Commit 928da0bf authored by Tomáš Oberhuber's avatar Tomáš Oberhuber
Browse files

Fixing tnlCublasWrapper.

parent fda03f94

src/core/cuda/CMakeLists.txt

+2 −1
Original line number Diff line number Diff line
@@ -2,7 +2,8 @@ set( headers cuda-prefix-sum.h 
             cuda-prefix-sum_impl.h

             cuda-reduction.h             

             cuda-reduction_impl.h

             reduction-operations.h )

             reduction-operations.h

             tnlCublasWrapper.h )



SET( CURRENT_DIR ${CMAKE_SOURCE_DIR}/src/core/cuda ) 

IF( BUILD_CUDA )

src/core/cuda/tnlCublasWrapper.h

+40 −2
Original line number Diff line number Diff line
@@ -18,17 +18,55 @@ 
#ifndef TNLCUBLASWARPER_H

#define	TNLCUBLASWARPER_H



#if defined HAVE_CUBLAS && defined HAVE_CUDA

#include <cublas_v2.h>

#endif



template< typename Real1, 

          typename Real2,

          typename Index >

class tnlCublasWrapper

{

    public:

        static bool sdot( const Real1* v1, const Real2* v2, const Index size, Real1& result)

        static bool dot( const Real1* v1, const Real2* v2, const Index size, Real1& result)

        {

            return false;

        }        

};



#if defined HAVE_CUBLAS && defined HAVE_CUDA



template< typename Index >

class tnlCublasWrapper< float, float, Index >

{

    public:

        static bool dot( const float* v1, const float* v2, const Index size, float& result)

        {



            cublasHandle_t handle;

            cublasCreate( &handle );

            cublasSdot( handle, size, v1, 1, v2, 1, &result );

            cublasDestroy( handle );

            cerr<< "~~~~~~~~~~~~~~~" << endl;

            return false;

        }        

};



template< typename Index >

class tnlCublasWrapper< double, double, Index >

{

    public:

        static bool dot( const double* v1, const double* v2, const Index size, double& result)

        {

            cublasHandle_t handle;

            cublasCreate( &handle );

            cublasDdot( handle, size, v1, 1, v2, 1, &result );

            cublasDestroy( handle );

            cerr<< "~~~~~~~~~~~~~~~" << endl;

            return false;

        }        

};

#endif            



#endif	/* TNLCUBLASWARPER_H */

src/core/vectors/tnlVectorOperationsCuda_impl.h

+5 −3
Original line number Diff line number Diff line
@@ -18,6 +18,7 @@ 
#ifndef TNLVECTOROPERATIONSCUDA_IMPL_H_

#define TNLVECTOROPERATIONSCUDA_IMPL_H_



#include <tnlConfig.h>

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

#include <core/cuda/tnlCublasWrapper.h>
@@ -350,10 +351,11 @@ typename Vector1 :: RealType tnlVectorOperations< tnlCuda > :: getScalarProduct( 
              cerr << "Vector names are " << v1. getName() << " and " << v2. getName() );



   Real result( 0 );

#ifdef HAVE_CUBLAS

#if defined HAVE_CUBLAS && defined HAVE_CUDA

   cerr << endl << "##############" << endl;

   if( tnlCublasWrapper< typename Vector1::RealType,

                         typename Vector2::RealType,

                         typename Vector1::IndexType >::sdot( v1.getData(), v1.getData(), v1.getSize(), result ) )

                         typename Vector1::IndexType >::dot( v1.getData(), v1.getData(), v1.getSize(), result ) )

       return result;

#endif

   tnlParallelReductionScalarProduct< Real, Index > operation;