From 4452baf3d295164877c5810a3867b1d2d79b04f3 Mon Sep 17 00:00:00 2001 From: Cathal Corbett Date: Fri, 13 May 2022 09:55:59 +0100 Subject: IVGCVSW-6260 ConstTensorsAsInput: Fully Connected Cl and Neon support. * IVGCVSW-6940 ConstTensorsAsInput: DepthwiseConvolution2d - Complete Neon and Cl Bug Fix * Bug fix to enable Cl and Neon Backend Compatibility ConstantTensorsAsInputs * Updated Cl and Neon FullyConnected workloads to handle constant weights and bias as inputs rather than reading from member variables. * Prevent non const weights and biases passing CL and NEON validate for Depthwise Convolution. Signed-off-by: Cathal Corbett Change-Id: I0f505ff5998a183152f843d0f6cc74327ba920e7 --- src/backends/cl/ClBackend.cpp | 9 ++-- src/backends/cl/ClBackend.hpp | 2 +- .../workloads/ClDepthwiseConvolutionWorkload.cpp | 19 ++++++- .../cl/workloads/ClFullyConnectedWorkload.cpp | 61 ++++++++++------------ .../cl/workloads/ClFullyConnectedWorkload.hpp | 5 -- 5 files changed, 53 insertions(+), 43 deletions(-) (limited to 'src/backends/cl') diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp index ed6f221511..47990d87dc 100644 --- a/src/backends/cl/ClBackend.cpp +++ b/src/backends/cl/ClBackend.cpp @@ -398,18 +398,19 @@ OptimizationViews ClBackend::OptimizeSubgraphView(const SubgraphView& subgraph, else if (base.GetType() == LayerType::FullyConnected) { FullyConnectedLayer* baseLayer = PolymorphicDowncast(&base); + FullyConnectedDescriptor descriptor = baseLayer->GetParameters(); + // As bias is optional only try to get TensorInfo from input if bias is enabled. Optional biases; - - if (baseLayer->GetParameters().m_BiasEnabled) + if (descriptor.m_BiasEnabled) { - biases = baseLayer->m_Bias->GetTensorInfo(); + biases = baseLayer->GetInputSlot(2).GetConnectedOutputSlot()->GetTensorInfo(); } arm_compute::Status status = ClFullyConnectedWorkloadValidate( baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), - baseLayer->m_Weight->GetTensorInfo(), + baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(), biases, baseLayer->GetParameters(), &activationDesc); diff --git a/src/backends/cl/ClBackend.hpp b/src/backends/cl/ClBackend.hpp index 99fe9069ff..ce56c3025c 100644 --- a/src/backends/cl/ClBackend.hpp +++ b/src/backends/cl/ClBackend.hpp @@ -27,7 +27,7 @@ const BackendCapabilities gpuAccCapabilities("GpuAcc", {"NonConstWeights", false}, {"AsyncExecution", false}, {"ProtectedContentAllocation", true}, - {"ConstantTensorsAsInputs", false}, + {"ConstantTensorsAsInputs", true}, {"PreImportIOTensors", false}, {"ExternallyManagedMemory", true}, {"MultiAxisPacking", false}, diff --git a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp b/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp index 9a4cad3ef0..3a972d3f39 100644 --- a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp +++ b/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp @@ -30,6 +30,15 @@ arm_compute::Status ClDepthwiseConvolutionWorkloadValidate(const TensorInfo& inp const Optional& biases, const ActivationDescriptor* activationDescriptor) { + // The CL implemented workload does support both const and non const + // weights. However, in the case of non const weights we'd have to call + // prepare or configure for each inference which we're not setup to do just yet. + if (!weights.IsConstant()) + { + return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR, + "ArmNN ClDepthwiseConv2dWorkload does not support non constant weights."}; + } + const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); @@ -47,14 +56,22 @@ arm_compute::Status ClDepthwiseConvolutionWorkloadValidate(const TensorInfo& inp std::tie(weightsPermuted, aclDepthMultiplier) = Convert1HWOTensorInfoToAcl(weights, input,descriptor.m_DataLayout); // Convert the weights into the compute library format - const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.m_DataLayout); + arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.m_DataLayout); + aclWeightsInfo.set_are_values_constant(weights.IsConstant()); arm_compute::TensorInfo aclBiasesInfo; arm_compute::TensorInfo* optionalAclBiasesInfo = nullptr; if (descriptor.m_BiasEnabled) { ARMNN_ASSERT(biases.has_value()); + // Same for bias as weights. We don't currently support non const. + if (!biases.value().IsConstant()) + { + return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR, + "ArmNN ClDepthwiseConv2dWorkload does not support non constant bias."}; + } aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); + aclBiasesInfo.set_are_values_constant(biases.value().IsConstant()); optionalAclBiasesInfo = &aclBiasesInfo; } diff --git a/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp b/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp index 017f4fff6b..c2da5f297a 100644 --- a/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp +++ b/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp @@ -23,22 +23,37 @@ arm_compute::Status ClFullyConnectedWorkloadValidate(const TensorInfo& input, const FullyConnectedDescriptor& descriptor, const ActivationDescriptor* activationDescriptor) { + // The CL implemented workload does support both const and non const + // weights. However, in the case of non const weights we'd have to call + // prepare or configure for each inference which we're not setup to do just yet. + if (!weights.IsConstant()) + { + return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR, + "Arm NN ClFullyConnectedWorkload does not support non constant weights."}; + } const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input); const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output); - const arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights); + arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights); + aclWeights.set_are_values_constant(weights.IsConstant()); arm_compute::TensorInfo aclBiases; arm_compute::TensorInfo* optionalAclBiases = nullptr; if (descriptor.m_BiasEnabled) { ARMNN_ASSERT(biases.has_value()); + // Same for bias as weights. We don't currently support non const. + if (!biases.value().IsConstant()) + { + return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR, + "Arm NN ClFullyConnectedWorkload does not support non constant bias."}; + } aclBiases = BuildArmComputeTensorInfo(biases.value()); + aclBiases.set_are_values_constant(biases.value().IsConstant()); optionalAclBiases = &aclBiases; } const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo = ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor, activationDescriptor); - return arm_compute::CLFullyConnectedLayer::validate(&aclInput, &aclWeights, optionalAclBiases, @@ -70,46 +85,34 @@ ClFullyConnectedWorkload::ClFullyConnectedWorkload( detailsInfo, this->GetGuid()); - m_WeightsTensor = std::make_unique(); - BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo()); + m_Data.ValidateInputsOutputs("ClFullyConnectedWorkload", descriptor.m_Parameters.GetNumInputs(), + 1); + arm_compute::ICLTensor& input = static_cast(m_Data.m_Inputs[0])->GetTensor(); + arm_compute::ICLTensor& output = static_cast(m_Data.m_Outputs[0])->GetTensor(); + arm_compute::ICLTensor& weights = PolymorphicDowncast(m_Data.m_Inputs[1])->GetTensor(); + + arm_compute::ICLTensor* bias = nullptr; if (m_Data.m_Parameters.m_BiasEnabled) { - m_BiasesTensor = std::make_unique(); - BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo()); + bias = &PolymorphicDowncast(m_Data.m_Inputs[2])->GetTensor(); } - m_Data.ValidateInputsOutputs("ClFullyConnectedWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast(m_Data.m_Outputs[0])->GetTensor(); - const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor); arm_compute::FullyConnectedLayerInfo fc_info = - ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor.m_Parameters, activationInfo); + ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor.m_Parameters, + activationInfo); { ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClFullyConnectedWorkload_configure"); m_FullyConnectedLayer.configure(clCompileContext, &input, - m_WeightsTensor.get(), - m_BiasesTensor.get(), + &weights, + bias, &output, fc_info); } - - InitializeArmComputeClTensorData(*m_WeightsTensor, m_Data.m_Weight); - - if (m_BiasesTensor) - { - InitializeArmComputeClTensorData(*m_BiasesTensor, m_Data.m_Bias); - } - - // Force Compute Library to perform the necessary copying and reshaping, after which - // delete all the input tensors that will no longer be needed - m_FullyConnectedLayer.prepare(); - FreeUnusedTensors(); } void ClFullyConnectedWorkload::Execute() const @@ -118,10 +121,4 @@ void ClFullyConnectedWorkload::Execute() const RunClFunction(m_FullyConnectedLayer, CHECK_LOCATION()); } -void ClFullyConnectedWorkload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_WeightsTensor); - FreeTensorIfUnused(m_BiasesTensor); -} - } //namespace armnn diff --git a/src/backends/cl/workloads/ClFullyConnectedWorkload.hpp b/src/backends/cl/workloads/ClFullyConnectedWorkload.hpp index 3ab9f986a8..214c7fcb8b 100644 --- a/src/backends/cl/workloads/ClFullyConnectedWorkload.hpp +++ b/src/backends/cl/workloads/ClFullyConnectedWorkload.hpp @@ -35,11 +35,6 @@ public: private: mutable arm_compute::CLFullyConnectedLayer m_FullyConnectedLayer; - - std::unique_ptr m_WeightsTensor; - std::unique_ptr m_BiasesTensor; - - void FreeUnusedTensors(); }; } //namespace armnn -- cgit v1.2.1