IVGCVSW-6806 Add Unidirectional Sequence Lstm support to Neon
* Corrected TensorInfo order for IsUnidirectionalSequenceLstmSupported
* outputStateOut TensorInfo is not optional.
* cellStateOut TensorInfo is not optional.
* TensorInfo Order matches other QLSTM/LSTM layers.
* Added missing parameters to UnidirectionalSequenceLstmOperator for
delegate.
* Added quantized UnidirectionalSequenceLstm support to Neon
!android-nn-driver:7457
Signed-off-by: Mike Kelly <mike.kelly@arm.com>
Change-Id: I26dde1bb96793dd25eb9081ca5ae5f63752288c4
diff --git a/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.cpp b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.cpp
new file mode 100644
index 0000000..dfbbb3c
--- /dev/null
+++ b/src/backends/neon/workloads/NeonUnidirectionalSequenceLstmWorkload.cpp
@@ -0,0 +1,879 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonUnidirectionalSequenceLstmWorkload.hpp"
+#include "NeonWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <armnn/utility/NumericCast.hpp>
+#include <armnnUtils/Permute.hpp>
+#include <neon/test/NeonWorkloadFactoryHelper.hpp>
+#include <backendsCommon/WorkloadUtils.hpp>
+
+#include "neon/NeonTensorHandle.hpp"
+
+namespace
+{
+
+unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int axis)
+{
+ return (numDimensions - axis) - 1;
+}
+} //namespace
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+NeonUnidirectionalSequenceLstmWorkload::NeonUnidirectionalSequenceLstmWorkload
+ (const UnidirectionalSequenceLstmQueueDescriptor& descriptor, const WorkloadInfo& info)
+ : NeonBaseWorkload<UnidirectionalSequenceLstmQueueDescriptor>(descriptor, info)
+{
+ // Report Profiling Details
+ ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonUnidirectionalSequenceLstmWorkload_Construct",
+ descriptor.m_Parameters,
+ info,
+ GetGuid());
+
+ // Input/Output tensors
+ const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+ arm_compute::ITensor& outputStateIn = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
+ const arm_compute::ITensor& cellStateIn = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
+
+ arm_compute::ITensor& outputStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+ arm_compute::ITensor& cellStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
+ arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
+
+ TensorInfo inputInfo = info.m_InputTensorInfos[0];
+ TensorInfo outputInfo = info.m_OutputTensorInfos[2];
+
+ TensorShape inputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetShape();
+ TensorShape outputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetShape();
+
+ unsigned int maxTime = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[0] : inputLayerShape[1];
+ unsigned int batchSize = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[1] : inputLayerShape[0];
+ unsigned int inputSize = inputLayerShape[2];
+ unsigned int outputSize = outputLayerShape[2];
+
+ const TensorShape timeMajorShapeInput({maxTime, batchSize, inputSize});
+ const TensorShape timeMajorShapeOutput({maxTime, batchSize, outputSize});
+
+ //
+ // Permute: performed if Unidirectional Sequence Layer inputs/outputs are in batch major format.
+ //
+ if (!m_Data.m_Parameters.m_TimeMajor)
+ {
+ std::unique_ptr<arm_compute::NEPermute> layer(new arm_compute::NEPermute());
+
+ TensorInfo permuteOutInfo = inputInfo;
+ permuteOutInfo.SetShape(timeMajorShapeInput);
+ BuildArmComputeTensor(m_PermuteFirstOut, permuteOutInfo);
+ armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_PermuteFirstOut);
+
+ // Permute to time major format.
+ layer->configure(&input, &m_PermuteFirstOut, arm_compute::PermutationVector(0U,2U,1U));
+ m_Permute1.reset(layer.release());
+ }
+
+ //
+ // Split and Concat Tensors
+ //
+ for (unsigned int i = 0; i < maxTime; ++i)
+ {
+ arm_compute::Tensor splitter_out;
+ arm_compute::Tensor concat_in;
+
+ auto splitterTensorInfo = inputInfo;
+ auto concatTensorInfo = outputInfo;
+ splitterTensorInfo.SetShape({batchSize, inputSize});
+ concatTensorInfo.SetShape({batchSize, outputSize});
+ BuildArmComputeTensor(splitter_out, splitterTensorInfo);
+ BuildArmComputeTensor(concat_in, concatTensorInfo);
+
+ armcomputetensorutils::InitialiseArmComputeTensorEmpty(splitter_out);
+ armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_in);
+
+ // append to std::vector<arm_compute::Tensor>
+ m_SplitterOutputsTensors.push_back(std::move(splitter_out));
+ m_ConcatInputsTensors.push_back(std::move(concat_in));
+ }
+
+ for (unsigned int i = 0; i < maxTime; ++i)
+ {
+ // append to std::vector<arm_compute::ITensor*>
+ m_SplitterOutputs.push_back(&m_SplitterOutputsTensors[i]);
+ m_ConcatInputs.push_back(&m_ConcatInputsTensors[i]);
+ }
+
+ //
+ // Split
+ //
+ unsigned int numberDimensions = 3;
+ unsigned int dimension = 0; // splitting on 0-dimension (i.e. maxTime dimension)
+
+ if (maxTime != 1) // ACL split does not work with only one element to split.
