2 * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 * Copyright (c) 2018-2020 ARM Limited.
20 * SPDX-License-Identifier: MIT
22 * Permission is hereby granted, free of charge, to any person obtaining a copy
23 * of this software and associated documentation files (the "Software"), to
24 * deal in the Software without restriction, including without limitation the
25 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
26 * sell copies of the Software, and to permit persons to whom the Software is
27 * furnished to do so, subject to the following conditions:
29 * The above copyright notice and this permission notice shall be included in all
30 * copies or substantial portions of the Software.
32 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
33 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
34 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
35 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
36 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
37 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
41 #include "arm_compute/runtime/CL/functions/CLArgMinMaxLayerEx.h"
43 #include "arm_compute/core/Error.h"
44 #include "arm_compute/core/TensorInfo.h"
45 #include "arm_compute/core/Types.h"
46 #include "arm_compute/core/Validate.h"
47 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
48 #include "arm_compute/runtime/Utils.h"
52 CLArgMinMaxLayerEx::CLArgMinMaxLayerEx(std::shared_ptr<IMemoryManager> memory_manager)
53 : _memory_group(std::move(memory_manager)), _results_vector(), _not_reshaped_output(),
54 _reduction_kernels_vector(), _reshape_kernel(), _num_of_stages(), _reduction_axis()
58 Status CLArgMinMaxLayerEx::validate(const ITensorInfo *input, int axis, const ITensorInfo *output,
59 const ReductionOperation &op)
61 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
62 ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX &&
63 op != ReductionOperation::ARG_IDX_MIN,
64 "Invalid reduction operation");
65 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= static_cast<int>(TensorShape::num_max_dimensions),
66 "Reduction axis greater than max number of dimensions");
67 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
68 const unsigned int num_of_stages =
69 calculate_number_of_stages_only_x_axis(input->dimension(0), axis);
71 DataType output_data_type = DataType::S32;
72 TensorInfo not_reshaped_output;
73 const auto input_num_channles = input->num_channels();
74 const auto input_qinfo = input->quantization_info();
76 if (output->total_size() != 0)
78 output_data_type = output->data_type();
79 const TensorInfo expected_output_shape = output->clone()->set_tensor_shape(
80 arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis,
82 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output_shape, output);
85 auto shape_before_reshape = input->tensor_shape();
86 shape_before_reshape.set(axis, 1);
87 auto initialize_tensorinfo = [](TensorInfo &ti, TensorShape shape, DataType data_type,
88 int num_channels, QuantizationInfo qinfo) {
89 ti.set_data_type(data_type)
90 .set_tensor_shape(shape)
91 .set_num_channels(num_channels)
92 .set_quantization_info(qinfo);
95 initialize_tensorinfo(not_reshaped_output, shape_before_reshape, output_data_type,
96 input_num_channles, input_qinfo);
98 if (num_of_stages == 1)
100 ARM_COMPUTE_RETURN_ON_ERROR(
101 CLArgMinMaxLayerKernelEx::validate(input, nullptr, ¬_reshaped_output, axis, op));
105 // Create temporary tensor infos
106 std::vector<TensorInfo> sums_vector(num_of_stages - 1);
108 // Create intermediate tensor info
109 TensorShape shape{input->tensor_shape()};
111 for (unsigned int i = 0; i < num_of_stages - 1; i++)
113 shape.set(0, ceil(shape.x() / 128.f));
114 sums_vector[i].set_data_type(input->data_type());
115 sums_vector[i].set_tensor_shape(shape);
116 sums_vector[i].set_num_channels(input->num_channels());
119 // Validate ReductionOperation only on first kernel
120 ARM_COMPUTE_RETURN_ON_ERROR(
121 CLArgMinMaxLayerKernelEx::validate(input, nullptr, &sums_vector[0], axis, op));
123 // Validate ReductionOperation on intermediate stages
124 for (unsigned int i = 1; i < num_of_stages - 1; ++i)
126 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernelEx::validate(input, &sums_vector[i - 1],
127 &sums_vector[i], axis, op));
130 // Validate ReductionOperation on the last stage
131 const unsigned int last_stage = num_of_stages - 1;
132 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernelEx::validate(
133 input, &sums_vector[last_stage - 1], ¬_reshaped_output, axis, op));
135 ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(¬_reshaped_output, output));
139 void CLArgMinMaxLayerEx::configure(const ICLTensor *input, int axis, ICLTensor *output,
140 const ReductionOperation &op)
142 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
143 _num_of_stages = calculate_number_of_stages_only_x_axis(input->info()->dimension(0), axis);
144 _reduction_axis = axis;
146 const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(
147 input->info()->tensor_shape(), axis, false);
148 DataType output_data_type = (output->info()->data_type() == DataType::UNKNOWN)
150 : output->info()->data_type();
151 auto_init_if_empty(*output->info(), input->info()
153 ->set_tensor_shape(output_shape)
154 .set_data_type(output_data_type)
156 .set_is_resizable(true));
158 // Configure reduction operation kernels
159 _reduction_kernels_vector.resize(_num_of_stages);
161 _memory_group.manage(&_not_reshaped_output);
162 // Create temporary tensors
163 if (_num_of_stages == 1)
165 // Force an early initialization for int64 output type
166 TensorShape output_shape{input->info()->tensor_shape()};
167 output_shape.set(axis, 1);
168 auto_init_if_empty(*_not_reshaped_output.info(), input->info()
170 ->set_tensor_shape(output_shape)
171 .set_data_type(output_data_type)
173 .set_is_resizable(true));
174 _not_reshaped_output.info()->set_tensor_shape(output_shape);
175 _reduction_kernels_vector[0].configure(input, nullptr, &_not_reshaped_output, axis, op);
179 _results_vector.resize(_num_of_stages - 1);
180 TensorShape shape{input->info()->tensor_shape()};
181 for (unsigned int i = 0; i < _num_of_stages - 1; i++)
183 shape.set(0, ceil(shape.x() / 128.f));
184 _results_vector[i].allocator()->init(
185 input->info()->clone()->set_tensor_shape(shape).set_data_type(output_data_type));
188 // Apply ReductionOperation only on first kernel
189 _memory_group.manage(&_results_vector[0]);
190 _reduction_kernels_vector[0].configure(input, nullptr, &_results_vector[0], axis, op);
192 // Apply ReductionOperation on intermediate stages
193 for (unsigned int i = 1; i < _num_of_stages - 1; ++i)
195 _memory_group.manage(&_results_vector[i]);
196 _reduction_kernels_vector[i].configure(input, &_results_vector[i - 1], &_results_vector[i],
198 _results_vector[i - 1].allocator()->allocate();
201 // Apply ReductionOperation on the last stage
202 const unsigned int last_stage = _num_of_stages - 1;
203 _reduction_kernels_vector[last_stage].configure(input, &_results_vector[last_stage - 1],
204 &_not_reshaped_output, axis, op);
205 _results_vector[last_stage - 1].allocator()->allocate();
207 _reshape_kernel.configure(&_not_reshaped_output, output);
208 _not_reshaped_output.allocator()->allocate();
211 void CLArgMinMaxLayerEx::run()
213 MemoryGroupResourceScope scope_mg(_memory_group);
215 for (unsigned int i = 0; i < _num_of_stages; ++i)
217 CLScheduler::get().enqueue(_reduction_kernels_vector[i], false);
219 CLScheduler::get().enqueue(_reshape_kernel, false);
221 } // namespace arm_compute