• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2018-2020 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/NEReduceMean.h"
25 
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
29 #include "arm_compute/runtime/NEON/NEScheduler.h"
30 #include "src/core/CPP/Validate.h"
31 #include "src/core/NEON/kernels/NEReductionOperationKernel.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 
34 namespace arm_compute
35 {
36 namespace
37 {
validate_config(const ITensorInfo * input,const Coordinates & reduction_axis,bool keep_dims,const ITensorInfo * output)38 Status validate_config(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
39 {
40     ARM_COMPUTE_UNUSED(keep_dims);
41     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
42     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
43     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8_SIGNED, DataType::QASYMM8, DataType::F16, DataType::F32);
44     ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() < 1);
45     ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() > input->num_dimensions());
46 
47     const unsigned int reduction_ops = reduction_axis.num_dimensions();
48     const int          input_dims    = input->num_dimensions();
49     Coordinates        axis_local    = reduction_axis;
50 
51     for(unsigned int i = 0; i < axis_local.num_dimensions(); ++i)
52     {
53         //axis: The dimensions to reduce. Must be in the range [-rank(input_tensor), rank(input_tensor)).
54         ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] < (-static_cast<int>(input->num_dimensions())));
55         ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] >= static_cast<int>(input->num_dimensions()));
56     }
57 
58     if(output->tensor_shape().total_size() != 0)
59     {
60         // Only validate if not using auto_init for the output tensor
61         TensorShape out_shape = input->tensor_shape();
62         // Validate output_shape only if not using auto_init
63         convert_negative_axis(axis_local, input_dims);
64         std::sort(axis_local.begin(), axis_local.begin() + reduction_ops);
65         for(unsigned int i = 0; i < reduction_ops; ++i)
66         {
67             ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] > 3);
68             ARM_COMPUTE_RETURN_ERROR_ON(static_cast<unsigned int>(axis_local[i]) > input->num_dimensions() - 1);
69             if(output->total_size() > 0 && keep_dims)
70             {
71                 ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(axis_local[i]) != 1);
72             }
73             if(keep_dims)
74             {
75                 out_shape.set(axis_local[i], 1);
76             }
77             else
78             {
79                 ARM_COMPUTE_RETURN_ERROR_ON(i > static_cast<unsigned int>(axis_local[i]));
80                 const unsigned int remove_index = axis_local[i] - i;
81                 ARM_COMPUTE_RETURN_ERROR_ON(remove_index >= out_shape.num_dimensions());
82                 out_shape.remove_dimension(remove_index);
83             }
84         }
85         const TensorInfo out_info = input->clone()->set_tensor_shape(out_shape);
86         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info);
87         const bool requant = is_data_type_quantized(input->data_type()) && input->quantization_info() != output->quantization_info();
88         if(requant)
89         {
90             TensorInfo input_no_quant(input->clone()->set_data_type(DataType::F32));
91             NEDequantizationLayer::validate(input, &input_no_quant);
92             TensorInfo output_no_quant(output->clone()->set_data_type(DataType::F32));
93             NEQuantizationLayer::validate(&output_no_quant, output);
94         }
95     }
96     return Status{};
97 }
98 } // namespace
99 
100 NEReduceMean::~NEReduceMean() = default;
101 
NEReduceMean(std::shared_ptr<IMemoryManager> memory_manager)102 NEReduceMean::NEReduceMean(std::shared_ptr<IMemoryManager> memory_manager)
103     : _memory_group(std::move(memory_manager)), _reduction_kernels(), _reduced_outs(), _reshape(), _dequant(), _requant(), _reduction_ops(), _keep_dims(), _do_requant(), _input_no_quant(),
104       _output_no_quant()
105 {
106 }
107 
validate(const ITensorInfo * input,const Coordinates & reduction_axis,bool keep_dims,const ITensorInfo * output)108 Status NEReduceMean::validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
109 {
110     return validate_config(input, reduction_axis, keep_dims, output);
