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/CL/functions/CLReduceMean.h"
25
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/Types.h"
29 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
30 #include "src/core/CL/CLValidate.h"
31 #include "src/core/CL/kernels/CLFillBorderKernel.h"
32 #include "src/core/CL/kernels/CLReductionOperationKernel.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34
35 namespace arm_compute
36 {
37 namespace
38 {
validate_config(const ITensorInfo * input,const Coordinates & reduction_axis,bool keep_dims,const ITensorInfo * output)39 Status validate_config(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
40 {
41 ARM_COMPUTE_UNUSED(keep_dims);
42 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
43 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
44 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
45 ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() < 1);
46 ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() > input->num_dimensions());
47
48 const unsigned int reduction_ops = reduction_axis.num_dimensions();
49 const int input_dims = input->num_dimensions();
50 Coordinates axis_local = reduction_axis;
51
52 for(unsigned int i = 0; i < axis_local.num_dimensions(); ++i)
53 {
54 //axis: The dimensions to reduce. Must be in the range [-rank(input_tensor), rank(input_tensor)).
55 ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] < (-static_cast<int>(input->num_dimensions())));
56 ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] >= static_cast<int>(input->num_dimensions()));
57 }
58
59 if(output->tensor_shape().total_size() != 0)
60 {
61 // Only validate if not using auto_init for the output tensor
62 TensorShape out_shape = input->tensor_shape();
63 // Validate output_shape only if not using auto_init
64 convert_negative_axis(axis_local, input_dims);
65 std::sort(axis_local.begin(), axis_local.begin() + reduction_ops);
66 for(unsigned int i = 0; i < reduction_ops; ++i)
67 {
68 ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] > 3);
69 ARM_COMPUTE_RETURN_ERROR_ON(static_cast<unsigned int>(axis_local[i]) > input->num_dimensions() - 1);
70 if(output->total_size() > 0 && keep_dims)
71 {
72 ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(axis_local[i]) != 1);
73 }
74 if(keep_dims)
75 {
76 out_shape.set(axis_local[i], 1);
77 }
78 else
79 {
80 ARM_COMPUTE_RETURN_ERROR_ON(i > static_cast<unsigned int>(axis_local[i]));
81 const unsigned int remove_index = axis_local[i] - i;
82 ARM_COMPUTE_RETURN_ERROR_ON(remove_index >= out_shape.num_dimensions());
83 out_shape.remove_dimension(remove_index);
84 }
85 }
86 const TensorInfo out_info = input->clone()->set_tensor_shape(out_shape);
87 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info);
88 const bool requant = is_data_type_quantized(input->data_type()) && input->quantization_info() != output->quantization_info();
89 if(requant)
90 {
91 TensorInfo input_no_quant(input->clone()->set_data_type(DataType::F32));
92 CLDequantizationLayer::validate(input, &input_no_quant);
93 TensorInfo output_no_quant(output->clone()->set_data_type(DataType::F32));
94 CLQuantizationLayer::validate(&output_no_quant, output);
95 }
96 }
97 return Status{};
98 }
99 }
100
CLReduceMean(std::shared_ptr<IMemoryManager> memory_manager)101 CLReduceMean::CLReduceMean(std::shared_ptr<IMemoryManager> memory_manager)
102 : _memory_group(std::move(memory_manager)), _reduction_kernels(), _reduced_outs(), _reshape(), _dequant(), _requant(), _reduction_ops(), _keep_dims(), _do_requant(), _input_no_quant(),
103 _output_no_quant()
104 {
105 }
106
configure(ICLTensor * input,const Coordinates & reduction_axis,bool keep_dims,ICLTensor * output)107 void CLReduceMean::configure(ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output)
108 {
109 configure(CLKernelLibrary::get().get_compile_context(), input, reduction_axis, keep_dims, output);
110 }
111
