Node interface made for duplicating a graph, for example in the case of multi-GPU computation. More...

#include "abstract_execution_pipeline.h"

Inheritance diagram for hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >:

Collaboration diagram for hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >:

Public Member Functions
	AbstractExecutionPipeline ()=delete
	Deleted default constructor, an execution pipeline need to be constructed with a compatible graph.

	AbstractExecutionPipeline (std::shared_ptr< Graph< GraphOutput, GraphInputs... >> graph, size_t const &numberGraphDuplications, bool iota=false)
	Constructor to set an execution pipeline with a graph and the total number of graphs. More...

	AbstractExecutionPipeline (std::shared_ptr< Graph< GraphOutput, GraphInputs... >> graph, size_t const &numberGraphDuplications, std::vector< int > const deviceIds, bool automaticStart=false)
	Constructor to set an execution pipeline with a graph, the total number of graphs, the device Ids, and set the execution to automatically start. More...

	AbstractExecutionPipeline (std::string_view const &name, std::shared_ptr< Graph< GraphOutput, GraphInputs... >> const &graph, size_t const &numberGraphDuplications, std::vector< int > const &deviceIds, bool automaticStart=false)
	Constructor to set an execution pipeline with a graph, the total number of graphs, the device Ids and set the execution to automatically start. More...

virtual	~AbstractExecutionPipeline ()=default
	Default destructor.

std::shared_ptr< core::CoreNode >	core () override
	Execution Pipeline accessor core. More...

std::shared_ptr< Graph< GraphOutput, GraphInputs... > > const &	graph () const
	Inner graph accessor. More...

Public Member Functions inherited from hh::behavior::Node
virtual std::string	extraPrintingInformation () const
	Adds node information to print in the dot file. More...

virtual bool	canTerminate ()
	Determine if the node can terminate. More...

Public Member Functions inherited from hh::behavior::SwitchRule< GraphInputs >
virtual bool	sendToGraph (std::shared_ptr< GraphInputs > &data, size_t const &graphId)=0
	Switch rule to determine if data should be sent to the graph graphId managed by an AbstractExecutionPipeline. More...

Private Attributes
std::shared_ptr< Graph< GraphOutput, GraphInputs... > >	graph_ = nullptr
	Original Graph that will be duplicated.

std::shared_ptr< core::CoreDefaultExecutionPipeline< GraphOutput, GraphInputs... > >	coreExecutionPipeline_ = nullptr
	Execution Pipeline core.

Additional Inherited Members
Public Types inherited from hh::behavior::MultiReceivers< GraphInputs... >
using	inputs_t = std::tuple< Inputs... >
	Tuple with the list of input types.

Public Types inherited from hh::behavior::Sender< GraphOutput >
using	output_t = GraphOutput
	Output Type.

Detailed Description

template<class GraphOutput, class ... GraphInputs>
class hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >

Node interface made for duplicating a graph, for example in the case of multi-GPU computation.

The duplicated graph will run concurrently, and their inside nodes is also duplicated. Each graph will be distinguished thanks to a device ID (deviceId), from 0, the original to n the number of duplications minus one.

Also, it is used as tool to enable multi-GPU computation, each graph will be bounded to a specific device id accessible from the graph's inside nodes.

Attention: The template parameters GraphOutput and GraphInputs have to exactly match to the managed graph.; There is no default constructor, because a graph is mandatory.

The code below is from Hedgehog tutorial 5.

template<class MatrixType>
class MultiGPUExecPipeline : public AbstractExecutionPipeline<
    MatrixBlockData<MatrixType, 'p', Order::Column>,
                             MatrixBlockData<MatrixType, 'a', Order::Column>,
                             MatrixBlockData<MatrixType, 'b', Order::Column>> {
private:
    size_t
        numberGraphDuplication_ = 0;
    public:
    MultiGPUExecPipeline(std::shared_ptr<
        Graph<MatrixBlockData<MatrixType, 'p', Order::Column>,
        MatrixBlockData<MatrixType, 'a', Order::Column>,
                         MatrixBlockData<MatrixType, 'b', Order::Column>>> const &graph,
    size_t const &numberGraphDuplications,
        std::vector<int> const &deviceIds) : AbstractExecutionPipeline<
        MatrixBlockData<MatrixType, 'p', Order::Column>,
    MatrixBlockData<MatrixType, 'a', Order::Column>,
    MatrixBlockData<MatrixType, 'b', Order::Column>>("Cuda Execution Pipeline", graph, numberGraphDuplications,
    deviceIds), numberGraphDuplication_(numberGraphDuplications) {}
    virtual ~MultiGPUExecPipeline() = default;
    bool sendToGraph(
        [[maybe_unused]]std::shared_ptr<MatrixBlockData<MatrixType, 'a', Order::Column>> &data,
    [[maybe_unused]]size_t const &graphId) override {
    return data->colIdx() % numberGraphDuplication_ == graphId;
    }
    bool sendToGraph(
        [[maybe_unused]]std::shared_ptr<MatrixBlockData<MatrixType, 'b', Order::Column>> &data,
    [[maybe_unused]]size_t const &graphId) override {
    return data->rowIdx() % numberGraphDuplication_ == graphId;
    }
};
// Main
// GPU Graph
auto cudaMatrixMultiplication = std::make_shared<CUDAComputationGraph<MatrixType>>(n, m, p, blockSize,
numberThreadProduct);
// Execution Pipeline
auto executionPipeline =
    std::make_shared<MultiGPUExecPipeline<MatrixType>>(cudaMatrixMultiplication, deviceIds.size(), deviceIds);

A data send to a specialized AbstractExecutionPipeline, before reaching a graph will first go through SwitchRule::sendToGraph that will send it to the graph's copy or not.

