portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-09 23:45:36 +00:00
Code Releases Activity

Merge branch 'develop' into feature/hadrons

Browse Source 
# Conflicts:
#	lib/communicator/Communicator_mpi3_leader.cc
#	lib/communicator/Communicator_shmem.cc
This commit is contained in:
Antonin Portelli 2018-01-11 18:50:15 +00:00
commit 7bb405e790
23 changed files with 1426 additions and 2838 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
TODO
View File

@ -1,20 +1,36 @@
TODO:
---------------
Large item work list:
Code item work list
a) namespaces & indentation
GRID_BEGIN_NAMESPACE();
GRID_END_NAMESPACE();
-- delete QCD namespace
b) GPU branch
- start branch
- Increase Macro use in core library support; prepare for change
- Audit volume of "device" code
- Virtual function audit
- Start port once Nvidia box is up
- Cut down volume of code for first port? How?
Physics item work list:
1)- BG/Q port and check ; Andrew says ok.
2)- Christoph's local basis expansion Lanczos
--
3a)- RNG I/O in ILDG/SciDAC (minor)
3b)- Precision conversion and sort out localConvert <-- partial/easy
3c)- Consistent linear solver flop count/rate -- PARTIAL, time but no flop/s yet
4)- Physical propagator interface
5)- Conserved currents
6)- Multigrid Wilson and DWF, compare to other Multigrid implementations
7)- HDCR resume
2)- Consistent linear solver flop count/rate -- PARTIAL, time but no flop/s yet
3)- Physical propagator interface
4)- Multigrid Wilson and DWF, compare to other Multigrid implementations
5)- HDCR resume
----------------------------
Recent DONE
-- RNG I/O in ILDG/SciDAC (minor)
-- Precision conversion and sort out localConvert <-- partial/easy
-- Conserved currents (Andrew)
-- Split grid
-- Christoph's local basis expansion Lanczos
-- MultiRHS with spread out extra dim -- Go through filesystem with SciDAC I/O ; <-- DONE ; bmark cori
-- Lanczos Remove DenseVector, DenseMatrix; Use Eigen instead. <-- DONE
-- GaugeFix into central location <-- DONE

12
benchmarks/Benchmark_comms.cc
View File

@ -106,7 +106,7 @@ int main (int argc, char ** argv)
for(int i=0;i<Nloop;i++){
double start=usecond();
std::vector<CartesianCommunicator::CommsRequest_t> requests;
std::vector<CommsRequest_t> requests;
ncomm=0;
for(int mu=0;mu<4;mu++){
@ -202,7 +202,7 @@ int main (int argc, char ** argv)
int recv_from_rank;
{
std::vector<CartesianCommunicator::CommsRequest_t> requests;
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFromBegin(requests,
(void *)&xbuf[mu][0],
@ -215,7 +215,7 @@ int main (int argc, char ** argv)
comm_proc = mpi_layout[mu]-1;
{
std::vector<CartesianCommunicator::CommsRequest_t> requests;
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFromBegin(requests,
(void *)&xbuf[mu+4][0],
@ -290,7 +290,7 @@ int main (int argc, char ** argv)
dbytes=0;
ncomm=0;
std::vector<CartesianCommunicator::CommsRequest_t> requests;
std::vector<CommsRequest_t> requests;
for(int mu=0;mu<4;mu++){
@ -383,7 +383,7 @@ int main (int argc, char ** argv)
for(int i=0;i<Nloop;i++){
double start=usecond();
std::vector<CartesianCommunicator::CommsRequest_t> requests;
std::vector<CommsRequest_t> requests;
dbytes=0;
ncomm=0;
for(int mu=0;mu<4;mu++){
@ -481,7 +481,7 @@ int main (int argc, char ** argv)
for(int i=0;i<Nloop;i++){
double start=usecond();
std::vector<CartesianCommunicator::CommsRequest_t> requests;
std::vector<CommsRequest_t> requests;
dbytes=0;
ncomm=0;

