Commit 3e785f01 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Merge branch 'ndarray' into 'develop' 

NDArray

See merge request !18
parents 84b7a213 05634323

src/Benchmarks/CMakeLists.txt

+1 −0
Original line number Diff line number Diff line 
add_subdirectory( HeatEquation )

add_subdirectory( BLAS )

add_subdirectory( NDArray )

add_subdirectory( SpMV )

add_subdirectory( DistSpMV )

add_subdirectory( LinearSolvers )

src/Benchmarks/NDArray/CMakeLists.txt

0 → 100644
+17 −0
Original line number Diff line number Diff line 
add_executable( tnl-benchmark-ndarray tnl-benchmark-ndarray.cpp )

target_compile_options( tnl-benchmark-ndarray PRIVATE ${CXX_TESTS_FLAGS} )

install( TARGETS tnl-benchmark-ndarray RUNTIME DESTINATION bin )



add_executable( tnl-benchmark-ndarray-boundary tnl-benchmark-ndarray-boundary.cpp )

target_compile_options( tnl-benchmark-ndarray-boundary PRIVATE ${CXX_TESTS_FLAGS} )

install( TARGETS tnl-benchmark-ndarray-boundary RUNTIME DESTINATION bin )



if( BUILD_CUDA )

   cuda_add_executable( tnl-benchmark-ndarray-cuda tnl-benchmark-ndarray-cuda.cu

                        OPTIONS ${CXX_TESTS_FLAGS} )

   install( TARGETS tnl-benchmark-ndarray-cuda RUNTIME DESTINATION bin )



   cuda_add_executable( tnl-benchmark-ndarray-boundary-cuda tnl-benchmark-ndarray-boundary-cuda.cu

                        OPTIONS ${CXX_TESTS_FLAGS} )

   install( TARGETS tnl-benchmark-ndarray-boundary-cuda RUNTIME DESTINATION bin )

endif()

src/Benchmarks/NDArray/tnl-benchmark-ndarray-boundary-cuda.cu

0 → 100644
+1 −0
Original line number Diff line number Diff line 
#include "tnl-benchmark-ndarray-boundary.h"

src/Benchmarks/NDArray/tnl-benchmark-ndarray-boundary.cpp

0 → 100644
+1 −0
Original line number Diff line number Diff line 
#include "tnl-benchmark-ndarray-boundary.h"

src/Benchmarks/NDArray/tnl-benchmark-ndarray-boundary.h

0 → 100644
+466 −0
Original line number Diff line number Diff line 
/***************************************************************************

                          tnl-benchmark-ndarray-boundary.h  -  description

                             -------------------

    begin                : Feb 9, 2019

    copyright            : (C) 2019 by Tomas Oberhuber et al.

    email                : tomas.oberhuber@fjfi.cvut.cz

 ***************************************************************************/



/* See Copyright Notice in tnl/Copyright */



// Implemented by: Jakub Klinkovsky



#pragma once



#include <TNL/Assert.h>

#include <TNL/Math.h>

#include <TNL/ParallelFor.h>



#include <TNL/Containers/NDArray.h>



#include "../Benchmarks.h"



using namespace TNL;

using namespace TNL::Benchmarks;

using namespace TNL::Containers;

using std::index_sequence;



using value_type = float;

//using index_type = std::size_t;

using index_type = unsigned;



template< typename Array >

void expect_eq_chunked( Array& a, Array& b )

{

   // TODO: use something like EXPECT_EQ

   TNL_ASSERT_EQ( a.getSize(), b.getSize(), "array sizes don't match" );

   if( a.getSize() != b.getSize() )

      return;



   using IndexType = typename Array::IndexType;



   const IndexType chunk_size = 4096;

   for( IndexType c = 0; c < (IndexType) roundUpDivision( a.getSize(), chunk_size ); c++ ) {

      const typename Array::IndexType this_chunk_size = TNL::min( chunk_size, a.getSize() - c * chunk_size );

      Array a_chunk( &a[ c * chunk_size ], this_chunk_size );

      Array b_chunk( &b[ c * chunk_size ], this_chunk_size );

      // TODO: use something like EXPECT_EQ

      TNL_ASSERT_EQ( a_chunk, b_chunk, "chunks are not equal" );