+ {
+ ViewsDescriptor splitterDesc(maxTime, numberDimensions);
+ unsigned int splitterDimSizes[3] = {1, batchSize, inputSize};
+ for (unsigned int outputIdx = 0u; outputIdx < maxTime; ++outputIdx)
+ {
+ splitterDesc.SetViewOriginCoord(outputIdx, dimension, splitterDimSizes[dimension] * outputIdx);
+ for (unsigned int dimIdx = 0u; dimIdx < numberDimensions; ++dimIdx)
+ {
+ splitterDesc.SetViewSize(outputIdx, dimIdx, splitterDimSizes[dimIdx]);
+ }
+ }
+
+ std::set<unsigned int> splitAxis = ComputeSplitAxis(splitterDesc, timeMajorShapeInput);
+
+ std::unique_ptr<arm_compute::NESplit> split_layer(new arm_compute::NESplit());
+ unsigned int aclAxisSplit = CalcAclAxis(splitterDesc.GetNumDimensions(),
+ *splitAxis.begin());
+ if (!m_Data.m_Parameters.m_TimeMajor)
+ {
+ split_layer->configure(&m_PermuteFirstOut, m_SplitterOutputs, aclAxisSplit);
+ } else
+ {
+ split_layer->configure(&input, m_SplitterOutputs, aclAxisSplit);
+ }
+
+ split_layer->prepare();
+ m_Splitter.reset(split_layer.release());
+ }
+
+ //
+ // Lstm
+ //
+ arm_compute::LSTMParams<arm_compute::ITensor> lstm_param;
+
+ lstm_param.set_cell_clip_params(descriptor.m_Parameters.m_ClippingThresCell);
+ lstm_param.set_projection_clip_params(descriptor.m_Parameters.m_ClippingThresProj);
+
+ lstm_param.set_matmul_scale_params(descriptor.m_Parameters.m_InputIntermediateScale,
+ descriptor.m_Parameters.m_ForgetIntermediateScale,
+ descriptor.m_Parameters.m_CellIntermediateScale,
+ descriptor.m_Parameters.m_OutputIntermediateScale);
+
+ lstm_param.set_hidden_state_params(descriptor.m_Parameters.m_HiddenStateZeroPoint,
+ descriptor.m_Parameters.m_HiddenStateScale);
+
+ m_InputToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
+
+ m_InputToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
+
+ m_InputToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
+
+ m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
+
+ m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
+
+ m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
+
+ m_ForgetGateBiasTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
+
+ m_CellBiasTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
+
+ m_OutputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
+
+ // for future reference: check the AndroidNN API for the logic here
+ if (!m_Data.m_Parameters.m_CifgEnabled)
+ {
+ m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
+
+ m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
+
+ m_CellToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ if (m_Data.m_CellToInputWeights != nullptr)
+ {
+ BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
+ }
+
+ m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
+ lstm_param.set_cifg_params(m_InputToInputWeightsTensor.get(),
+ m_RecurrentToInputWeightsTensor.get(),
+ m_Data.m_CellToInputWeights ? m_CellToInputWeightsTensor.get() : nullptr,
+ m_InputGateBiasTensor.get());
+ }
+
+ if (m_Data.m_Parameters.m_ProjectionEnabled)
+ {
+ m_ProjectionWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());
+
+ m_ProjectionBiasTensor = std::make_unique<arm_compute::Tensor>();
+ if (m_Data.m_ProjectionBias != nullptr)
+ {
+ BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
+ }
+
+ lstm_param.set_projection_params(m_ProjectionWeightsTensor.get(),
+ m_Data.m_ProjectionBias ? m_ProjectionBiasTensor.get() : nullptr);
+ }
+
+ if (m_Data.m_Parameters.m_PeepholeEnabled)
+ {
+ m_CellToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());
+
+ m_CellToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());
+
+ lstm_param.set_peephole_params(m_CellToForgetWeightsTensor.get(), m_CellToOutputWeightsTensor.get());
+ }
+
+ if (m_Data.m_Parameters.m_LayerNormEnabled)
+ {
+ m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ if (!m_Data.m_Parameters.m_CifgEnabled)
+ {
+ BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
+ }
+
+ m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
+
+ m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
+
+ m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
+ BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());
+
+ auto inputNormWeightTensor = m_Data.m_Parameters.m_CifgEnabled ? nullptr : m_InputLayerNormWeightsTensor.get();
+ lstm_param.set_layer_normalization_params(inputNormWeightTensor,
+ m_ForgetLayerNormWeightsTensor.get(),
+ m_CellLayerNormWeightsTensor.get(),
+ m_OutputLayerNormWeightsTensor.get());
+ }
+
+ for (unsigned int i = 0; i != maxTime; ++i)
+ {
+ // Set LSTM input and output ITensors depending on:
+ // input format (timeMajor) & number of LSTM batches (maxTime).