22
configure.ac
View File

@ -337,15 +337,11 @@ case ${ac_PRECISION} in
esac
###################### Shared memory allocation technique under MPI3
AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmget|shmopen|hugetlbfs],
AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmopen|hugetlbfs],
[Select SHM allocation technique])],[ac_SHM=${enable_shm}],[ac_SHM=shmopen])
case ${ac_SHM} in
shmget)
AC_DEFINE([GRID_MPI3_SHMGET],[1],[GRID_MPI3_SHMGET] )
;;
shmopen)
AC_DEFINE([GRID_MPI3_SHMOPEN],[1],[GRID_MPI3_SHMOPEN] )
;;
@ -367,7 +363,7 @@ AC_ARG_ENABLE([shmpath],[AC_HELP_STRING([--enable-shmpath=path],
AC_DEFINE_UNQUOTED([GRID_SHM_PATH],["$ac_SHMPATH"],[Path to a hugetlbfs filesystem for MMAPing])
############### communication type selection
AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto|mpi3|mpi3-auto|shmem],
AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto],
[Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
case ${ac_COMMS} in
@ -375,22 +371,10 @@ case ${ac_COMMS} in
AC_DEFINE([GRID_COMMS_NONE],[1],[GRID_COMMS_NONE] )
comms_type='none'
;;
mpi3*)
mpi*)
AC_DEFINE([GRID_COMMS_MPI3],[1],[GRID_COMMS_MPI3] )
comms_type='mpi3'
;;
mpit)
AC_DEFINE([GRID_COMMS_MPIT],[1],[GRID_COMMS_MPIT] )
comms_type='mpit'
;;
mpi*)
AC_DEFINE([GRID_COMMS_MPI],[1],[GRID_COMMS_MPI] )
comms_type='mpi'
;;
shmem)
AC_DEFINE([GRID_COMMS_SHMEM],[1],[GRID_COMMS_SHMEM] )
comms_type='shmem'
;;
*)
AC_MSG_ERROR([${ac_COMMS} unsupported --enable-comms option]);
;;

18
lib/Makefile.am
View File

@ -1,28 +1,18 @@
extra_sources=
extra_headers=
if BUILD_COMMS_MPI
extra_sources+=communicator/Communicator_mpi.cc
extra_sources+=communicator/Communicator_base.cc
endif
if BUILD_COMMS_MPI3
extra_sources+=communicator/Communicator_mpi3.cc
extra_sources+=communicator/Communicator_base.cc
endif
if BUILD_COMMS_MPIT
extra_sources+=communicator/Communicator_mpit.cc
extra_sources+=communicator/Communicator_base.cc
endif
if BUILD_COMMS_SHMEM
extra_sources+=communicator/Communicator_shmem.cc
extra_sources+=communicator/Communicator_base.cc
extra_sources+=communicator/SharedMemoryMPI.cc
extra_sources+=communicator/SharedMemory.cc
endif
if BUILD_COMMS_NONE
extra_sources+=communicator/Communicator_none.cc
extra_sources+=communicator/Communicator_base.cc
extra_sources+=communicator/SharedMemoryNone.cc
extra_sources+=communicator/SharedMemory.cc
endif
if BUILD_HDF5

1
lib/communicator/Communicator.h
View File

@ -28,6 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_COMMUNICATOR_H
#define GRID_COMMUNICATOR_H
#include <Grid/communicator/SharedMemory.h>
#include <Grid/communicator/Communicator_base.h>
#endif

