xref: /aosp_15_r20/external/eigen/unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Mehdi Goli    Codeplay Software Ltd.
5 // Ralph Potter  Codeplay Software Ltd.
6 // Luke Iwanski  Codeplay Software Ltd.
7 // Contact: <[email protected]>
8 //
9 // This Source Code Form is subject to the terms of the Mozilla
10 // Public License v. 2.0. If a copy of the MPL was not distributed
11 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
12 
13 /*****************************************************************
14  * TensorReductionSycl.h
15  *
16  * \brief:
17  *  This is the specialization of the reduction operation. Two phase reduction approach
18  * is used since the GPU does not have Global Synchronization for global memory among
19  * different work-group/thread block. To solve the problem, we need to create two kernels
20  * to reduce the data, where the first kernel reduce the data locally and each local
21  * workgroup/thread-block save the input data into global memory. In the second phase (global reduction)
22  * one work-group uses one work-group/thread-block to reduces the intermediate data into one single element.
23  * Here is an NVIDIA presentation explaining the optimized two phase reduction algorithm on GPU:
24  * https://developer.download.nvidia.com/assets/cuda/files/reduction.pdf
25  *
26  *****************************************************************/
27 
28 #ifndef UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP
29 #define UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP
30 namespace Eigen {
31 namespace TensorSycl {
32 namespace internal {
33 
34 template <typename Op, typename CoeffReturnType, typename Index, bool Vectorizable>
35 struct OpDefiner {
36   typedef typename Vectorise<CoeffReturnType, Eigen::SyclDevice, Vectorizable>::PacketReturnType PacketReturnType;
37   typedef Op type;
get_opOpDefiner38   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE type get_op(Op &op) { return op; }
39 
finalise_opOpDefiner40   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType finalise_op(const PacketReturnType &accumulator,
41                                                                             const Index &) {
42     return accumulator;
43   }
44 };
45 
46 template <typename CoeffReturnType, typename Index>
47 struct OpDefiner<Eigen::internal::MeanReducer<CoeffReturnType>, CoeffReturnType, Index, false> {
48   typedef Eigen::internal::SumReducer<CoeffReturnType> type;
49   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE type get_op(Eigen::internal::MeanReducer<CoeffReturnType> &) {
50     return type();
51   }
52 
53   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType finalise_op(const CoeffReturnType &accumulator,
54                                                                            const Index &scale) {
55     ::Eigen::internal::scalar_quotient_op<CoeffReturnType> quotient_op;
56     return quotient_op(accumulator, CoeffReturnType(scale));
57   }
58 };
59 
60 template <typename CoeffReturnType, typename Index>
61 struct OpDefiner<Eigen::internal::MeanReducer<CoeffReturnType>, CoeffReturnType, Index, true> {
62   typedef typename Vectorise<CoeffReturnType, Eigen::SyclDevice, true>::PacketReturnType PacketReturnType;
63   typedef Eigen::internal::SumReducer<CoeffReturnType> type;
64   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE type get_op(Eigen::internal::MeanReducer<CoeffReturnType> &) {
65     return type();
66   }
67 
68   static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType finalise_op(const PacketReturnType &accumulator,
69                                                                             const Index &scale) {
70     return ::Eigen::internal::pdiv(accumulator, ::Eigen::internal::pset1<PacketReturnType>(CoeffReturnType(scale)));
71   }
72 };
73 
74 template <typename CoeffReturnType, typename OpType, typename InputAccessor, typename OutputAccessor, typename Index,
75           Index local_range>
76 struct SecondStepFullReducer {
77   typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
78       LocalAccessor;
79   typedef OpDefiner<OpType, CoeffReturnType, Index, true> OpDef;
80   typedef typename OpDef::type Op;
81   LocalAccessor scratch;
82   InputAccessor aI;
83   OutputAccessor outAcc;
84   Op op;
85   SecondStepFullReducer(LocalAccessor scratch_, InputAccessor aI_, OutputAccessor outAcc_, OpType op_)
86       : scratch(scratch_), aI(aI_), outAcc(outAcc_), op(OpDef::get_op(op_)) {}
87 
88   void operator()(cl::sycl::nd_item<1> itemID) {
89     // Our empirical research shows that the best performance will be achieved
90     // when there is only one element per thread to reduce in the second step.