+ arm_compute::ITensor* outputLSTM;
+ arm_compute::ITensor* inputLSTM;
+
+ // If there is only one LSTM time major batch, we will not concat OR permute.
+ // Set input of LSTM to be first input ITensor.
+ // Set output of LSTM to be final output ITensor.
+ // LSTM input/output cannot be > 2 dimensions so need to resize its TensorInfo.
+ if (maxTime == 1 && m_Data.m_Parameters.m_TimeMajor)
+ {
+ TensorShape inputShape = GetTensorShape(input.info()->tensor_shape(), 1U);
+ TensorShape outputShape = GetTensorShape(output.info()->tensor_shape(), 1U);
+
+ TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
+ TensorShape outputShapeShrink({outputShape[1], outputShape[2]});
+
+ auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
+ auto acl_output_shape_shrink = BuildArmComputeTensorShape(outputShapeShrink);
+
+ input.info()->set_tensor_shape(acl_input_shape_shrink);
+ inputLSTM = const_cast<arm_compute::ITensor*>(&input);
+
+ output.info()->set_tensor_shape(acl_output_shape_shrink);
+ outputLSTM = &output;
+ }
+ // If there is only one LSTM batch major batch, we will not concat, only permute.
+ // Set input of LSTM to be output of initial permute.
+ // Set output of LSTM to be first element of m_ConcatInputs & use that value later in permute.
+ // LSTM output cannot be > 2 dimensions so need to resize its TensorInfo.
+ else if (maxTime == 1 && !m_Data.m_Parameters.m_TimeMajor)
+ {
+ TensorShape inputShape = GetTensorShape(m_PermuteFirstOut.info()->tensor_shape(), 1U);
+ TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
+ auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
+ m_PermuteFirstOut.info()->set_tensor_shape(acl_input_shape_shrink);
+ inputLSTM = &m_PermuteFirstOut;
+
+ outputLSTM = const_cast<arm_compute::ITensor*>(m_ConcatInputs[i]);
+ }
+ // Batch major AND/OR 2+ LSTM batches so will use concat AND/OR permute later on.
+ else
+ {
+ inputLSTM = m_SplitterOutputs[i];
+ outputLSTM = const_cast<arm_compute::ITensor*>(m_ConcatInputs[i]);
+ }
+
+ std::unique_ptr<arm_compute::NEQLSTMLayer> lstm_layer(new arm_compute::NEQLSTMLayer());
+
+ lstm_layer->configure(inputLSTM,
+ m_InputToForgetWeightsTensor.get(),
+ m_InputToCellWeightsTensor.get(),
+ m_InputToOutputWeightsTensor.get(),
+ m_RecurrentToForgetWeightsTensor.get(),
+ m_RecurrentToCellWeightsTensor.get(),
+ m_RecurrentToOutputWeightsTensor.get(),
+ m_ForgetGateBiasTensor.get(),
+ m_CellBiasTensor.get(),
+ m_OutputGateBiasTensor.get(),
+ &cellStateIn,
+ &outputStateIn,
+ &cellStateOut,
+ &outputStateOut,
+ outputLSTM,
+ lstm_param);
+
+ m_Layers.emplace_back(std::move(lstm_layer));
+ }
+
+ InitializeArmComputeTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights);
+ InitializeArmComputeTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights);
+ InitializeArmComputeTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights);
+ InitializeArmComputeTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
+ InitializeArmComputeTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights);
+ InitializeArmComputeTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
+ InitializeArmComputeTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
+ InitializeArmComputeTensorData(*m_CellBiasTensor, m_Data.m_CellBias);
+ InitializeArmComputeTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
+
+ if (!m_Data.m_Parameters.m_CifgEnabled)
+ {
+ InitializeArmComputeTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
+ InitializeArmComputeTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
+ if (m_Data.m_CellToInputWeights != nullptr)
+ {
+ InitializeArmComputeTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights);
+ }
+ InitializeArmComputeTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
+ }
+
+ if (m_Data.m_Parameters.m_ProjectionEnabled)
+ {
+ InitializeArmComputeTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights);
+ if (m_Data.m_ProjectionBias != nullptr)
+ {
+ InitializeArmComputeTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias);
+ }
+ }
+
+ if (m_Data.m_Parameters.m_PeepholeEnabled)
+ {
+ InitializeArmComputeTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights);
+ InitializeArmComputeTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights);
+ }
+
+ if (m_Data.m_Parameters.m_LayerNormEnabled)
+ {
+ if (!m_Data.m_Parameters.m_CifgEnabled)
+ {
+ InitializeArmComputeTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights);
+ }
+ InitializeArmComputeTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
+ InitializeArmComputeTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights);
+ InitializeArmComputeTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
+ }
+
+ // Force Compute Library to perform the necessary copying and reshaping.
