Create an optimized version of the network.
1619 if (backendPreferences.empty())
1621 throw InvalidArgumentException(
"Invoked Optimize with no backends specified");
1624 if (options.m_ReduceFp32ToFp16 && options.m_ReduceFp32ToBf16)
1626 throw InvalidArgumentException(
"BFloat16 and Float16 optimization cannot be enabled at the same time.");
1629 std::unique_ptr<Graph> graph = std::make_unique<Graph>(inNetwork.pNetworkImpl->GetGraph());
1637 Graph& optGraph = optNetObjPtr->pOptimizedNetworkImpl->GetGraph();
1642 optGraph.InferTensorInfos();
1646 using namespace optimizations;
1652 optGraph.InferTensorInfos();
1678 if (options.m_ReduceFp32ToFp16)
1687 if (options.m_ReduceFp32ToBf16)
1693 BackendSettings backendSettings(backendPreferences, deviceSpec);
1694 if (backendSettings.GetAvailablePreferredBackends().empty())
1696 std::stringstream failureMsg;
1697 failureMsg <<
"None of the preferred backends " << backendPreferences
1698 <<
" are supported. Current platform provides " << backendSettings.m_SupportedBackends;
1700 throw InvalidArgumentException(failureMsg.str());
1704 TensorHandleFactoryRegistry tensorHandleFactoryRegistry;
1710 OptimizationResult assignBackendsResult =
AssignBackends(optNetObjPtr->pOptimizedNetworkImpl.get(),
1715 if (assignBackendsResult.m_Error)
1718 throw InvalidArgumentException(
"Failed to assign a backend to each layer");
1728 options.m_ModelOptions,
1730 if (backendOptimizationResult.m_Error)
1733 throw InvalidArgumentException(
"Failed to apply the backend-specific optimizations");
1738 if (options.m_Debug)
1746 tensorHandleFactoryRegistry,
1747 options.m_ImportEnabled,
1749 if (strategyResult.m_Error)
1756 optGraph.AddCompatibilityLayers(backends, tensorHandleFactoryRegistry);
1763 for (
auto&& chosenBackend : backendSettings.m_SelectedBackends)
1766 auto backendPtr = factoryFun();
1770 auto backendSpecificOptimizations = backendPtr->GetOptimizations();
1773 if (!backendSpecificOptimizations.empty())
1775 Optimizer::Pass(optNetObjPtr->pOptimizedNetworkImpl->GetGraph(), backendSpecificOptimizations);
OptimizeForConnection< Layer, PermuteLayer, SquashEqualSiblingsImpl< PermuteLayer > > SquashEqualPermuteSiblings
void ReportError(const std::string &errorMessage, Optional< std::vector< std::string > &> errorMessages)
OptimizationResult AssignBackends(OptimizedNetworkImpl *optNetObjPtr, BackendSettings &backendSettings, Graph::Iterator &firstLayer, Graph::Iterator &lastLayer, Optional< std::vector< std::string > &> errMessages)
FactoryFunction GetFactory(const BackendId &id) const
OptimizeForConnection< PermuteLayer, PermuteLayer, OptimizeInversePermutesImpl< PermuteLayer > > OptimizeInversePermutes
OptimizeForExclusiveConnection< PadLayer, Convolution2dLayer, pad_fold::FoldPadIntoConvolution2dImpl > FoldPadIntoConvolution2d
Optimizer::Optimizations MakeOptimizations(Args &&... args)
OptimizeForConnection< TransposeLayer, TransposeLayer, OptimizeInversePermutesImpl< TransposeLayer > > OptimizeInverseTransposes
OptimizeForExclusiveConnection< PadLayer, DepthwiseConvolution2dLayer, pad_fold::FoldPadIntoDepthwiseConvolution2dImpl > FoldPadIntoDepthwiseConvolution2d
OptimizeForConnection< TransposeLayer, BatchToSpaceNdLayer, PermuteAndBatchToSpaceAsDepthToSpaceImpl< TransposeLayer > > TransposeAndBatchToSpaceAsDepthToSpace
OptimizeForExclusiveConnection< DepthwiseConvolution2dLayer, BatchNormalizationLayer, FuseBatchNorm< DepthwiseConvolution2dLayer, armnn::DataType::Float32 > > FuseBatchNormIntoDepthwiseConvolution2DFloat32
