Implmenting the prefix-sum on the GPU. (da35b794) · Commits · TNL / tnl-dev

src/core/cuda/cuda-prefix-sum.h

+10 −0

Original line number	Diff line number	Diff line
		@@ -18,7 +18,17 @@
		#ifndef CUDA_PREFIX_SUM_H_
		#define CUDA_PREFIX_SUM_H_

		enum enumPrefixSumType { exclusivePrefixSum = 0,
		inclusivePrefixSum };

		template< typename DataType,
		template< typename T > class Operation,
		typename Index >
		bool cudaPrefixSum( const Index size,
		const Index blockSize,
		const DataType *deviceInput,
		DataType* deviceOutput,
		const enumPrefixSumType prefixSumType = inclusivePrefixSum );


		#endif /* CUDA_PREFIX_SUM_H_ */

src/implementation/core/cuda/cuda-prefix-sum_impl.h

+3 −3

Original line number	Diff line number	Diff line
		@@ -287,11 +287,11 @@ bool cudaGridPrefixSum( enumPrefixSumType prefixSumType,
		template< typename DataType,
		template< typename T > class Operation,
		typename Index >
		bool cudaPrefixSum( const enumPrefixSumType prefixSumType,
		const Index size,
		bool cudaPrefixSum( const Index size,
		const Index blockSize,
		const DataType *deviceInput,
		DataType* deviceOutput )
		DataType* deviceOutput,
		const enumPrefixSumType prefixSumType )
		{
		/****
		* Compute the number of grids

src/implementation/core/vectors/tnlVectorOperationsCuda_impl.h

+12 −2

Original line number	Diff line number	Diff line
		@@ -18,6 +18,8 @@
		#ifndef TNLVECTOROPERATIONSCUDA_IMPL_H_
		#define TNLVECTOROPERATIONSCUDA_IMPL_H_

		#include <core/cuda/cuda-prefix-sum.h>

		template< typename Vector >
		void tnlVectorOperations< tnlCuda >::addElement( Vector& v,
		const typename Vector::IndexType i,
		@@ -563,7 +565,11 @@ void tnlVectorOperations< tnlCuda >::computePrefixSum( Vector& v,
		typename Vector::IndexType begin,
		typename Vector::IndexType end )
		{
		typedef typename Vector::IndexType Index;
		cudaPrefixSum( end - begin,
		256,
		v.getData()[ begin ],
		v.getData()[ begin ],
		inclusivePrefixSum );
		}

		template< typename Vector >
		@@ -571,7 +577,11 @@ void tnlVectorOperations< tnlCuda >::computeExclusivePrefixSum( Vector& v,
		typename Vector::IndexType begin,
		typename Vector::IndexType end )
		{

		cudaPrefixSum( end - begin,
		256,
		v.getData()[ begin ],
		v.getData()[ begin ],
		exclusivePrefixSum );
		}

		#ifdef TEMPLATE_EXPLICIT_INSTANTIATION

tests/unit-tests/core/vectors/tnlVectorOperationsTester.h

+45 −72

Original line number	Diff line number	Diff line
		@@ -34,6 +34,9 @@ template< typename Real, typename Device >
		class tnlVectorOperationsTester : public CppUnit :: TestCase
		{
		public:

		typedef CppUnit::TestCaller< tnlVectorOperationsTester< Real, Device > > TestCallerType;

		tnlVectorOperationsTester(){};

		virtual
		@@ -44,78 +47,26 @@ class tnlVectorOperationsTester : public CppUnit :: TestCase
		CppUnit :: TestSuite* suiteOfTests = new CppUnit :: TestSuite( "tnlVectorOperationsTester" );
		CppUnit :: TestResult result;

