From 5f8068294d4acae46259644e9118b6369b349f6b Mon Sep 17 00:00:00 2001
From: Martin Bauer <martin.bauer@fau.de>
Date: Tue, 30 May 2017 13:39:01 +0200
Subject: [PATCH] CUDA: removed explicit instantiation of GPUFields
- easier Python export
---
apps/pythonmodule/PythonModule.cpp | 7 +-
src/cuda/FieldIndexing.h | 1 +
...FieldIndexing.cpp => FieldIndexing.impl.h} | 5 -
src/cuda/FieldIndexing3D.h | 1 +
...dIndexing3D.cpp => FieldIndexing3D.impl.h} | 5 -
src/cuda/FieldIndexingXYZ.h | 1 +
...ndexingXYZ.cpp => FieldIndexingXYZ.impl.h} | 3 -
src/cuda/GPUField.h | 1 +
src/cuda/{GPUField.cpp => GPUField.impl.h} | 4 -
src/cuda/GPUTypesExplicitInstantiation.h | 8 --
src/cuda/python/Exports.h | 2 +-
src/cuda/python/Exports.impl.h | 109 +++++++++++++++---
12 files changed, 102 insertions(+), 45 deletions(-)
rename src/cuda/{FieldIndexing.cpp => FieldIndexing.impl.h} (98%)
rename src/cuda/{FieldIndexing3D.cpp => FieldIndexing3D.impl.h} (98%)
rename src/cuda/{FieldIndexingXYZ.cpp => FieldIndexingXYZ.impl.h} (97%)
rename src/cuda/{GPUField.cpp => GPUField.impl.h} (99%)
delete mode 100644 src/cuda/GPUTypesExplicitInstantiation.h
diff --git a/apps/pythonmodule/PythonModule.cpp b/apps/pythonmodule/PythonModule.cpp
index a648d08c..82fab2e0 100644
--- a/apps/pythonmodule/PythonModule.cpp
+++ b/apps/pythonmodule/PythonModule.cpp
@@ -67,7 +67,6 @@ typedef bmpl::vector<
Field<walberla::uint32_t,1>
> FieldTypes;
-typedef bmpl::vector<double, float, int, uint8_t, uint16_t>CudaFieldTypes;
typedef bmpl::vector<
GhostLayerField<walberla::real_t,1>,
@@ -118,7 +117,11 @@ struct InitObject
pythonManager->addExporterFunction( timeloop::exportModuleToPython );
#ifdef WALBERLA_BUILD_WITH_CUDA
- pythonManager->addExporterFunction( cuda::exportModuleToPython<CudaFieldTypes> );
+ using walberla::cuda::GPUField;
+ typedef bmpl::vector<GPUField<double>, GPUField<float>, GPUField<int>, GPUField<uint8_t>, GPUField<uint16_t> > GPUFields;
+
+ pythonManager->addExporterFunction( cuda::exportModuleToPython<GPUFields> );
+ pythonManager->addBlockDataConversion<GPUFields>();
#endif
python_coupling::initWalberlaForPythonModule();
diff --git a/src/cuda/FieldIndexing.h b/src/cuda/FieldIndexing.h
index 653a5de2..7ed089f4 100644
--- a/src/cuda/FieldIndexing.h
+++ b/src/cuda/FieldIndexing.h
@@ -91,4 +91,5 @@ namespace cuda {
} // namespace cuda
} // namespace walberla
+#include "FieldIndexing.impl.h"
diff --git a/src/cuda/FieldIndexing.cpp b/src/cuda/FieldIndexing.impl.h
similarity index 98%
rename from src/cuda/FieldIndexing.cpp
rename to src/cuda/FieldIndexing.impl.h
index 413bbe1a..c4837d3c 100644
--- a/src/cuda/FieldIndexing.cpp
+++ b/src/cuda/FieldIndexing.impl.h
@@ -20,7 +20,6 @@
//======================================================================================================================
#include "FieldIndexing.h"
-#include "GPUTypesExplicitInstantiation.h"
#include "GPUField.h"
#include "core/cell/CellInterval.h"
@@ -224,10 +223,6 @@ FieldIndexing<T> FieldIndexing<T>::all ( const GPUField<T> & f, const cell::Cell
-
-GPU_CLASS_TEMPLATE_INSTANTIATION( FieldIndexing )
-
-
} // namespace cuda
} // namespace walberla
diff --git a/src/cuda/FieldIndexing3D.h b/src/cuda/FieldIndexing3D.h
index 0dbe9756..c6637ec6 100644
--- a/src/cuda/FieldIndexing3D.h
+++ b/src/cuda/FieldIndexing3D.h
@@ -103,3 +103,4 @@ namespace cuda {
} // namespace walberla
+#include "FieldIndexing.impl.h"
\ No newline at end of file
diff --git a/src/cuda/FieldIndexing3D.cpp b/src/cuda/FieldIndexing3D.impl.h
similarity index 98%
rename from src/cuda/FieldIndexing3D.cpp
rename to src/cuda/FieldIndexing3D.impl.h