   }

}



template< typename Array >

void expect_eq( Array& a, Array& b )

{

   if( std::is_same< typename Array::DeviceType, TNL::Devices::Cuda >::value ) {

      typename Array::HostType a_host, b_host;

      a_host = a;

      b_host = b;

      expect_eq_chunked( a_host, b_host );

   }

   else {

      expect_eq_chunked( a, b );

   }

}



template< typename Device >

const char* performer()

{

   if( std::is_same< Device, Devices::Host >::value )

      return "CPU";

   else if( std::is_same< Device, Devices::Cuda >::value )

      return "GPU";

   else

      return "unknown";

}



void reset() {}



// NOTE: having the sizes as function parameters keeps the compiler from treating them

// as "compile-time constants" and thus e.g. optimizing the 1D iterations with memcpy



template< typename Device >

void benchmark_1D( Benchmark& benchmark, index_type size = 500000000 )

{

   NDArray< value_type,

            SizesHolder< index_type, 0 >,

            std::make_index_sequence< 1 >,

            Device > a, b;

   a.setSizes( size );

   b.setSizes( size );

   a.getStorageArray().setValue( -1 );

   b.getStorageArray().setValue( 1 );



   auto a_view = a.getView();

   auto b_view = b.getView();



   auto f = [&]() {

      a.forBoundary( [=] __cuda_callable__ ( index_type i ) mutable { a_view( i ) = b_view( i ); } );

      a.forInternal( [=] __cuda_callable__ ( index_type i ) mutable { a_view( i ) = b_view( i ); } );

   };



   const double datasetSize = 2 * size * sizeof(value_type) / oneGB;

   benchmark.setOperation( "1D", datasetSize );

   benchmark.time< Device >( reset, performer< Device >(), f );



   expect_eq( a.getStorageArray(), b.getStorageArray() );

}



template< typename Device >

void benchmark_2D( Benchmark& benchmark, index_type size = 22333 )

{

   NDArray< value_type,

            SizesHolder< index_type, 0, 0 >,

            std::make_index_sequence< 2 >,

            Device > a, b;

   a.setSizes( size, size );

   b.setSizes( size, size );

   a.getStorageArray().setValue( -1 );

   b.getStorageArray().setValue( 1 );



   auto a_view = a.getView();

   auto b_view = b.getView();



   auto f = [&]() {

      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j ) mutable { a_view( i, j ) = b_view( i, j ); } );

      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j ) mutable { a_view( i, j ) = b_view( i, j ); } );

   };



   const double datasetSize = 2 * std::pow( size, 2 ) * sizeof(value_type) / oneGB;

   benchmark.setOperation( "2D", datasetSize );

   benchmark.time< Device >( reset, performer< Device >(), f );



   expect_eq( a.getStorageArray(), b.getStorageArray() );

}



template< typename Device >

void benchmark_3D( Benchmark& benchmark, index_type size = 800 )

{

   NDArray< value_type,

            SizesHolder< index_type, 0, 0, 0 >,

            std::make_index_sequence< 3 >,

            Device > a, b;

   a.setSizes( size, size, size );

   b.setSizes( size, size, size );

   a.getStorageArray().setValue( -1 );

   b.getStorageArray().setValue( 1 );



   auto a_view = a.getView();

   auto b_view = b.getView();



   auto f = [&]() {

      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j, index_type k ) mutable { a_view( i, j, k ) = b_view( i, j, k ); } );

      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j, index_type k ) mutable { a_view( i, j, k ) = b_view( i, j, k ); } );

   };



   const double datasetSize = 2 * std::pow( size, 3 ) * sizeof(value_type) / oneGB;

   benchmark.setOperation( "3D", datasetSize );

   benchmark.time< Device >( reset, performer< Device >(), f );



   expect_eq( a.getStorageArray(), b.getStorageArray() );

}



// TODO: implement general ParallelBoundaryExecutor

//template< typename Device >

//void benchmark_4D( Benchmark& benchmark, index_type size = 150 )

//{

//   NDArray< value_type,

//            SizesHolder< index_type, 0, 0, 0, 0 >,

//            std::make_index_sequence< 4 >,

//            Device > a, b;

