xmp_reduce_acc.c File Reference

#include <stdio.h>
#include <stdarg.h>
#include <stdlib.h>
#include "xmp_internal.h"

Include dependency graph for xmp_reduce_acc.c:

Macros
#define	MPI_PORTABLE_PLATFORM_H

Functions
void	_XMP_reduce_gpu_NODES_ENTIRE (_XMP_nodes_t *nodes, void *addr, int count, int datatype, int op)
void	_XMP_reduce_gpu_CLAUSE (void *data_addr, int count, int datatype, int op)
void	_XMP_reduce_acc_NODES_ENTIRE (_XMP_nodes_t *nodes, void *data_addr, int count, int datatype, int op)
void	_XMP_reduce_acc_FLMM_NODES_ENTIRE (_XMP_nodes_t *nodes, void *addr, int count, int datatype, int op, int num_locs,...)
void	_XMP_reduce_acc_CLAUSE (void *data_addr, int count, int datatype, int op)
void	_XMP_reduce_acc_FLMM_CLAUSE (void *data_addr, int count, int datatype, int op, int num_locs,...)

Macro Definition Documentation

MPI_PORTABLE_PLATFORM_H

#define MPI_PORTABLE_PLATFORM_H

Function Documentation

_XMP_reduce_acc_CLAUSE()

void _XMP_reduce_acc_CLAUSE	(	void *	data_addr,
		int	count,
		int	datatype,
		int	op
	)

{
  set_comm_mode();
 
  if(comm_mode >= 1){
    _XMP_reduce_CLAUSE(data_addr, count, datatype, op);
  }else{
#ifdef _XMP_TCA
    //
#else
    _XMP_reduce_gpu_CLAUSE(data_addr, count, datatype, op);
#endif
 
  }
}

_XMP_reduce_acc_FLMM_CLAUSE()

void _XMP_reduce_acc_FLMM_CLAUSE	(	void *	data_addr,
		int	count,
		int	datatype,
		int	op,
		int	num_locs,
			...
	)

{
  _XMP_fatal("_XMP_reduce_acc_FLMM_CLAUSE is unimplemented");
}

Here is the call graph for this function:

_XMP_reduce_acc_FLMM_NODES_ENTIRE()

void _XMP_reduce_acc_FLMM_NODES_ENTIRE	(	_XMP_nodes_t *	nodes,
		void *	addr,
		int	count,
		int	datatype,
		int	op,
		int	num_locs,
			...
	)

{
  _XMP_fatal("_XMP_reduce_acc_FLMM_NODES_ENTIRE is unimplemented");
}

Here is the call graph for this function:

_XMP_reduce_acc_NODES_ENTIRE()

void _XMP_reduce_acc_NODES_ENTIRE	(	_XMP_nodes_t *	nodes,
		void *	data_addr,
		int	count,
		int	datatype,
		int	op
	)

{
  set_comm_mode();
 
  if(comm_mode >= 1){
    _XMP_reduce_NODES_ENTIRE(nodes, data_addr, count, datatype, op);
  }else{
#ifdef _XMP_TCA
    set_hybrid_comm();
    if (enable_hybrid) {
      _XMP_reduce_hybrid_NODES_ENTIRE(nodes, data_addr, count, datatype, op);
    } else {
      _XMP_reduce_tca_NODES_ENTIRE(nodes, data_addr, count, datatype, op);
    }
#else
    _XMP_reduce_gpu_NODES_ENTIRE(nodes, data_addr, count, datatype, op);
#endif
  }
}

_XMP_reduce_gpu_CLAUSE()

void _XMP_reduce_gpu_CLAUSE	(	void *	data_addr,
		int	count,
		int	datatype,
		int	op
	)

                                                                             {
  // setup information
  MPI_Datatype mpi_datatype = MPI_DATATYPE_NULL;
  size_t datatype_size = 0;
  MPI_Op mpi_op;
  _XMP_setup_reduce_type(&mpi_datatype, &datatype_size, datatype);
  _XMP_setup_reduce_op(&mpi_op, op);
 
  size_t size = datatype_size * count;
  void *host_buf = _XMP_alloc(size);
 
  // copy dev to host
  //e = cudaMemcpy(host_buf, dev_addr, size, cudaMemcpyDeviceToHost);
  _XACC_memory_read(host_buf, dev_addr, 0, size, _XACC_QUEUE_NULL, true);
 
  // reduce
  MPI_Allreduce(MPI_IN_PLACE, host_buf, count, mpi_datatype, mpi_op, *((MPI_Comm *)(_XMP_get_execution_nodes())->comm));
 
  // copy host to dev
  //e = cudaMemcpy(dev_addr, host_buf, size, cudaMemcpyHostToDevice);
  _XACC_memory_write(dev_addr, 0, host_buf, size, _XACC_QUEUE_NULL, true);
 
  _XMP_free(host_buf);
}

_XMP_reduce_gpu_NODES_ENTIRE()

void _XMP_reduce_gpu_NODES_ENTIRE	(	_XMP_nodes_t *	nodes,
		void *	addr,
		int	count,
		int	datatype,
		int	op
	)

{
  if (count == 0) {
    return; // FIXME not good implementation
  }
  if (!nodes->is_member) {
    return;
  }
 
  // setup information
  MPI_Datatype mpi_datatype = MPI_DATATYPE_NULL;
  size_t datatype_size = 0;
  MPI_Op mpi_op;
  _XMP_setup_reduce_type(&mpi_datatype, &datatype_size, datatype);
  _XMP_setup_reduce_op(&mpi_op, op);
 
  size_t size = datatype_size * count;
  void *host_buf = _XMP_alloc(size);
 
  // copy dev to host
  //e = cudaMemcpy(host_buf, dev_addr, size, cudaMemcpyDeviceToHost);
  _XACC_memory_read(host_buf, dev_addr, 0, size, _XACC_QUEUE_NULL, true);
 
  MPI_Allreduce(MPI_IN_PLACE, host_buf, count, mpi_datatype, mpi_op, *((MPI_Comm *)nodes->comm));
 
  // copy host to dev
  //e = cudaMemcpy(dev_addr, host_buf, size, cudaMemcpyHostToDevice);
  _XACC_memory_write(dev_addr, 0, host_buf, size, _XACC_QUEUE_NULL, true);
 
  _XMP_free(host_buf);
}