288
lib/communicator/Communicator_base.cc
View File

@ -36,33 +36,9 @@ namespace Grid {
///////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////
void * CartesianCommunicator::ShmCommBuf;
uint64_t CartesianCommunicator::MAX_MPI_SHM_BYTES = 1024LL*1024LL*1024LL;
CartesianCommunicator::CommunicatorPolicy_t
CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
int CartesianCommunicator::nCommThreads = -1;
int CartesianCommunicator::Hugepages = 0;
/////////////////////////////////
// Alloc, free shmem region
/////////////////////////////////
void *CartesianCommunicator::ShmBufferMalloc(size_t bytes){
// bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
void *ptr = (void *)heap_top;
heap_top += bytes;
heap_bytes+= bytes;
if (heap_bytes >= MAX_MPI_SHM_BYTES) {
std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
std::cout<< " Current value is " << (MAX_MPI_SHM_BYTES/(1024*1024)) <<std::endl;
assert(heap_bytes<MAX_MPI_SHM_BYTES);
}
return ptr;
}
void CartesianCommunicator::ShmBufferFreeAll(void) {
heap_top =(size_t)ShmBufferSelf();
heap_bytes=0;
}
/////////////////////////////////
// Grid information queries
@ -96,269 +72,5 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
GlobalSumVector((double *)c,2*N);
}
#if defined( GRID_COMMS_MPI) || defined (GRID_COMMS_MPIT) || defined (GRID_COMMS_MPI3)
void CartesianCommunicator::AllToAll(int dim,void *in,void *out,uint64_t words,uint64_t bytes)
{
std::vector<int> row(_ndimension,1);
assert(dim>=0 && dim<_ndimension);
// Split the communicator
row[dim] = _processors[dim];
int me;
CartesianCommunicator Comm(row,*this,me);
Comm.AllToAll(in,out,words,bytes);
}
void CartesianCommunicator::AllToAll(void *in,void *out,uint64_t words,uint64_t bytes)
{
// MPI is a pain and uses "int" arguments
// 64*64*64*128*16 == 500Million elements of data.
// When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
// (Turns up on 32^3 x 64 Gparity too)
MPI_Datatype object;
int iwords;
int ibytes;
iwords = words;
ibytes = bytes;
assert(words == iwords); // safe to cast to int ?
assert(bytes == ibytes); // safe to cast to int ?
MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
MPI_Type_commit(&object);
MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
MPI_Type_free(&object);
}
#endif
#if defined( GRID_COMMS_MPI) || defined (GRID_COMMS_MPIT)
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)
{
_ndimension = processors.size();
int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
std::vector<int> parent_processor_coor(_ndimension,0);
std::vector<int> parent_processors (_ndimension,1);
// Can make 5d grid from 4d etc...
int pad = _ndimension-parent_ndimension;
for(int d=0;d<parent_ndimension;d++){
parent_processor_coor[pad+d]=parent._processor_coor[d];
parent_processors [pad+d]=parent._processors[d];
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// split the communicator
//////////////////////////////////////////////////////////////////////////////////////////////////////
int Nparent;
MPI_Comm_size(parent.communicator,&Nparent);
int childsize=1;
for(int d=0;d<processors.size();d++) {
childsize *= processors[d];
}
int Nchild = Nparent/childsize;
assert (childsize * Nchild == Nparent);
std::vector<int> ccoor(_ndimension); // coor within subcommunicator
std::vector<int> scoor(_ndimension); // coor of split within parent
std::vector<int> ssize(_ndimension); // coor of split within parent
for(int d=0;d<_ndimension;d++){
ccoor[d] = parent_processor_coor[d] % processors[d];
scoor[d] = parent_processor_coor[d] / processors[d];
ssize[d] = parent_processors[d] / processors[d];
}
int crank; // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
// Mpi uses the reverse Lexico convention to us
Lexicographic::IndexFromCoorReversed(ccoor,crank,processors);
Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);
MPI_Comm comm_split;
if ( Nchild > 1 ) {
if(0){
std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"] ";
for(int d=0;d<parent._ndimension;d++) std::cout << parent._processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" child grid["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< parent._ndimension <<"] ";
for(int d=0;d<parent._ndimension;d++) std::cout << parent._processor_coor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new split "<< srank<<" scoor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << scoor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << ccoor[d] << " ";
