Commit 0dd215e8 authored by Jakub Klinkovský's avatar Jakub Klinkovský Committed by Jakub Klinkovský
Browse files

DistributedNDArray: added forOverlaps method

parent c6221549

src/TNL/Containers/DistributedNDArray.h

+24 −8
Original line number Diff line number Diff line
@@ -258,13 +258,13 @@ public: 
   template< typename Device2 = DeviceType, typename Func >

   void forLocalInternal( Func f ) const

   {

      // add dynamic sizes

      // add overlaps to dynamic sizes

      LocalBeginsType begins;

      __ndarray_impl::SetSizesAddHelper< 1, LocalBeginsType, SizesHolderType, Overlaps >::add( begins, localBegins, false );

      __ndarray_impl::SetSizesAddOverlapsHelper< LocalBeginsType, SizesHolderType, Overlaps >::add( begins, localBegins );



      // subtract dynamic sizes

      // subtract overlaps from dynamic sizes

      SizesHolderType ends;

      __ndarray_impl::SetSizesSubtractHelper< 1, SizesHolderType, SizesHolderType, Overlaps >::subtract( ends, localEnds, false );

      __ndarray_impl::SetSizesSubtractOverlapsHelper< SizesHolderType, SizesHolderType, Overlaps >::subtract( ends, localEnds );



      __ndarray_impl::ExecutorDispatcher< PermutationType, Device2 > dispatch;

      dispatch( begins, ends, f );
@@ -274,18 +274,34 @@ public: 
   template< typename Device2 = DeviceType, typename Func >

   void forLocalBoundary( Func f ) const

   {

      // add dynamic sizes

      // add overlaps to dynamic sizes

      LocalBeginsType skipBegins;

      __ndarray_impl::SetSizesAddHelper< 1, LocalBeginsType, SizesHolderType, Overlaps >::add( skipBegins, localBegins, false );

      __ndarray_impl::SetSizesAddOverlapsHelper< LocalBeginsType, SizesHolderType, Overlaps >::add( skipBegins, localBegins );



      // subtract dynamic sizes

      // subtract overlaps from dynamic sizes

      SizesHolderType skipEnds;

      __ndarray_impl::SetSizesSubtractHelper< 1, SizesHolderType, SizesHolderType, Overlaps >::subtract( skipEnds, localEnds, false );

      __ndarray_impl::SetSizesSubtractOverlapsHelper< SizesHolderType, SizesHolderType, Overlaps >::subtract( skipEnds, localEnds );



      __ndarray_impl::BoundaryExecutorDispatcher< PermutationType, Device2 > dispatch;

      dispatch( localBegins, skipBegins, skipEnds, localEnds, f );

   }



   // iterate over elements of overlaps (if all overlaps are 0, it has no effect)

   template< typename Device2 = DeviceType, typename Func >

   void forOverlaps( Func f ) const

   {

      // subtract overlaps from dynamic sizes

      LocalBeginsType begins;

      __ndarray_impl::SetSizesSubtractOverlapsHelper< LocalBeginsType, SizesHolderType, Overlaps >::subtract( begins, localBegins );



      // add overlaps to dynamic sizes

      SizesHolderType ends;

      __ndarray_impl::SetSizesAddOverlapsHelper< SizesHolderType, SizesHolderType, Overlaps >::add( ends, localEnds );



      __ndarray_impl::BoundaryExecutorDispatcher< PermutationType, Device2 > dispatch;

      dispatch( begins, localBegins, localEnds, ends, f );

   }





   // extra methods

src/TNL/Containers/DistributedNDArrayView.h

+24 −8
Original line number Diff line number Diff line
@@ -287,13 +287,13 @@ public: 
   template< typename Device2 = DeviceType, typename Func >

   void forLocalInternal( Func f ) const

   {

      // add dynamic sizes

      // add overlaps to dynamic sizes

      LocalBeginsType begins;

      __ndarray_impl::SetSizesAddHelper< 1, LocalBeginsType, SizesHolderType, Overlaps >::add( begins, localBegins, false );

      __ndarray_impl::SetSizesAddOverlapsHelper< LocalBeginsType, SizesHolderType, Overlaps >::add( begins, localBegins );



      // subtract dynamic sizes

      // subtract overlaps from dynamic sizes

      SizesHolderType ends;

      __ndarray_impl::SetSizesSubtractHelper< 1, SizesHolderType, SizesHolderType, Overlaps >::subtract( ends, localEnds, false );