91     // in this step the second step reduction time is almost negligible.
92     // Hence, in the second step of reduction the input size is fixed to the
93     // local size, thus, there is only one element read per thread. The
94     // algorithm must be changed if the number of reduce per thread in the
95     // second step is greater than 1. Otherwise, the result will be wrong.
96     const Index localid = itemID.get_local_id(0);
97     auto aInPtr = aI.get_pointer() + localid;
98     auto aOutPtr = outAcc.get_pointer();
99     CoeffReturnType *scratchptr = scratch.get_pointer();
100     CoeffReturnType accumulator = *aInPtr;
101 
102     scratchptr[localid] = op.finalize(accumulator);
103     for (Index offset = itemID.get_local_range(0) / 2; offset > 0; offset /= 2) {
104       itemID.barrier(cl::sycl::access::fence_space::local_space);
105       if (localid < offset) {
106         op.reduce(scratchptr[localid + offset], &accumulator);
107         scratchptr[localid] = op.finalize(accumulator);
108       }
109     }
110     if (localid == 0) *aOutPtr = op.finalize(accumulator);
111   }
112 };
113 
114 // Full reduction first phase. In this version the vectorization is true and the reduction accept
115 // any generic reducerOp  e.g( max, min, sum, mean, iamax, iamin, etc ).
116 template <typename Evaluator, typename OpType, typename Evaluator::Index local_range>
117 class FullReductionKernelFunctor {
118  public:
119   typedef typename Evaluator::CoeffReturnType CoeffReturnType;
120   typedef typename Evaluator::Index Index;
121   typedef OpDefiner<OpType, typename Evaluator::CoeffReturnType, Index,
122                     (Evaluator::ReducerTraits::PacketAccess & Evaluator::InputPacketAccess)>
123       OpDef;
124 
125   typedef typename OpDef::type Op;
126   typedef typename Evaluator::EvaluatorPointerType EvaluatorPointerType;
127   typedef typename Evaluator::PacketReturnType PacketReturnType;
128   typedef
129       typename ::Eigen::internal::conditional<(Evaluator::ReducerTraits::PacketAccess & Evaluator::InputPacketAccess),
130                                               PacketReturnType, CoeffReturnType>::type OutType;
131   typedef cl::sycl::accessor<OutType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
132       LocalAccessor;
133   LocalAccessor scratch;
134   Evaluator evaluator;
135   EvaluatorPointerType final_output;
136   Index rng;
137   Op op;
138 
139   FullReductionKernelFunctor(LocalAccessor scratch_, Evaluator evaluator_, EvaluatorPointerType final_output_,
140                              Index rng_, OpType op_)
141       : scratch(scratch_), evaluator(evaluator_), final_output(final_output_), rng(rng_), op(OpDef::get_op(op_)) {}
142 
143   void operator()(cl::sycl::nd_item<1> itemID) { compute_reduction(itemID); }
144 
145   template <bool Vect = (Evaluator::ReducerTraits::PacketAccess & Evaluator::InputPacketAccess)>
146   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<Vect>::type compute_reduction(
147       const cl::sycl::nd_item<1> &itemID) {
148     auto output_ptr = final_output.get_pointer();
149     Index VectorizedRange = (rng / Evaluator::PacketSize) * Evaluator::PacketSize;
150     Index globalid = itemID.get_global_id(0);
151     Index localid = itemID.get_local_id(0);
152     Index step = Evaluator::PacketSize * itemID.get_global_range(0);
153     Index start = Evaluator::PacketSize * globalid;
154     // vectorizable parts
155     PacketReturnType packetAccumulator = op.template initializePacket<PacketReturnType>();
156     for (Index i = start; i < VectorizedRange; i += step) {
157       op.template reducePacket<PacketReturnType>(evaluator.impl().template packet<Unaligned>(i), &packetAccumulator);
158     }
159     globalid += VectorizedRange;