std::cout<<std::endl;
}
int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
assert(ierr==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Declare victory
//////////////////////////////////////////////////////////////////////////////////////////////////////
// std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
// << Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
} else {
comm_split=parent.communicator;
srank = 0;
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Set up from the new split communicator
//////////////////////////////////////////////////////////////////////////////////////////////////////
InitFromMPICommunicator(processors,comm_split);
if(0){
std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
for(int d=0;d<processors.size();d++){
std::cout << d<< " " << _processor_coor[d] <<" " << ccoor[d]<<std::endl;
}
}
for(int d=0;d<processors.size();d++){
assert(_processor_coor[d] == ccoor[d] );
}
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Take an MPI_Comm and self assemble
//////////////////////////////////////////////////////////////////////////////////////////////////////
void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
{
_ndimension = processors.size();
_processor_coor.resize(_ndimension);
/////////////////////////////////
// Count the requested nodes
/////////////////////////////////
_Nprocessors=1;
_processors = processors;
for(int i=0;i<_ndimension;i++){
_Nprocessors*=_processors[i];
}
std::vector<int> periodic(_ndimension,1);
MPI_Cart_create(communicator_base, _ndimension,&_processors[0],&periodic[0],0,&communicator);
MPI_Comm_rank(communicator,&_processor);
MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
if ( 0 && (communicator_base != communicator_world) ) {
std::cout << "InitFromMPICommunicator Cartesian communicator created with a non-world communicator"<<std::endl;
std::cout << " new communicator rank "<<_processor<< " coor ["<<_ndimension<<"] ";
for(int d=0;d<_processors.size();d++){
std::cout << _processor_coor[d]<<" ";
}
std::cout << std::endl;
}
int Size;
MPI_Comm_size(communicator,&Size);
#if defined(GRID_COMMS_MPIT) || defined (GRID_COMMS_MPI3)
communicator_halo.resize (2*_ndimension);
for(int i=0;i<_ndimension*2;i++){
MPI_Comm_dup(communicator,&communicator_halo[i]);
}
#endif
assert(Size==_Nprocessors);
}
#endif
#if defined( GRID_COMMS_MPI) || defined (GRID_COMMS_MPIT)
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
InitFromMPICommunicator(processors,communicator_world);
}
#endif
#if !defined( GRID_COMMS_MPI3)
int CartesianCommunicator::NodeCount(void) { return ProcessorCount();};
int CartesianCommunicator::RankCount(void) { return ProcessorCount();};
#endif
#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPIT)
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes, int dir)
{
std::vector<CommsRequest_t> list;
// Discard the "dir"
SendToRecvFromBegin (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
SendToRecvFromComplete(list);
return 2.0*bytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes, int dir)
{
// Discard the "dir"
SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
return 2.0*bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
SendToRecvFromComplete(waitall);
}
#endif
#if !defined( GRID_COMMS_MPI3)
void CartesianCommunicator::StencilBarrier(void){};
commVector<uint8_t> CartesianCommunicator::ShmBufStorageVector;
void *CartesianCommunicator::ShmBufferSelf(void) { return ShmCommBuf; }
void *CartesianCommunicator::ShmBuffer(int rank) {
return NULL;
}
void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) {
return NULL;
}
void CartesianCommunicator::ShmInitGeneric(void){
#if 1
int mmap_flag =0;
#ifdef MAP_ANONYMOUS
mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
#endif
#ifdef MAP_ANON
mmap_flag = mmap_flag| MAP_SHARED | MAP_ANON;
#endif
#ifdef MAP_HUGETLB
if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
#endif
ShmCommBuf =(void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag, -1, 0);
if (ShmCommBuf == (void *)MAP_FAILED) {
perror("mmap failed ");
exit(EXIT_FAILURE);
}
#ifdef MADV_HUGEPAGE
if (!Hugepages ) madvise(ShmCommBuf,MAX_MPI_SHM_BYTES,MADV_HUGEPAGE);
#endif
#else
ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
ShmCommBuf=(void *)&ShmBufStorageVector[0];
#endif
bzero(ShmCommBuf,MAX_MPI_SHM_BYTES);
}
#endif
}