      __ndarray_impl::SetSizesSubtractOverlapsHelper< SizesHolderType, SizesHolderType, Overlaps >::subtract( ends, localEnds );



      __ndarray_impl::ExecutorDispatcher< PermutationType, Device2 > dispatch;

      dispatch( begins, ends, f );
@@ -303,18 +303,34 @@ public: 
   template< typename Device2 = DeviceType, typename Func >

   void forLocalBoundary( Func f ) const

   {

      // add dynamic sizes

      // add overlaps to dynamic sizes

      LocalBeginsType skipBegins;

      __ndarray_impl::SetSizesAddHelper< 1, LocalBeginsType, SizesHolderType, Overlaps >::add( skipBegins, localBegins, false );

      __ndarray_impl::SetSizesAddOverlapsHelper< LocalBeginsType, SizesHolderType, Overlaps >::add( skipBegins, localBegins );



      // subtract dynamic sizes

      // subtract overlaps from dynamic sizes

      SizesHolderType skipEnds;

      __ndarray_impl::SetSizesSubtractHelper< 1, SizesHolderType, SizesHolderType, Overlaps >::subtract( skipEnds, localEnds, false );

      __ndarray_impl::SetSizesSubtractOverlapsHelper< SizesHolderType, SizesHolderType, Overlaps >::subtract( skipEnds, localEnds );



      __ndarray_impl::BoundaryExecutorDispatcher< PermutationType, Device2 > dispatch;

      dispatch( localBegins, skipBegins, skipEnds, localEnds, f );

   }



   // iterate over elements of overlaps (if all overlaps are 0, it has no effect)

   template< typename Device2 = DeviceType, typename Func >

   void forOverlaps( Func f ) const

   {

      // subtract overlaps from dynamic sizes

      LocalBeginsType begins;

      __ndarray_impl::SetSizesSubtractOverlapsHelper< LocalBeginsType, SizesHolderType, Overlaps >::subtract( begins, localBegins );



      // add overlaps to dynamic sizes

      SizesHolderType ends;

      __ndarray_impl::SetSizesAddOverlapsHelper< SizesHolderType, SizesHolderType, Overlaps >::add( ends, localEnds );



      __ndarray_impl::BoundaryExecutorDispatcher< PermutationType, Device2 > dispatch;

      dispatch( begins, localBegins, localEnds, ends, f );

   }



protected:

   NDArrayView localView;

   CommunicationGroup group = Communicator::NullGroup;

src/TNL/Containers/ndarray/SizesHolder.h

+1 −2
Original line number Diff line number Diff line
@@ -37,7 +37,7 @@ public: 


   void setSize( LevelTag, Index newSize )

   {

      TNL_ASSERT( newSize == 0, );

      TNL_ASSERT_EQ( newSize, 0, "Dynamic size for a static dimension must be 0." );

   }



   __cuda_callable__
@@ -60,7 +60,6 @@ public: 


   void setSize( LevelTag, Index size )

   {

      TNL_ASSERT( size >= 0, );

      this->size = size;

   }

src/TNL/Containers/ndarray/SizesHolderHelpers.h

+77 −37
Original line number Diff line number Diff line
@@ -128,7 +128,7 @@ void assertIndicesInBounds( const SizesHolder& sizes, Index&& i, IndexTypes&&... 
   // shouldn't be declared when compiling without assertions

   constexpr std::size_t level = SizesHolder::getDimension() - sizeof...(indices) - 1;

   const auto size = sizes.template getSize< level >();

   TNL_ASSERT_LT( i, size, "Input error - some index is out of bounds." );

   TNL_ASSERT_LT( i, (Index) size, "Input error - some index is out of bounds." );

#endif

   assertIndicesInBounds( sizes, std::forward< IndexTypes >( indices )... );

}
@@ -156,14 +156,14 @@ void assertIndicesInRange( const SizesHolder1& begins, const SizesHolder2& ends, 
   constexpr std::size_t level = SizesHolder1::getDimension() - sizeof...(indices) - 1;

   const auto begin = begins.template getSize< level >();

   const auto end = ends.template getSize< level >();

   TNL_ASSERT_LE( begin - get<level>( overlaps ), i, "Input error - some index is below the lower bound." );

   TNL_ASSERT_LT( i, end + get<level>( overlaps ), "Input error - some index is above the upper bound." );

