Utility for implementing parallel algorithms based on colored ranges. More...

#include <dune/assembler/parallel/coloredrangeexecutor.hh>

Public Member Functions
	ColoredRangeExecutor (ColoredRange &coloredRange, std::size_t threadCount, bool useWorkStealing=true)

template<class ThreadImpl >
void	operator() (ThreadImpl &&threadImpl) const

Static Public Attributes
static constexpr bool	isParallel = true

Detailed Description

template<class ColoredRange>
class Dune::Assembler::Experimental::ColoredRangeExecutor< ColoredRange >

Utility for implementing parallel algorithms based on colored ranges.

Template Parameters

ColoredRange Type of the colored range

This allows to execute algorithms containing loops of the underlying range. The algorithm is executed in parallel threads while forEach loops over the range internally the coloring.

This requires that the colored range passed is overlap-free with respect to the given algorithm, in the sense that the parallel thread do not produce race conditions when processing the values stored in a color concurrently. Furthermore the additional code in the threads must also not contain race conditions.

Attention: Data that should be thread-local must be captured by value in the algorithm callback passed to the executor. Otherwise concurrent threads will access a single instance of this data.

Example: Assume that accessElementData(entity) writes some date associated to the entity and that the partition is overlap-free wrt this function. Then a parallel algorithm can e.g. take the form

auto executor = ColoredRangeExecutor(coloredRange, threadCount);

executor([&, threadLocalData](auto&& context) { doSomeSetupForThisThread(); context.forEach(range, [&](const auto& entity) { accessElementData(entity); }); context.barrier().arrive_and_wait(); if (context.threadIndex() == 1) std::cout << "Hello from thread 1!" << std::endl; context.barrier().arrive_and_wait(); doSomeParallelComputation(); context.barrier().arrive_and_wait(); doSomeMoreParallelComputation(); });

Constructor & Destructor Documentation

◆ ColoredRangeExecutor()

template<class ColoredRange >

Dune::Assembler::Experimental::ColoredRangeExecutor< ColoredRange >::ColoredRangeExecutor	(	ColoredRange &	coloredRange,
		std::size_t	threadCount,
		bool	useWorkStealing = `true`
	)