//   a.setSizes( size, size, size, size );

//   b.setSizes( size, size, size, size );

//   a.getStorageArray().setValue( -1 );

//   b.getStorageArray().setValue( 1 );

//

//   auto a_view = a.getView();

//   auto b_view = b.getView();

//

//   auto f = [&]() {

//      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l ) mutable { a_view( i, j, k, l ) = b_view( i, j, k, l ); } );

//      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l ) mutable { a_view( i, j, k, l ) = b_view( i, j, k, l ); } );

//   };

//

//   const double datasetSize = 2 * std::pow( size, 4 ) * sizeof(value_type) / oneGB;

//   benchmark.setOperation( "4D", datasetSize );

//   benchmark.time< Device >( reset, performer< Device >(), f );

//

//   expect_eq( a.getStorageArray(), b.getStorageArray() );

//}

//

//template< typename Device >

//void benchmark_5D( Benchmark& benchmark, index_type size = 56 )

//{

//   NDArray< value_type,

//            SizesHolder< index_type, 0, 0, 0, 0, 0 >,

//            std::make_index_sequence< 5 >,

//            Device > a, b;

//   a.setSizes( size, size, size, size, size );

//   b.setSizes( size, size, size, size, size );

//   a.getStorageArray().setValue( -1 );

//   b.getStorageArray().setValue( 1 );

//

//   auto a_view = a.getView();

//   auto b_view = b.getView();

//

//   auto f = [&]() {

//      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l, index_type m ) mutable { a_view( i, j, k, l, m ) = b_view( i, j, k, l, m ); } );

//      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l, index_type m ) mutable { a_view( i, j, k, l, m ) = b_view( i, j, k, l, m ); } );

//   };

//

//   const double datasetSize = 2 * std::pow( size, 5 ) * sizeof(value_type) / oneGB;

//   benchmark.setOperation( "5D", datasetSize );

//   benchmark.time< Device >( reset, performer< Device >(), f );

//

//   expect_eq( a.getStorageArray(), b.getStorageArray() );

//}

//

//template< typename Device >

//void benchmark_6D( Benchmark& benchmark, index_type size = 28 )

//{

//   NDArray< value_type,

//            SizesHolder< index_type, 0, 0, 0, 0, 0, 0 >,

//            std::make_index_sequence< 6 >,

//            Device > a, b;

//   a.setSizes( size, size, size, size, size, size );

//   b.setSizes( size, size, size, size, size, size );

//   a.getStorageArray().setValue( -1 );

//   b.getStorageArray().setValue( 1 );

//

//   auto a_view = a.getView();

//   auto b_view = b.getView();

//

//   auto f = [&]() {

//      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l, index_type m, index_type n ) mutable { a_view( i, j, k, l, m, n ) = b_view( i, j, k, l, m, n ); } );

//      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l, index_type m, index_type n ) mutable { a_view( i, j, k, l, m, n ) = b_view( i, j, k, l, m, n ); } );

//   };

//

//   const double datasetSize = 2 * std::pow( size, 6 ) * sizeof(value_type) / oneGB;

//   benchmark.setOperation( "6D", datasetSize );

//   benchmark.time< Device >( reset, performer< Device >(), f );

//

//   expect_eq( a.getStorageArray(), b.getStorageArray() );

//}





template< typename Device >

void benchmark_2D_perm( Benchmark& benchmark, index_type size = 22333 )

{

   NDArray< value_type,

            SizesHolder< index_type, 0, 0 >,

            std::index_sequence< 1, 0 >,

            Device > a, b;

   a.setSizes( size, size );

   b.setSizes( size, size );

   a.getStorageArray().setValue( -1 );

   b.getStorageArray().setValue( 1 );



   auto a_view = a.getView();

   auto b_view = b.getView();



   auto f = [&]() {

      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j ) mutable { a_view( i, j ) = b_view( i, j ); } );

      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j ) mutable { a_view( i, j ) = b_view( i, j ); } );

   };



   const double datasetSize = 2 * std::pow( size, 2 ) * sizeof(value_type) / oneGB;

   benchmark.setOperation( "2D permuted", datasetSize );

   benchmark.time< Device >( reset, performer< Device >(), f );



   expect_eq( a.getStorageArray(), b.getStorageArray() );