   TNL_ASSERT_LE( begin - (decltype(begin)) get<level>( overlaps ), i, "Input error - some index is below the lower bound." );

   TNL_ASSERT_LT( i, end + (decltype(end)) get<level>( overlaps ), "Input error - some index is above the upper bound." );

#endif

   assertIndicesInRange( begins, ends, overlaps, std::forward< IndexTypes >( indices )... );

}





// helper for the forInternal method

// helper for the forInternal and forBoundary methods (NDArray and DistributedNDArray)

template< std::size_t ConstValue,

          typename TargetHolder,

          typename SourceHolder,
@@ -172,15 +172,10 @@ template< std::size_t ConstValue, 
struct SetSizesSubtractHelper

{

   static void subtract( TargetHolder& target,

                         const SourceHolder& source,

                         bool negateOverlaps = true )

                         const SourceHolder& source )

   {

      if( source.template getStaticSize< level >() == 0 ) {

         if( negateOverlaps )

      if( source.template getStaticSize< level >() == 0 )

         target.template setSize< level >( source.template getSize< level >() - ConstValue * ! get< level >( Overlaps{} ) );

         else

            target.template setSize< level >( source.template getSize< level >() - ConstValue * !! get< level >( Overlaps{} ) );

      }

      SetSizesSubtractHelper< ConstValue, TargetHolder, SourceHolder, Overlaps, level - 1 >::subtract( target, source );

   }

};
@@ -192,20 +187,15 @@ template< std::size_t ConstValue, 
struct SetSizesSubtractHelper< ConstValue, TargetHolder, SourceHolder, Overlaps, 0 >

{

   static void subtract( TargetHolder& target,

                         const SourceHolder& source,

                         bool negateOverlaps = true )

                         const SourceHolder& source )

   {

      if( source.template getStaticSize< 0 >() == 0 ) {

         if( negateOverlaps )

      if( source.template getStaticSize< 0 >() == 0 )

         target.template setSize< 0 >( source.template getSize< 0 >() - ConstValue * ! get< 0 >( Overlaps{} ) );

         else

            target.template setSize< 0 >( source.template getSize< 0 >() - ConstValue * !! get< 0 >( Overlaps{} ) );

      }

   }

};





// helper for the forInternal method (DistributedNDArray)

// helper for the forInternal and forBoundary methods (DistributedNDArray)

template< std::size_t ConstValue,

          typename TargetHolder,

          typename SourceHolder,
@@ -214,15 +204,10 @@ template< std::size_t ConstValue, 
struct SetSizesAddHelper

{

   static void add( TargetHolder& target,

                    const SourceHolder& source,

                    bool negateOverlaps = true )

                    const SourceHolder& source )

   {

      if( source.template getStaticSize< level >() == 0 ) {

         if( negateOverlaps )

      if( source.template getStaticSize< level >() == 0 )

         target.template setSize< level >( source.template getSize< level >() + ConstValue * ! get< level >( Overlaps{} ) );

         else

            target.template setSize< level >( source.template getSize< level >() + ConstValue * !! get< level >( Overlaps{} ) );

      }

      SetSizesAddHelper< ConstValue, TargetHolder, SourceHolder, Overlaps, level - 1 >::add( target, source );

   }

};
@@ -234,15 +219,70 @@ template< std::size_t ConstValue, 
struct SetSizesAddHelper< ConstValue, TargetHolder, SourceHolder, Overlaps, 0 >

{

   static void add( TargetHolder& target,

                    const SourceHolder& source,

                    bool negateOverlaps = true )

                    const SourceHolder& source )

   {

      if( source.template getStaticSize< 0 >() == 0 ) {

         if( negateOverlaps )

      if( source.template getStaticSize< 0 >() == 0 )

         target.template setSize< 0 >( source.template getSize< 0 >() + ConstValue * ! get< 0 >( Overlaps{} ) );

         else

            target.template setSize< 0 >( source.template getSize< 0 >() + ConstValue * !! get< 0 >( Overlaps{} ) );

   }

};





// helper for the forLocalInternal, forLocalBoundary and forOverlaps methods (DistributedNDArray)

template< typename TargetHolder,

          typename SourceHolder,

          typename Overlaps = make_constant_index_sequence< TargetHolder::getDimension(), 0 >,

          std::size_t level = TargetHolder::getDimension() - 1 >

struct SetSizesSubtractOverlapsHelper

{

   static void subtract( TargetHolder& target,

                         const SourceHolder& source )