index 5a797a23..896f7e1d 100644
--- a/src/cuda/FieldIndexing3D.cpp
+++ b/src/cuda/FieldIndexing3D.impl.h
@@ -20,7 +20,6 @@
//======================================================================================================================
#include "FieldIndexing3D.h"
-#include "GPUTypesExplicitInstantiation.h"
#include "GPUField.h"
#include "core/cell/CellInterval.h"
@@ -165,10 +164,6 @@ FieldIndexing3D<T> FieldIndexing3D<T>::intervalXYZ( const GPUField<T> & f, const
-
-GPU_CLASS_TEMPLATE_INSTANTIATION( FieldIndexing3D )
-
-
} // namespace cuda
} // namespace walberla
diff --git a/src/cuda/FieldIndexingXYZ.h b/src/cuda/FieldIndexingXYZ.h
index 2c25975e..18a6e264 100644
--- a/src/cuda/FieldIndexingXYZ.h
+++ b/src/cuda/FieldIndexingXYZ.h
@@ -77,3 +77,4 @@ template< typename T> class GPUField;
} // namespace walberla
+#include "FieldIndexingXYZ.impl.h"
\ No newline at end of file
diff --git a/src/cuda/FieldIndexingXYZ.cpp b/src/cuda/FieldIndexingXYZ.impl.h
similarity index 97%
rename from src/cuda/FieldIndexingXYZ.cpp
rename to src/cuda/FieldIndexingXYZ.impl.h
index 8cc0bd63..c8ec561f 100644
--- a/src/cuda/FieldIndexingXYZ.cpp
+++ b/src/cuda/FieldIndexingXYZ.impl.h
@@ -20,7 +20,6 @@
//======================================================================================================================
#include "FieldIndexingXYZ.h"
-#include "GPUTypesExplicitInstantiation.h"
#include "GPUField.h"
#include "core/cell/CellInterval.h"
@@ -114,8 +113,6 @@ FieldIndexingXYZ<T> FieldIndexingXYZ<T>::withGhostLayerXYZ( const GPUField<T> &
}
-GPU_CLASS_TEMPLATE_INSTANTIATION( FieldIndexingXYZ )
-
} // namespace cuda
} // namespace walberla
diff --git a/src/cuda/GPUField.h b/src/cuda/GPUField.h
index aa059b53..437fe1c9 100755
--- a/src/cuda/GPUField.h
+++ b/src/cuda/GPUField.h
@@ -149,3 +149,4 @@ namespace cuda {
} // namespace walberla
+#include "GPUField.impl.h"
\ No newline at end of file
diff --git a/src/cuda/GPUField.cpp b/src/cuda/GPUField.impl.h
similarity index 99%
rename from src/cuda/GPUField.cpp
rename to src/cuda/GPUField.impl.h
index fe7c3ed9..b6fe3f8a 100644
--- a/src/cuda/GPUField.cpp
+++ b/src/cuda/GPUField.impl.h
@@ -21,7 +21,6 @@
#include "GPUField.h"
#include "ErrorChecking.h"
-#include "GPUTypesExplicitInstantiation.h"
#include "core/logging/Logging.h"
@@ -249,9 +248,6 @@ void GPUField<T>::swapDataPointers( GPUField<T> & other )
-GPU_CLASS_TEMPLATE_INSTANTIATION( GPUField )
-
-
} // namespace cuda
} // namespace walberla
diff --git a/src/cuda/GPUTypesExplicitInstantiation.h b/src/cuda/GPUTypesExplicitInstantiation.h
deleted file mode 100644
index bdc4b584..00000000
--- a/src/cuda/GPUTypesExplicitInstantiation.h
+++ /dev/null
@@ -1,8 +0,0 @@
-#define GPU_CLASS_TEMPLATE_INSTANTIATION(ClassName)\
- template class ClassName< double >;\
- template class ClassName< float >;\
- template class ClassName< int >;\
- template class ClassName< uint8_t >;\
- template class ClassName< uint16_t >;
-
-
diff --git a/src/cuda/python/Exports.h b/src/cuda/python/Exports.h
index f4cb588a..a2990cfc 100644
--- a/src/cuda/python/Exports.h
+++ b/src/cuda/python/Exports.h
@@ -30,7 +30,7 @@ namespace walberla {
namespace cuda {
- template<typename DataTypes >
+ template<typename GpuFields >
void exportModuleToPython();
diff --git a/src/cuda/python/Exports.impl.h b/src/cuda/python/Exports.impl.h
index 8cd714ed..1d2e2dc2 100644
--- a/src/cuda/python/Exports.impl.h
+++ b/src/cuda/python/Exports.impl.h
@@ -38,6 +38,7 @@
#include <boost/type_traits/is_unsigned.hpp>
#include <iostream>
+#include <cuda/communication/GPUPackInfo.h>