160     // non vectorizable parts
161     for (Index i = globalid; i < rng; i += itemID.get_global_range(0)) {
162       op.template reducePacket<PacketReturnType>(
163           ::Eigen::TensorSycl::internal::PacketWrapper<PacketReturnType, Evaluator::PacketSize>::convert_to_packet_type(
164               evaluator.impl().coeff(i), op.initialize()),
165           &packetAccumulator);
166     }
167     scratch[localid] = packetAccumulator =
168         OpDef::finalise_op(op.template finalizePacket<PacketReturnType>(packetAccumulator), rng);
169     // reduction parts // Local size is always power of 2
170     EIGEN_UNROLL_LOOP
171     for (Index offset = local_range / 2; offset > 0; offset /= 2) {
172       itemID.barrier(cl::sycl::access::fence_space::local_space);
173       if (localid < offset) {
174         op.template reducePacket<PacketReturnType>(scratch[localid + offset], &packetAccumulator);
175         scratch[localid] = op.template finalizePacket<PacketReturnType>(packetAccumulator);
176       }
177     }
178     if (localid == 0) {
179       output_ptr[itemID.get_group(0)] =
180           op.finalizeBoth(op.initialize(), op.template finalizePacket<PacketReturnType>(packetAccumulator));
181     }
182   }
183 
184   template <bool Vect = (Evaluator::ReducerTraits::PacketAccess & Evaluator::InputPacketAccess)>
185   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename ::Eigen::internal::enable_if<!Vect>::type compute_reduction(
186       const cl::sycl::nd_item<1> &itemID) {
187     auto output_ptr = final_output.get_pointer();
188     Index globalid = itemID.get_global_id(0);
189     Index localid = itemID.get_local_id(0);
190     // vectorizable parts
191     CoeffReturnType accumulator = op.initialize();
192     // non vectorizable parts
193     for (Index i = globalid; i < rng; i += itemID.get_global_range(0)) {
194       op.reduce(evaluator.impl().coeff(i), &accumulator);
195     }
196     scratch[localid] = accumulator = OpDef::finalise_op(op.finalize(accumulator), rng);
197 
198     // reduction parts. the local size is always power of 2
199     EIGEN_UNROLL_LOOP
200     for (Index offset = local_range / 2; offset > 0; offset /= 2) {
201       itemID.barrier(cl::sycl::access::fence_space::local_space);
202       if (localid < offset) {
203         op.reduce(scratch[localid + offset], &accumulator);
204         scratch[localid] = op.finalize(accumulator);
205       }
206     }
207     if (localid == 0) {
208       output_ptr[itemID.get_group(0)] = op.finalize(accumulator);
209     }
210   }
211 };
212 
213 template <typename Evaluator, typename OpType>
214 class GenericNondeterministicReducer {
215  public:
216   typedef typename Evaluator::CoeffReturnType CoeffReturnType;
217   typedef typename Evaluator::EvaluatorPointerType EvaluatorPointerType;
218   typedef typename Evaluator::Index Index;
219   typedef OpDefiner<OpType, CoeffReturnType, Index, false> OpDef;
220   typedef typename OpDef::type Op;
221   template <typename Scratch>
222   GenericNondeterministicReducer(Scratch, Evaluator evaluator_, EvaluatorPointerType output_accessor_, OpType functor_,
223                        Index range_, Index num_values_to_reduce_)
224       : evaluator(evaluator_),
225         output_accessor(output_accessor_),
226         functor(OpDef::get_op(functor_)),
227         range(range_),
228         num_values_to_reduce(num_values_to_reduce_) {}
229 
230   void operator()(cl::sycl::nd_item<1> itemID) {
231     auto output_accessor_ptr = output_accessor.get_pointer();
232     /// const cast added as a naive solution to solve the qualifier drop error
233     Index globalid = static_cast<Index>(itemID.get_global_linear_id());
234     if (globalid < range) {
