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
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 #include "support/MemorySupport.h"
37
38 namespace arm_compute
39 {
CLArgMinMaxLayer(std::shared_ptr<IMemoryManager> memory_manager)40 CLArgMinMaxLayer::CLArgMinMaxLayer(std::shared_ptr<IMemoryManager> memory_manager)
41 : _memory_group(std::move(memory_manager)), _results_vector(), _not_reshaped_output(), _reduction_kernels_vector(), _reshape(), _num_of_stages(), _reduction_axis()
42 {
43 }
44
45 CLArgMinMaxLayer::~CLArgMinMaxLayer() = default;
46
validate(const ITensorInfo * input,int axis,const ITensorInfo * output,const ReductionOperation & op)47 Status CLArgMinMaxLayer::validate(const ITensorInfo *input, int axis, const ITensorInfo *output, const ReductionOperation &op)
48 {
49 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
50 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
51 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S32, DataType::F16, DataType::F32);
52 ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX && op != ReductionOperation::ARG_IDX_MIN, "Invalid reduction operation");
53 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= static_cast<int>(TensorShape::num_max_dimensions), "Reduction axis greater than max number of dimensions");
54 ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
55 const unsigned int num_of_stages = utils::calculate_number_of_stages_only_x_axis(input->dimension(0), axis);
56
57 DataType output_data_type = DataType::S32;
58 TensorInfo not_reshaped_output;
59 const auto input_num_channles = input->num_channels();
60 const auto input_qinfo = input->quantization_info();
61
62 if(output->total_size() != 0)
63 {
64 output_data_type = output->data_type();
65 const TensorInfo expected_output_shape = output->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis, false));
66 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output_shape, output);
67 }
68
69 auto shape_before_reshape = input->tensor_shape();
70 shape_before_reshape.set(axis, 1);
71 auto initialize_tensorinfo = [](TensorInfo & ti, TensorShape shape, DataType data_type, int num_channels, QuantizationInfo qinfo)
72 {
73 ti.set_data_type(data_type).set_tensor_shape(shape).set_num_channels(num_channels).set_quantization_info(qinfo);
74 };
75
76 initialize_tensorinfo(not_reshaped_output, shape_before_reshape, output_data_type, input_num_channles, input_qinfo);
77
78 if(num_of_stages == 1)
79 {
80 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, ¬_reshaped_output, axis, op));
81 }
82 else
83 {
84 // Create temporary tensor infos
85 std::vector<TensorInfo> sums_vector(num_of_stages - 1);
86
87 // Create intermediate tensor info
88 TensorShape shape{ input->tensor_shape() };
89
90 for(unsigned int i = 0; i < num_of_stages - 1; i++)
91 {
92 shape.set(0, ceil(shape.x() / 128.f));
93 sums_vector[i].set_data_type(input->data_type());
94 sums_vector[i].set_tensor_shape(shape);
95 sums_vector[i].set_num_channels(input->num_channels());
96 }
97
98 // Validate ReductionOperation only on first kernel
99 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, &sums_vector[0], axis, op));
100
101 // Validate ReductionOperation on intermediate stages
102 for(unsigned int i = 1; i < num_of_stages - 1; ++i)
103 {
104 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[i - 1], &sums_vector[i], axis, op));
105 }
106
107 // Validate ReductionOperation on the last stage
108 const unsigned int last_stage = num_of_stages - 1;
109 ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[last_stage - 1], ¬_reshaped_output, axis, op));
110 }
111 ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(¬_reshaped_output, output));
112 return Status{};
113 }
114
configure(const ICLTensor * input,int axis,ICLTensor * output,const ReductionOperation & op)115 void CLArgMinMaxLayer::configure(const ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op)
116 {
117 configure(CLKernelLibrary::get().get_compile_context(), input, axis, output, op);
118 }
119
configure(const CLCompileContext & compile_context,const ICLTensor * input,int axis,ICLTensor * output,const ReductionOperation & op)120 void CLArgMinMaxLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op)
121 {
122 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
123 _num_of_stages = utils::calculate_number_of_stages_only_x_axis(input->info()->dimension(0), axis);
124 _reduction_axis = axis;
125
126 const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, false);
127 DataType output_data_type = (output->info()->data_type() == DataType::UNKNOWN) ? DataType::S32 : output->info()->data_type();
128 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));
129
130 // Configure reduction operation kernels
131 _reduction_kernels_vector.reserve(_num_of_stages);
132
133 auto add_reduction_kernel = [this, &compile_context, axis, op](const ICLTensor * input, const ICLTensor * prev_output, ICLTensor * output)
134 {
135 _reduction_kernels_vector.emplace_back(support::cpp14::make_unique<CLArgMinMaxLayerKernel>());
136 _reduction_kernels_vector.back()->configure(compile_context, input, prev_output, output, axis, op);
137 };
138
139 _memory_group.manage(&_not_reshaped_output);
140 // Create temporary tensors
141 if(_num_of_stages == 1)
142 {
143 add_reduction_kernel(input, nullptr, &_not_reshaped_output);
144 }
145 else
146 {
147 _results_vector.resize(_num_of_stages - 1);
148 TensorShape shape{ input->info()->tensor_shape() };
149 for(unsigned int i = 0; i < _num_of_stages - 1; i++)
150 {
151 shape.set(0, ceil(shape.x() / 128.f));
152 _results_vector[i].allocator()->init(input->info()->clone()->set_tensor_shape(shape).set_data_type(output_data_type));
153 }
154
155 // Apply ReductionOperation only on first kernel
156 _memory_group.manage(&_results_vector[0]);
157 add_reduction_kernel(input, nullptr, &_results_vector[0]);
158
159 // Apply ReductionOperation on intermediate stages
160 for(unsigned int i = 1; i < _num_of_stages - 1; ++i)
161 {
162 _memory_group.manage(&_results_vector[i]);
163 add_reduction_kernel(input, &_results_vector[i - 1], &_results_vector[i]);
164 _results_vector[i - 1].allocator()->allocate();
165 }
166
167 // Apply ReductionOperation on the last stage
168 const unsigned int last_stage = _num_of_stages - 1;
169 add_reduction_kernel(input, &_results_vector[last_stage - 1], &_not_reshaped_output);
170 _results_vector[last_stage - 1].allocator()->allocate();
171 }
172 _reshape.configure(compile_context, &_not_reshaped_output, output);
173 _not_reshaped_output.allocator()->allocate();
174 }
175
run()176 void CLArgMinMaxLayer::run()
177 {
178 MemoryGroupResourceScope scope_mg(_memory_group);
179
180 for(unsigned int i = 0; i < _num_of_stages; ++i)
181 {
182 CLScheduler::get().enqueue(*_reduction_kernels_vector[i], false);
183 }
184 _reshape.run();
185 }
186 } // namespace arm_compute
187