xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/NEON/functions/NEReductionOperation.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2021 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/NEReductionOperation.h"
25 
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
28 #include "arm_compute/runtime/NEON/NEScheduler.h"
29 #include "src/common/utils/Log.h"
30 #include "src/core/NEON/kernels/NEReductionOperationKernel.h"
31 #include "src/core/helpers/AutoConfiguration.h"
32 
33 namespace arm_compute
34 {
35 namespace
36 {
37 /** Define dimension to split the window
38  *
39  * @param[in] axis Reduction axis
40  *
41  * @return The dimension to split the window
42  */
reduction_window_split_dimension(unsigned int axis)43 size_t reduction_window_split_dimension(unsigned int axis)
44 {
45     switch(axis)
46     {
47         case 0:
48             return Window::DimY;
49         case 1:
50         case 2:
51         case 3:
52             return Window::DimX;
53         default:
54             ARM_COMPUTE_ERROR("Unsupported reduction axis");
55     }
56 }
57 } // namespace
58 
59 NEReductionOperation::~NEReductionOperation() = default;
60 
NEReductionOperation(std::shared_ptr<IMemoryManager> memory_manager)61 NEReductionOperation::NEReductionOperation(std::shared_ptr<IMemoryManager> memory_manager)
62     : _memory_group(memory_manager), _reduction_kernel(), _reshape(), _output_internal(), _window_split(0), _reduction_axis(), _is_reshape_required(false)
63 {
64 }
65 
validate(const ITensorInfo * input,const ITensorInfo * output,unsigned int axis,ReductionOperation op,bool keep_dims)66 Status NEReductionOperation::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op, bool keep_dims)
67 {
68     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
69     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
70 
71     const auto is_reshape_required = !keep_dims;
72 
73     auto *output_internal = output;
74 
75     TensorInfo info_before_reshape;
76 
77     if(is_reshape_required)
78     {
79         const TensorInfo expected_output_shape = output->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis, keep_dims));
80         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output_shape, output);
81 
82         auto shape_before_reshape = input->tensor_shape();
83         shape_before_reshape.set(axis, 1);
84 
85         const auto input_num_channles = input->num_channels();
86         const auto input_qinfo        = input->quantization_info();
87         const auto is_arg_min_max     = (op == ReductionOperation::ARG_IDX_MAX) || (op == ReductionOperation::ARG_IDX_MIN);
88         const auto output_data_type   = is_arg_min_max ? DataType::S32 : output->data_type();
89 
90         info_before_reshape.set_data_type(output_data_type).set_tensor_shape(shape_before_reshape).set_num_channels(input_num_channles).set_quantization_info(input_qinfo);
91 
92         output_internal = &info_before_reshape;
93     }
94 
95     ARM_COMPUTE_RETURN_ON_ERROR(NEReductionOperationKernel::validate(input, output_internal, axis, op));
96 
97     if(is_reshape_required)
98     {
99         ARM_COMPUTE_RETURN_ON_ERROR(NEReshapeLayer::validate(output_internal, output));
100     }
101 
102     return Status{};
103 }
104 
configure(ITensor * input,ITensor * output,unsigned int axis,ReductionOperation op,bool keep_dims)105 void NEReductionOperation::configure(ITensor *input, ITensor *output, unsigned int axis, ReductionOperation op, bool keep_dims)
106 {
107     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
108     ARM_COMPUTE_LOG_PARAMS(input, output, axis, op, keep_dims);
109 
110     _is_reshape_required = !keep_dims;
111 
112     auto      *output_internal = output;
113     const auto is_arg_min_max  = (op == ReductionOperation::ARG_IDX_MAX) || (op == ReductionOperation::ARG_IDX_MIN);
114 
115     if(_is_reshape_required)
116     {
117         const auto output_internal_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis);
118         const auto output_external_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, false);
119         const auto output_data_type      = is_arg_min_max ? DataType::S32 : input->info()->data_type();
120         const auto num_channels          = input->info()->num_channels();
121         const auto qinfo                 = input->info()->quantization_info();
122 
123         _output_internal.allocator()->init(input->info()->clone()->set_data_type(output_data_type).set_tensor_shape(output_internal_shape).reset_padding().set_is_resizable(true).set_num_channels(
124                                                num_channels).set_quantization_info(qinfo));
125         _memory_group.manage(&_output_internal);
126         output_internal = &_output_internal;
127         auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(output_data_type).set_tensor_shape(output_external_shape).reset_padding().set_is_resizable(true));
128     }
129 
130     ARM_COMPUTE_ERROR_THROW_ON(NEReductionOperation::validate(input->info(), output->info(), axis, op, keep_dims));
131 
132     // Configure reduction kernel
133     _reduction_kernel = std::make_unique<NEReductionOperationKernel>();
134     _reduction_kernel->configure(input, output_internal, axis, op);
135     _window_split   = reduction_window_split_dimension(axis);
136     _reduction_axis = axis;
137 
138     if(_is_reshape_required)
139     {
140         _reshape.configure(output_internal, output);
141         _output_internal.allocator()->allocate();
142     }
143 }
144 
run()145 void NEReductionOperation::run()
146 {
147     MemoryGroupResourceScope scope_mg(_memory_group);
148     NEScheduler::get().schedule(_reduction_kernel.get(), _window_split);
149     if(_is_reshape_required)
150     {
151         _reshape.run();
152     }
153 }
154 } // namespace arm_compute
155