1 /*
2 * Copyright (c) 2018-2021 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
25 #include "arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h"
26
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/core/Types.h"
30 #include "arm_compute/core/Validate.h"
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "src/core/CL/CLValidate.h"
33 #include "src/core/CL/kernels/CLArgMinMaxLayerKernel.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/runtime/Utils.h"
36
37 #include "src/common/utils/Log.h"
38
39 namespace arm_compute
40 {
CLArgMinMaxLayer(std::shared_ptr<IMemoryManager> memory_manager)41 CLArgMinMaxLayer::CLArgMinMaxLayer(std::shared_ptr<IMemoryManager> memory_manager)
42 : _memory_group(std::move(memory_manager)), _results_vector(), _not_reshaped_output(), _reduction_kernels_vector(), _reshape(), _num_of_stages(), _reduction_axis()
43 {
44 }
45
46 CLArgMinMaxLayer::~CLArgMinMaxLayer() = default;
47
validate(const ITensorInfo * input,int axis,const ITensorInfo * output,const ReductionOperation & op)48 Status CLArgMinMaxLayer::validate(const ITensorInfo *input, int axis, const ITensorInfo *output, const ReductionOperation &op)
49 {
50 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
51 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
52 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S32, DataType::F16, DataType::F32);
53 ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX && op != ReductionOperation::ARG_IDX_MIN, "Invalid reduction operation");
54 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= static_cast<int>(TensorShape::num_max_dimensions), "Reduction axis greater than max number of dimensions");
55 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
56 const unsigned int num_of_stages = utils::calculate_number_of_stages_only_x_axis(input->dimension(0), axis);
57
58 DataType output_data_type = DataType::S32;
59 TensorInfo not_reshaped_output;
60 const auto input_num_channles = input->num_channels();
61 const auto input_qinfo = input->quantization_info();
62
63 if(output->total_size() != 0)
64 {
65 output_data_type = output->data_type();
66 const TensorInfo expected_output_shape = output->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis, false));
67 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output_shape, output);
68 }
69
70 auto shape_before_reshape = input->tensor_shape();
71 shape_before_reshape.set(axis, 1);
72 auto initialize_tensorinfo = [](TensorInfo & ti, TensorShape shape, DataType data_type, int num_channels, QuantizationInfo qinfo)
73 {
74 ti.set_data_type(data_type).set_tensor_shape(shape).set_num_channels(num_channels).set_quantization_info(qinfo);
75 };
76
77 initialize_tensorinfo(not_reshaped_output, shape_before_reshape, output_data_type, input_num_channles, input_qinfo);
78
79 if(num_of_stages == 1)
80 {
81 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, ¬_reshaped_output, axis, op));
82 }
83 else
84 {
85 // Create temporary tensor infos
86 std::vector<TensorInfo> sums_vector(num_of_stages - 1);
87
88 // Create intermediate tensor info
89 TensorShape shape{ input->tensor_shape() };
90
91 for(unsigned int i = 0; i < num_of_stages - 1; i++)
92 {
93 shape.set(0, ceil(shape.x() / 128.f));
94 sums_vector[i].set_data_type(input->data_type());
95 sums_vector[i].set_tensor_shape(shape);
96 sums_vector[i].set_num_channels(input->num_channels());
97 }
98
99 // Validate ReductionOperation only on first kernel
100 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, &sums_vector[0], axis, op));
101
102 // Validate ReductionOperation on intermediate stages
103 for(unsigned int i = 1; i < num_of_stages - 1; ++i)
104 {
105 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[i - 1], &sums_vector[i], axis, op));
106 }
107
108 // Validate ReductionOperation on the last stage
109 const unsigned int last_stage = num_of_stages - 1;
110 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[last_stage - 1], ¬_reshaped_output, axis, op));
111 }
112 ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(¬_reshaped_output, output));
113 return Status{};
114 }
115
configure(const ICLTensor * input,int axis,ICLTensor * output,const ReductionOperation & op)116 void CLArgMinMaxLayer::configure(const ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op)
117 {
118 configure(CLKernelLibrary::get().get_compile_context(), input, axis, output, op);
119 }
120
configure(const CLCompileContext & compile_context,const ICLTensor * input,int axis,ICLTensor * output,const ReductionOperation & op)121 void CLArgMinMaxLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op)
122 {
123 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
124 ARM_COMPUTE_LOG_PARAMS(input, axis, output, op);
125
126 _num_of_stages = utils::calculate_number_of_stages_only_x_axis(input->info()->dimension(0), axis);
127 _reduction_axis = axis;
128
129 const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, false);
130 DataType output_data_type = (output->info()->data_type() == DataType::UNKNOWN) ? DataType::S32 : output->info()->data_type();
131 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true));
132
133 // Configure reduction operation kernels
134 _reduction_kernels_vector.reserve(_num_of_stages);
135
136 auto add_reduction_kernel = [this, &compile_context, axis, op](const ICLTensor * input, const ICLTensor * prev_output, ICLTensor * output)
137 {
138 _reduction_kernels_vector.emplace_back(std::make_unique<CLArgMinMaxLayerKernel>());
139 _reduction_kernels_vector.back()->configure(compile_context, input, prev_output, output, axis, op);
140 };
141
142 _memory_group.manage(&_not_reshaped_output);
143 // Create temporary tensors
144 if(_num_of_stages == 1)
145 {
146 add_reduction_kernel(input, nullptr, &_not_reshaped_output);
147 }
148 else
149 {
150 _results_vector.resize(_num_of_stages - 1);
151 TensorShape shape{ input->info()->tensor_shape() };
152 for(unsigned int i = 0; i < _num_of_stages - 1; i++)
153 {
154 shape.set(0, ceil(shape.x() / 128.f));
155 _results_vector[i].allocator()->init(input->info()->clone()->set_tensor_shape(shape).set_data_type(output_data_type));
156 }
157
158 // Apply ReductionOperation only on first kernel
159 _memory_group.manage(&_results_vector[0]);
160 add_reduction_kernel(input, nullptr, &_results_vector[0]);
161
162 // Apply ReductionOperation on intermediate stages
163 for(unsigned int i = 1; i < _num_of_stages - 1; ++i)
164 {
165 _memory_group.manage(&_results_vector[i]);
166 add_reduction_kernel(input, &_results_vector[i - 1], &_results_vector[i]);
167 _results_vector[i - 1].allocator()->allocate();
168 }
169
170 // Apply ReductionOperation on the last stage
171 const unsigned int last_stage = _num_of_stages - 1;
172 add_reduction_kernel(input, &_results_vector[last_stage - 1], &_not_reshaped_output);
173 _results_vector[last_stage - 1].allocator()->allocate();
174 }
175 _reshape.configure(compile_context, &_not_reshaped_output, output);
176 _not_reshaped_output.allocator()->allocate();
177 }
178
run()179 void CLArgMinMaxLayer::run()
180 {
181 MemoryGroupResourceScope scope_mg(_memory_group);
182
183 for(unsigned int i = 0; i < _num_of_stages; ++i)
184 {
185 CLScheduler::get().enqueue(*_reduction_kernels_vector[i], false);
186 }
187 _reshape.run();
188 }
189 } // namespace arm_compute
190