Virtual functions: SwitchRule::sendToGraph for each input data type

Template Parameters

GraphOutput	Graph output data type
GraphInputs	Graph inputs data type

Definition at line 97 of file abstract_execution_pipeline.h.

Constructor & Destructor Documentation

◆ AbstractExecutionPipeline() [1/3]

template<class GraphOutput, class ... GraphInputs>

hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >::AbstractExecutionPipeline	(	std::shared_ptr< Graph< GraphOutput, GraphInputs... >>	graph,
		size_t const &	numberGraphDuplications,
		bool	iota = `false`
	)

inline

Constructor to set an execution pipeline with a graph and the total number of graphs.

If iota is set to true, the device ids are initialized from 0 to numberGraphDuplications - 1, if iota is set to false, the device Ids are all set to 0.

Parameters

graph	Inside Base Graph
numberGraphDuplications	Total number of graphs inside the execution pipeline (numberGraphDuplications - 1 duplications)
iota	If true the device Id are set from 0 to numberGraphDuplications - 1 (using std::iota), else all device Ids are set to 0.

Definition at line 123 of file abstract_execution_pipeline.h.

◆ AbstractExecutionPipeline() [2/3]

template<class GraphOutput, class ... GraphInputs>

hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >::AbstractExecutionPipeline	(	std::shared_ptr< Graph< GraphOutput, GraphInputs... >>	graph,
		size_t const &	numberGraphDuplications,
		std::vector< int > const	deviceIds,
		bool	automaticStart = `false`
	)

inline

Constructor to set an execution pipeline with a graph, the total number of graphs, the device Ids, and set the execution to automatically start.

Parameters

graph	Inside Base Graph
numberGraphDuplications	Total number of graphs inside the execution pipeline (numberGraphDuplications - 1 duplications)
deviceIds	Vector of Ids, corresponding to device Ids for each inside graphs
automaticStart	If True, the execution pipeline will be automatically started by sending nullptr to the SwitchRule::sendToGraph, else not

Definition at line 143 of file abstract_execution_pipeline.h.

◆ AbstractExecutionPipeline() [3/3]

template<class GraphOutput, class ... GraphInputs>

hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >::AbstractExecutionPipeline	(	std::string_view const &	name,
		std::shared_ptr< Graph< GraphOutput, GraphInputs... >> const &	graph,
		size_t const &	numberGraphDuplications,
		std::vector< int > const &	deviceIds,
		bool	automaticStart = `false`
	)

inline

Constructor to set an execution pipeline with a graph, the total number of graphs, the device Ids and set the execution to automatically start.

Parameters

name	Execution pipeline's name
graph	Inside Base Graph
numberGraphDuplications	Total number of graphs inside the execution pipeline (numberGraphDuplications - 1 duplications)
deviceIds	Vector of Ids, corresponding to device Ids for each inside graphs
automaticStart	If True, the execution pipeline will be automatically started by sending nullptr to the SwitchRule::sendToGraph, else not

Definition at line 163 of file abstract_execution_pipeline.h.

Member Function Documentation

◆ core()

template<class GraphOutput, class ... GraphInputs>

std::shared_ptr<core::CoreNode> hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >::core ( )

inlineoverridevirtual

Execution Pipeline accessor core.

Returns: Execution Pipeline's core

Implements hh::behavior::Node.

Definition at line 183 of file abstract_execution_pipeline.h.

◆ graph()

template<class GraphOutput, class ... GraphInputs>

std::shared_ptr<Graph<GraphOutput, GraphInputs...> > const& hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >::graph ( ) const

inline

Inner graph accessor.

Returns: Inner graph

Definition at line 187 of file abstract_execution_pipeline.h.

Here is the caller graph for this function:

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

/home/tjb3/work/hedgehog-nist/hedgehog/api/abstract_execution_pipeline.h

Public Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

template<class GraphOutput, class ... GraphInputs> class hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >

Constructor & Destructor Documentation

◆ AbstractExecutionPipeline() [1/3]

◆ AbstractExecutionPipeline() [2/3]

◆ AbstractExecutionPipeline() [3/3]

Member Function Documentation

◆ core()

◆ graph()

template<class GraphOutput, class ... GraphInputs>
class hh::AbstractExecutionPipeline< GraphOutput, GraphInputs >