235       CoeffReturnType accum = functor.initialize();
236       Eigen::internal::GenericDimReducer<Evaluator::NumReducedDims - 1, Evaluator, Op>::reduce(
237           evaluator, evaluator.firstInput(globalid), functor, &accum);
238       output_accessor_ptr[globalid] = OpDef::finalise_op(functor.finalize(accum), num_values_to_reduce);
239     }
240   }
241 
242  private:
243   Evaluator evaluator;
244   EvaluatorPointerType output_accessor;
245   Op functor;
246   Index range;
247   Index num_values_to_reduce;
248 };
249 
250 enum class reduction_dim { inner_most, outer_most };
251 // default is preserver
252 template <typename Evaluator, typename OpType, typename PannelParameters, reduction_dim rt>
253 struct PartialReductionKernel {
254   typedef typename Evaluator::CoeffReturnType CoeffReturnType;
255   typedef typename Evaluator::EvaluatorPointerType EvaluatorPointerType;
256   typedef typename Evaluator::Index Index;
257   typedef OpDefiner<OpType, CoeffReturnType, Index, false> OpDef;
258   typedef typename OpDef::type Op;
259   typedef cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
260       ScratchAcc;
261   ScratchAcc scratch;
262   Evaluator evaluator;
263   EvaluatorPointerType output_accessor;
264   Op op;
265   const Index preserve_elements_num_groups;
266   const Index reduce_elements_num_groups;
267   const Index num_coeffs_to_preserve;
268   const Index num_coeffs_to_reduce;
269 
270   PartialReductionKernel(ScratchAcc scratch_, Evaluator evaluator_, EvaluatorPointerType output_accessor_, OpType op_,
271                          const Index preserve_elements_num_groups_, const Index reduce_elements_num_groups_,
272                          const Index num_coeffs_to_preserve_, const Index num_coeffs_to_reduce_)
273       : scratch(scratch_),
274         evaluator(evaluator_),
275         output_accessor(output_accessor_),
276         op(OpDef::get_op(op_)),
277         preserve_elements_num_groups(preserve_elements_num_groups_),
278         reduce_elements_num_groups(reduce_elements_num_groups_),
279         num_coeffs_to_preserve(num_coeffs_to_preserve_),
280         num_coeffs_to_reduce(num_coeffs_to_reduce_) {}
281 
282   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void element_wise_reduce(Index globalRId, Index globalPId,
283                                                                  CoeffReturnType &accumulator) {
284     if (globalPId >= num_coeffs_to_preserve) {
285       return;
286     }
287     Index global_offset = rt == reduction_dim::outer_most ? globalPId + (globalRId * num_coeffs_to_preserve)
288                                                           : globalRId + (globalPId * num_coeffs_to_reduce);
289     Index localOffset = globalRId;
290 
291     const Index per_thread_local_stride = PannelParameters::LocalThreadSizeR * reduce_elements_num_groups;
292     const Index per_thread_global_stride =
293         rt == reduction_dim::outer_most ? num_coeffs_to_preserve * per_thread_local_stride : per_thread_local_stride;
294     for (Index i = globalRId; i < num_coeffs_to_reduce; i += per_thread_local_stride) {
295       op.reduce(evaluator.impl().coeff(global_offset), &accumulator);
296       localOffset += per_thread_local_stride;
297       global_offset += per_thread_global_stride;
298     }
299   }
300   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(cl::sycl::nd_item<1> itemID) {
301     const Index linearLocalThreadId = itemID.get_local_id(0);
302     Index pLocalThreadId = rt == reduction_dim::outer_most ? linearLocalThreadId % PannelParameters::LocalThreadSizeP
303                                                            : linearLocalThreadId / PannelParameters::LocalThreadSizeR;
304     Index rLocalThreadId = rt == reduction_dim::outer_most ? linearLocalThreadId / PannelParameters::LocalThreadSizeP