+ // After which delete all the input tensors that will no longer be needed.
+ for (uint32_t i = 0; i < m_Layers.size(); ++i)
+ {
+ m_Layers[i]->prepare();
+ }
+
+ //
+ // Concat
+ //
+
+ // Expand dimensions of LSTM outputs adding one empty dimension to fit concatenate inputs.
+ TensorShape shape = GetTensorShape(m_ConcatInputs[0]->info()->tensor_shape(), 1U);
+ TensorShape shapeExpandTimeMajor({1, shape[0], shape[1]});
+ TensorShape shapeExpandBatchMajor({shape[0], 1, shape[1]});
+
+ if (maxTime != 1) // ACL concat does not work with only one element to concatenate.
+ {
+ for (unsigned int i = 0; i < maxTime; ++i)
+ {
+ m_ConcatInputs[i]->info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
+ }
+ ConcatDescriptor concatDescriptor(maxTime, numberDimensions); // maxTime = num inputs (aka. number of views).
+
+ for (unsigned int inputIdx = 0u; inputIdx < maxTime; ++inputIdx)
+ {
+ concatDescriptor.SetViewOriginCoord(inputIdx, dimension, inputIdx);
+ concatDescriptor.SetConcatAxis(dimension);
+ }
+ m_Concat.reset(new arm_compute::NEConcatenateLayer());
+
+ unsigned int aclAxisConcat = CalcAclAxis(concatDescriptor.GetNumDimensions(), concatDescriptor.GetConcatAxis());
+ if (!m_Data.m_Parameters.m_TimeMajor)
+ {
+ TensorInfo concatOutputTensorInfo = outputInfo;
+ concatOutputTensorInfo.SetShape(timeMajorShapeOutput);
+ BuildArmComputeTensor(concat_out, concatOutputTensorInfo);
+ armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_out);
+
+ m_Concat->configure(m_ConcatInputs, &concat_out, aclAxisConcat);
+ }
+ else
+ {
+ m_Concat->configure(m_ConcatInputs, &output, aclAxisConcat);
+ }
+
+ m_Concat->prepare();
+ }
+ // If only one LSTM batch, we do not concat and/or permute.
+ // Must ensure final output info is expanded to correct batch major dimensions.
+ else
+ {
+ if (!m_Data.m_Parameters.m_TimeMajor)
+ {
+ output.info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandBatchMajor));
+ }
+ else
+ {
+ output.info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
+ }
+ }
+
+ //
+ // Permute: only done if input/output are in batch major format.
+ //
+ if (!m_Data.m_Parameters.m_TimeMajor)
+ {
+ // Output now time major. Permute output back to batch major.