namespace walberla {
namespace cuda {
@@ -66,11 +67,9 @@ namespace internal {
struct GpuFieldExporter
{
- template< typename DataType>
- void operator() ( DataType )
+ template< typename GpuField_T>
+ void operator() ( python_coupling::NonCopyableWrap<GpuField_T> )
{
- typedef GPUField<DataType> GpuField_T;
-
using namespace boost::python;
class_<GpuField_T, shared_ptr<GpuField_T>, boost::noncopyable>( "GpuField", no_init )
@@ -122,11 +121,11 @@ namespace internal {
layout_( layout), type_( type ), usePitchedMem_( usePitchedMem ) , resultPointer_( resultPointer )
{}
- template< typename T>
- void operator() ( T )
+ template< typename GpuField_T>
+ void operator() ( python_coupling::NonCopyableWrap<GpuField_T> )
{
using namespace boost::python;
-
+ typedef typename GpuField_T::value_type T;
if( python_coupling::isCppEqualToPythonType<T>( (PyTypeObject *)type_.ptr() ) )
{
*resultPointer_ = object( make_shared< GPUField<T> >( xs_,ys_,zs_, fs_, gl_, layout_, usePitchedMem_ ) );
@@ -145,7 +144,7 @@ namespace internal {
shared_ptr<boost::python::object> resultPointer_;
};
- template<typename DataTypes>
+ template<typename GpuFields>
boost::python::object createPythonGpuField( boost::python::list size,
boost::python::object type,
uint_t ghostLayers,
@@ -168,7 +167,7 @@ namespace internal {
auto result = make_shared<boost::python::object>();
CreateFieldExporter exporter( xSize,ySize, zSize, fSize, ghostLayers, layout, type, usePitchedMem, result );
- boost::mpl::for_each< DataTypes > ( exporter );
+ python_coupling::for_each_noncopyable_type< GpuFields >( exporter );
if ( *result == object() )
{
@@ -198,9 +197,10 @@ namespace internal {
gl_(gl),layout_( layout), type_( type ), usePitchedMem_(usePitchedMem), found_(false)
{}
- template< typename T>
- void operator() ( T )
+ template< typename GpuField_T>
+ void operator() ( python_coupling::NonCopyableWrap<GpuField_T> )
{
+ typedef typename GpuField_T::value_type T;
if( python_coupling::isCppEqualToPythonType<T>( (PyTypeObject *)type_.ptr() ) )
{
WALBERLA_ASSERT(!found_);
@@ -221,7 +221,7 @@ namespace internal {
bool found_;
};
- template<typename DataTypes>
+ template<typename GpuFields>
void addToStorage( const shared_ptr<StructuredBlockStorage> & blocks, const std::string & name,
boost::python::object type, uint_t fs, uint_t gl, Layout layout, bool usePitchedMem )
{
@@ -234,7 +234,7 @@ namespace internal {
auto result = make_shared<boost::python::object>();
AddToStorageExporter exporter( blocks, name, fs, gl, layout, type, usePitchedMem );
- boost::mpl::for_each<DataTypes>( boost::ref(exporter) );
+ python_coupling::for_each_noncopyable_type<GpuFields>( boost::ref(exporter) );
if ( ! exporter.successful() ) {
PyErr_SetString( PyExc_ValueError, "Adding Field failed.");
@@ -243,29 +243,101 @@ namespace internal {
}
+ //===================================================================================================================
+ //
+ // createPackInfo Export
+ //
+ //===================================================================================================================
+
+ template< typename GPUField_T >
+ boost::python::object createGPUPackInfoToObject( BlockDataID bdId, uint_t numberOfGhostLayers )
+ {
+ using cuda::communication::GPUPackInfo;
+ if ( numberOfGhostLayers > 0 )
+ return boost::python::object( make_shared< GPUPackInfo<GPUField_T> >( bdId, numberOfGhostLayers ) );
+ else
+ return boost::python::object( make_shared< GPUPackInfo<GPUField_T> >( bdId ) );
+ }
+
+ FunctionExporterClass( createGPUPackInfoToObject, boost::python::object( BlockDataID, uint_t ) );