#define ARMNN_NO_DEPRECATE_WARN_BEGIN
static void Pass(Graph &graph, const Optimizations &optimizations)
OptimizeForExclusiveConnection< DepthwiseConvolution2dLayer, BatchNormalizationLayer, FuseBatchNorm< DepthwiseConvolution2dLayer, armnn::DataType::Float16 > > FuseBatchNormIntoDepthwiseConvolution2DFloat16
OptimizeForExclusiveConnection< Convolution2dLayer, BatchNormalizationLayer, FuseBatchNorm< Convolution2dLayer, armnn::DataType::Float16 > > FuseBatchNormIntoConvolution2DFloat16
OptimizeForExclusiveConnection< Convolution2dLayer, BatchNormalizationLayer, FuseBatchNorm< Convolution2dLayer, armnn::DataType::Float32 > > FuseBatchNormIntoConvolution2DFloat32
BackendRegistry & BackendRegistryInstance()
OptimizeForConnection< Layer, ReshapeLayer, SquashEqualSiblingsImpl< ReshapeLayer > > SquashEqualReshapeSiblings
OptimizeForConnection< Layer, TransposeLayer, MoveTransposeUpImpl > MoveTransposeUp
LayerList::const_iterator Iterator
OptimizeForType< Layer, AddDebugImpl > InsertDebugLayer
OptimizeForConnection< ReshapeLayer, ReshapeLayer, OptimizeConsecutiveReshapesImpl > OptimizeConsecutiveReshapes
OptimizeForConnection< ConvertFp16ToFp32Layer, ConvertFp32ToFp16Layer, OptimizeInverseConversionsImpl > OptimizeInverseConversionsFp16
OptimizeForConnection< PermuteLayer, BatchToSpaceNdLayer, PermuteAndBatchToSpaceAsDepthToSpaceImpl< PermuteLayer > > PermuteAndBatchToSpaceAsDepthToSpace
OptimizeForConnection< Layer, PermuteLayer, MovePermuteUpImpl > MovePermuteUp
ConvertConstants< Float32ToFloat16, IsFloat16Layer > ConvertConstantsFloatToHalf
OptimizeForType< TransposeLayer, TransposeAsReshapeImpl > TransposeAsReshape
#define ARMNN_NO_DEPRECATE_WARN_END
OptimizationResult ApplyBackendOptimizations(OptimizedNetworkImpl *optNetObjPtr, BackendSettings &backendSettings, BackendsMap &backends, const ModelOptions &modelOptions, Optional< std::vector< std::string > &> errMessages)
Validate all output shapes.
std::unique_ptr< IOptimizedNetwork, void(*)(IOptimizedNetwork *network)> IOptimizedNetworkPtr
#define ARMNN_ASSERT(COND)
OptimizeForType< PermuteLayer, PermuteAsReshapeImpl > PermuteAsReshape
OptimizeForConnection< Layer, TransposeLayer, SquashEqualSiblingsImpl< TransposeLayer > > SquashEqualTransposeSiblings
ConvertConstants< Float16ToFloat32, IsFloat32Layer > ConvertConstantsHalfToFloat
BackendsMap CreateSupportedBackends(TensorHandleFactoryRegistry &handleFactoryRegistry, BackendSettings &backendSettings)
OptimizeForConnection< ConvertFp32ToFp16Layer, ConvertFp16ToFp32Layer, OptimizeInverseConversionsImpl > OptimizeInverseConversionsFp32
OptimizeForExclusiveConnection< PadLayer, Pooling2dLayer, pad_fold::FoldPadIntoPooling2dImpl > FoldPadIntoPooling2d
Infer missing output shapes and validate all output shapes.
OptimizeForType< Layer, ConvertFp32NetworkToFp16Impl > Fp32NetworkToFp16Converter
OptimizationResult SelectTensorHandleStrategy(Graph &optGraph, BackendsMap &backends, TensorHandleFactoryRegistry ®istry, bool importEnabled, Optional< std::vector< std::string > &> errMessages)
IOptimizedNetwork(const IOptimizedNetwork &other, const ModelOptions &modelOptions)
OptimizeForType< Layer, AddBroadcastReshapeLayerImpl > AddBroadcastReshapeLayer
static void Destroy(IOptimizedNetwork *network)
OptimizeForType< Layer, ConvertFp32NetworkToBf16Impl > Fp32NetworkToBf16Converter
std::map< BackendId, std::unique_ptr< class IBackendInternal > > BackendsMap
OptimizeForType< FullyConnectedLayer, RedirectMembersToConstantInputsImpl > RedirectMembersToConstantInputs