305                                                            : linearLocalThreadId % PannelParameters::LocalThreadSizeR;
306     const Index pGroupId = rt == reduction_dim::outer_most ? itemID.get_group(0) % preserve_elements_num_groups
307                                                            : itemID.get_group(0) / reduce_elements_num_groups;
308     const Index rGroupId = rt == reduction_dim::outer_most ? itemID.get_group(0) / preserve_elements_num_groups
309                                                            : itemID.get_group(0) % reduce_elements_num_groups;
310 
311     Index globalPId = pGroupId * PannelParameters::LocalThreadSizeP + pLocalThreadId;
312     const Index globalRId = rGroupId * PannelParameters::LocalThreadSizeR + rLocalThreadId;
313     auto scratchPtr = scratch.get_pointer().get();
314     auto outPtr =
315         output_accessor.get_pointer() + (reduce_elements_num_groups > 1 ? rGroupId * num_coeffs_to_preserve : 0);
316     CoeffReturnType accumulator = op.initialize();
317 
318     element_wise_reduce(globalRId, globalPId, accumulator);
319 
320     accumulator = OpDef::finalise_op(op.finalize(accumulator), num_coeffs_to_reduce);
321     scratchPtr[pLocalThreadId + rLocalThreadId * (PannelParameters::LocalThreadSizeP + PannelParameters::BC)] =
322         accumulator;
323     if (rt == reduction_dim::inner_most) {
324       pLocalThreadId = linearLocalThreadId % PannelParameters::LocalThreadSizeP;
325       rLocalThreadId = linearLocalThreadId / PannelParameters::LocalThreadSizeP;
326       globalPId = pGroupId * PannelParameters::LocalThreadSizeP + pLocalThreadId;
327     }
328 
329     /* Apply the reduction operation between the current local
330      * id and the one on the other half of the vector. */
331     auto out_scratch_ptr =
332         scratchPtr + (pLocalThreadId + (rLocalThreadId * (PannelParameters::LocalThreadSizeP + PannelParameters::BC)));
333     itemID.barrier(cl::sycl::access::fence_space::local_space);
334     if (rt == reduction_dim::inner_most) {
335       accumulator = *out_scratch_ptr;
336     }
337     // The Local LocalThreadSizeR is always power of 2
338     EIGEN_UNROLL_LOOP
339     for (Index offset = PannelParameters::LocalThreadSizeR >> 1; offset > 0; offset >>= 1) {
340       if (rLocalThreadId < offset) {
341         op.reduce(out_scratch_ptr[(PannelParameters::LocalThreadSizeP + PannelParameters::BC) * offset], &accumulator);
342         // The result has already been divided for mean reducer in the
343         // previous reduction so no need to divide furthermore
344         *out_scratch_ptr = op.finalize(accumulator);
345       }
346       /* All threads collectively read from global memory into local.
347        * The barrier ensures all threads' IO is resolved before
348        * execution continues (strictly speaking, all threads within
349        * a single work-group - there is no co-ordination between
350        * work-groups, only work-items). */
351       itemID.barrier(cl::sycl::access::fence_space::local_space);
352     }
353 
354     if (rLocalThreadId == 0 && (globalPId < num_coeffs_to_preserve)) {
355       outPtr[globalPId] = op.finalize(accumulator);
356     }
357   }
358 };
359 
360 template <typename OutScalar, typename Index, typename InputAccessor, typename OutputAccessor, typename OpType>
361 struct SecondStepPartialReduction {
362   typedef OpDefiner<OpType, OutScalar, Index, false> OpDef;
363   typedef typename OpDef::type Op;
364   typedef cl::sycl::accessor<OutScalar, 1, cl::sycl::access::mode::read_write, cl::sycl::access::target::local>
365       ScratchAccessor;
366   InputAccessor input_accessor;
367   OutputAccessor output_accessor;
368   Op op;
