From 920d7c1820fe783fd0465abd9bfd4948bbbf1e2f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Tom=C3=A1=C5=A1=20Oberhuber?= <oberhuber.tomas@gmail.com>
Date: Fri, 21 Dec 2018 21:47:40 +0100
Subject: [PATCH] Fixed memory bandwidth in traversers benchmark.
---
.../Traversers/GridTraversersBenchmark.h | 63 +++++++++++++++----
.../Traversers/tnl-benchmark-traversers.h | 14 ++---
src/TNL/Meshes/GridDetails/Grid2D.h | 4 +-
src/TNL/Meshes/GridDetails/Grid2D_impl.h | 2 +-
src/TNL/Meshes/GridDetails/Grid3D.h | 4 +-
5 files changed, 64 insertions(+), 23 deletions(-)
diff --git a/src/Benchmarks/Traversers/GridTraversersBenchmark.h b/src/Benchmarks/Traversers/GridTraversersBenchmark.h
index 3302c4cb94..6f1019deba 100644
--- a/src/Benchmarks/Traversers/GridTraversersBenchmark.h
+++ b/src/Benchmarks/Traversers/GridTraversersBenchmark.h
@@ -16,6 +16,10 @@
#include <TNL/Devices/Host.h>
#include <TNL/Devices/Cuda.h>
#include <TNL/Containers/Vector.h>
+#include <TNL/Meshes/Grid.h>
+#include <TNL/Meshes/GridEntityConfig.h>
+#include <TNL/Meshes/Traverser.h>
+#include <TNL/Functions/MeshFunction.h>
namespace TNL {
namespace Benchmarks {
@@ -35,26 +39,52 @@ class GridTraversersBenchmark< 1, Device, Real, Index >
public:
using Vector = Containers::Vector< Real, Device, Index >;
+ using Grid = Meshes::Grid< 1, Real, Device, Index >;
+ using Coordinates = typename Grid::CoordinatesType;
+ using MeshFunction = Functions::MeshFunction< Grid >;
+ using Cell = typename Grid::EntityType< 1, Meshes::GridEntityNoStencilStorage >;
+ using Traverser = Meshes::Traverser< Grid, Cell >;
GridTraversersBenchmark( Index size )
- :v( size ), size( size )
- {}
+ :v( size ), size( size ), grid( size )
+ {
+ }
- void writeOne()
+ void writeOneUsingParallelFor()
{
auto f = [] __cuda_callable__ ( Index i, Real* data )
{
- data[ i ] = i;
+ data[ i ] = 1.0;
};
ParallelFor< Device >::exec( ( Index ) 0, size, f, v.getData() );
}
+ void writeOneUsingTraverser()
+ {
+ class EntitiesProcessor
+ {
+
+ };
+
+ class UserData
+ {
+
+ };
+
+ Traverser traverser;
+ /*traverser.template processAllEntities< UserData, EntitiesProcessor >
+ ( meshPointer,
+ userData );*/
+
+ }
+
protected:
Index size;
Vector v;
+ Grid grid;
};
@@ -66,16 +96,20 @@ class GridTraversersBenchmark< 2, Device, Real, Index >
public:
using Vector = Containers::Vector< Real, Device, Index >;
+ using Grid = Meshes::Grid< 2, Real, Device, Index >;
+ using Coordinates = typename Grid::CoordinatesType;
GridTraversersBenchmark( Index size )
- :size( size ), v( size * size ) { }
+ :size( size ), v( size * size ), grid( size, size )
+ {
+ }
- void writeOne()
+ void writeOneUsingParallelFor()
{
Index _size = this->size;
auto f = [=] __cuda_callable__ ( Index i, Index j, Real* data )
{
- data[ i * _size + j ] = i + j;
+ data[ i * _size + j ] = 1.0;
};
ParallelFor2D< Device >::exec( ( Index ) 0,
@@ -88,8 +122,8 @@ class GridTraversersBenchmark< 2, Device, Real, Index >
protected:
Index size;
-
Vector v;
+ Grid grid;
};
@@ -101,16 +135,22 @@ class GridTraversersBenchmark< 3, Device, Real, Index >
public:
using Vector = Containers::Vector< Real, Device, Index >;
+ using Grid = Meshes::Grid< 3, Real, Device, Index >;
+ using Coordinates = typename Grid::CoordinatesType;