119
lib/communicator/Communicator_base.h
View File

@ -32,117 +32,33 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
///////////////////////////////////
// Processor layout information
///////////////////////////////////
#ifdef GRID_COMMS_MPI
#include <mpi.h>
#endif
#ifdef GRID_COMMS_MPI3
#include <mpi.h>
#endif
#ifdef GRID_COMMS_MPIT
#include <mpi.h>
#endif
#ifdef GRID_COMMS_SHMEM
#include <mpp/shmem.h>
#endif
#include <Grid/communicator/SharedMemory.h>
namespace Grid {
class CartesianCommunicator {
public:
class CartesianCommunicator : public SharedMemory {
public:
////////////////////////////////////////////
// Isend/Irecv/Wait, or Sendrecv blocking
// Policies
////////////////////////////////////////////
enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
static CommunicatorPolicy_t CommunicatorPolicy;
static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
///////////////////////////////////////////
// Up to 65536 ranks per node adequate for now
// 128MB shared memory for comms enought for 48^4 local vol comms
// Give external control (command line override?) of this
///////////////////////////////////////////
static const int MAXLOG2RANKSPERNODE = 16;
static uint64_t MAX_MPI_SHM_BYTES;
static int nCommThreads;
// use explicit huge pages
static int Hugepages;
////////////////////////////////////////////
// Communicator should know nothing of the physics grid, only processor grid.
////////////////////////////////////////////
int _Nprocessors; // How many in all
std::vector<int> _processors; // Which dimensions get relayed out over processors lanes.
int _processor; // linear processor rank
std::vector<int> _processor_coor; // linear processor coordinate
unsigned long _ndimension;
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
static MPI_Comm communicator_world;
MPI_Comm communicator;
std::vector<MPI_Comm> communicator_halo;
typedef MPI_Request CommsRequest_t;
#else
typedef int CommsRequest_t;
#endif
////////////////////////////////////////////////////////////////////
// Helper functionality for SHM Windows common to all other impls
////////////////////////////////////////////////////////////////////
// Longer term; drop this in favour of a master / slave model with
// cartesian communicator on a subset of ranks, slave ranks controlled
// by group leader with data xfer via shared memory
////////////////////////////////////////////////////////////////////
#ifdef GRID_COMMS_MPI3
static int ShmRank;
static int ShmSize;
static int GroupRank;
static int GroupSize;
static int WorldRank;
static int WorldSize;
std::vector<int> WorldDims;
std::vector<int> GroupDims;
std::vector<int> ShmDims;
std::vector<int> GroupCoor;
std::vector<int> ShmCoor;
std::vector<int> WorldCoor;
static std::vector<int> GroupRanks;
static std::vector<int> MyGroup;
static int ShmSetup;
static MPI_Win ShmWindow;
static MPI_Comm ShmComm;
std::vector<int> LexicographicToWorldRank;
static std::vector<void *> ShmCommBufs;
#else
static void ShmInitGeneric(void);
static commVector<uint8_t> ShmBufStorageVector;
#endif
/////////////////////////////////
// Grid information and queries
// Implemented in Communicator_base.C
/////////////////////////////////
static void * ShmCommBuf;
size_t heap_top;
size_t heap_bytes;
void *ShmBufferSelf(void);
void *ShmBuffer(int rank);
void *ShmBufferTranslate(int rank,void * local_p);
void *ShmBufferMalloc(size_t bytes);
void ShmBufferFreeAll(void) ;
unsigned long _ndimension;
static Grid_MPI_Comm communicator_world;
Grid_MPI_Comm communicator;
std::vector<Grid_MPI_Comm> communicator_halo;
////////////////////////////////////////////////
// Must call in Grid startup
@ -158,13 +74,13 @@ class CartesianCommunicator {
virtual ~CartesianCommunicator();
private:
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPIT) || defined (GRID_COMMS_MPI3)
////////////////////////////////////////////////
// Private initialise from an MPI communicator
// Can use after an MPI_Comm_split, but hidden from user so private
////////////////////////////////////////////////
void InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base);
#endif
void InitFromMPICommunicator(const std::vector<int> &processors, Grid_MPI_Comm communicator_base);
public:
@ -182,8 +98,6 @@ class CartesianCommunicator {
const std::vector<int> & ThisProcessorCoor(void) ;
const std::vector<int> & ProcessorGrid(void) ;
int ProcessorCount(void) ;
int NodeCount(void) ;
int RankCount(void) ;
////////////////////////////////////////////////////////////////////////////////
// very VERY rarely (Log, serial RNG) we need world without a grid
@ -271,15 +185,10 @@ class CartesianCommunicator {
template<class T> void AllToAll(int dim,std::vector<T> &in, std::vector<T> &out){
assert(dim>=0);
assert(dim<_ndimension);
int numnode = _processors[dim];
// std::cerr << " AllToAll in.size() "<<in.size()<<std::endl;
// std::cerr << " AllToAll out.size() "<<out.size()<<std::endl;
assert(in.size()==out.size());
int numnode = _processors[dim];
uint64_t bytes=sizeof(T);
uint64_t words=in.size()/numnode;
// std:: cout << "AllToAll buffer size "<< in.size()*sizeof(T)<<std::endl;
// std:: cout << "AllToAll datum bytes "<< bytes<<std::endl;
// std:: cout << "AllToAll datum count "<< words<<std::endl;
assert(numnode * words == in.size());
assert(words < (1ULL<<31));
AllToAll(dim,(void *)&in[0],(void *)&out[0],words,bytes);

