TosaRefBackend.cpp

Go to the documentation of this file.

//
// Copyright © 2022-2025 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
 
#include "TosaRefBackend.hpp"
#include "TosaRefBackendId.hpp"
#include "TosaRefWorkloadFactory.hpp"
#include "TosaRefLayerSupport.hpp"
#include "TosaRefTensorHandleFactory.hpp"
 
#include <tosaCommon/TosaMappings.hpp>
#include <armnn/BackendRegistry.hpp>
#include <armnn/backends/IBackendContext.hpp>
#include <armnn/backends/IMemoryManager.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <backendsCommon/DefaultAllocator.hpp>
#include <backendsCommon/SubgraphUtils.hpp>
 
#include <Optimizer.hpp>
 
namespace armnn
{
 
// Utility function to construct a valid Deleter for TosaSerializationHandler ptrs passed back to ArmNN
template <typename T>
void DeleteAsType(const void* const blob)
{
    delete static_cast<const T*>(blob);
}
 
const BackendId& TosaRefBackend::GetIdStatic()
{
    static const BackendId s_Id{TosaRefBackendId()};
    return s_Id;
}
 
IBackendInternal::IWorkloadFactoryPtr TosaRefBackend::CreateWorkloadFactory(
    const IBackendInternal::IMemoryManagerSharedPtr& memoryManager) const
{
    return std::make_unique<TosaRefWorkloadFactory>(PolymorphicPointerDowncast<TosaRefMemoryManager>(memoryManager));
}
 
IBackendInternal::IWorkloadFactoryPtr TosaRefBackend::CreateWorkloadFactory(
    class TensorHandleFactoryRegistry& tensorHandleFactoryRegistry) const
{
    auto memoryManager = std::make_shared<TosaRefMemoryManager>();
 
    tensorHandleFactoryRegistry.RegisterMemoryManager(memoryManager);
 
    auto factory = std::make_unique<TosaRefTensorHandleFactory>(memoryManager);
    // Register copy and import factory pair
    tensorHandleFactoryRegistry.RegisterCopyAndImportFactoryPair(factory->GetId(), factory->GetId());
    // Register the factory
    tensorHandleFactoryRegistry.RegisterFactory(std::move(factory));
 
    return std::make_unique<TosaRefWorkloadFactory>(PolymorphicPointerDowncast<TosaRefMemoryManager>(memoryManager));
}
 
IBackendInternal::IBackendContextPtr TosaRefBackend::CreateBackendContext(const IRuntime::CreationOptions&) const
{
    return IBackendContextPtr{};
}
 
IBackendInternal::IBackendProfilingContextPtr TosaRefBackend::CreateBackendProfilingContext(
    const IRuntime::CreationOptions&, IBackendProfilingPtr&)
{
    return IBackendProfilingContextPtr{};
}
 
IBackendInternal::IMemoryManagerUniquePtr TosaRefBackend::CreateMemoryManager() const
{
    return std::make_unique<TosaRefMemoryManager>();
}
 
IBackendInternal::ILayerSupportSharedPtr TosaRefBackend::GetLayerSupport() const
{
    static ILayerSupportSharedPtr layerSupport{new TosaRefLayerSupport};
    return layerSupport;
}
 
OptimizationViews TosaRefBackend::OptimizeSubgraphView(const SubgraphView& subgraph,
                                                       const ModelOptions& modelOptions) const
{
    OptimizationViews optimizationViews(modelOptions);
 
    auto handler = std::make_unique<TosaSerializationHandler>();
 
    std::vector<std::string> graphInputs;
    std::vector<std::string> graphOutputs;
    std::vector<TosaSerializationOperator*> operators;
    std::vector<TosaSerializationTensor*> tensors;
    OpBlockSequencer<Layer, TosaSerializationBasicBlock> sequencer;
 
    // These sets are created to check the duplication of tensor and input
    std::unordered_set<std::string> graphInputsSet;
    std::unordered_set<std::string> uniqueTensorNamesSet;
 
    auto it = subgraph.begin();
    while (it != subgraph.end())
    {
        Layer& base = *(PolymorphicDowncast<Layer*>(*it));
        ++it;
 
        if (base.GetType() == LayerType::Input ||
            base.GetType() == LayerType::Output)
        {
            continue;
        }
 
