Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 1 | // |
Teresa Charlin | 588cbdf | 2022-01-19 15:55:37 +0000 | [diff] [blame] | 2 | // Copyright © 2017 Arm Ltd and Contributors. All rights reserved. |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "NeonSplitterWorkload.hpp" |
| 7 | |
Matthew Bentham | 5e98b01 | 2020-01-24 23:11:43 +0000 | [diff] [blame] | 8 | #include <arm_compute/runtime/NEON/functions/NESplit.h> |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 9 | |
| 10 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
| 11 | #include <aclCommon/ArmComputeUtils.hpp> |
Jan Eilers | bb446e5 | 2020-04-02 13:56:54 +0100 | [diff] [blame] | 12 | #include <armnn/utility/PolymorphicDowncast.hpp> |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 13 | #include <armnn/backends/TensorHandle.hpp> |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 14 | #include <neon/NeonTensorHandle.hpp> |
| 15 | |
Matthew Bentham | 5e98b01 | 2020-01-24 23:11:43 +0000 | [diff] [blame] | 16 | #include "NeonWorkloadUtils.hpp" |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 17 | |
| 18 | namespace armnn |
| 19 | { |
| 20 | |
| 21 | using namespace armcomputetensorutils; |
| 22 | |
| 23 | namespace |
| 24 | { |
| 25 | unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int splitAxis) |
| 26 | { |
| 27 | return (numDimensions - splitAxis) - 1; |
| 28 | } |
| 29 | |
| 30 | } //namespace |
| 31 | |
| 32 | arm_compute::Status NeonSplitterWorkloadValidate(const TensorInfo& input, |
| 33 | const std::vector<std::reference_wrapper<TensorInfo>>& outputs, |
| 34 | unsigned int splitAxis) |
| 35 | { |
| 36 | const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); |
| 37 | |
| 38 | size_t numOutputs = outputs.size(); |
| 39 | |
| 40 | std::vector<arm_compute::TensorInfo> aclOutputs; |
| 41 | aclOutputs.reserve(numOutputs); |
| 42 | |
| 43 | std::vector<arm_compute::ITensorInfo*> aclOutputPtr; |
| 44 | aclOutputPtr.reserve(numOutputs); |
| 45 | |
| 46 | for (size_t i = 0u; i < outputs.size(); ++i) |
| 47 | { |
| 48 | aclOutputs.emplace_back(BuildArmComputeTensorInfo(outputs[i])); |
| 49 | aclOutputPtr.emplace_back(&aclOutputs.back()); |
| 50 | } |
| 51 | |
| 52 | unsigned int aclAxis = CalcAclAxis(input.GetNumDimensions(), splitAxis); |
| 53 | return arm_compute::NESplit::validate(&aclInputInfo, aclOutputPtr, aclAxis); |
| 54 | } |
| 55 | |
| 56 | NeonSplitterWorkload::NeonSplitterWorkload(const SplitterQueueDescriptor& descriptor, const WorkloadInfo& info) |
Teresa Charlin | 588cbdf | 2022-01-19 15:55:37 +0000 | [diff] [blame] | 57 | : NeonBaseWorkload<SplitterQueueDescriptor>(descriptor, info) |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 58 | { |
Keith Davis | 2d0679f | 2021-08-05 11:35:00 +0100 | [diff] [blame] | 59 | // Report Profiling Details |
| 60 | ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonSplitterWorkload_Construct", |
| 61 | descriptor.m_Parameters, |
| 62 | info, |
| 63 | this->GetGuid()); |
| 64 | |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 65 | bool allOutputsAreSubtensors = true; |
| 66 | |
| 67 | // Check that all outputs are sub-tensors |
| 68 | for (auto output : m_Data.m_Outputs) |
| 69 | { |
| 70 | if (output && !output->GetParent()) |
| 71 | { |
| 72 | // Non sub-tensor input found so we need to execute the split function |
| 73 | allOutputsAreSubtensors = false; |
| 74 | break; |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | if (allOutputsAreSubtensors) |
| 79 | { |
| 80 | // Can skip configuring the split function since it's not executed |
| 81 | return; |
| 82 | } |
| 83 | |
Jan Eilers | bb446e5 | 2020-04-02 13:56:54 +0100 | [diff] [blame] | 84 | arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 85 | |
| 86 | std::vector<arm_compute::ITensor *> aclOutputs; |
| 87 | for (auto output : m_Data.m_Outputs) |
| 88 | { |
Jan Eilers | 3c9e045 | 2020-04-10 13:00:44 +0100 | [diff] [blame] | 89 | arm_compute::ITensor& aclOutput = PolymorphicPointerDowncast<IAclTensorHandle>(output)->GetTensor(); |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 90 | aclOutputs.emplace_back(&aclOutput); |
| 91 | } |
| 92 | |
| 93 | // Create the layer function |
Matthew Bentham | 5e98b01 | 2020-01-24 23:11:43 +0000 | [diff] [blame] | 94 | std::unique_ptr<arm_compute::NESplit> layer(new arm_compute::NESplit()); |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 95 | |
| 96 | // Configure input and output tensors |
| 97 | std::set<unsigned int> splitAxis = ComputeSplitAxis(descriptor.m_Parameters, m_Data.m_Inputs[0]->GetShape()); |
| 98 | if (splitAxis.size() != 1) |
| 99 | { |
| 100 | throw InvalidArgumentException("Cannot derive split axis from SplitterDescriptor"); |
| 101 | } |
| 102 | |
| 103 | unsigned int aclAxis = CalcAclAxis(descriptor.m_Parameters.GetNumDimensions(), *splitAxis.begin()); |
Matthew Bentham | 5e98b01 | 2020-01-24 23:11:43 +0000 | [diff] [blame] | 104 | layer->configure(&input, aclOutputs, aclAxis); |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 105 | |
| 106 | // Prepare |
Matthew Bentham | 5e98b01 | 2020-01-24 23:11:43 +0000 | [diff] [blame] | 107 | layer->prepare(); |
| 108 | m_Layer.reset(layer.release()); |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 109 | } |
| 110 | |
| 111 | void NeonSplitterWorkload::Execute() const |
| 112 | { |
| 113 | if (m_Layer) |
| 114 | { |
Keith Davis | 2d0679f | 2021-08-05 11:35:00 +0100 | [diff] [blame] | 115 | ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonSplitterWorkload_Execute", this->GetGuid()); |
Narumol Prangnawarat | 15eb583 | 2019-05-20 15:31:05 +0100 | [diff] [blame] | 116 | m_Layer->run(); |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | } //namespace armnn |
| 121 | |