222
lib/communicator/Communicator_mpi.cc
View File

@ -1,222 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/Communicator_mpi.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <Grid/GridQCDcore.h>
#include <Grid/qcd/action/ActionCore.h>
#include <mpi.h>
namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Comm CartesianCommunicator::communicator_world;
// Should error check all MPI calls.
void CartesianCommunicator::Init(int *argc, char ***argv) {
int flag;
int provided;
MPI_Initialized(&flag); // needed to coexist with other libs apparently
if ( !flag ) {
MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
if ( provided != MPI_THREAD_MULTIPLE ) {
QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
}
}
MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
ShmInitGeneric();
}
CartesianCommunicator::~CartesianCommunicator()
{
int MPI_is_finalised;
MPI_Finalized(&MPI_is_finalised);
if (communicator && !MPI_is_finalised)
MPI_Comm_free(&communicator);
}
void CartesianCommunicator::GlobalSum(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(float &f){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(float *f,int N)
{
int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(double &d)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(double *d,int N)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
{
int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
assert(ierr==0);
}
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
{
int rank;
int ierr=MPI_Cart_rank (communicator, &coor[0], &rank);
assert(ierr==0);
return rank;
}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
{
coor.resize(_ndimension);
int ierr=MPI_Cart_coords (communicator, rank, _ndimension,&coor[0]);
assert(ierr==0);
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFrom(void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
std::vector<CommsRequest_t> reqs(0);
SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
SendToRecvFromComplete(reqs);
}
void CartesianCommunicator::SendRecvPacket(void *xmit,
void *recv,
int sender,
int receiver,
int bytes)
{
MPI_Status stat;
assert(sender != receiver);
int tag = sender;
if ( _processor == sender ) {
MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
}
if ( _processor == receiver ) {
MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
}
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
int myrank = _processor;
int ierr;
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
MPI_Request xrq;
MPI_Request rrq;
ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
assert(ierr==0);
list.push_back(xrq);
list.push_back(rrq);
} else {
// Give the CPU to MPI immediately; can use threads to overlap optionally
ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
recv,bytes,MPI_CHAR,from, from,
communicator,MPI_STATUS_IGNORE);
assert(ierr==0);
}
}
void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
{
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
int nreq=list.size();
std::vector<MPI_Status> status(nreq);
int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
assert(ierr==0);
}
}
void CartesianCommunicator::Barrier(void)
{
int ierr = MPI_Barrier(communicator);
assert(ierr==0);
}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
{
int ierr=MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator);
assert(ierr==0);
}
///////////////////////////////////////////////////////
// Should only be used prior to Grid Init finished.
// Check for this?
///////////////////////////////////////////////////////
int CartesianCommunicator::RankWorld(void){
int r;
MPI_Comm_rank(communicator_world,&r);
return r;
}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
{
int ierr= MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator_world);
assert(ierr==0);
}
}

