Commit fe479759 authored by Tomáš Oberhuber's avatar Tomáš Oberhuber
Browse files

Debugging the CUDA code.

parent be9600f2

CMakeLists.txt

+4 −4
Original line number Diff line number Diff line
@@ -18,8 +18,8 @@ if( CMAKE_BUILD_TYPE STREQUAL "Debug") 
    set( LIBRARY_OUTPUT_PATH ${PROJECT_SOURCE_DIR}/Debug/lib )

    set( EXECUTABLE_OUTPUT_PATH ${PROJECT_SOURCE_DIR}/Debug/bin )

    set( debugExt -dbg )

    AddCompilerFlag( "-std=gnu++0x -DTEMPLATE_EXPLICIT_INSTANTIATION" )

    #AddCompilerFlag( "-DHAVE_NOT_CXX11 -DTEMPLATE_EXPLICIT_INSTANTIATION" )

    #AddCompilerFlag( "-std=gnu++0x -DTEMPLATE_EXPLICIT_INSTANTIATION" )

    AddCompilerFlag( "-DHAVE_NOT_CXX11 -DTEMPLATE_EXPLICIT_INSTANTIATION" )

else()

    set( PROJECT_BUILD_PATH ${PROJECT_SOURCE_DIR}/Release/src )

    set( PROJECT_TESTS_PATH ${PROJECT_SOURCE_DIR}/Release/tests )
@@ -27,8 +27,8 @@ else() 
    set( EXECUTABLE_OUTPUT_PATH ${PROJECT_SOURCE_DIR}/Release/bin)

    #set( CXX_OPTIMIZE_FLAGS "-O3")

    OptimizeForArchitecture()

    AddCompilerFlag( "-std=gnu++0x -O3 -DNDEBUG -DTEMPLATE_EXPLICIT_INSTANTIATION" )

    #AddCompilerFlag( "-DHAVE_NOT_CXX11 -O3 -DNDEBUG -DTEMPLATE_EXPLICIT_INSTANTIATION" )

    #AddCompilerFlag( "-std=gnu++0x -O3 -DNDEBUG -DTEMPLATE_EXPLICIT_INSTANTIATION" )

    AddCompilerFlag( "-DHAVE_NOT_CXX11 -O3 -DNDEBUG -DTEMPLATE_EXPLICIT_INSTANTIATION" )

endif()



#####

buildAll

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


TARGET=TNL

INSTALL_PREFIX=${HOME}/local

WITH_CUDA=no

WITH_CUDA=yes

WITH_CUSPARSE=no

CUDA_ARCHITECTURE=2.0

VERBOSE=1

src/core/tnlVector.h

+0 −16
Original line number Diff line number Diff line
@@ -133,20 +133,4 @@ class tnlVector : public tnlArray< Real, Device, Index > 


#include <implementation/core/tnlVector_impl.h>



#ifdef TEMPLATE_EXPLICIT_INSTANTIATION



extern template class tnlVector< float, tnlHost, int >;

extern template class tnlVector< double, tnlHost, int >;

extern template class tnlVector< float, tnlHost, long int >;

extern template class tnlVector< double, tnlHost, long int >;



#ifdef HAVE_CUDA

extern template class tnlVector< float, tnlCuda, int >;

extern template class tnlVector< double, tnlCuda, int >;

extern template class tnlVector< float, tnlCuda, long int >;

extern template class tnlVector< double, tnlCuda, long int >;

#endif



#endif



#endif /* TNLVECTOR_H_ */

src/implementation/core/CMakeLists.txt

+9 −1
Original line number Diff line number Diff line
@@ -16,7 +16,13 @@ SET( headers cuda-long-vector-kernels.h 
             tnlVector_impl.h )



SET( CURRENT_DIR ${CMAKE_SOURCE_DIR}/src/implementation/core )

IF( BUILD_CUDA )

   set( tnl_implementation_core_CUDA_SOURCES

        ${CURRENT_DIR}/tnlArray_impl.cu

        ${CURRENT_DIR}/tnlVector_impl.cu )

ENDIF()    

set( tnl_implementation_core_SOURCES

     ${tnl_implementation_core_CUDA_SOURCES}

     ${CURRENT_DIR}/tnlArray_impl.cpp

     ${CURRENT_DIR}/tnlSharedArray_impl.cpp

     ${CURRENT_DIR}/tnlMultiArray_impl.cpp
@@ -35,5 +41,7 @@ set( tnl_implementation_core_SOURCES 
     ${CURRENT_DIR}/mpi-supp.cpp     

    PARENT_SCOPE )

    

    

        

INSTALL( FILES ${headers} DESTINATION include/tnl-${tnlVersion}/implementation/core )

src/implementation/core/cuda-long-vector-kernels.h

+1 −0
Original line number Diff line number Diff line
@@ -22,6 +22,7 @@ 
#include <cuda.h>

#endif

#include <iostream>

#include <core/tnlVector.h>



using namespace std;