+ std::unique_ptr<arm_compute::NEPermute> layer(new arm_compute::NEPermute());
+ if (maxTime != 1)
+ {
+ layer->configure(&concat_out, &output, arm_compute::PermutationVector(0U, 2U, 1U));
+ }
+ else
+ {
+ layer->configure(m_ConcatInputs[0], &output, arm_compute::PermutationVector(0U, 2U, 1U));
+ }
+ m_Permute2.reset(layer.release());
+ }
+
+ FreeUnusedTensors();
+}
+
+void NeonUnidirectionalSequenceLstmWorkload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonUnidirectionalSequenceLstmWorkload_Execute", GetGuid());
+ if (m_Permute1)
+ {
+ m_Permute1->run();
+ }
+ if (m_Splitter)
+ {
+ m_Splitter->run();
+ }
+ for (uint32_t i = 0; i < m_Layers.size(); ++i)
+ {
+ m_Layers[i]->run();
+ }
+ if (m_Concat)
+ {
+ m_Concat->run();
+ }
+ if (m_Permute2)
+ {
+ m_Permute2->run();
+ }
+}
+
+arm_compute::Status
+NeonUnidirectionalSequenceLstmWorkloadValidate(const TensorInfo& input,
+ const TensorInfo& outputStateIn,
+ const TensorInfo& cellStateIn,
+ const TensorInfo& outputStateOut,
+ const TensorInfo& cellStateOut,
+ const TensorInfo& output,
+ const UnidirectionalSequenceLstmDescriptor& descriptor,
+ const LstmInputParamsInfo& paramsInfo)
+{
+ TensorShape inputLayerShape = input.GetShape();
+ TensorShape outputLayerShape = output.GetShape();
+
+ unsigned int maxTime = descriptor.m_TimeMajor ? inputLayerShape[0] : inputLayerShape[1];
+ unsigned int batchSize = descriptor.m_TimeMajor ? inputLayerShape[1] : inputLayerShape[0];
+ unsigned int inputSize = inputLayerShape[2];
+ unsigned int outputSize = outputLayerShape[2];
+
+ const TensorShape timeMajorShapeInput({maxTime, batchSize, inputSize});
+ const TensorShape timeMajorShapeOutput({maxTime, batchSize, outputSize});
+
+ arm_compute::Status statusPermute1 = arm_compute::Status(arm_compute::ErrorCode::OK,
+ "Permute1 status");
+ arm_compute::Status statusSplit = arm_compute::Status(arm_compute::ErrorCode::OK,
+ "Split status");
+ arm_compute::Status statusLSTM = arm_compute::Status(arm_compute::ErrorCode::OK,
+ "LSTM status");
+ arm_compute::Status statusConcat = arm_compute::Status(arm_compute::ErrorCode::OK,
+ "Concat status");
+ arm_compute::Status statusPermute2 = arm_compute::Status(arm_compute::ErrorCode::OK,
+ "Permute2 status");
+
+ const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+ const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+ //
+ // Permute validate
+ //
+ TensorInfo permuteOutInfo = TensorInfo(input);
+ arm_compute::TensorInfo aclPermuteOutInfo = armcomputetensorutils::BuildArmComputeTensorInfo(permuteOutInfo);
+ if (!descriptor.m_TimeMajor)
+ {
+ statusPermute1 = arm_compute::NEPermute::validate(&aclInputInfo,
+ &aclPermuteOutInfo,
+ arm_compute::PermutationVector(0U, 2U, 1U));
+ }
+
+ //
+ // Split and Concat Tensors validate
+ //
+ std::vector<arm_compute::TensorInfo> splitterOutputsTensorInfos;
+ std::vector<arm_compute::TensorInfo> concatInputsTensorInfos;
+ std::vector<arm_compute::ITensorInfo*> splitterOutputsTensorInfosPtr;
+ std::vector<const arm_compute::ITensorInfo*> concatInputsTensorInfosPtr;
+ splitterOutputsTensorInfos.reserve(maxTime);
+ concatInputsTensorInfos.reserve(maxTime);
+ for (unsigned int i = 0; i < maxTime; ++i)
+ {
+ arm_compute::TensorInfo splitter_out;
+ arm_compute::TensorInfo concat_in;
+
+ auto splitterTensorInfo = TensorInfo(input);
+ auto concatTensorInfo = TensorInfo(output);
+ splitterTensorInfo.SetShape({batchSize, inputSize});
+ concatTensorInfo.SetShape({batchSize, outputSize});
+
+ arm_compute::TensorInfo aclSplitterTensorInfo
+ = armcomputetensorutils::BuildArmComputeTensorInfo(splitterTensorInfo);
+ arm_compute::TensorInfo aclConcatTensorInfo
+ = armcomputetensorutils::BuildArmComputeTensorInfo(concatTensorInfo);
+
+ splitterOutputsTensorInfos.emplace_back(aclSplitterTensorInfo);
+ concatInputsTensorInfos.emplace_back(aclConcatTensorInfo);
+ splitterOutputsTensorInfosPtr.emplace_back(&splitterOutputsTensorInfos[i]);
+ concatInputsTensorInfosPtr.emplace_back(&concatInputsTensorInfos[i]);
+ }
+
+ //
+ // Split validate
+ //
+ unsigned int numberDimensions = 3;
+ unsigned int dimension = 0; // splitting on 0-dimension (i.e. maxTime dimension)
+ unsigned int aclAxisSplit = CalcAclAxis(numberDimensions, dimension);
+
+ if (maxTime != 1) // ACL split does not work with only one element to split.
