diff options
Diffstat (limited to 'src/backends/neon/workloads/NeonMergerWorkload.cpp')
-rw-r--r-- | src/backends/neon/workloads/NeonMergerWorkload.cpp | 52 |
1 files changed, 33 insertions, 19 deletions
diff --git a/src/backends/neon/workloads/NeonMergerWorkload.cpp b/src/backends/neon/workloads/NeonMergerWorkload.cpp index be096b4b25..64d4d93d97 100644 --- a/src/backends/neon/workloads/NeonMergerWorkload.cpp +++ b/src/backends/neon/workloads/NeonMergerWorkload.cpp @@ -11,12 +11,20 @@ #include <backendsCommon/CpuTensorHandle.hpp> #include <neon/NeonTensorHandle.hpp> -#include <arm_compute/runtime/NEON/functions/NEConcatenateLayer.h> + namespace armnn { using namespace armcomputetensorutils; +namespace +{ +size_t CalcAxis(const armnn::MergerDescriptor& desc) +{ + return (desc.GetNumDimensions() - desc.GetConcatAxis()) - 1; +} +} //namespace + arm_compute::Status NeonMergerWorkloadValidate(const std::vector<const TensorInfo*>& inputs, const TensorInfo& output, const MergerDescriptor& descriptor) @@ -25,60 +33,66 @@ arm_compute::Status NeonMergerWorkloadValidate(const std::vector<const TensorInf std::vector<arm_compute::TensorInfo> aclInputs; for (const TensorInfo* input : inputs) { - arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW); - aclInputs.emplace_back(aclInputInfo); + arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW); + aclInputs.emplace_back(aclInputInfo); } const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - arm_compute::DataLayoutDimension aclAxis = arm_compute::DataLayoutDimension::WIDTH; - std::vector<arm_compute::ITensorInfo*> aclInputPtrs; for (arm_compute::ITensorInfo& input : aclInputs) { aclInputPtrs.emplace_back(&input); } + size_t aclAxis = CalcAxis(descriptor); return arm_compute::NEConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis); - } NeonMergerWorkload::NeonMergerWorkload( const MergerQueueDescriptor& descriptor, const WorkloadInfo& info) : BaseWorkload<MergerQueueDescriptor>(descriptor, info) { - m_Execute = true; + bool allInputsAreSubtensors = true; - unsigned int innerAxisOrder = descriptor.m_Parameters.GetNumDimensions() - descriptor.m_Parameters.GetConcatAxis(); + // Check that all inputs are sub-tensors + for (auto input : descriptor.m_Inputs) + { + if (!input->GetParent()) + { + // Non sub-tensor input found so we need to execute the merger function + allInputsAreSubtensors = false; + break; + } + } - if (innerAxisOrder != 1) + if (allInputsAreSubtensors) { - m_Execute = false; + // Can skip configuring the merger function since it's not executed return; } std::vector<arm_compute::ITensor *> aclInputs; - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(armnn::DataLayout::NCHW); for (auto input : m_Data.m_Inputs) { arm_compute::ITensor& aclInput = boost::polymorphic_pointer_downcast<INeonTensorHandle>(input)->GetTensor(); - aclInput.info()->set_data_layout(aclDataLayout); aclInputs.emplace_back(&aclInput); } arm_compute::ITensor& output = boost::polymorphic_pointer_downcast<INeonTensorHandle>( - m_Data.m_Outputs[0])->GetTensor(); - output.info()->set_data_layout(aclDataLayout); + m_Data.m_Outputs[0])->GetTensor(); - arm_compute::DataLayoutDimension aclAxis = arm_compute::DataLayoutDimension::WIDTH; + // Create the layer function + m_Layer.reset(new arm_compute::NEConcatenateLayer()); - auto layer = std::make_unique<arm_compute::NEConcatenateLayer>(); - layer->configure(aclInputs, &output, aclAxis); - m_Layer.reset(layer.release()); + // Configure input and output tensors + size_t aclAxis = CalcAxis(descriptor.m_Parameters); + m_Layer->configure(aclInputs, &output, aclAxis); + // Prepare m_Layer->prepare(); } void NeonMergerWorkload::Execute() const { - if (m_Execute) + if (m_Layer) { ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonMergerWorkload_Execute"); m_Layer->run(); |