Commit 23d906d6 authored by Vít Hanousek's avatar Vít Hanousek
Browse files

Add Cuda-Aware-MPI-copy simple sample program to test build system.

parent 63c9f306

CMakeLists.txt

+7 −3
Original line number Diff line number Diff line
@@ -92,15 +92,19 @@ endif() 
if( ${CXX_COMPILER_NAME} STREQUAL "mpic++" )

   message( "MPI compiler detected."    )

   set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DHAVE_MPI" )

   set( CUDA_HOST_COMPILER "mpic++" )



endif()



####

# Check for MPI

# Check for MPI -- not working

#

#find_package( MPI )

#if( MPI_CXX_FOUND )

   # set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DHAVE_MPI" )

   # message( "MPI headers found -- ${MPI_CXX_INCLUDE_PATH}")

   # message( "MPI link flags  -- ${MPI_CXX_LINK_FLAGS}")

   # message( "MPI libraries-- ${MPI_CXX_LIBRARIES}")

#endif()



#####

tests/mpi/CMakeLists.txt

+5 −0
Original line number Diff line number Diff line
@@ -28,3 +28,8 @@ 
	#TARGET_COMPILE_DEFINITIONS( tnlMeshFuncttionEvaluateTestXY PUBLIC "-DDIMENSION=3" )

	#TARGET_COMPILE_DEFINITIONS( tnlMeshFuncttionEvaluateTestXY PUBLIC "-DXDISTR -DYDISTR" )



IF( BUILD_CUDA )

    CUDA_ADD_EXECUTABLE( mpi-gpu-test ${headers} mpi-gpu.cu )

    #TARGET_LINK_LIBRARIES( mpi-gpu-test ${CPPUNIT_LIBRARIES}

     #                                               tnl )   

ENDIF( BUILD_CUDA )

tests/mpi/mpi-gpu.cu

0 → 100644
+79 −0
Original line number Diff line number Diff line 
#include <iostream>



using namespace std;





#if defined(HAVE_MPI) && defined(HAVE_CUDA)



#include <cuda_runtime.h>

#include <mpi.h>

 

__global__ void SetKernel(float *deviceData, float value)

{

    // Just a dummy kernel

    const int idx = blockIdx.x * blockDim.x + threadIdx.x;

    deviceData[idx] = value;

}



double sum(float * data, int count)

{

    double sum=0;

    for(int i=0;i<count;i++)

        sum+=data[i];



    return sum;

}





int main(int argc, char **argv)

{

    MPI_Init(&argc, &argv);



    int rank;

    MPI_Comm_rank(MPI_COMM_WORLD, &rank);



    int blockSize = 256; //počet threadů v v bloku

    int gridSize = 1000; //počet bloků v gridu -> musí být menší než maxGridSize v cudeGetDeviceProperties



    int dataCount=blockSize*gridSize;



    float * deviceData=NULL;

    cudaMalloc((void **)&deviceData, dataCount * sizeof(float));



    if(rank==0)

    {

        cout << rank<<": "<<"Setup GPU alocated array to 1" << endl;

        SetKernel<<< gridSize,blockSize >>>(deviceData,1.0f);

        cout << rank<<": "<<" Sending GPU data " <<endl;

        MPI_Send((void*)deviceData, dataCount, MPI_FLOAT, 1, 1, MPI_COMM_WORLD);

    }

    

    if(rank==1) 

    {

        cout << rank<<": "<<" Reciving GPU data " <<endl;

        MPI_Recv((void*) deviceData, dataCount, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, MPI_STATUS_IGNORE);



        float *data = new float[dataCount];

        cout << rank<<": "<<" Copying data from GPU to CPU " <<endl;

        cudaMemcpy( (void*) data, (void*)deviceData, dataCount*sizeof(float),  cudaMemcpyDeviceToHost);    

        cout << rank<<": "<<" Computin Sum on CPU " <<endl;

        cout << rank<<": "<< "sum:" << sum(data,dataCount) << endl;

        delete [] data;

    }



    cudaFree(deviceData);



    MPI_Finalize();

return 0;

}



#else



int main(void)

{

    cout << "CUDA or MPI missing...." <<endl;

}



#endif