+ {
+ if (!descriptor.m_TimeMajor)
+ {
+ statusSplit = arm_compute::NESplit::validate(&aclPermuteOutInfo,
+ splitterOutputsTensorInfosPtr,
+ aclAxisSplit);
+ } else
+ {
+ statusSplit = arm_compute::NESplit::validate(&aclInputInfo, splitterOutputsTensorInfosPtr, aclAxisSplit);
+ }
+ }
+
+ //
+ // LSTM validate
+ //
+
+ arm_compute::LSTMParams<arm_compute::ITensorInfo> lstm_params_info;
+
+ const TensorInfo& scratchBuffer = TensorInfo(cellStateIn.GetShape(), input.GetDataType());
+
+ lstm_params_info.set_cell_clip_params(descriptor.m_ClippingThresCell);
+ lstm_params_info.set_projection_clip_params(descriptor.m_ClippingThresProj);
+ // The inputs and outputs
+ const arm_compute::TensorInfo aclOutputStateInInfo = BuildArmComputeTensorInfo(outputStateIn);
+ const arm_compute::TensorInfo aclCellStateInInfo = BuildArmComputeTensorInfo(cellStateIn);
+ const arm_compute::TensorInfo aclScratchBufferInfo = BuildArmComputeTensorInfo(scratchBuffer);
+ const arm_compute::TensorInfo aclOutputStateOutInfo = BuildArmComputeTensorInfo(outputStateOut);
+ const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut);
+
+ // Basic parameters
+ const arm_compute::TensorInfo aclInputToForgetWeightsInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetInputToForgetWeights());
+ const arm_compute::TensorInfo aclInputToCellWeightsInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetInputToCellWeights());
+ const arm_compute::TensorInfo aclInputToOutputWeightsInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetInputToOutputWeights());
+ const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToForgetWeights());
+ const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToCellWeights());
+ const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToOutputWeights());
+ const arm_compute::TensorInfo aclForgetGateBiasInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetForgetGateBias());
+ const arm_compute::TensorInfo aclCellBiasInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetCellBias());
+ const arm_compute::TensorInfo aclOutputGateBiasInfo
+ = BuildArmComputeTensorInfo(paramsInfo.GetOutputGateBias());
+
+ arm_compute::TensorInfo aclInputToInputWeightsInfo;
+ arm_compute::TensorInfo aclRecurrentToInputWeightsInfo;
+ arm_compute::TensorInfo aclCellToInputWeightsInfo;
+ arm_compute::TensorInfo aclInputGateBiasInfo;
+ arm_compute::TensorInfo aclProjectionWeightsInfo;
+ arm_compute::TensorInfo aclProjectionBiasInfo;
+ arm_compute::TensorInfo aclCellToForgetWeightsInfo;
+ arm_compute::TensorInfo aclCellToOutputWeightsInfo;
+
+ arm_compute::TensorInfo aclInputLayerNormWeightsInfo;
+ arm_compute::TensorInfo aclForgetLayerNormWeightsInfo;
+ arm_compute::TensorInfo aclCellLayerNormWeightsInfo;
+ arm_compute::TensorInfo aclOutputLayerNormWeightsInfo;
+
+ if (!descriptor.m_CifgEnabled)
+ {
+ if (descriptor.m_PeepholeEnabled)
+ {
+ aclCellToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToInputWeights());
+ }
+ aclInputToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputToInputWeights());
+ aclRecurrentToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToInputWeights());
+ aclInputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias());
+
+ lstm_params_info.set_cifg_params(&aclInputToInputWeightsInfo,
+ &aclRecurrentToInputWeightsInfo,
+ descriptor.m_PeepholeEnabled ? &aclCellToInputWeightsInfo : nullptr,
+ &aclInputGateBiasInfo);
+ }
+
+ if (descriptor.m_ProjectionEnabled)
+ {
+ if (paramsInfo.m_ProjectionBias != nullptr)
+ {
+ aclProjectionBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionBias());
+ }
+ aclProjectionWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionWeights());
+
+ lstm_params_info.set_projection_params(&aclProjectionWeightsInfo,
+ paramsInfo.m_ProjectionBias ? &aclProjectionBiasInfo : nullptr);
+ }
+
+ if (descriptor.m_PeepholeEnabled)
+ {
+ aclCellToForgetWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToForgetWeights());
+ aclCellToOutputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToOutputWeights());
+
+ lstm_params_info.set_peephole_params(&aclCellToForgetWeightsInfo, &aclCellToOutputWeightsInfo);
+ }
+
+ if (descriptor.m_LayerNormEnabled)
+ {
+ if (!descriptor.m_CifgEnabled)
+ {
+ aclInputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputLayerNormWeights());
+ }
+ aclForgetLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetForgetLayerNormWeights());
+ aclCellLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellLayerNormWeights());
+ aclOutputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputLayerNormWeights());
+
+ lstm_params_info.set_layer_normalization_params(descriptor.m_CifgEnabled ? nullptr :
+ &aclInputLayerNormWeightsInfo,
+ &aclForgetLayerNormWeightsInfo,
+ &aclCellLayerNormWeightsInfo,
+ &aclOutputLayerNormWeightsInfo);
+ }
+
+ lstm_params_info.set_matmul_scale_params(descriptor.m_InputIntermediateScale,
+ descriptor.m_ForgetIntermediateScale,
+ descriptor.m_CellIntermediateScale,
+ descriptor.m_OutputIntermediateScale);
+
+ lstm_params_info.set_hidden_state_params(descriptor.m_HiddenStateZeroPoint, descriptor.m_HiddenStateScale);
+
+ for (unsigned int i = 0; i != maxTime; ++i)
+ {
+
+ // Set LSTM input and output ITensors depending on:
+ // input format (timeMajor) & number of LSTM batches (maxTime).