        TosaSerializationBasicBlock* mappings = GetTosaMappingFromLayer(&base);
 
        // Loop through inputs to see if there are any graph inputs, if so save them.
        // If it's an input to the graph "input" can be found in the string.
        for (const std::string& blockInputName : mappings->GetInputs())
        {
            if ((blockInputName.find("input") != std::string::npos) && !graphInputsSet.count(blockInputName))
            {
                graphInputs.push_back(blockInputName);
                graphInputsSet.insert(blockInputName);
            }
        }
 
        // Loop through outputs to see if there are any graph outputs, if so save them.
        // If it's an output to the graph "output" can be found in the string.
        for (const std::string& blockOutputName : mappings->GetOutputs())
        {
            if (blockOutputName.find("output") != std::string::npos)
            {
                graphOutputs.push_back(blockOutputName);
            }
        }
 
        sequencer.Add(&base, mappings);
    }
 
    for (auto & pair : sequencer.Finish())
    {
        auto blockOperators = pair.block->GetOperators();
        operators.insert(operators.end(), blockOperators.begin(), blockOperators.end());
 
        auto blockTensors = pair.block->GetTensors();
 
        // Checking for duplicate tensors, to handle special cases where same input is
        // added to different layer in the architecture.
        for (const auto& tensor : blockTensors)
        {
            const std::string& name = tensor->GetName();
 
            // If the input tensor is already present, we don't need to add to the list
            if (!uniqueTensorNamesSet.count(name))
            {
                uniqueTensorNamesSet.insert(name);
                tensors.push_back(tensor);
            }
        }
    }
 
    // Add all mappings to main block.
    auto* block = new TosaSerializationBasicBlock("main", "main", operators, tensors, graphInputs, graphOutputs);
 
    std::vector<TosaSerializationBasicBlock*> blocks;
    blocks.emplace_back(block);
 
    // Add blocks to the main region.
    auto* region = new TosaSerializationRegion("main", blocks);
    handler->GetRegions().emplace_back(region);
 
    auto compiledBlob =
            std::make_unique<PreCompiledObjectPtr>(handler.release(), DeleteAsType<TosaSerializationHandler>);
 
    IConnectableLayer* preCompiledLayer = optimizationViews.GetINetwork()->AddPrecompiledLayer(
            PreCompiledDescriptor(subgraph.GetNumInputSlots(), subgraph.GetNumOutputSlots()),
            std::move(*compiledBlob),
            armnn::Optional<BackendId>(GetId()),
            "TOSA_Pre_Compiled_Layer");
 
    // Copy the output tensor infos from sub-graph
    for (unsigned int i = 0; i < subgraph.GetNumOutputSlots(); i++)
    {
        preCompiledLayer->GetOutputSlot(i).SetTensorInfo(subgraph.GetIOutputSlot(i)->GetTensorInfo());
    }
 
    optimizationViews.AddSubstitution({ std::move(subgraph), SubgraphView(preCompiledLayer) });
    return optimizationViews;
}
 
 
std::vector<ITensorHandleFactory::FactoryId> TosaRefBackend::GetHandleFactoryPreferences() const
{
    return std::vector<ITensorHandleFactory::FactoryId> { TosaRefTensorHandleFactory::GetIdStatic() };
}
 
void TosaRefBackend::RegisterTensorHandleFactories(class TensorHandleFactoryRegistry& registry)
{
    auto memoryManager = std::make_shared<TosaRefMemoryManager>();
 
    registry.RegisterMemoryManager(memoryManager);
 
    auto factory = std::make_unique<TosaRefTensorHandleFactory>(memoryManager);
 
    // Register copy and import factory pair
    registry.RegisterCopyAndImportFactoryPair(factory->GetId(), factory->GetId());
    // Register the factory
    registry.RegisterFactory(std::move(factory));
}
 
std::unique_ptr<ICustomAllocator> TosaRefBackend::GetDefaultAllocator() const
{
    return std::make_unique<DefaultAllocator>();
}
 
BackendCapabilities TosaRefBackend::GetCapabilities() const
{
    return BackendCapabilities ("TosaRef",
                                {
                                        {"NonConstWeights", true},
                                        {"ConstantTensorsAsInputs", true}
                                });
}
 
} // namespace armnn