747
lib/communicator/Communicator_mpi3.cc
View File

@ -26,89 +26,20 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <mpi.h>
#include <semaphore.h>
#include <fcntl.h>
#include <unistd.h>
#include <limits.h>
#include <sys/types.h>
#include <sys/ipc.h>
#include <sys/shm.h>
#include <sys/mman.h>
#include <zlib.h>
#ifdef HAVE_NUMAIF_H
#include <numaif.h>
#endif
#include <Grid/communicator/SharedMemory.h>
namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
int CartesianCommunicator::ShmSetup = 0;
Grid_MPI_Comm CartesianCommunicator::communicator_world;
int CartesianCommunicator::ShmRank;
int CartesianCommunicator::ShmSize;
int CartesianCommunicator::GroupRank;
int CartesianCommunicator::GroupSize;
int CartesianCommunicator::WorldRank;
int CartesianCommunicator::WorldSize;
MPI_Comm CartesianCommunicator::communicator_world;
MPI_Comm CartesianCommunicator::ShmComm;
MPI_Win CartesianCommunicator::ShmWindow;
std::vector<int> CartesianCommunicator::GroupRanks;
std::vector<int> CartesianCommunicator::MyGroup;
std::vector<void *> CartesianCommunicator::ShmCommBufs;
int CartesianCommunicator::NodeCount(void) { return GroupSize;};
int CartesianCommunicator::RankCount(void) { return WorldSize;};
#undef FORCE_COMMS
void *CartesianCommunicator::ShmBufferSelf(void)
////////////////////////////////////////////
// First initialise of comms system
////////////////////////////////////////////
void CartesianCommunicator::Init(int *argc, char ***argv)
{
return ShmCommBufs[ShmRank];
}
void *CartesianCommunicator::ShmBuffer(int rank)
{
int gpeer = GroupRanks[rank];
#ifdef FORCE_COMMS
return NULL;
#endif
if (gpeer == MPI_UNDEFINED){
return NULL;
} else {
return ShmCommBufs[gpeer];
}
}
void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p)
{
static int count =0;
int gpeer = GroupRanks[rank];
assert(gpeer!=ShmRank); // never send to self
assert(rank!=WorldRank);// never send to self
#ifdef FORCE_COMMS
return NULL;
#endif
if (gpeer == MPI_UNDEFINED){
return NULL;
} else {
uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
return (void *) remote;
}
}
void CartesianCommunicator::Init(int *argc, char ***argv) {
int flag;
int provided;
// mtrace();
MPI_Initialized(&flag); // needed to coexist with other libs apparently
if ( !flag ) {
@ -119,487 +50,202 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
Grid_quiesce_nodes();
MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
MPI_Comm_rank(communicator_world,&WorldRank);
MPI_Comm_size(communicator_world,&WorldSize);
if ( WorldRank == 0 ) {
std::cout << GridLogMessage<< "Initialising MPI "<< WorldRank <<"/"<<WorldSize <<std::endl;
}
/////////////////////////////////////////////////////////////////////
// Split into groups that can share memory
/////////////////////////////////////////////////////////////////////
MPI_Comm_split_type(communicator_world, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
MPI_Comm_rank(ShmComm ,&ShmRank);
MPI_Comm_size(ShmComm ,&ShmSize);
GroupSize = WorldSize/ShmSize;
/////////////////////////////////////////////////////////////////////
// find world ranks in our SHM group (i.e. which ranks are on our node)
/////////////////////////////////////////////////////////////////////
MPI_Group WorldGroup, ShmGroup;
MPI_Comm_group (communicator_world, &WorldGroup);
MPI_Comm_group (ShmComm, &ShmGroup);
std::vector<int> world_ranks(WorldSize);
GroupRanks.resize(WorldSize);
for(int r=0;r<WorldSize;r++) world_ranks[r]=r;
MPI_Group_translate_ranks (WorldGroup,WorldSize,&world_ranks[0],ShmGroup, &GroupRanks[0]);
///////////////////////////////////////////////////////////////////
// Identify who is in my group and noninate the leader
///////////////////////////////////////////////////////////////////
int g=0;
MyGroup.resize(ShmSize);
for(int rank=0;rank<WorldSize;rank++){
if(GroupRanks[rank]!=MPI_UNDEFINED){
assert(g<ShmSize);
MyGroup[g++] = rank;
}
}
std::sort(MyGroup.begin(),MyGroup.end(),std::less<int>());
int myleader = MyGroup[0];
std::vector<int> leaders_1hot(WorldSize,0);
std::vector<int> leaders_group(GroupSize,0);
leaders_1hot [ myleader ] = 1;
///////////////////////////////////////////////////////////////////
// global sum leaders over comm world
///////////////////////////////////////////////////////////////////
int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,communicator_world);
assert(ierr==0);
///////////////////////////////////////////////////////////////////
// find the group leaders world rank
///////////////////////////////////////////////////////////////////
int group=0;
for(int l=0;l<WorldSize;l++){
if(leaders_1hot[l]){
leaders_group[group++] = l;
}
}
///////////////////////////////////////////////////////////////////
// Identify the rank of the group in which I (and my leader) live
///////////////////////////////////////////////////////////////////
GroupRank=-1;
for(int g=0;g<GroupSize;g++){
if (myleader == leaders_group[g]){
GroupRank=g;
}
}
assert(GroupRank!=-1);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// allocate the shared window for our group
