Michalis Spyrou | 7930db4 | 2018-11-22 17:36:28 +0000 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2018-2020 Arm Limited. |
Michalis Spyrou | 7930db4 | 2018-11-22 17:36:28 +0000 | [diff] [blame] | 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 | |
Michalis Spyrou | 7930db4 | 2018-11-22 17:36:28 +0000 | [diff] [blame] | 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" |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 31 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 32 | #include "src/core/CL/CLValidate.h" |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 33 | #include "src/core/CL/kernels/CLArgMinMaxLayerKernel.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 34 | #include "src/core/helpers/AutoConfiguration.h" |
| 35 | #include "src/runtime/Utils.h" |
Michalis Spyrou | 7930db4 | 2018-11-22 17:36:28 +0000 | [diff] [blame] | 36 | |
| 37 | namespace arm_compute |
| 38 | { |
Sang-Hoon Park | 2697fd8 | 2019-10-15 16:49:24 +0100 | [diff] [blame] | 39 | CLArgMinMaxLayer::CLArgMinMaxLayer(std::shared_ptr<IMemoryManager> memory_manager) |
Michalis Spyrou | 2aad21a | 2020-07-02 12:43:53 +0100 | [diff] [blame] | 40 | : _memory_group(std::move(memory_manager)), _results_vector(), _not_reshaped_output(), _reduction_kernels_vector(), _reshape(), _num_of_stages(), _reduction_axis() |
Michalis Spyrou | 7930db4 | 2018-11-22 17:36:28 +0000 | [diff] [blame] | 41 | { |
Sang-Hoon Park | 2697fd8 | 2019-10-15 16:49:24 +0100 | [diff] [blame] | 42 | } |
| 43 | |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 44 | CLArgMinMaxLayer::~CLArgMinMaxLayer() = default; |
| 45 | |
Michalis Spyrou | 7930db4 | 2018-11-22 17:36:28 +0000 | [diff] [blame] | 46 | Status CLArgMinMaxLayer::validate(const ITensorInfo *input, int axis, const ITensorInfo *output, const ReductionOperation &op) |
| 47 | { |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 48 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
Sheri Zhang | c5b6d88 | 2020-06-26 14:46:59 +0100 | [diff] [blame] | 49 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); |
| 50 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S32, DataType::F16, DataType::F32); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 51 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX && op != ReductionOperation::ARG_IDX_MIN, "Invalid reduction operation"); |
| 52 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= static_cast<int>(TensorShape::num_max_dimensions), "Reduction axis greater than max number of dimensions"); |
| 53 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis"); |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 54 | const unsigned int num_of_stages = utils::calculate_number_of_stages_only_x_axis(input->dimension(0), axis); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 55 | |
| 56 | DataType output_data_type = DataType::S32; |
| 57 | TensorInfo not_reshaped_output; |
| 58 | const auto input_num_channles = input->num_channels(); |
| 59 | const auto input_qinfo = input->quantization_info(); |
| 60 | |
| 61 | if(output->total_size() != 0) |
| 62 | { |
| 63 | output_data_type = output->data_type(); |
| 64 | const TensorInfo expected_output_shape = output->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis, false)); |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output_shape, output); |
| 66 | } |
| 67 | |
| 68 | auto shape_before_reshape = input->tensor_shape(); |
| 69 | shape_before_reshape.set(axis, 1); |
| 70 | auto initialize_tensorinfo = [](TensorInfo & ti, TensorShape shape, DataType data_type, int num_channels, QuantizationInfo qinfo) |
| 71 | { |
| 72 | ti.set_data_type(data_type).set_tensor_shape(shape).set_num_channels(num_channels).set_quantization_info(qinfo); |
| 73 | }; |
| 74 | |
| 75 | initialize_tensorinfo(not_reshaped_output, shape_before_reshape, output_data_type, input_num_channles, input_qinfo); |
| 76 | |
| 77 | if(num_of_stages == 1) |
| 78 | { |
| 79 | ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, ¬_reshaped_output, axis, op)); |
| 80 | } |
| 81 | else |
| 82 | { |
| 83 | // Create temporary tensor infos |
| 84 | std::vector<TensorInfo> sums_vector(num_of_stages - 1); |
| 85 | |
| 86 | // Create intermediate tensor info |
| 87 | TensorShape shape{ input->tensor_shape() }; |
| 88 | |
| 89 | for(unsigned int i = 0; i < num_of_stages - 1; i++) |
| 90 | { |
| 91 | shape.set(0, ceil(shape.x() / 128.f)); |
| 92 | sums_vector[i].set_data_type(input->data_type()); |
| 93 | sums_vector[i].set_tensor_shape(shape); |
| 94 | sums_vector[i].set_num_channels(input->num_channels()); |
| 95 | } |
| 96 | |
| 97 | // Validate ReductionOperation only on first kernel |
| 98 | ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, &sums_vector[0], axis, op)); |
| 99 | |
| 100 | // Validate ReductionOperation on intermediate stages |
| 101 | for(unsigned int i = 1; i < num_of_stages - 1; ++i) |
| 102 | { |
| 103 | ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[i - 1], &sums_vector[i], axis, op)); |
| 104 | } |
| 105 | |
| 106 | // Validate ReductionOperation on the last stage |
