Matthew Jackson | 87f65ea | 2019-08-01 10:01:34 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | #include "NeonStackWorkload.hpp" |
| 6 | #include "NeonWorkloadUtils.hpp" |
| 7 | |
| 8 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
Jan Eilers | 3c9e045 | 2020-04-10 13:00:44 +0100 | [diff] [blame] | 9 | #include <armnn/utility/PolymorphicDowncast.hpp> |
Matthew Jackson | 87f65ea | 2019-08-01 10:01:34 +0100 | [diff] [blame] | 10 | #include <backendsCommon/CpuTensorHandle.hpp> |
| 11 | #include <neon/NeonTensorHandle.hpp> |
| 12 | |
| 13 | #include <boost/numeric/conversion/cast.hpp> |
Matthew Jackson | 87f65ea | 2019-08-01 10:01:34 +0100 | [diff] [blame] | 14 | |
| 15 | namespace armnn |
| 16 | { |
| 17 | using namespace armcomputetensorutils; |
| 18 | |
| 19 | namespace |
| 20 | { |
| 21 | int CalcAxis(const unsigned int axis, const unsigned int inputDimensions) |
| 22 | { |
| 23 | const int intAxis = boost::numeric_cast<int>(axis); |
| 24 | return boost::numeric_cast<int>(inputDimensions) - intAxis; |
| 25 | } |
| 26 | } //namespace |
| 27 | |
| 28 | arm_compute::Status NeonStackWorkloadValidate(const std::vector<const TensorInfo*>& inputs, |
| 29 | const TensorInfo& output, |
| 30 | const StackDescriptor& descriptor) |
| 31 | { |
| 32 | std::vector<arm_compute::TensorInfo> aclInputs; |
| 33 | for (const TensorInfo* input : inputs) |
| 34 | { |
| 35 | arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW); |
| 36 | aclInputs.emplace_back(aclInputInfo); |
| 37 | } |
| 38 | |
| 39 | std::vector<arm_compute::ITensorInfo*> aclInputPtrs; |
| 40 | for (arm_compute::ITensorInfo& input : aclInputs) |
| 41 | { |
| 42 | aclInputPtrs.emplace_back(&input); |
| 43 | } |
| 44 | |
| 45 | const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); |
| 46 | int aclAxis = CalcAxis(descriptor.m_Axis, descriptor.m_InputShape.GetNumDimensions()); |
| 47 | return arm_compute::NEStackLayer::validate(aclInputPtrs, aclAxis, &aclOutputInfo); |
| 48 | } |
| 49 | |
| 50 | NeonStackWorkload::NeonStackWorkload(const StackQueueDescriptor& descriptor, const WorkloadInfo& info) |
| 51 | : BaseWorkload<StackQueueDescriptor>(descriptor, info) |
| 52 | { |
| 53 | std::vector<arm_compute::ITensor*> aclInputs; |
| 54 | for (auto input : m_Data.m_Inputs) |
| 55 | { |
Jan Eilers | 3c9e045 | 2020-04-10 13:00:44 +0100 | [diff] [blame] | 56 | arm_compute::ITensor& aclInput = PolymorphicPointerDowncast<IAclTensorHandle>(input)->GetTensor(); |
Matthew Jackson | 87f65ea | 2019-08-01 10:01:34 +0100 | [diff] [blame] | 57 | aclInputs.emplace_back(&aclInput); |
| 58 | } |
Jan Eilers | 3c9e045 | 2020-04-10 13:00:44 +0100 | [diff] [blame] | 59 | arm_compute::ITensor& output = PolymorphicPointerDowncast<IAclTensorHandle>( |
Matthew Jackson | 87f65ea | 2019-08-01 10:01:34 +0100 | [diff] [blame] | 60 | m_Data.m_Outputs[0])->GetTensor(); |
| 61 | |
| 62 | m_Layer.reset(new arm_compute::NEStackLayer()); |
| 63 | int aclAxis = CalcAxis(descriptor.m_Parameters.m_Axis, descriptor.m_Parameters.m_InputShape.GetNumDimensions()); |
| 64 | m_Layer->configure(aclInputs, aclAxis, &output); |
| 65 | } |
| 66 | |
| 67 | void NeonStackWorkload::Execute() const |
| 68 | { |
| 69 | if (m_Layer) |
| 70 | { |
| 71 | ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonStackWorkload_Execute"); |
| 72 | m_Layer->run(); |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | } //namespace armnn |