369   const Index num_coeffs_to_preserve;
370   const Index num_coeffs_to_reduce;
371 
372   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE SecondStepPartialReduction(ScratchAccessor, InputAccessor input_accessor_,
373                                                                    OutputAccessor output_accessor_, OpType op_,
374                                                                    const Index num_coeffs_to_preserve_,
375                                                                    const Index num_coeffs_to_reduce_)
376       : input_accessor(input_accessor_),
377         output_accessor(output_accessor_),
378         op(OpDef::get_op(op_)),
379         num_coeffs_to_preserve(num_coeffs_to_preserve_),
380         num_coeffs_to_reduce(num_coeffs_to_reduce_) {}
381 
382   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(cl::sycl::nd_item<1> itemID) {
383     const Index globalId = itemID.get_global_id(0);
384 
385     if (globalId >= num_coeffs_to_preserve) return;
386 
387     auto in_ptr = input_accessor.get_pointer() + globalId;
388 
389     OutScalar accumulator = op.initialize();
390 // num_coeffs_to_reduce is not bigger that 256
391     for (Index i = 0; i < num_coeffs_to_reduce; i++) {
392       op.reduce(*in_ptr, &accumulator);
393       in_ptr += num_coeffs_to_preserve;
394     }
395     output_accessor.get_pointer()[globalId] = op.finalize(accumulator);
396   }
397 };  // namespace internal
398 
399 template <typename Index, Index LTP, Index LTR, bool BC_>
400 struct ReductionPannel {
401   static EIGEN_CONSTEXPR Index LocalThreadSizeP = LTP;
402   static EIGEN_CONSTEXPR Index LocalThreadSizeR = LTR;
403   static EIGEN_CONSTEXPR bool BC = BC_;
404 };
405 
406 template <typename Self, typename Op, TensorSycl::internal::reduction_dim rt>
407 struct PartialReducerLauncher {
408   typedef typename Self::EvaluatorPointerType EvaluatorPointerType;
409   typedef typename Self::CoeffReturnType CoeffReturnType;
410   typedef typename Self::Storage Storage;
411   typedef typename Self::Index Index;
412   typedef ReductionPannel<typename Self::Index, EIGEN_SYCL_LOCAL_THREAD_DIM0, EIGEN_SYCL_LOCAL_THREAD_DIM1, true>
413       PannelParameters;
414 
415   typedef PartialReductionKernel<Self, Op, PannelParameters, rt> SyclReducerKerneType;
416 
417   static bool run(const Self &self, const Op &reducer, const Eigen::SyclDevice &dev, EvaluatorPointerType output,
418                   Index num_coeffs_to_reduce, Index num_coeffs_to_preserve) {
419     Index roundUpP = roundUp(num_coeffs_to_preserve, PannelParameters::LocalThreadSizeP);
420 
421     // getPowerOfTwo makes sure local range is power of 2 and <=
422     // maxSyclThreadPerBlock this will help us to avoid extra check on the
423     // kernel
424     static_assert(!((PannelParameters::LocalThreadSizeP * PannelParameters::LocalThreadSizeR) &
425                     (PannelParameters::LocalThreadSizeP * PannelParameters::LocalThreadSizeR - 1)),
426                   "The Local thread size must be a power of 2 for the reduction "
427                   "operation");
428 
429     EIGEN_CONSTEXPR Index localRange = PannelParameters::LocalThreadSizeP * PannelParameters::LocalThreadSizeR;
430     // In this step, we force the code not to be more than 2-step reduction:
431     // Our empirical research shows that if each thread reduces at least 64
432     // elemnts individually, we get better performance. However, this can change
433     // on different platforms. In this step we force the code not to be
434     // morthan step reduction: Our empirical research shows that for inner_most
435     // dim reducer, it is better to have 8 group in a reduce dimension for sizes
436     // > 1024 to achieve the best performance.
