xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #include "arm_compute/runtime/NEON/functions/NEWinogradConvolutionLayer.h"
25 
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/ITensorPack.h"
28 #include "arm_compute/core/Utils.h"
29 #include "arm_compute/core/Validate.h"
30 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
31 #include "src/core/CPP/Validate.h"
32 #include "src/core/helpers/MemoryHelpers.h"
33 #include "src/cpu/kernels/CpuWinogradConv2dKernel.h"
34 #include "src/cpu/operators/CpuWinogradConv2d.h"
35 
36 #include "src/core/NEON/kernels/convolution/common/utils.hpp"
37 
38 namespace arm_compute
39 {
40 using namespace arm_compute::experimental;
41 
42 struct NEWinogradConvolutionLayer::Impl
43 {
44     MemoryGroup                             memory_group{};
45     std::unique_ptr<cpu::CpuWinogradConv2d> op{ nullptr };
46     ITensorPack                             run_pack{};
47     ITensorPack                             prep_pack{};
48     WorkspaceData<Tensor>                   workspace{};
49     experimental::MemoryRequirements        aux_mem_req{};
50     const ITensor                          *original_weights{ nullptr };
51     bool                                    is_prepared{ false };
52     bool                                    is_activationlayer_enabled{ false };
53     DataLayout                              data_layout{};
54 };
55 
NEWinogradConvolutionLayer(const std::shared_ptr<IMemoryManager> & memory_manager)56 NEWinogradConvolutionLayer::NEWinogradConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager)
57     : _impl(std::make_unique<Impl>())
58 {
59     _impl->memory_group = MemoryGroup(std::move(memory_manager));
60 }
61 
62 NEWinogradConvolutionLayer::~NEWinogradConvolutionLayer() = default;
63 
configure(const ITensor * input,const ITensor * weights,const ITensor * biases,ITensor * output,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info,bool enable_fast_math)64 void NEWinogradConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info,
65                                            bool enable_fast_math)
66 {
67     _impl->original_weights = weights;
68     _impl->op               = std::make_unique<cpu::CpuWinogradConv2d>();
69     _impl->op->configure(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info, act_info, enable_fast_math);
70 
71     _impl->aux_mem_req = _impl->op->workspace();
72     _impl->run_pack    = { { ACL_SRC_0, input }, { ACL_SRC_1, weights }, { ACL_SRC_2, biases }, { ACL_DST, output } };
73     _impl->prep_pack   = { { ACL_SRC_1, weights }, { ACL_SRC_2, biases } };
74     _impl->workspace   = manage_workspace<Tensor>(_impl->aux_mem_req, _impl->memory_group, _impl->run_pack, _impl->prep_pack);
75 }
76 
run()77 void NEWinogradConvolutionLayer::run()
78 {
79     prepare();
80 
81     MemoryGroupResourceScope scope_mg(_impl->memory_group);
82     _impl->op->run(_impl->run_pack);
83 }
84 
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * output,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info,bool enable_fast_math)85 Status NEWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
86                                             const ActivationLayerInfo &act_info, bool enable_fast_math)
87 {
88     return cpu::CpuWinogradConv2d::validate(input, weights, biases, output, conv_info, act_info, enable_fast_math);
89 }
90 
prepare()91 void NEWinogradConvolutionLayer::prepare()
92 {
93     if(!_impl->is_prepared)
94     {
95         _impl->op->prepare(_impl->prep_pack);
96         _impl->original_weights->mark_as_unused();
97 
98         // Release temporary tensors that are only used in prepare stage
99         release_temporaries<Tensor>(_impl->aux_mem_req, _impl->workspace);
100 
101         _impl->is_prepared = true;
102     }
103 }
104 } // namespace arm_compute
105