Added check of the benchmark results. (2220c328) · Commits · TNL / tnl-dev

src/Benchmarks/Benchmarks.h

+28 −14

Original line number	Diff line number	Diff line
		@@ -202,33 +202,35 @@ public:
		BenchmarkResult & result )
		{
		result.time = std::numeric_limits<double>::quiet_NaN();
		FunctionTimer< Device > functionTimer;
		try {
		if( verbose > 1 ) {
		// run the monitor main loop
		Solvers::SolverMonitorThread monitor_thread( monitor );
		if( this->timing )
		if( this->reset )
		result.time = FunctionTimer< Device, true >::timeFunction( compute, reset, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< true >( compute, reset, loops, minTime, verbose, monitor );
		else
		result.time = FunctionTimer< Device, true >::timeFunction( compute, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< true >( compute, loops, minTime, verbose, monitor );
		else
		if( this->reset )
		result.time = FunctionTimer< Device, false >::timeFunction( compute, reset, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< false >( compute, reset, loops, minTime, verbose, monitor );
		else
		result.time = FunctionTimer< Device, false >::timeFunction( compute, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< false >( compute, loops, minTime, verbose, monitor );
		}
		else {
		if( this->timing )
		if( this->reset )
		result.time = FunctionTimer< Device, true >::timeFunction( compute, reset, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< true >( compute, reset, loops, minTime, verbose, monitor );
		else
		result.time = FunctionTimer< Device, true >::timeFunction( compute, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< true >( compute, loops, minTime, verbose, monitor );
		else
		if( this->reset )
		result.time = FunctionTimer< Device, false >::timeFunction( compute, reset, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< false >( compute, reset, loops, minTime, verbose, monitor );
		else
		result.time = FunctionTimer< Device, false >::timeFunction( compute, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< false >( compute, loops, minTime, verbose, monitor );
		}
		this->performedLoops = functionTimer.getPerformedLoops();
		}
		catch ( const std::exception& e ) {
		std::cerr << "timeFunction failed due to a C++ exception with description: " << e.what() << std::endl;
		@@ -269,24 +271,25 @@ public:
		BenchmarkResult & result )
		{
		result.time = std::numeric_limits<double>::quiet_NaN();
		FunctionTimer< Device > functionTimer;
		try {
		if( verbose > 1 ) {
		// run the monitor main loop
		Solvers::SolverMonitorThread monitor_thread( monitor );
		if( this->timing )
		result.time = FunctionTimer< Device, true >::timeFunction( compute, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< true >( compute, loops, minTime, verbose, monitor );
		else
		result.time = FunctionTimer< Device, false >::timeFunction( compute, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< false >( compute, loops, minTime, verbose, monitor );
		}
		else {
		if( this->timing )
		result.time = FunctionTimer< Device, true >::timeFunction( compute, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< true >( compute, loops, minTime, verbose, monitor );
		else
		result.time = FunctionTimer< Device, false >::timeFunction( compute, loops, minTime, verbose, monitor );
		result.time = functionTimer. template timeFunction< false >( compute, loops, minTime, verbose, monitor );
		}
		}
		catch ( const std::exception& e ) {
		std::cerr << "timeFunction failed due to a C++ exception with description: " << e.what() << std::endl;
		std::cerr << "Function timer failed due to a C++ exception with description: " << e.what() << std::endl;
		}

		result.bandwidth = datasetSize / result.time;
		@@ -320,6 +323,7 @@ public:
		// each computation has 3 subcolumns
		const int colspan = 3 * numberOfComputations;
		writeErrorMessage( msg, colspan );
		std::cerr << msg << std::endl;
		}

		using Logging::save;
		@@ -330,8 +334,18 @@ public:
		return monitor;
		}

		int getPerformedLoops() const
		{
		return this->performedLoops;
		}

		bool isResetingOn() const
		{
		return reset;
		}

		protected:
		int loops = 1;
		int loops = 1, performedLoops = 0;
		double minTime = 0.0;
		double datasetSize = 0.0;
		double baseTime = 0.0;

src/Benchmarks/FunctionTimer.h

+17 −10

Original line number	Diff line number	Diff line
		@@ -22,17 +22,17 @@ namespace TNL {
		namespace Benchmarks {


		template< typename Device,
		bool timing >
		template< typename Device >
		class FunctionTimer
		{
		public:
		using DeviceType = Device;

		template< typename ComputeFunction,
		template< bool timing,
		typename ComputeFunction,
		typename ResetFunction,
		typename Monitor = TNL::Solvers::IterativeSolverMonitor< double, int > >
		static double
		double
		timeFunction( ComputeFunction compute,
		ResetFunction reset,
		int maxLoops,
		@@ -52,7 +52,6 @@ class FunctionTimer
		reset();
		compute();

		int loops;
		// If we do not perform reset function and don't need
		// the monitor, the timer is not interrupted after each loop.
		if( ! performReset && verbose < 2 )
		@@ -85,7 +84,6 @@ class FunctionTimer
		{
		// abuse the monitor's "time" for loops
		monitor.setTime( loops + 1 );

		reset();