+ arm_compute::ITensorInfo* outputLSTM;
+ arm_compute::ITensorInfo* inputLSTM;
+
+ // If there is only one LSTM time major batch, we will not concat OR permute.
+ // Set input of LSTM to be first input ITensor.
+ // Set output of LSTM to be final output ITensor.
+ // LSTM input/output cannot be > 2 dimensions so need to resize its TensorInfo.
+ if (maxTime == 1 && !descriptor.m_TimeMajor)
+ {
+ TensorShape inputShape = GetTensorShape(aclInputInfo.tensor_shape(), 1U);
+ TensorShape outputShape = GetTensorShape(aclOutputInfo.tensor_shape(), 1U);
+
+ TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
+ TensorShape outputShapeShrink({outputShape[1], outputShape[2]});
+
+ auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
+ auto acl_output_shape_shrink = BuildArmComputeTensorShape(outputShapeShrink);
+
+ const_cast<arm_compute::TensorInfo*>(&aclInputInfo)->set_tensor_shape(acl_input_shape_shrink);
+ inputLSTM = const_cast<arm_compute::TensorInfo*>(&aclInputInfo);
+
+ const_cast<arm_compute::TensorInfo*>(&aclOutputInfo)->set_tensor_shape(acl_output_shape_shrink);
+ outputLSTM = const_cast<arm_compute::TensorInfo*>(&aclOutputInfo);
+ }
+ // If there is only one LSTM batch major batch, we will not concat, only permute.
+ // Set input of LSTM to be output of initial permute.
+ // Set output of LSTM to be first element of m_ConcatInputs & use that value later in permute.
+ // LSTM output cannot be > 2 dimensions so need to resize its TensorInfo.
+ else if (maxTime == 1 && !descriptor.m_TimeMajor)
+ {
+ TensorShape inputShape = GetTensorShape(aclPermuteOutInfo.tensor_shape(), 1U);
+ TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
+ auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
+ aclPermuteOutInfo.set_tensor_shape(acl_input_shape_shrink);
+ inputLSTM = &aclPermuteOutInfo;
+
+ outputLSTM = const_cast<arm_compute::ITensorInfo*>(concatInputsTensorInfosPtr[i]);
+ }
+ // Batch major AND/OR 2+ LSTM batches so will use concat AND/OR permute later on.
+ else
+ {
+ inputLSTM = splitterOutputsTensorInfosPtr[i];
+ outputLSTM = const_cast<arm_compute::ITensorInfo*>(concatInputsTensorInfosPtr[i]);
+ }
+
+ statusLSTM = arm_compute::NEQLSTMLayer::validate(inputLSTM,
+ &aclInputToForgetWeightsInfo,
+ &aclInputToCellWeightsInfo,
+ &aclInputToOutputWeightsInfo,
+ &aclRecurrentToForgetWeightsInfo,
+ &aclRecurrentToCellWeightsInfo,
+ &aclRecurrentToOutputWeightsInfo,
+ &aclForgetGateBiasInfo,
+ &aclCellBiasInfo,
+ &aclOutputGateBiasInfo,
+ &aclCellStateInInfo,
+ &aclOutputStateInInfo,
+ &aclCellStateOutInfo,
+ &aclOutputStateOutInfo,
+ outputLSTM,
+ lstm_params_info);
+ }
+
+ //
+ // Concat validate
+ //
+
+ // Expand dimensions of LSTM outputs adding one empty dimension to fit concatenate inputs.
+ TensorShape shape = GetTensorShape(concatInputsTensorInfosPtr[0]->tensor_shape(), 1U);
+ TensorShape shapeExpandTimeMajor({1, shape[0], shape[1]});
+ TensorShape shapeExpandBatchMajor({shape[0], 1, shape[1]});
+
+ TensorInfo concatOutputTensorInfo = TensorInfo(output);
+ concatOutputTensorInfo.SetShape(timeMajorShapeOutput);
+ arm_compute::TensorInfo aclConcatOutputTensorInfo= BuildArmComputeTensorInfo(concatOutputTensorInfo);
+
+ if (maxTime != 1) // ACL concat does not work with only one element to concatenate.