437     const Index reductionPerThread = 64;
438     Index cu = dev.getPowerOfTwo(dev.getNumSyclMultiProcessors(), true);
439     const Index pNumGroups = roundUpP / PannelParameters::LocalThreadSizeP;
440     Index rGroups = (cu + pNumGroups - 1) / pNumGroups;
441     const Index rNumGroups = num_coeffs_to_reduce > reductionPerThread * localRange ? std::min(rGroups, localRange) : 1;
442     const Index globalRange = pNumGroups * rNumGroups * localRange;
443 
444     EIGEN_CONSTEXPR Index scratchSize =
445         PannelParameters::LocalThreadSizeR * (PannelParameters::LocalThreadSizeP + PannelParameters::BC);
446     auto thread_range = cl::sycl::nd_range<1>(cl::sycl::range<1>(globalRange), cl::sycl::range<1>(localRange));
447     if (rNumGroups > 1) {
448       CoeffReturnType *temp_pointer = static_cast<CoeffReturnType *>(
449           dev.allocate_temp(num_coeffs_to_preserve * rNumGroups * sizeof(CoeffReturnType)));
450       EvaluatorPointerType temp_accessor = dev.get(temp_pointer);
451       dev.template unary_kernel_launcher<CoeffReturnType, SyclReducerKerneType>(
452           self, temp_accessor, thread_range, scratchSize, reducer, pNumGroups, rNumGroups, num_coeffs_to_preserve,
453           num_coeffs_to_reduce);
454 
455       typedef SecondStepPartialReduction<CoeffReturnType, Index, EvaluatorPointerType, EvaluatorPointerType, Op>
456           SecondStepPartialReductionKernel;
457 
458       dev.template unary_kernel_launcher<CoeffReturnType, SecondStepPartialReductionKernel>(
459           temp_accessor, output,
460           cl::sycl::nd_range<1>(cl::sycl::range<1>(pNumGroups * localRange), cl::sycl::range<1>(localRange)), Index(1),
461           reducer, num_coeffs_to_preserve, rNumGroups);
462 
463       self.device().deallocate_temp(temp_pointer);
464     } else {
465       dev.template unary_kernel_launcher<CoeffReturnType, SyclReducerKerneType>(
466           self, output, thread_range, scratchSize, reducer, pNumGroups, rNumGroups, num_coeffs_to_preserve,
467           num_coeffs_to_reduce);
468     }
469     return false;
470   }
471 };
472 }  // namespace internal
473 }  // namespace TensorSycl
474 
475 namespace internal {
476 
477 template <typename Self, typename Op, bool Vectorizable>
478 struct FullReducer<Self, Op, Eigen::SyclDevice, Vectorizable> {
479   typedef typename Self::CoeffReturnType CoeffReturnType;
480   typedef typename Self::EvaluatorPointerType EvaluatorPointerType;
481   static EIGEN_CONSTEXPR bool HasOptimizedImplementation = true;
482   static EIGEN_CONSTEXPR int PacketSize = Self::PacketAccess ? Self::PacketSize : 1;
483   static void run(const Self &self, Op &reducer, const Eigen::SyclDevice &dev, EvaluatorPointerType data) {
484     typedef typename conditional<Self::PacketAccess, typename Self::PacketReturnType, CoeffReturnType>::type OutType;
485     static_assert(!((EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1) &
486                     (EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1 - 1)),
487                   "The Local thread size must be a power of 2 for the reduction "
488                   "operation");
489     EIGEN_CONSTEXPR Index local_range = EIGEN_SYCL_LOCAL_THREAD_DIM0 * EIGEN_SYCL_LOCAL_THREAD_DIM1;
490 
491     typename Self::Index inputSize = self.impl().dimensions().TotalSize();
492     // In this step we force the code not to be more than 2-step reduction:
493     // Our empirical research shows that if each thread reduces at least 512
494     // elemnts individually, we get better performance.