| 107 | const unsigned int last_stage = num_of_stages - 1; |
| 108 | ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[last_stage - 1], ¬_reshaped_output, axis, op)); |
| 109 | } |
Michalis Spyrou | 2aad21a | 2020-07-02 12:43:53 +0100 | [diff] [blame] | 110 | ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(¬_reshaped_output, output)); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 111 | return Status{}; |
| 112 | } |
| 113 | |
| 114 | void CLArgMinMaxLayer::configure(const ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op) |
| 115 | { |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 116 | configure(CLKernelLibrary::get().get_compile_context(), input, axis, output, op); |
| 117 | } |
| 118 | |
| 119 | void CLArgMinMaxLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op) |
| 120 | { |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 121 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 122 | _num_of_stages = utils::calculate_number_of_stages_only_x_axis(input->info()->dimension(0), axis); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 123 | _reduction_axis = axis; |
| 124 | |
| 125 | const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, false); |
| 126 | DataType output_data_type = (output->info()->data_type() == DataType::UNKNOWN) ? DataType::S32 : output->info()->data_type(); |
| 127 | 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)); |
| 128 | |
| 129 | // Configure reduction operation kernels |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 130 | _reduction_kernels_vector.reserve(_num_of_stages); |
| 131 | |
| 132 | auto add_reduction_kernel = [this, &compile_context, axis, op](const ICLTensor * input, const ICLTensor * prev_output, ICLTensor * output) |
| 133 | { |
Georgios Pinitas | 40f51a6 | 2020-11-21 03:04:18 +0000 | [diff] [blame] | 134 | _reduction_kernels_vector.emplace_back(std::make_unique<CLArgMinMaxLayerKernel>()); |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 135 | _reduction_kernels_vector.back()->configure(compile_context, input, prev_output, output, axis, op); |
| 136 | }; |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 137 | |
| 138 | _memory_group.manage(&_not_reshaped_output); |
| 139 | // Create temporary tensors |
| 140 | if(_num_of_stages == 1) |
| 141 | { |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 142 | add_reduction_kernel(input, nullptr, &_not_reshaped_output); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 143 | } |
| 144 | else |
| 145 | { |
| 146 | _results_vector.resize(_num_of_stages - 1); |
| 147 | TensorShape shape{ input->info()->tensor_shape() }; |
| 148 | for(unsigned int i = 0; i < _num_of_stages - 1; i++) |
| 149 | { |
| 150 | shape.set(0, ceil(shape.x() / 128.f)); |
| 151 | _results_vector[i].allocator()->init(input->info()->clone()->set_tensor_shape(shape).set_data_type(output_data_type)); |
| 152 | } |
| 153 | |
| 154 | // Apply ReductionOperation only on first kernel |
| 155 | _memory_group.manage(&_results_vector[0]); |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 156 | add_reduction_kernel(input, nullptr, &_results_vector[0]); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 157 | |
| 158 | // Apply ReductionOperation on intermediate stages |
| 159 | for(unsigned int i = 1; i < _num_of_stages - 1; ++i) |
| 160 | { |
| 161 | _memory_group.manage(&_results_vector[i]); |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 162 | add_reduction_kernel(input, &_results_vector[i - 1], &_results_vector[i]); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 163 | _results_vector[i - 1].allocator()->allocate(); |
| 164 | } |
| 165 | |
| 166 | // Apply ReductionOperation on the last stage |
| 167 | const unsigned int last_stage = _num_of_stages - 1; |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 168 | add_reduction_kernel(input, &_results_vector[last_stage - 1], &_not_reshaped_output); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 169 | _results_vector[last_stage - 1].allocator()->allocate(); |
| 170 | } |
Michalis Spyrou | 2aad21a | 2020-07-02 12:43:53 +0100 | [diff] [blame] | 171 | _reshape.configure(compile_context, &_not_reshaped_output, output); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 172 | _not_reshaped_output.allocator()->allocate(); |
Sang-Hoon Park | 2697fd8 | 2019-10-15 16:49:24 +0100 | [diff] [blame] | 173 | } |
| 174 | |
| 175 | void CLArgMinMaxLayer::run() |
| 176 | { |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 177 | MemoryGroupResourceScope scope_mg(_memory_group); |
| 178 | |
| 179 | for(unsigned int i = 0; i < _num_of_stages; ++i) |
| 180 | { |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 181 | CLScheduler::get().enqueue(*_reduction_kernels_vector[i], false); |
Manuel Bottini | 7b9998d | 2019-10-21 17:59:07 +0100 | [diff] [blame] | 182 | } |
Michalis Spyrou | 2aad21a | 2020-07-02 12:43:53 +0100 | [diff] [blame] | 183 | _reshape.run(); |
Michalis Spyrou | 7930db4 | 2018-11-22 17:36:28 +0000 | [diff] [blame] | 184 | } |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 185 | } // namespace arm_compute |