GridTraversersBenchmark( Index size )
- : size( size ), v( size * size * size ) {}
+ : size( size ),
+ v( size * size * size ),
+ grid( size, size, size )
+ {
+ }
- void writeOne()
+ void writeOneUsingParallelFor()
{
Index _size = this->size;
auto f = [=] __cuda_callable__ ( Index i, Index j, Index k, Real* data )
{
- data[ ( i * _size + j ) * _size + k ] = i + j + k;
+ data[ ( i * _size + j ) * _size + k ] = 1.0;
};
ParallelFor3D< Device >::exec( ( Index ) 0,
@@ -126,6 +166,7 @@ class GridTraversersBenchmark< 3, Device, Real, Index >
Index size;
Vector v;
+ Grid grid;
};
diff --git a/src/Benchmarks/Traversers/tnl-benchmark-traversers.h b/src/Benchmarks/Traversers/tnl-benchmark-traversers.h
index 9b69a31639..c6349f5960 100644
--- a/src/Benchmarks/Traversers/tnl-benchmark-traversers.h
+++ b/src/Benchmarks/Traversers/tnl-benchmark-traversers.h
@@ -54,20 +54,20 @@ bool runBenchmark( const Config::ParameterContainer& parameters,
Benchmark::MetadataColumns(
{ {"size", convertToString( size ) }, } ) );
- auto hostWriteOne = [&] ()
+ auto hostWriteOneUsingParallelFor = [&] ()
{
- hostTraverserBenchmark.writeOne();
+ hostTraverserBenchmark.writeOneUsingParallelFor();
};
- auto cudaWriteOne = [&] ()
+ auto cudaWriteOneUsingParallelFor = [&] ()
{
- cudaTraverserBenchmark.writeOne();
+ cudaTraverserBenchmark.writeOneUsingParallelFor();
};
- benchmark.setOperation( "writeOne", size * sizeof( Real ) );
- benchmark.time( reset, "CPU", hostWriteOne );
+ benchmark.setOperation( "write 1 using parallel for", size * sizeof( Real ) / oneGB );
+ benchmark.time( reset, "CPU", hostWriteOneUsingParallelFor );
#ifdef HAVE_CUDA
- benchmark.time( reset, "GPU", cudaWriteOne );
+ benchmark.time( reset, "GPU", cudaWriteOneUsingParallelFor );
#endif
}
diff --git a/src/TNL/Meshes/GridDetails/Grid2D.h b/src/TNL/Meshes/GridDetails/Grid2D.h
index 896b615486..f2dbebc5c7 100644
--- a/src/TNL/Meshes/GridDetails/Grid2D.h
+++ b/src/TNL/Meshes/GridDetails/Grid2D.h
@@ -82,8 +82,8 @@ class Grid< 2, Real, Device, Index > : public Object
/**
* \brief Sets the size of dimensions.
- * \param xSize Size of dimesion x.
- * \param ySize Size of dimesion y.
+ * \param xSize Size of dimension x.
+ * \param ySize Size of dimension y.
*/
void setDimensions( const Index xSize, const Index ySize );
diff --git a/src/TNL/Meshes/GridDetails/Grid2D_impl.h b/src/TNL/Meshes/GridDetails/Grid2D_impl.h
index 49ad910353..41e05d8b5a 100644
--- a/src/TNL/Meshes/GridDetails/Grid2D_impl.h
+++ b/src/TNL/Meshes/GridDetails/Grid2D_impl.h
@@ -43,7 +43,7 @@ Grid< 2, Real, Device, Index >::Grid( const Index xSize, const Index ySize )
: numberOfCells( 0 ),
numberOfNxFaces( 0 ),
numberOfNyFaces( 0 ),
- numberOfFaces( 0 ),
+ numberOfFaces( 0 ),
numberOfVertices( 0 ),
distGrid(nullptr)
{
diff --git a/src/TNL/Meshes/GridDetails/Grid3D.h b/src/TNL/Meshes/GridDetails/Grid3D.h
index 3ddd44735d..617efe7f31 100644
--- a/src/TNL/Meshes/GridDetails/Grid3D.h
+++ b/src/TNL/Meshes/GridDetails/Grid3D.h
@@ -57,8 +57,8 @@ class Grid< 3, Real, Device, Index > : public Object
* \brief See Grid1D::Grid().
*/
Grid();
-
- Grid( const Index xSize, const Index ySize, const Index zSize );
+
+ Grid( const Index xSize, const Index ySize, const Index zSize );
/**
* \brief See Grid1D::getType().
--
GitLab