}



template< typename Device >

void benchmark_3D_perm( Benchmark& benchmark, index_type size = 800 )

{

   NDArray< value_type,

            SizesHolder< index_type, 0, 0, 0 >,

            std::index_sequence< 2, 1, 0 >,

            Device > a, b;

   a.setSizes( size, size, size );

   b.setSizes( size, size, size );

   a.getStorageArray().setValue( -1 );

   b.getStorageArray().setValue( 1 );



   auto a_view = a.getView();

   auto b_view = b.getView();



   auto f = [&]() {

      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j, index_type k ) mutable { a_view( i, j, k ) = b_view( i, j, k ); } );

      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j, index_type k ) mutable { a_view( i, j, k ) = b_view( i, j, k ); } );

   };



   const double datasetSize = 2 * std::pow( size, 3 ) * sizeof(value_type) / oneGB;

   benchmark.setOperation( "3D permuted", datasetSize );

   benchmark.time< Device >( reset, performer< Device >(), f );



   expect_eq( a.getStorageArray(), b.getStorageArray() );

}



// TODO: implement general ParallelBoundaryExecutor

//template< typename Device >

//void benchmark_4D_perm( Benchmark& benchmark, index_type size = 150 )

//{

//   NDArray< value_type,

//            SizesHolder< index_type, 0, 0, 0, 0 >,

//            std::index_sequence< 3, 2, 1, 0 >,

//            Device > a, b;

//   a.setSizes( size, size, size, size );

//   b.setSizes( size, size, size, size );

//   a.getStorageArray().setValue( -1 );

//   b.getStorageArray().setValue( 1 );

//

//   auto a_view = a.getView();

//   auto b_view = b.getView();

//

//   auto f = [&]() {

//      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l ) mutable { a_view( i, j, k, l ) = b_view( i, j, k, l ); } );

//      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l ) mutable { a_view( i, j, k, l ) = b_view( i, j, k, l ); } );

//   };

//

//   const double datasetSize = 2 * std::pow( size, 4 ) * sizeof(value_type) / oneGB;

//   benchmark.setOperation( "4D permuted", datasetSize );

//   benchmark.time< Device >( reset, performer< Device >(), f );

//

//   expect_eq( a.getStorageArray(), b.getStorageArray() );

//}

//

//template< typename Device >

//void benchmark_5D_perm( Benchmark& benchmark, index_type size = 56 )

//{

//   NDArray< value_type,

//            SizesHolder< index_type, 0, 0, 0, 0, 0 >,

//            std::index_sequence< 4, 3, 2, 1, 0 >,

//            Device > a, b;

//   a.setSizes( size, size, size, size, size );

//   b.setSizes( size, size, size, size, size );

//   a.getStorageArray().setValue( -1 );

//   b.getStorageArray().setValue( 1 );

//

//   auto a_view = a.getView();

//   auto b_view = b.getView();

//

//   auto f = [&]() {

//      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l, index_type m ) mutable { a_view( i, j, k, l, m ) = b_view( i, j, k, l, m ); } );

//      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l, index_type m ) mutable { a_view( i, j, k, l, m ) = b_view( i, j, k, l, m ); } );

//   };

//

//   const double datasetSize = 2 * std::pow( size, 5 ) * sizeof(value_type) / oneGB;

//   benchmark.setOperation( "5D permuted", datasetSize );

//   benchmark.time< Device >( reset, performer< Device >(), f );

//

//   expect_eq( a.getStorageArray(), b.getStorageArray() );

//}

//

//template< typename Device >

//void benchmark_6D_perm( Benchmark& benchmark, index_type size = 28 )

//{

//   NDArray< value_type,

//            SizesHolder< index_type, 0, 0, 0, 0, 0, 0 >,

//            std::index_sequence< 5, 4, 3, 2, 1, 0 >,

//            Device > a, b;

//   a.setSizes( size, size, size, size, size, size );

//   b.setSizes( size, size, size, size, size, size );

//   a.getStorageArray().setValue( -1 );

//   b.getStorageArray().setValue( 1 );

//

//   auto a_view = a.getView();