+
+ template< typename GpuFields>
+ boost::python::object createPackInfo( const shared_ptr<StructuredBlockStorage> & bs,
+ const std::string & blockDataName, uint_t numberOfGhostLayers )
+ {
+ using cuda::communication::GPUPackInfo;
+
+ auto bdId = python_coupling::blockDataIDFromString( *bs, blockDataName );
+ if ( bs->begin() == bs->end() ) {
+ // if no blocks are on this field an arbitrary PackInfo can be returned
+ return createGPUPackInfoToObject< GPUField<real_t> > ( bdId, numberOfGhostLayers );
+ }
+
+ IBlock * firstBlock = & ( * bs->begin() );
+ python_coupling::Dispatcher<GpuFields, Exporter_createGPUPackInfoToObject > dispatcher( firstBlock );
+ return dispatcher( bdId )( bdId, numberOfGhostLayers ) ;
+ }
+
+
+ //===================================================================================================================
+ //
+ // createMPIDatatypeInfo
+ //
+ //===================================================================================================================
+
+
+ template< typename GpuField_T >
+ boost::python::object createMPIDatatypeInfoToObject( BlockDataID bdId, uint_t numberOfGhostLayers )
+ {
+ using field::communication::UniformMPIDatatypeInfo;
+ if ( numberOfGhostLayers > 0 )
+ return boost::python::object( make_shared< UniformMPIDatatypeInfo<GpuField_T> >( bdId, numberOfGhostLayers ) );
+ else
+ return boost::python::object( make_shared< UniformMPIDatatypeInfo<GpuField_T> >( bdId ) );
+ }
+
+ FunctionExporterClass( createMPIDatatypeInfoToObject, boost::python::object( BlockDataID, uint_t ) );
+
+ template< typename GpuFields>
+ boost::python::object createMPIDatatypeInfo( const shared_ptr<StructuredBlockStorage> & bs,
+ const std::string & blockDataName,
+ uint_t numberOfGhostLayers)
+ {
+ auto bdId = python_coupling::blockDataIDFromString( *bs, blockDataName );
+ if ( bs->begin() == bs->end() ) {
+ // if no blocks are on this field an arbitrary MPIDatatypeInfo can be returned
+ return createMPIDatatypeInfoToObject< GPUField<real_t> > ( bdId, numberOfGhostLayers );
+ }
+
+ IBlock * firstBlock = & ( * bs->begin() );
+ python_coupling::Dispatcher<GpuFields, Exporter_createMPIDatatypeInfoToObject > dispatcher( firstBlock );
+ return dispatcher( bdId )( bdId, numberOfGhostLayers );
+ }
+
+
} // namespace internal
-template<typename FieldTypes >
+template<typename GpuFields >
void exportModuleToPython()
{
python_coupling::ModuleScope fieldModule( "cuda" );
using namespace boost::python;
- boost::mpl::for_each<FieldTypes>( internal::GpuFieldExporter() );
+ python_coupling::for_each_noncopyable_type<GpuFields>( internal::GpuFieldExporter() );
- def( "createGpuField", &internal::createPythonGpuField<FieldTypes>, ( ( arg("size") ),
+ def( "createGpuField", &internal::createPythonGpuField<GpuFields>, ( ( arg("size") ),
( arg("type") ),
( arg("ghostLayers") = uint_t(1) ),
( arg("layout") = field::zyxf),
( arg("usePitchedMem") = true ) ) );
- def( "addGpuFieldToStorage", &internal::addToStorage<FieldTypes>, ( ( arg("blocks") ),
+ def( "addGpuFieldToStorage", &internal::addToStorage<GpuFields>, ( ( arg("blocks") ),
( arg("name") ),
( arg("type") ),
( arg("fSize") = 1 ),
@@ -273,6 +345,9 @@ void exportModuleToPython()
( arg("layout") = field::zyxf ),
( arg("usePitchedMem") = object() ) ) );
+ def( "createMPIDatatypeInfo",&internal::createMPIDatatypeInfo<GpuFields>, ( arg("blocks"), arg("blockDataName"), arg("numberOfGhostLayers" ) =0 ) );
+ def( "createPackInfo", &internal::createPackInfo<GpuFields>, ( arg("blocks"), arg("blockDataName"), arg("numberOfGhostLayers" ) =0 ) );
+
}
--
GitLab