1
0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00

Reasonably fast version

This commit is contained in:
azusayamaguchi 2016-10-21 15:17:39 +01:00
parent f22317748f
commit 09fd5c43a7

View File

@ -28,6 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
/* END LEGAL */
#include "Grid.h"
#include <mpi.h>
#include <numaif.h>
namespace Grid {
@ -182,6 +183,15 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
ShmCommBuf = 0;
ierr = MPI_Win_allocate_shared(MAX_MPI_SHM_BYTES,1,MPI_INFO_NULL,ShmComm,&ShmCommBuf,&ShmWindow);
assert(ierr==0);
for(uint64_t page=0;page<MAX_MPI_SHM_BYTES;page+=4096){
void *pages = (void *) ( page + ShmCommBuf );
int status;
int flags=MPOL_MF_MOVE_ALL;
int nodes=1; // numa domain == MCDRAM
unsigned long count=1;
ierr= move_pages(0,count, &pages,&nodes,&status,flags);
if (ierr && (page==0)) perror("numa relocate command failed");
}
MPI_Win_lock_all (MPI_MODE_NOCHECK, ShmWindow);
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -367,10 +377,10 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
char *to_ptr = (char *)ShmCommBufs[gdest];
char *from_ptr = (char *)ShmCommBufs[ShmRank];
int small = (bytes<MAX_MPI_SHM_BYTES) || (shm_mode==0);
int small = (bytes<MAX_MPI_SHM_BYTES);
#ifndef SHM_USE_BCOPY
typedef double T;
typedef long long T __attribute__ ((__vector_size__ (16)));
int words = bytes/sizeof(T);
assert(((size_t)bytes &(sizeof(T)-1))==0);
// assert(((size_t)xmit &(sizeof(T)-1))==0);