+ {
+ for (unsigned int i = 0; i < maxTime; ++i)
+ {
+ auto acl_shape_expand = BuildArmComputeTensorShape(shapeExpandTimeMajor);
+ concatInputsTensorInfos[i].set_tensor_shape(acl_shape_expand);
+ }
+
+ unsigned int aclAxisConcat = CalcAclAxis(numberDimensions, dimension);
+ if (!descriptor.m_TimeMajor)
+ {
+ statusConcat = arm_compute::NEConcatenateLayer::validate(concatInputsTensorInfosPtr,
+ &aclConcatOutputTensorInfo,
+ aclAxisConcat);
+ }
+ else
+ {
+ statusConcat = arm_compute::NEConcatenateLayer::validate(concatInputsTensorInfosPtr,
+ &aclOutputInfo,
+ aclAxisConcat);
+ }
+ }
+ // If only one LSTM batch, we do not concat and/or permute.
+ // Must ensure final output info is expanded to correct batch major dimensions.
+ else
+ {
+ if (!descriptor.m_TimeMajor)
+ {
+ const_cast<arm_compute::TensorInfo*>(&aclInputInfo)->set_tensor_shape(
+ BuildArmComputeTensorShape(shapeExpandBatchMajor));
+ }
+ else
+ {
+ const_cast<arm_compute::TensorInfo*>(&aclInputInfo)->set_tensor_shape(
+ BuildArmComputeTensorShape(shapeExpandTimeMajor));
+ }
+ }
+
+ //
+ // Permute validate
+ //
+ if (!descriptor.m_TimeMajor)
+ {
+ // Output now time major. Permute output back to batch major.
+ if (maxTime != 1)
+ {
+ statusPermute2 = arm_compute::NEPermute::validate(&aclConcatOutputTensorInfo,
+ &aclOutputInfo,
+ arm_compute::PermutationVector(0U, 2U, 1U));
+ }
+ else
+ {
+ statusPermute2 = arm_compute::NEPermute::validate(concatInputsTensorInfosPtr[0],
+ &aclOutputInfo,
+ arm_compute::PermutationVector(0U, 2U, 1U));
+ }
+ }
+
+ auto okCode = arm_compute::ErrorCode::OK;
+ if (statusPermute1.error_code() == okCode &&
+ statusSplit.error_code() == okCode &&
+ statusLSTM .error_code() == okCode &&
+ statusConcat.error_code() == okCode &&
+ statusPermute2.error_code() == okCode)
+ {
+ return arm_compute::Status(arm_compute::ErrorCode::OK,
+ "All Unidirectional Sequence LSTM layer validate status OK.");
+ }
+ else
+ {
+ return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR,
+ "Unidirectional Sequence LSTM layer validate status failed.");
+ }
+}
+
+void NeonUnidirectionalSequenceLstmWorkload::FreeUnusedTensors()
+{
+ FreeTensorIfUnused(m_InputToInputWeightsTensor);
+ FreeTensorIfUnused(m_InputToForgetWeightsTensor);
+ FreeTensorIfUnused(m_InputToCellWeightsTensor);
+ FreeTensorIfUnused(m_InputToOutputWeightsTensor);
+ FreeTensorIfUnused(m_RecurrentToInputWeightsTensor);
+ FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor);
+ FreeTensorIfUnused(m_RecurrentToCellWeightsTensor);
+ FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor);
+ FreeTensorIfUnused(m_CellToInputWeightsTensor);
+ FreeTensorIfUnused(m_CellToForgetWeightsTensor);
+ FreeTensorIfUnused(m_CellToOutputWeightsTensor);
+ FreeTensorIfUnused(m_InputGateBiasTensor);
+ FreeTensorIfUnused(m_ForgetGateBiasTensor);
+ FreeTensorIfUnused(m_CellBiasTensor);
+ FreeTensorIfUnused(m_OutputGateBiasTensor);
+ FreeTensorIfUnused(m_ProjectionWeightsTensor);
+ FreeTensorIfUnused(m_ProjectionBiasTensor);
+ FreeTensorIfUnused(m_InputLayerNormWeightsTensor);
+ FreeTensorIfUnused(m_ForgetLayerNormWeightsTensor);
+ FreeTensorIfUnused(m_CellLayerNormWeightsTensor);
+ FreeTensorIfUnused(m_OutputLayerNormWeightsTensor);
+}
+
+} //namespace armnn