Ejemplos de gaspi_gpu_write en C++ (Cpp)

Ejemplo n.º 1

Archivo: read_write_gpu_gpu.c Proyecto: mckellyln/GPI-2

int main(int argc, char *argv[])
{
  gaspi_rank_t numranks, myrank;
  gaspi_rank_t rankSend;
  gaspi_size_t segSize;
  const  gaspi_offset_t localOff_r= 0;
  const gaspi_offset_t remOff_r = 0;
  const  gaspi_offset_t localOff_w = _128MB / 2 ;
  const gaspi_offset_t remOff_w = _128MB / 2;
  gaspi_number_t queueSize, qmax;
  const gaspi_size_t commSize = _8MB;
  int i;
  gaspi_gpu_t gpus[8]; 
  gaspi_gpu_num nGPUs;


  TSUITE_INIT(argc, argv);

  ASSERT (gaspi_proc_init(GASPI_BLOCK));

  ASSERT (gaspi_proc_num(&numranks));
  ASSERT (gaspi_proc_rank(&myrank));
  ASSERT (gaspi_init_GPUs());
  ASSERT (gaspi_number_of_GPUs(&nGPUs));
  ASSERT (gaspi_GPU_ids(gpus));

  ASSERT (gaspi_segment_create(0, _128MB, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED|GASPI_MEM_GPU));

  ASSERT( gaspi_segment_size(0, myrank, &segSize));

  ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));

  ASSERT (gaspi_queue_size_max(&qmax));

  for(i = 0; i < 100; i++ )
  {
    for(rankSend = 0; rankSend < numranks; rankSend++)
    {
      if(rankSend == myrank)
        continue;

      gaspi_printf("partner rank: %d - %lu bytes (%d)\n", rankSend, commSize, i);

      ASSERT (gaspi_queue_size(1, &queueSize));
      if (queueSize > qmax - 24)
        ASSERT (gaspi_wait(1, GASPI_BLOCK));

      ASSERT (gaspi_read(0, localOff_r, rankSend, 0,  remOff_r,  commSize, 1, GASPI_BLOCK));
    }
  }
  for(i = 0; i < 100; i++ )
  {
    for(rankSend = 0; rankSend < numranks; rankSend++)
    {
      if(rankSend == myrank)
        continue;
      ASSERT (gaspi_gpu_write(0, localOff_r, rankSend, 0,  remOff_r,  commSize, 1, GASPI_BLOCK));
    }
  }

  ASSERT (gaspi_wait(1, GASPI_BLOCK));

  ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));

  ASSERT (gaspi_proc_term(GASPI_BLOCK));

  return EXIT_SUCCESS;
}

Ejemplo n.º 2

Archivo: write_gpu_all.c Proyecto: cc-hpc-itwm/GPI-2

//create GPU segemnts on all nodes
//and then tranfer data between all segments
int main(int argc, char *argv[])
{
  TSUITE_INIT(argc, argv);
  ASSERT (gaspi_proc_init(GASPI_BLOCK));

  //need the barrier to make sn is up
  ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));

  gaspi_rank_t rank, nprocs, i,j;
  gaspi_number_t seg_max;

  gaspi_gpu_id_t gpus[8];
  gaspi_number_t nGPUs;

  ASSERT(gaspi_proc_num(&nprocs));
  ASSERT (gaspi_proc_rank(&rank));
  ASSERT(gaspi_gpu_init());
  seg_max = 1;
  ASSERT (gaspi_gpu_number(&nGPUs));
  ASSERT (gaspi_gpu_ids(gpus));

  for(i = 0; i < nGPUs; i++)
    {
      cudaSetDevice(gpus[i]);
      ASSERT (gaspi_segment_create(i, _128MB, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED|GASPI_MEM_GPU));
    }

  ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));

  gaspi_number_t queueSize;

  int rankSend = (rank + 1) % nprocs;

  for (j =0; j<nGPUs; j++)
    {

      //sleep(1);
      if (gaspi_gpu_write(j, //seg
			  1024, //local
			  rankSend, //rank
			  j, //seg rem
			  1024, //remote
			  32768, //size
			  1, //queue
			  GASPI_BLOCK) != GASPI_SUCCESS)
	{
	  gaspi_queue_size(1, &queueSize);
	  gaspi_printf (" failed with i = %d queue %u\n", j, queueSize);
	  exit(-1);

	}
    }

  ASSERT(gaspi_wait(GASPI_GROUP_ALL, GASPI_BLOCK));

  ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));

  for (i =0; i<nGPUs; i++)
    {
      ASSERT (gaspi_segment_delete(i));
    }

  ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));

  ASSERT (gaspi_proc_term(GASPI_BLOCK));

  return EXIT_SUCCESS;
}