		// Explicit synchronization of the CUDA device
		@@ -104,15 +102,17 @@ class FunctionTimer
		timer.stop();
		}
		}
		std::cerr << loops << std::endl;
		if( timing )
		return timer.getRealTime() / ( double ) loops;
		else
		return std::numeric_limits<double>::quiet_NaN();
		}

		template< typename ComputeFunction,
		template< bool timing,
		typename ComputeFunction,
		typename Monitor = TNL::Solvers::IterativeSolverMonitor< double, int > >
		static double
		double
		timeFunction( ComputeFunction compute,
		int maxLoops,
		const double& minTime,
		@@ -120,8 +120,15 @@ class FunctionTimer
		Monitor && monitor = Monitor() )
		{
		auto noReset = [] () {};
		return timeFunction( compute, noReset, maxLoops, minTime, verbose, monitor, false );
		return timeFunction< timing >( compute, noReset, maxLoops, minTime, verbose, monitor, false );
		}

		int getPerformedLoops() const
		{
		return this->loops;
		}
		protected:
		int loops;
		};

		} // namespace Benchmarks

src/Benchmarks/Traversers/GridTraversersBenchmark_1D.h

+9 −1

Original line number	Diff line number	Diff line
		@@ -54,12 +54,12 @@ class GridTraversersBenchmark< 1, Device, Real, Index >
		userData( this->u )
		{
		v_data = v.getData();
		u->getData().bind( v );
		}

		void reset()
		{
		v.setValue( 0.0 );
		u->getData().setValue( 0.0 );
		};

		void addOneUsingPureC()
		@@ -146,6 +146,14 @@ class GridTraversersBenchmark< 1, Device, Real, Index >
		size );*/
		}

		bool checkAddOne( int loops, bool reseting )
		{
		std::cout << loops << " -> " << v << std::endl;
		if( reseting )
		return v.containsOnlyValue( 1.0 );
		return v.containsOnlyValue( ( Real ) loops );
		}

		void traverseUsingPureC()
		{
		if( std::is_same< Device, Devices::Host >::value )

src/Benchmarks/Traversers/GridTraversersBenchmark_2D.h

+9 −2

Original line number	Diff line number	Diff line
		@@ -52,12 +52,12 @@ class GridTraversersBenchmark< 2, Device, Real, Index >
		userData( u )
		{
		v_data = v.getData();
		u->getData().bind( v );
		}

		void reset()
		{
		v.setValue( 0.0 );
		u->getData().setValue( 0.0 );
		};

		void addOneUsingPureC()
		@@ -71,7 +71,7 @@ class GridTraversersBenchmark< 2, Device, Real, Index >
		else // Device == Devices::Cuda
		{
		#ifdef HAVE_CUDA
		dim3 blockSize( 256 ), blocksCount, gridsCount;
		dim3 blockSize( 16, 16 ), blocksCount, gridsCount;
		Devices::Cuda::setupThreads(
		blockSize,
		blocksCount,
		@@ -183,6 +183,13 @@ class GridTraversersBenchmark< 2, Device, Real, Index >
		}*/
		}

		bool checkAddOne( int loops, bool reseting )
		{
		if( reseting )
		return v.containsOnlyValue( 1.0 );
		return v.containsOnlyValue( ( Real ) loops );
		}

		void traverseUsingPureC()
		{
		if( std::is_same< Device, Devices::Host >::value )

src/Benchmarks/Traversers/GridTraversersBenchmark_3D.h

+9 −3

Original line number	Diff line number	Diff line
		@@ -58,12 +58,12 @@ class GridTraversersBenchmark< 3, Device, Real, Index >
		userData( u )
		{
		v_data = v.getData();
		u->getData().bind( v );
		}

		void reset()
		{
		v.setValue( 0.0 );
		u->getData().setValue( 0.0 );
		};

		void addOneUsingPureC()
		@@ -78,7 +78,7 @@ class GridTraversersBenchmark< 3, Device, Real, Index >
		else // Device == Devices::Cuda
		{
		#ifdef HAVE_CUDA
		dim3 blockSize( 256 ), blocksCount, gridsCount;
		dim3 blockSize( 32, 4, 2 ), blocksCount, gridsCount;
		Devices::Cuda::setupThreads(
		blockSize,
		blocksCount,
		@@ -174,13 +174,19 @@ class GridTraversersBenchmark< 3, Device, Real, Index >
		f, v.getData() );
		}


		void addOneUsingTraverser()
		{
		traverser.template processAllEntities< UserDataType, AddOneEntitiesProcessorType >
		( grid, userData );
		}

		bool checkAddOne( int loops, bool reseting )
		{
		if( reseting )
		return v.containsOnlyValue( 1.0 );
		return v.containsOnlyValue( ( Real ) loops );
		}

		void traverseUsingPureC()
		{
		if( std::is_same< Device, Devices::Host >::value )