From 737d9ff58b348b11234b6c2363390607d576177d Mon Sep 17 00:00:00 2001 From: Ferran Balaguer Date: Thu, 1 Aug 2019 09:58:08 +0100 Subject: IVGCVSW-3342 Add CL backend support for Quantized_LSTM (16bit cell state) !android-nn-driver:1685 Signed-off-by: Ferran Balaguer Signed-off-by: Matthew Bentham Change-Id: I17278562f72d4b77e22c3af25bf7199b9150a765 --- src/backends/cl/workloads/CMakeLists.txt | 2 + .../cl/workloads/ClQuantizedLstmWorkload.cpp | 158 +++++++++++++++++++++ .../cl/workloads/ClQuantizedLstmWorkload.hpp | 48 +++++++ src/backends/cl/workloads/ClWorkloads.hpp | 1 + 4 files changed, 209 insertions(+) create mode 100644 src/backends/cl/workloads/ClQuantizedLstmWorkload.cpp create mode 100644 src/backends/cl/workloads/ClQuantizedLstmWorkload.hpp (limited to 'src/backends/cl/workloads') diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt index 49a8b177d0..f62600b983 100644 --- a/src/backends/cl/workloads/CMakeLists.txt +++ b/src/backends/cl/workloads/CMakeLists.txt @@ -56,6 +56,8 @@ list(APPEND armnnClBackendWorkloads_sources ClPooling2dWorkload.hpp ClPreluWorkload.cpp ClPreluWorkload.hpp + ClQuantizedLstmWorkload.cpp + ClQuantizedLstmWorkload.hpp ClQuantizeWorkload.cpp ClQuantizeWorkload.hpp ClReshapeWorkload.cpp diff --git a/src/backends/cl/workloads/ClQuantizedLstmWorkload.cpp b/src/backends/cl/workloads/ClQuantizedLstmWorkload.cpp new file mode 100644 index 0000000000..76a6694153 --- /dev/null +++ b/src/backends/cl/workloads/ClQuantizedLstmWorkload.cpp @@ -0,0 +1,158 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "ClQuantizedLstmWorkload.hpp" +#include "ClWorkloadUtils.hpp" + +#include +#include +#include + +namespace armnn +{ + +using namespace armcomputetensorutils; + +arm_compute::Status ClQuantizedLstmWorkloadValidate(const TensorInfo& input, const TensorInfo& previousCellStateIn, + const TensorInfo& previousOutputIn, const TensorInfo& cellStateOut, + const TensorInfo& output, + const QuantizedLstmInputParamsInfo& paramsInfo) +{ + // Inputs + const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); + const arm_compute::TensorInfo aclPreviousCellStateInInfo = BuildArmComputeTensorInfo(previousCellStateIn); + const arm_compute::TensorInfo aclPreviousOutputInInfo = BuildArmComputeTensorInfo(previousOutputIn); + + // Outputs + const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut); + const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); + + // Basic parameters + const arm_compute::TensorInfo aclInputToInputWeightsInfo + = BuildArmComputeTensorInfo(paramsInfo.get_InputToInputWeights()); + const arm_compute::TensorInfo aclInputToForgetWeightsInfo + = BuildArmComputeTensorInfo(paramsInfo.get_InputToForgetWeights()); + const arm_compute::TensorInfo aclInputToCellWeightsInfo + = BuildArmComputeTensorInfo(paramsInfo.get_InputToCellWeights()); + const arm_compute::TensorInfo aclInputToOutputWeightsInfo + = BuildArmComputeTensorInfo(paramsInfo.get_InputToOutputWeights()); + const arm_compute::TensorInfo aclRecurrentToInputWeightsInfo + = BuildArmComputeTensorInfo(paramsInfo.get_RecurrentToInputWeights()); + const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo + = BuildArmComputeTensorInfo(paramsInfo.get_RecurrentToForgetWeights()); + const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo + = BuildArmComputeTensorInfo(paramsInfo.get_RecurrentToCellWeights()); + const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo + = BuildArmComputeTensorInfo(paramsInfo.get_RecurrentToOutputWeights()); + const arm_compute::TensorInfo aclInputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.get_InputGateBias()); + const arm_compute::TensorInfo aclForgetGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.get_ForgetGateBias()); + const arm_compute::TensorInfo aclCellBiasInfo = BuildArmComputeTensorInfo(paramsInfo.get_CellBias()); + const arm_compute::TensorInfo aclOutputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.get_OutputGateBias()); + + return arm_compute::CLLSTMLayerQuantized::validate(&aclInputInfo, &aclInputToInputWeightsInfo, + &aclInputToForgetWeightsInfo, &aclInputToCellWeightsInfo, + &aclInputToOutputWeightsInfo, &aclRecurrentToInputWeightsInfo, + &aclRecurrentToForgetWeightsInfo, &aclRecurrentToCellWeightsInfo, + &aclRecurrentToOutputWeightsInfo, &aclInputGateBiasInfo, + &aclForgetGateBiasInfo, &aclCellBiasInfo, &aclOutputGateBiasInfo, + &aclPreviousCellStateInInfo, &aclPreviousOutputInInfo, + &aclCellStateOutInfo, &aclOutputInfo); +} + +ClQuantizedLstmWorkload::ClQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor &descriptor, + const WorkloadInfo &info): + BaseWorkload(descriptor, info) +{ + m_InputToInputWeightsTensor = std::make_unique(); + BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo()); + + m_InputToForgetWeightsTensor = std::make_unique(); + BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo()); + + m_InputToCellWeightsTensor = std::make_unique(); + BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo()); + + m_InputToOutputWeightsTensor = std::make_unique(); + BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo()); + + m_RecurrentToInputWeightsTensor = std::make_unique(); + BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo()); + + m_RecurrentToForgetWeightsTensor = std::make_unique(); + BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo()); + + m_RecurrentToCellWeightsTensor = std::make_unique(); + BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo()); + + m_RecurrentToOutputWeightsTensor = std::make_unique(); + BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo()); + + m_InputGateBiasTensor = std::make_unique(); + BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo()); + + m_ForgetGateBiasTensor = std::make_unique(); + BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo()); + + m_CellBiasTensor = std::make_unique(); + BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo()); + + m_OutputGateBiasTensor = std::make_unique(); + BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo()); + + const arm_compute::ICLTensor& inputTensor = static_cast(m_Data.m_Inputs[0])->GetTensor(); + arm_compute::ICLTensor& cellStateInTensor = static_cast(m_Data.m_Inputs[1])->GetTensor(); + const arm_compute::ICLTensor& outputStateInTensor = static_cast(m_Data.m_Inputs[2])->GetTensor(); + + arm_compute::ICLTensor& cellStateOutTensor = static_cast(m_Data.m_Outputs[0])->GetTensor(); + arm_compute::ICLTensor& outputStateOutTensor = static_cast(m_Data.m_Outputs[1])->GetTensor(); + + m_QuantizedLstmLayer.configure(&inputTensor, m_InputToInputWeightsTensor.get(), m_InputToForgetWeightsTensor.get(), + m_InputToCellWeightsTensor.get(), m_InputToOutputWeightsTensor.get(), + m_RecurrentToInputWeightsTensor.get(), m_RecurrentToForgetWeightsTensor.get(), + m_RecurrentToCellWeightsTensor.get(), m_RecurrentToOutputWeightsTensor.get(), + m_InputGateBiasTensor.get(), m_ForgetGateBiasTensor.get(), m_CellBiasTensor.get(), + m_OutputGateBiasTensor.get(), &cellStateInTensor, &outputStateInTensor, + &cellStateOutTensor, &outputStateOutTensor); + + InitializeArmComputeClTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights); + InitializeArmComputeClTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights); + InitializeArmComputeClTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights); + InitializeArmComputeClTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights); + InitializeArmComputeClTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights); + InitializeArmComputeClTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights); + InitializeArmComputeClTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights); + InitializeArmComputeClTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights); + InitializeArmComputeClTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias); + InitializeArmComputeClTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias); + InitializeArmComputeClTensorData(*m_CellBiasTensor, m_Data.m_CellBias); + InitializeArmComputeClTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias); + + m_QuantizedLstmLayer.prepare(); + FreeUnusedTensors(); +} + +void ClQuantizedLstmWorkload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT_CL("ClQuantizedLstmWorkload_Execute"); + RunClFunction(m_QuantizedLstmLayer, CHECK_LOCATION()); +} + +void ClQuantizedLstmWorkload::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_InputGateBiasTensor); + FreeTensorIfUnused(m_ForgetGateBiasTensor); + FreeTensorIfUnused(m_CellBiasTensor); + FreeTensorIfUnused(m_OutputGateBiasTensor); +} + +} // namespace armnn \ No newline at end of file diff --git a/src/backends/cl/workloads/ClQuantizedLstmWorkload.hpp b/src/backends/cl/workloads/ClQuantizedLstmWorkload.hpp new file mode 100644 index 0000000000..c7d83755c7 --- /dev/null +++ b/src/backends/cl/workloads/ClQuantizedLstmWorkload.hpp @@ -0,0 +1,48 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include +#include + +#include + +namespace armnn +{ + +arm_compute::Status ClQuantizedLstmWorkloadValidate(const TensorInfo& input, const TensorInfo& previousCellStateIn, + const TensorInfo& previousOutputIn, const TensorInfo& cellStateOut, + const TensorInfo& output, + const QuantizedLstmInputParamsInfo& paramsInfo); + +class ClQuantizedLstmWorkload : public BaseWorkload +{ +public: + ClQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor& descriptor, const WorkloadInfo& info); + void Execute() const override; + +private: + mutable arm_compute::CLLSTMLayerQuantized m_QuantizedLstmLayer; + + std::unique_ptr m_InputToInputWeightsTensor; + std::unique_ptr m_InputToForgetWeightsTensor; + std::unique_ptr m_InputToCellWeightsTensor; + std::unique_ptr m_InputToOutputWeightsTensor; + std::unique_ptr m_RecurrentToInputWeightsTensor; + std::unique_ptr m_RecurrentToForgetWeightsTensor; + std::unique_ptr m_RecurrentToCellWeightsTensor; + std::unique_ptr m_RecurrentToOutputWeightsTensor; + std::unique_ptr m_InputGateBiasTensor; + std::unique_ptr m_ForgetGateBiasTensor; + std::unique_ptr m_CellBiasTensor; + std::unique_ptr m_OutputGateBiasTensor; + + void FreeUnusedTensors(); +}; + +} //namespace armnn + + diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp index 03dffc4edc..1af30ffb34 100644 --- a/src/backends/cl/workloads/ClWorkloads.hpp +++ b/src/backends/cl/workloads/ClWorkloads.hpp @@ -29,6 +29,7 @@ #include "ClPooling2dWorkload.hpp" #include "ClPreluWorkload.hpp" #include "ClQuantizeWorkload.hpp" +#include "ClQuantizedLstmWorkload.hpp" #include "ClReshapeWorkload.hpp" #include "ClResizeWorkload.hpp" #include "ClSoftmaxFloatWorkload.hpp" -- cgit v1.2.1