495     const Index reductionPerThread = 2048;
496     // const Index num_work_group =
497     Index reductionGroup = dev.getPowerOfTwo(
498         (inputSize + (reductionPerThread * local_range - 1)) / (reductionPerThread * local_range), true);
499     const Index num_work_group = std::min(reductionGroup, local_range);
500     // 1
501     // ? local_range
502     // : 1);
503     const Index global_range = num_work_group * local_range;
504 
505     auto thread_range = cl::sycl::nd_range<1>(cl::sycl::range<1>(global_range), cl::sycl::range<1>(local_range));
506     typedef TensorSycl::internal::FullReductionKernelFunctor<Self, Op, local_range> reduction_kernel_t;
507     if (num_work_group > 1) {
508       CoeffReturnType *temp_pointer =
509           static_cast<CoeffReturnType *>(dev.allocate_temp(num_work_group * sizeof(CoeffReturnType)));
510       typename Self::EvaluatorPointerType tmp_global_accessor = dev.get(temp_pointer);
511       dev.template unary_kernel_launcher<OutType, reduction_kernel_t>(self, tmp_global_accessor, thread_range,
512                                                                       local_range, inputSize, reducer);
513 
514       typedef TensorSycl::internal::SecondStepFullReducer<CoeffReturnType, Op, EvaluatorPointerType,
515                                                           EvaluatorPointerType, Index, local_range>
516           GenericRKernel;
517       dev.template unary_kernel_launcher<CoeffReturnType, GenericRKernel>(
518           tmp_global_accessor, data,
519           cl::sycl::nd_range<1>(cl::sycl::range<1>(num_work_group), cl::sycl::range<1>(num_work_group)), num_work_group,
520           reducer);
521 
522       dev.deallocate_temp(temp_pointer);
523     } else {
524       dev.template unary_kernel_launcher<OutType, reduction_kernel_t>(self, data, thread_range, local_range, inputSize,
525                                                                       reducer);
526     }
527   }
528 };
529 // vectorizable inner_most most dim preserver
530 // col reduction
531 template <typename Self, typename Op>
532 struct OuterReducer<Self, Op, Eigen::SyclDevice> {
533   static EIGEN_CONSTEXPR bool HasOptimizedImplementation = true;
534 
535   static bool run(const Self &self, const Op &reducer, const Eigen::SyclDevice &dev,
536                   typename Self::EvaluatorPointerType output, typename Self::Index num_coeffs_to_reduce,
537                   typename Self::Index num_coeffs_to_preserve) {
538     return ::Eigen::TensorSycl::internal::PartialReducerLauncher<
539         Self, Op, ::Eigen::TensorSycl::internal::reduction_dim::outer_most>::run(self, reducer, dev, output,
540                                                                                  num_coeffs_to_reduce,
541                                                                                  num_coeffs_to_preserve);
542   }
543 };
544 // row reduction
545 template <typename Self, typename Op>
546 struct InnerReducer<Self, Op, Eigen::SyclDevice> {
547   static EIGEN_CONSTEXPR bool HasOptimizedImplementation = true;
548 
549   static bool run(const Self &self, const Op &reducer, const Eigen::SyclDevice &dev,
550                   typename Self::EvaluatorPointerType output, typename Self::Index num_coeffs_to_reduce,
551                   typename Self::Index num_coeffs_to_preserve) {
552     return ::Eigen::TensorSycl::internal::PartialReducerLauncher<
553         Self, Op, ::Eigen::TensorSycl::internal::reduction_dim::inner_most>::run(self, reducer, dev, output,
554                                                                                  num_coeffs_to_reduce,
555                                                                                  num_coeffs_to_preserve);
556   }
557 };
558 
559 // ArmgMax uses this kernel for partial reduction//
560 // TODO(@mehdi.goli) come up with a better kernel
561 // generic partial reduction
562 template <typename Self, typename Op>
563 struct GenericReducer<Self, Op, Eigen::SyclDevice> {
564   static EIGEN_CONSTEXPR bool HasOptimizedImplementation = false;
565   static bool run(const Self &self, const Op &reducer, const Eigen::SyclDevice &dev,
566                   typename Self::EvaluatorPointerType output, typename Self::Index num_values_to_reduce,
567                   typename Self::Index num_coeffs_to_preserve) {
568     typename Self::Index range, GRange, tileSize;
569     dev.parallel_for_setup(num_coeffs_to_preserve, tileSize, range, GRange);
570 
571     dev.template unary_kernel_launcher<typename Self::CoeffReturnType,
572                                        TensorSycl::internal::GenericNondeterministicReducer<Self, Op>>(
573         self, output, cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), Index(1),
574         reducer, range, (num_values_to_reduce != 0) ? num_values_to_reduce : static_cast<Index>(1));
575     return false;
576   }
577 };
578 
579 }  // namespace internal
580 }  // namespace Eigen
581 
582 #endif  // UNSUPPORTED_EIGEN_CXX11_SRC_TENSOR_TENSOR_REDUCTION_SYCL_HPP
583