DepthConcatSubTensorMutator Class Reference

Mutation pass to optimize depth concatenation operations by using sub-tensors. More...

#include <DepthConcatSubTensorMutator.h>

Collaboration diagram for DepthConcatSubTensorMutator:

Public Member Functions
virtual void	mutate (Graph &g) override
	Walk the graph and perform a specific mutation. More...
MutationType	type () const override
	Returns mutation type. More...
const char *	name () override
	Returns mutator name. More...
Public Member Functions inherited from IGraphMutator
virtual	~IGraphMutator ()=default
	Virtual Destructor. More...

Additional Inherited Members
Public Types inherited from IGraphMutator
enum	MutationType { IR, Backend }
	Mutation type. More...

Detailed Description

Mutation pass to optimize depth concatenation operations by using sub-tensors.

Warning

Always run as one of the last mutation pass as optimizations might change the parent of sub-tensors.

Definition at line 37 of file DepthConcatSubTensorMutator.h.

Member Function Documentation

mutate()

void mutate ( Graph &  g )

overridevirtual

Walk the graph and perform a specific mutation.

Parameters

[in,out]

g

Graph to walk and mutate

Implements IGraphMutator.

Definition at line 50 of file DepthConcatSubTensorMutator.cpp.

{
    // Early exit if no Concatenation layers exist in graph
    if(g.nodes(NodeType::ConcatenateLayer).empty())
    {
        return;
    }
 
    // Perform topological sort
    std::vector<NodeID> topological_sorted_node_ids = dfs(g);
 
    // Should be in reverse order of execution
    for(auto &node_id : arm_compute::utils::iterable::reverse_iterate(topological_sorted_node_ids))
    {
        INode *node = g.node(node_id);
        if(node != nullptr && node->type() == NodeType::ConcatenateLayer && node->output(0) != nullptr)
        {
            // Get output tensor
            auto output_tensor = node->output(0);
 
            // Check concatenation axis (Sub-tensor optimization is supported for concatenation axis >=2)
            auto *concat_node = arm_compute::utils::cast::polymorphic_downcast<ConcatenateLayerNode *>(node);
            if(output_tensor == nullptr || get_dimension_idx(output_tensor->desc().layout, concat_node->concatenation_axis()) < 2)
            {
                continue;
            }
 
            // Check that all tensor have the same target, valid inputs and same quantization info
            bool is_valid = std::all_of(node->input_edges().cbegin(), node->input_edges().cend(),
                                        [&](const EdgeID & eid)
            {
                return (g.edge(eid) != nullptr) && (g.edge(eid)->tensor() != nullptr) && (g.edge(eid)->tensor()->desc().target == output_tensor->desc().target)
                       && (g.edge(eid)->tensor()->desc().quant_info == output_tensor->desc().quant_info);
            });
 
            // Create subtensors
            if(is_valid && is_target_supported(output_tensor->desc().target))
            {
                ARM_COMPUTE_LOG_GRAPH_VERBOSE("Using sub-tensors for the node with ID : "
                                              << node->id() << " and name : " << node->name() << std::endl);
                // Create sub-tensor handles
                unsigned depth = 0;
                for(unsigned int i = 0; i < node->input_edges().size(); ++i)
                {
                    auto       input_tensor = node->input(i);
                    const auto input_shape  = input_tensor->desc().shape;
 
                    backends::IDeviceBackend      &backend = backends::BackendRegistry::get().get_backend(input_tensor->desc().target);
                    std::unique_ptr<ITensorHandle> handle  = backend.create_subtensor(output_tensor->handle(), input_shape, Coordinates(0, 0, depth), false);
                    input_tensor->set_handle(std::move(handle));
 
                    depth += input_shape.z();
                }
 
                auto *dc_node = arm_compute::utils::cast::polymorphic_downcast<ConcatenateLayerNode *>(node);
                dc_node->set_enabled(false);
            }
        }
    }
}

References ARM_COMPUTE_LOG_GRAPH_VERBOSE, arm_compute::graph::ConcatenateLayer, IDeviceBackend::create_subtensor(), arm_compute::graph::dfs(), BackendRegistry::get(), BackendRegistry::get_backend(), arm_compute::graph::get_dimension_idx(), INode::id(), INode::input(), INode::input_edges(), arm_compute::test::validation::input_shape, arm_compute::graph::is_target_supported(), arm_compute::test::validation::is_valid, INode::name(), Graph::node(), Graph::nodes(), INode::output(), arm_compute::utils::iterable::reverse_iterate(), and INode::type().

name()

const char * name ( )

overridevirtual

Returns mutator name.

Returns

Mutator name

Implements IGraphMutator.

Definition at line 40 of file DepthConcatSubTensorMutator.cpp.

{
    return "DepthConcatSubTensorMutator";
}

type()

IGraphMutator::MutationType type ( ) const

overridevirtual

Returns mutation type.

Returns

Mutation type enumeration

Implements IGraphMutator.

Definition at line 45 of file DepthConcatSubTensorMutator.cpp.

{
    return IGraphMutator::MutationType::Backend;
}

References IGraphMutator::Backend.

---

The documentation for this class was generated from the following files:

• arm_compute/graph/mutators/DepthConcatSubTensorMutator.h
• src/graph/mutators/DepthConcatSubTensorMutator.cpp