketReturnType& x) const {
704 this->m_impl.template writePacket<StoreMode>(this->srcCoeffOuterChipping(index), x);
705 }
706};
707
708} // end namespace Eigen
709
710#endif // EIGEN_TENSOR_TENSOR_CHIPPING_H
The tensor base class.
Definition TensorForwardDeclarations.h:69
Definition TensorChipping.h:155
Namespace containing all symbols from the Eigen library.
The tensor evaluator class.
Definition TensorEvaluator.h:47