   {

      if( source.template getStaticSize< level >() == 0 )

         target.template setSize< level >( source.template getSize< level >() - get< level >( Overlaps{} ) );

      SetSizesSubtractOverlapsHelper< TargetHolder, SourceHolder, Overlaps, level - 1 >::subtract( target, source );

   }

};



template< typename TargetHolder,

          typename SourceHolder,

          typename Overlaps >

struct SetSizesSubtractOverlapsHelper< TargetHolder, SourceHolder, Overlaps, 0 >

{

   static void subtract( TargetHolder& target,

                         const SourceHolder& source )

   {

      if( source.template getStaticSize< 0 >() == 0 )

         target.template setSize< 0 >( source.template getSize< 0 >() - get< 0 >( Overlaps{} ) );

   }

};





// helper for the forLocalInternal, forLocalBoundary and forOverlaps methods (DistributedNDArray)

template< typename TargetHolder,

          typename SourceHolder,

          typename Overlaps = make_constant_index_sequence< TargetHolder::getDimension(), 0 >,

          std::size_t level = TargetHolder::getDimension() - 1 >

struct SetSizesAddOverlapsHelper

{

   static void add( TargetHolder& target,

                    const SourceHolder& source )

   {

      if( source.template getStaticSize< level >() == 0 )

         target.template setSize< level >( source.template getSize< level >() + get< level >( Overlaps{} ) );

      SetSizesAddOverlapsHelper< TargetHolder, SourceHolder, Overlaps, level - 1 >::add( target, source );

   }

};



template< typename TargetHolder,

          typename SourceHolder,

          typename Overlaps >

struct SetSizesAddOverlapsHelper< TargetHolder, SourceHolder, Overlaps, 0 >

{

   static void add( TargetHolder& target,

                    const SourceHolder& source )

   {

      if( source.template getStaticSize< 0 >() == 0 )

         target.template setSize< 0 >( source.template getSize< 0 >() + get< 0 >( Overlaps{} ) );

   }

};

src/UnitTests/Containers/ndarray/CMakeLists.txt

+11 −0
Original line number Diff line number Diff line
@@ -30,12 +30,23 @@ if( ${BUILD_MPI} ) 
      CUDA_ADD_EXECUTABLE( DistributedNDArrayTest DistributedNDArrayTest.cu

                           OPTIONS ${CXX_TESTS_FLAGS} )

      TARGET_LINK_LIBRARIES( DistributedNDArrayTest ${GTEST_BOTH_LIBRARIES} )



      CUDA_ADD_EXECUTABLE( DistributedNDArrayOverlapsTest DistributedNDArrayOverlapsTest.cu

                           OPTIONS ${CXX_TESTS_FLAGS} )

      TARGET_LINK_LIBRARIES( DistributedNDArrayOverlapsTest ${GTEST_BOTH_LIBRARIES} )

   else()

      ADD_EXECUTABLE( DistributedNDArrayTest DistributedNDArrayTest.cpp )

      TARGET_COMPILE_OPTIONS( DistributedNDArrayTest PRIVATE ${CXX_TESTS_FLAGS} )

      TARGET_LINK_LIBRARIES( DistributedNDArrayTest ${GTEST_BOTH_LIBRARIES} )



      ADD_EXECUTABLE( DistributedNDArrayOverlapsTest DistributedNDArrayOverlapsTest.cpp )

      TARGET_COMPILE_OPTIONS( DistributedNDArrayOverlapsTest PRIVATE ${CXX_TESTS_FLAGS} )

      TARGET_LINK_LIBRARIES( DistributedNDArrayOverlapsTest ${GTEST_BOTH_LIBRARIES} )

   endif()



   SET( mpi_test_parameters -np 4 -H localhost:4 "${EXECUTABLE_OUTPUT_PATH}/DistributedNDArrayTest${CMAKE_EXECUTABLE_SUFFIX}" )

   ADD_TEST( NAME DistributedNDArrayTest COMMAND "mpirun" ${mpi_test_parameters})



   SET( mpi_test_parameters -np 4 -H localhost:4 "${EXECUTABLE_OUTPUT_PATH}/DistributedNDArrayOverlapsTest${CMAKE_EXECUTABLE_SUFFIX}" )

   ADD_TEST( NAME DistributedNDArrayOverlapsTest COMMAND "mpirun" ${mpi_test_parameters})

endif()