Commit 505d0b68 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Optimized OpenMP thread counts for reduction and multireduction

parent 32c69a11

src/TNL/Containers/Algorithms/Multireduction.hpp

+45 −43
Original line number Diff line number Diff line
@@ -50,8 +50,9 @@ reduce( const Result zero, 
   const int blocks = size / block_size;



#ifdef HAVE_OPENMP

   if( TNL::Devices::Host::isOMPEnabled() && blocks >= 2 )

#pragma omp parallel

   if( Devices::Host::isOMPEnabled() && blocks >= 2 ) {

      const int threads = TNL::min( blocks, Devices::Host::getMaxThreadsCount() );

#pragma omp parallel num_threads(threads)

      {

         // first thread initializes the result array

         #pragma omp single nowait
@@ -105,6 +106,7 @@ reduce( const Result zero, 
               result[ k ] = reduction( result[ k ], r[ 4 * k ] );

         }

      }

   }

   else {

#endif

      if( blocks > 1 ) {

src/TNL/Containers/Algorithms/Reduction.hpp

+6 −4
Original line number Diff line number Diff line
@@ -53,10 +53,11 @@ reduce( const Index size, 
   const int blocks = size / block_size;



#ifdef HAVE_OPENMP

   if( TNL::Devices::Host::isOMPEnabled() && size >= 2 * block_size ) {

   if( Devices::Host::isOMPEnabled() && blocks >= 2 ) {

      // global result variable

      Result result = zero;

#pragma omp parallel

      const int threads = TNL::min( blocks, Devices::Host::getMaxThreadsCount() );

#pragma omp parallel num_threads(threads)

      {

         // initialize array for thread-local results

         Result r[ 4 ] = { zero, zero, zero, zero  };
@@ -145,10 +146,11 @@ reduceWithArgument( const Index size, 
   const int blocks = size / block_size;



#ifdef HAVE_OPENMP

   if( TNL::Devices::Host::isOMPEnabled() && size >= 2 * block_size ) {

   if( Devices::Host::isOMPEnabled() && blocks >= 2 ) {

      // global result variable

      std::pair< Index, Result > result( -1, zero );

#pragma omp parallel

      const int threads = TNL::min( blocks, Devices::Host::getMaxThreadsCount() );

#pragma omp parallel num_threads(threads)

      {

         // initialize array for thread-local results

         Index arg[ 4 ] = { 0, 0, 0, 0 };