111 }
112 
configure(ITensor * input,const Coordinates & reduction_axis,bool keep_dims,ITensor * output)113 void NEReduceMean::configure(ITensor *input, const Coordinates &reduction_axis, bool keep_dims, ITensor *output)
114 {
115     // Perform validate step
116     ARM_COMPUTE_ERROR_THROW_ON(NEReduceMean::validate(input->info(), reduction_axis, keep_dims, output->info()));
117     // Output auto inizialitation if not yet initialized
118     const TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_reduce_mean_shape(input->info(), reduction_axis, keep_dims);
119     auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
120 
121     _do_requant    = is_data_type_quantized(input->info()->data_type()) && input->info()->quantization_info() != output->info()->quantization_info();
122     _reduction_ops = reduction_axis.num_dimensions();
123     _reduction_kernels.resize(_reduction_ops);
124     _reduced_outs.resize(_reduction_ops - (keep_dims ? 1 : 0));
125     _keep_dims = keep_dims;
126 
127     ITensor *tmp_input  = input;
128     ITensor *tmp_output = output;
129     if(_do_requant)
130     {
131         _memory_group.manage(&_input_no_quant);
132         _memory_group.manage(&_output_no_quant);
133         TensorInfo output_no_quant_info = input->info()->clone()->set_tensor_shape(output_shape);
134         output_no_quant_info.set_data_type(DataType::F32);
135         auto_init_if_empty(*_output_no_quant.info(), output_no_quant_info);
136         auto_init_if_empty(*_input_no_quant.info(), input->info()->clone()->set_data_type(DataType::F32));
137         _dequant.configure(input, &_input_no_quant);
138         tmp_input  = &_input_no_quant;
139         tmp_output = &_output_no_quant;
140     }
141 
142     Coordinates axis_local = reduction_axis;
143     const int   input_dims = tmp_input->info()->num_dimensions();
144 
145     convert_negative_axis(axis_local, input_dims);
146 
147     // Perform reduction for every axis
148     for(int i = 0; i < _reduction_ops; ++i)
149     {
150         TensorShape out_shape = i == 0 ? tmp_input->info()->tensor_shape() : (&_reduced_outs[i - 1])->info()->tensor_shape();
151         out_shape.set(axis_local[i], 1);
152         auto in = (i == 0) ? tmp_input : (&_reduced_outs[i - 1]);
153 
154         if(i == _reduction_ops - 1 && keep_dims)
155         {
156             _reduction_kernels[i].configure(in, tmp_output, axis_local[i], ReductionOperation::MEAN_SUM);
157         }
158         else
159         {
160             _reduced_outs[i].allocator()->init(TensorInfo(out_shape, tmp_input->info()->num_channels(), tmp_input->info()->data_type(), tmp_input->info()->quantization_info()));
161             _memory_group.manage(&_reduced_outs[i]);
162             _reduction_kernels[i].configure(in, &_reduced_outs[i], axis_local[i], ReductionOperation::MEAN_SUM);
163         }
164     }
165 
166     // Allocate intermediate tensors
167     for(int i = 0; i < _reduction_ops - (keep_dims ? 1 : 0); ++i)
168     {
169         _reduced_outs[i].allocator()->allocate();
170     }
171 
172     // Configure reshape layer if we want to drop the dimensions
173     if(!keep_dims)
174     {
175         TensorShape out_shape = tmp_input->info()->tensor_shape();
176         // We have to sort the reduction axis vectors in order for remove_dimension
177         // to work properly
178         std::sort(axis_local.begin(), axis_local.begin() + _reduction_ops);
179         for(int i = 0; i < _reduction_ops; ++i)
180         {
181             out_shape.remove_dimension(axis_local[i] - i);
182         }
183         auto_init_if_empty(*tmp_output->info(), tmp_input->info()->clone()->set_tensor_shape(out_shape));
184         _reshape.configure(&_reduced_outs[_reduction_ops - 1], tmp_output);
185     }
186     if(_do_requant)
187     {
188         _requant.configure(&_output_no_quant, output);
189         _input_no_quant.allocator()->allocate();
190         _output_no_quant.allocator()->allocate();
191     }
192 }
193 
run()194 void NEReduceMean::run()
195 {
196     MemoryGroupResourceScope scope_mg(_memory_group);
197     if(_do_requant)
198     {
199         _dequant.run();
200     }
201     for(auto &kernel : _reduction_kernels)
202     {
203         kernel.run();
204     }
205     if(!_keep_dims)
206     {
207         _reshape.run();
208     }
209     if(_do_requant)
210     {
211         _requant.run();
212     }
213 }
214 } // namespace arm_compute
215