Commit 38c8c2ef authored by David M. Rogers's avatar David M. Rogers
Browse files

Added gpu functionality to source.

parent 2c2b6c84

CMakeLists.txt

+1 −1
Original line number Diff line number Diff line
@@ -33,7 +33,7 @@ find_package(MPI REQUIRED) 


##########  TARGETS  ################################################

add_library(mpiwrap src/wrapper.cc)

add_executable(mpitest src/bcast.cc)

add_executable(mpitest src/allreduce.cc)

#####################################################################



# configured header file stashing options

README.md

+2 −2
Original line number Diff line number Diff line
@@ -18,8 +18,8 @@ to "out of the box". 


## CUDA/HIP Kernel Calling Layer



Configuring with ``-DCMAKE_CUDA_ARCHITECTURE=70`` enables CUDA,

and configuring with ``-DCMAKE_HIP_ARCHITECTURE=gfx908`` enables HIP.

Configuring with ``-DCMAKE_CUDA_ARCHITECTURES=70`` enables CUDA,

and configuring with ``-DCMAKE_HIP_ARCHITECTURES=gfx908`` enables HIP.

The source code is identical for both, because the HIP

interface uses the simple cuda2hip.h header from

the [Quip project](https://github.com/twhite-cray/quip).

cmake/setup_cuda.cmake

+1 −1
Original line number Diff line number Diff line 
enable_language(CUDA)

#cmake_minimum_required(VERSION 3.17)

if(${CMAKE_VERSION} VERSION_LESS "3.17")

    message(FATAL_ERROR "Compilation for CUDA requires CMake 3.17 or later.")

endif()

enable_language(CUDA)



message(STATUS "Setting up CUDA")

cmake/setup_hip.cmake

+1 −1
Original line number Diff line number Diff line 
enable_language(HIP)

#cmake_minimum_required(VERSION 3.21.3)

if(${CMAKE_VERSION} VERSION_LESS "3.21.3")

    message(FATAL_ERROR "Compilation for HIP requires CMake 3.21.3 or later.")

endif()

enable_language(HIP)



message(STATUS "Setting up HIP using ROCM_ROOT = ${ROCM_ROOT}")

src/allreduce.cc

0 → 100644
+48 −0
Original line number Diff line number Diff line 
#include <mpiwrap.hh>

#include <stdio.h>



#ifdef ENABLE_GPU

#include <cuda_runtime.h>



#define CUDACHECK(cmd) do {                         \

  cudaError_t e = cmd;                              \

  if( e != cudaSuccess ) {                          \

    printf("Failed: Cuda error %s:%d '%s'\n",       \

        __FILE__,__LINE__,cudaGetErrorString(e));   \

    exit(EXIT_FAILURE);                             \

  }                                                 \

} while(0)

#define devMalloc(ptr,size) CUDACHECK( cudaMalloc(&(void *)ptr, size) )

#define devFree(ptr) CUDACHECK( cudaFree(ptr) )

#else



#define ERRCHECK(cmd) do {                          \

  int e = cmd;                                      \

  if( e != 0 ) {                                    \

    printf("Failed: error %s:%d '%d'\n",            \

        __FILE__,__LINE__,e);                       \

    exit(EXIT_FAILURE);                             \

  }                                                 \

} while(0)

#define devMalloc(ptr,size) ERRCHECK( (ptr = reinterpret_cast<decltype(ptr)>(malloc(size))) == nullptr )

#define devFree(ptr) free( ptr )

#endif



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

    MPIH mpi(&argc, &argv);

    double *src, *dst;

    int count = 1024;



    devMalloc(src, count*sizeof(double));

    devMalloc(dst, count*sizeof(double));



    if(mpi.rank == 0)

        printf("Broadcasting %d doubles from root.\n", count);

    //MPI_Bcast(src, count, MPI_DOUBLE, 0, mpi.comm);

    MPI_Allreduce(src, dst, count, MPI_DOUBLE, MPI_SUM, mpi.comm);



    devFree(src);

    devFree(dst);



    return 0;

}