Commit 994ee56d authored by Tomáš Oberhuber's avatar Tomáš Oberhuber
Browse files

Tuning Ellpack formats.

parent 670f1892

src/matrices/tnlEllpackMatrix_impl.h

+14 −4
Original line number Diff line number Diff line
@@ -687,11 +687,21 @@ class tnlEllpackMatrixDeviceDependentCode< tnlHost > 
                                 const InVector& inVector,

                                 OutVector& outVector )

      {

#ifdef HAVE_OPENMP

#pragma omp parallel for

#endif           

//#ifdef HAVE_OPENMP

//#pragma omp parallel for

//#endif           

//         for( Index row = 0; row < matrix.getRows(); row ++ )

//            outVector[ row ] = matrix.rowVectorProduct( row, inVector );

         Index col;

         for( Index row = 0; row < matrix.getRows(); row ++ )

            outVector[ row ] = matrix.rowVectorProduct( row, inVector );

         {

            outVector[ row ] = 0.0;

            const Index rowEnd = ( row + 1 ) * matrix.rowLengths;

            for( Index i = row * matrix.rowLengths; i < rowEnd; i++ )

               if( ( col = matrix.columnIndexes[ i ] ) < matrix.columns )

                  outVector[ row ] += matrix.values[ i ] * inVector[ col ];

         }



      }

};

tests/benchmarks/tnl-cuda-benchmarks.h

+5 −4
Original line number Diff line number Diff line
@@ -252,6 +252,7 @@ int main( int argc, char* argv[] ) 
   int elements( 0 );

   for( int row = 0; row < size; row++ )

   {

      if( row % 100 == 0 )

         cout << "Row " << row << "/" << size << "     \r" << flush;

      int col = Max( 0, row - elementsPerRow / 2 );   

      for( int element = 0; element < elementsPerRow; element++ )
@@ -266,7 +267,7 @@ int main( int argc, char* argv[] ) 
   }

   cout << endl;

   setCudaTestMatrix< DeviceMatrix >( deviceMatrix, elementsPerRow );

   datasetSize = loops * elements * sizeof( double ) / oneGB;

   datasetSize = loops * elements * ( 2 * sizeof( Real ) + sizeof( int ) ) / oneGB;

   hostVector.setValue( 1.0 );

   deviceVector.setValue( 1.0 );

   cout << "Benchmarking SpMV on CPU: ";
@@ -275,7 +276,7 @@ int main( int argc, char* argv[] ) 
      hostMatrix.vectorProduct( hostVector, hostVector2 );

   timer.stop();

   double hostTime = timer.getTime();

   bandwidth = 2 * datasetSize / loops / timer.getTime();

   bandwidth = datasetSize / timer.getTime();

   cout << timer.getTime() << " => " << bandwidth << " GB/s" << endl;

   

   cout << "Benchmarking SpMV on GPU: ";
@@ -293,7 +294,7 @@ int main( int argc, char* argv[] ) 
      //      cerr << " " << i;

         

   }

   bandwidth = 2 * datasetSize / loops / timer.getTime();

   bandwidth = datasetSize / timer.getTime();

   cout << timer.getTime() << " => " << bandwidth << " GB/s" << " speedup " << hostTime / timer.getTime() << endl;

   

   return EXIT_SUCCESS;