		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorMaxTest",
		&tnlVectorOperationsTester :: getVectorMaxTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorMinTest",
		&tnlVectorOperationsTester :: getVectorMinTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorAbsMaxTest",
		&tnlVectorOperationsTester :: getVectorAbsMaxTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorAbsMinTest",
		&tnlVectorOperationsTester :: getVectorAbsMinTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorLpNormTest",
		&tnlVectorOperationsTester :: getVectorLpNormTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorSumTest",
		&tnlVectorOperationsTester :: getVectorSumTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorDifferenceMaxTest",
		&tnlVectorOperationsTester :: getVectorDifferenceMaxTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorDifferenceMinTest",
		&tnlVectorOperationsTester :: getVectorDifferenceMinTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorDifferenceAbsMaxTest",
		&tnlVectorOperationsTester :: getVectorDifferenceAbsMaxTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorDifferenceAbsMinTest",
		&tnlVectorOperationsTester :: getVectorDifferenceAbsMinTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorDifferenceLpNormTest",
		&tnlVectorOperationsTester :: getVectorDifferenceLpNormTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getVectorDifferenceSumTest",
		&tnlVectorOperationsTester :: getVectorDifferenceSumTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"vectorScalarMultiplicationTest",
		&tnlVectorOperationsTester :: vectorScalarMultiplicationTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"getSclaraProductTest",
		&tnlVectorOperationsTester :: getVectorScalarProductTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"alphaXPlusYTest",
		&tnlVectorOperationsTester :: alphaXPlusYTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"alphaXPlusBetaYTest",
		&tnlVectorOperationsTester :: alphaXPlusBetaYTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"alphaXPlusBetaZTest",
		&tnlVectorOperationsTester :: alphaXPlusBetaZTest )
		);
		suiteOfTests -> addTest( new CppUnit :: TestCaller< tnlVectorOperationsTester >(
		"alphaXPlusBetaZPlusYTest",
		&tnlVectorOperationsTester :: alphaXPlusBetaZPlusYTest )
		);
		suiteOfTests -> addTest( new TestCallerType( "getVectorMaxTest", &tnlVectorOperationsTester::getVectorMaxTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorMinTest", &tnlVectorOperationsTester::getVectorMinTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorAbsMaxTest", &tnlVectorOperationsTester::getVectorAbsMaxTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorAbsMinTest", &tnlVectorOperationsTester::getVectorAbsMinTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorLpNormTest", &tnlVectorOperationsTester::getVectorLpNormTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorSumTest", &tnlVectorOperationsTester::getVectorSumTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorDifferenceMaxTest", &tnlVectorOperationsTester::getVectorDifferenceMaxTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorDifferenceMinTest", &tnlVectorOperationsTester::getVectorDifferenceMinTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorDifferenceAbsMaxTest", &tnlVectorOperationsTester::getVectorDifferenceAbsMaxTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorDifferenceAbsMinTest", &tnlVectorOperationsTester::getVectorDifferenceAbsMinTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorDifferenceLpNormTest", &tnlVectorOperationsTester::getVectorDifferenceLpNormTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getVectorDifferenceSumTest", &tnlVectorOperationsTester::getVectorDifferenceSumTest ) );
		suiteOfTests -> addTest( new TestCallerType( "vectorScalarMultiplicationTest", &tnlVectorOperationsTester::vectorScalarMultiplicationTest ) );
		suiteOfTests -> addTest( new TestCallerType( "getSclaraProductTest", &tnlVectorOperationsTester::getVectorScalarProductTest ) );
		suiteOfTests -> addTest( new TestCallerType( "alphaXPlusYTest", &tnlVectorOperationsTester::alphaXPlusYTest ) );
		suiteOfTests -> addTest( new TestCallerType( "alphaXPlusBetaYTest", &tnlVectorOperationsTester::alphaXPlusBetaYTest ) );
		suiteOfTests -> addTest( new TestCallerType( "alphaXPlusBetaZTest", &tnlVectorOperationsTester::alphaXPlusBetaZTest ) );
		suiteOfTests -> addTest( new TestCallerType( "alphaXPlusBetaZPlusYTest", &tnlVectorOperationsTester::alphaXPlusBetaZPlusYTest ) );
		suiteOfTests -> addTest( new TestCallerType( "prefixSumTest", &tnlVectorOperationsTester::prefixSumTest ) );
		suiteOfTests -> addTest( new TestCallerType( "exclusivePrefixSumTest", &tnlVectorOperationsTester::exclusivePrefixSumTest ) );
		return suiteOfTests;
		};

		@@ -418,6 +369,28 @@ class tnlVectorOperationsTester : public CppUnit :: TestCase
		for( int i = 0; i < size; i ++ )
		CPPUNIT_ASSERT( y.getElement( i ) == -1.0 + 3.0 * i );
		};

		void prefixSumTest()
		{
		const int size( 65536 );
		tnlVector< Real, Device > v;

		setOnesSequence( v );
		v.computePrefixSum();
		for( int i = 0; i < size; i++ )
		CPPUNIT_ASSERT( v.getElement( i ) == i );
		};

		void exclusivePrefixSumTest()
		{
		const int size( 65536 );
		tnlVector< Real, Device > v;

		setOnesSequence( v );
		v.computePrefixSum();
		for( int i = 0; i < size; i++ )
		CPPUNIT_ASSERT( v.getElement( i ) == i-1 );
		};
		};

		#else