configure(const CLCompileContext & compile_context,ICLTensor * input,const Coordinates & reduction_axis,bool keep_dims,ICLTensor * output)112 void CLReduceMean::configure(const CLCompileContext &compile_context, ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output)
113 {
114 // Perform validate step
115 ARM_COMPUTE_ERROR_THROW_ON(CLReduceMean::validate(input->info(), reduction_axis, keep_dims, output->info()));
116 // Output auto inizialitation if not yet initialized
117 const TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_reduce_mean_shape(input->info(), reduction_axis, keep_dims);
118 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
119
120 _do_requant = is_data_type_quantized(input->info()->data_type()) && input->info()->quantization_info() != output->info()->quantization_info();
121 _reduction_ops = reduction_axis.num_dimensions();
122 _reduction_kernels.resize(_reduction_ops);
123 _reduced_outs.resize(_reduction_ops - (keep_dims ? 1 : 0));
124 _keep_dims = keep_dims;
125
126 ICLTensor *tmp_input = input;
127 ICLTensor *tmp_output = output;
128 if(_do_requant)
129 {
130 _memory_group.manage(&_input_no_quant);
131 _memory_group.manage(&_output_no_quant);
132 TensorInfo output_no_quant_info = input->info()->clone()->set_tensor_shape(output_shape);
133 output_no_quant_info.set_data_type(DataType::F32);
134 auto_init_if_empty(*_output_no_quant.info(), output_no_quant_info);
135 auto_init_if_empty(*_input_no_quant.info(), input->info()->clone()->set_data_type(DataType::F32));
136 _dequant.configure(compile_context, input, &_input_no_quant);
137 tmp_input = &_input_no_quant;
138 tmp_output = &_output_no_quant;
139 }
140
141 Coordinates axis_local = reduction_axis;
142 const int input_dims = tmp_input->info()->num_dimensions();
143
144 convert_negative_axis(axis_local, input_dims);
145
146 // Perform reduction for every axis
147 for(int i = 0; i < _reduction_ops; ++i)
148 {
149 TensorShape out_shape = i == 0 ? tmp_input->info()->tensor_shape() : (&_reduced_outs[i - 1])->info()->tensor_shape();
150 out_shape.set(axis_local[i], 1);
151 auto in = (i == 0) ? tmp_input : (&_reduced_outs[i - 1]);
152
153 if(i == _reduction_ops - 1 && keep_dims)
154 {
155 _reduction_kernels[i].configure(compile_context, in, tmp_output, axis_local[i], ReductionOperation::MEAN_SUM);
156 }
157 else
158 {
159 _reduced_outs[i].allocator()->init(TensorInfo(out_shape, tmp_input->info()->num_channels(), tmp_input->info()->data_type(), tmp_input->info()->quantization_info()));
160 _memory_group.manage(&_reduced_outs[i]);
161 _reduction_kernels[i].configure(compile_context, in, &_reduced_outs[i], axis_local[i], ReductionOperation::MEAN_SUM);
162 }
163 }
164
165 // Allocate intermediate tensors
166 for(int i = 0; i < _reduction_ops - (keep_dims ? 1 : 0); ++i)
167 {
168 _reduced_outs[i].allocator()->allocate();
169 }
170
171 // Configure reshape layer if we want to drop the dimensions
172 if(!_keep_dims)
173 {
174 TensorShape out_shape = tmp_input->info()->tensor_shape();
175
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(compile_context, &_reduced_outs[_reduction_ops - 1], tmp_output);
185 }
186 if(_do_requant)
187 {
188 _requant.configure(compile_context, &_output_no_quant, output);
189 _input_no_quant.allocator()->allocate();
190 _output_no_quant.allocator()->allocate();
191 }
192 }
193
validate(const ITensorInfo * input,const Coordinates & reduction_axis,bool keep_dims,const ITensorInfo * output)194 Status CLReduceMean::validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
195 {
196 return validate_config(input, reduction_axis, keep_dims, output);
197 }
198
run()199 void CLReduceMean::run()
200 {
201 MemoryGroupResourceScope scope_mg(_memory_group);
202
203 if(_do_requant)
204 {
205 _dequant.run();
206 }
207 for(auto &kernel : _reduction_kernels)
208 {
209 kernel.run();
210 }
211 if(!_keep_dims)
212 {
213 _reshape.run();
214 }
215 if(_do_requant)
216 {
217 _requant.run();
218 }
219 }
220 } // namespace arm_compute
221