//   auto b_view = b.getView();

//

//   auto f = [&]() {

//      a.forBoundary( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l, index_type m, index_type n ) mutable { a_view( i, j, k, l, m, n ) = b_view( i, j, k, l, m, n ); } );

//      a.forInternal( [=] __cuda_callable__ ( index_type i, index_type j, index_type k, index_type l, index_type m, index_type n ) mutable { a_view( i, j, k, l, m, n ) = b_view( i, j, k, l, m, n ); } );

//   };

//

//   const double datasetSize = 2 * std::pow( size, 6 ) * sizeof(value_type) / oneGB;

//   benchmark.setOperation( "6D permuted", datasetSize );

//   benchmark.time< Device >( reset, performer< Device >(), f );

//

//   expect_eq( a.getStorageArray(), b.getStorageArray() );

//}



template< typename Device >

void run_benchmarks( Benchmark& benchmark )

{

   benchmark_1D< Device >( benchmark );

   benchmark_2D< Device >( benchmark );

   benchmark_3D< Device >( benchmark );

//   benchmark_4D< Device >( benchmark );

//   benchmark_5D< Device >( benchmark );

//   benchmark_6D< Device >( benchmark );

   benchmark_2D_perm< Device >( benchmark );

   benchmark_3D_perm< Device >( benchmark );

//   benchmark_4D_perm< Device >( benchmark );

//   benchmark_5D_perm< Device >( benchmark );

//   benchmark_6D_perm< Device >( benchmark );

}



void setupConfig( Config::ConfigDescription & config )

{

   config.addDelimiter( "Benchmark settings:" );

   config.addEntry< String >( "log-file", "Log file name.", "tnl-benchmark-ndarray-boundary.log");

   config.addEntry< String >( "output-mode", "Mode for opening the log file.", "overwrite" );

   config.addEntryEnum( "append" );

   config.addEntryEnum( "overwrite" );

   config.addEntry< int >( "loops", "Number of iterations for every computation.", 10 );

   config.addEntry< int >( "verbose", "Verbose mode.", 1 );

   config.addEntry< String >( "devices", "Run benchmarks on these devices.", "all" );

   config.addEntryEnum( "all" );

   config.addEntryEnum( "host" );

   #ifdef HAVE_CUDA

   config.addEntryEnum( "cuda" );

   #endif



   config.addDelimiter( "Device settings:" );

   Devices::Host::configSetup( config );

   Devices::Cuda::configSetup( config );

}



int main( int argc, char* argv[] )

{

   Config::ParameterContainer parameters;

   Config::ConfigDescription conf_desc;



   setupConfig( conf_desc );



   if( ! parseCommandLine( argc, argv, conf_desc, parameters ) ) {

      conf_desc.printUsage( argv[ 0 ] );

      return EXIT_FAILURE;

   }



   if( ! Devices::Host::setup( parameters ) ||

       ! Devices::Cuda::setup( parameters ) )

      return EXIT_FAILURE;



   const String & logFileName = parameters.getParameter< String >( "log-file" );

   const String & outputMode = parameters.getParameter< String >( "output-mode" );

   const int loops = parameters.getParameter< int >( "loops" );

   const int verbose = parameters.getParameter< int >( "verbose" );



   // open log file

   auto mode = std::ios::out;

   if( outputMode == "append" )

       mode |= std::ios::app;

   std::ofstream logFile( logFileName.getString(), mode );



   // init benchmark and common metadata

   Benchmark benchmark( loops, verbose );



   // prepare global metadata

   Benchmark::MetadataMap metadata = getHardwareMetadata();



   const String devices = parameters.getParameter< String >( "devices" );

   if( devices == "all" || devices == "host" )

      run_benchmarks< Devices::Host >( benchmark );

#ifdef HAVE_CUDA

   if( devices == "all" || devices == "cuda" )

      run_benchmarks< Devices::Cuda >( benchmark );

#endif



   if( ! benchmark.save( logFile ) ) {

      std::cerr << "Failed to write the benchmark results to file '" << parameters.getParameter< String >( "log-file" ) << "'." << std::endl;

      return EXIT_FAILURE;

   }



   return EXIT_SUCCESS;

}