//////////////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Barrier(ShmComm);
ShmCommBuf = 0;
ShmCommBufs.resize(ShmSize);
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbf and others map filesystems as mappable huge pages
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMMMAP
char shm_name [NAME_MAX];
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
sprintf(shm_name,GRID_SHM_PATH "/Grid_mpi3_shm_%d_%d",GroupRank,r);
//sprintf(shm_name,"/var/lib/hugetlbfs/group/wheel/pagesize-2MB/" "Grid_mpi3_shm_%d_%d",GroupRank,r);
// printf("Opening file %s \n",shm_name);
int fd=open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd == -1) {
printf("open %s failed\n",shm_name);
perror("open hugetlbfs");
exit(0);
}
int mmap_flag = MAP_SHARED ;
#ifdef MAP_POPULATE
mmap_flag|=MAP_POPULATE;
#endif
#ifdef MAP_HUGETLB
if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
#endif
void *ptr = (void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag,fd, 0);
if ( ptr == (void *)MAP_FAILED ) {
printf("mmap %s failed\n",shm_name);
perror("failed mmap"); assert(0);
}
assert(((uint64_t)ptr&0x3F)==0);
ShmCommBufs[r] =ptr;
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////
// POSIX SHMOPEN ; as far as I know Linux does not allow EXPLICIT HugePages with this case
// tmpfs (Larry Meadows says) does not support explicit huge page, and this is used for
// the posix shm virtual file system
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMOPEN
char shm_name [NAME_MAX];
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",GroupRank,r);
shm_unlink(shm_name);
int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd < 0 ) { perror("failed shm_open"); assert(0); }
ftruncate(fd, size);
int mmap_flag = MAP_SHARED;
#ifdef MAP_POPULATE
mmap_flag |= MAP_POPULATE;
#endif
#ifdef MAP_HUGETLB
if (Hugepages) mmap_flag |= MAP_HUGETLB;
#endif
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
if ( ptr == (void * )MAP_FAILED ) { perror("failed mmap"); assert(0); }
assert(((uint64_t)ptr&0x3F)==0);
// Experiments; Experiments; Try to force numa domain on the shm segment if we have numaif.h
#if 0
//#ifdef HAVE_NUMAIF_H
int status;
int flags=MPOL_MF_MOVE;
#ifdef KNL
int nodes=1; // numa domain == MCDRAM
// Find out if in SNC2,SNC4 mode ?
#else
int nodes=r; // numa domain == MPI ID
#endif
unsigned long count=1;
for(uint64_t page=0;page<size;page+=4096){
void *pages = (void *) ( page + (uint64_t)ptr );
uint64_t *cow_it = (uint64_t *)pages; *cow_it = 1;
ierr= move_pages(0,count, &pages,&nodes,&status,flags);
if (ierr && (page==0)) perror("numa relocate command failed");
}
#endif
ShmCommBufs[r] =ptr;
}
}
MPI_Barrier(ShmComm);
if ( ShmRank != 0 ) {
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES ;
sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",GroupRank,r);
int fd=shm_open(shm_name,O_RDWR,0666);
if ( fd<0 ) { perror("failed shm_open"); assert(0); }
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if ( ptr == MAP_FAILED ) { perror("failed mmap"); assert(0); }
assert(((uint64_t)ptr&0x3F)==0);
ShmCommBufs[r] =ptr;
}
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////
// SHMGET SHMAT and SHM_HUGETLB flag
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMGET
std::vector<int> shmids(ShmSize);
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
key_t key = IPC_PRIVATE;
int flags = IPC_CREAT | SHM_R | SHM_W;
#ifdef SHM_HUGETLB
if (Hugepages) flags|=SHM_HUGETLB;
#endif
if ((shmids[r]= shmget(key,size, flags)) ==-1) {
int errsv = errno;
printf("Errno %d\n",errsv);
printf("key %d\n",key);
printf("size %lld\n",size);
printf("flags %d\n",flags);
perror("shmget");
exit(1);
} else {
printf("shmid: 0x%x\n", shmids[r]);
}
}
}
MPI_Barrier(ShmComm);
MPI_Bcast(&shmids[0],ShmSize*sizeof(int),MPI_BYTE,0,ShmComm);
MPI_Barrier(ShmComm);
for(int r=0;r<ShmSize;r++){
ShmCommBufs[r] = (uint64_t *)shmat(shmids[r], NULL,0);
if (ShmCommBufs[r] == (uint64_t *)-1) {
perror("Shared memory attach failure");
shmctl(shmids[r], IPC_RMID, NULL);
exit(2);
}
printf("shmaddr: %p\n", ShmCommBufs[r]);
}
MPI_Barrier(ShmComm);
// Mark for clean up
for(int r=0;r<ShmSize;r++){
shmctl(shmids[r], IPC_RMID,(struct shmid_ds *)NULL);
}
MPI_Barrier(ShmComm);
#endif
ShmCommBuf = ShmCommBufs[ShmRank];
MPI_Barrier(ShmComm);
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
uint64_t * check = (uint64_t *) ShmCommBufs[r];
check[0] = GroupRank;
check[1] = r;
check[2] = 0x5A5A5A;
}
}
MPI_Barrier(ShmComm);
for(int r=0;r<ShmSize;r++){
uint64_t * check = (uint64_t *) ShmCommBufs[r];
assert(check[0]==GroupRank);
assert(check[1]==r);
assert(check[2]==0x5A5A5A);
}
MPI_Barrier(ShmComm);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// Verbose for now
//////////////////////////////////////////////////////////////////////////////////////////////////////////
if (WorldRank == 0){
std::cout<<GridLogMessage<< "Grid MPI-3 configuration: detected ";
std::cout<< WorldSize << " Ranks " ;
std::cout<< GroupSize << " Nodes " ;
std::cout<< " with "<< ShmSize << " ranks-per-node "<<std::endl;
std::cout<<GridLogMessage <<"Grid MPI-3 configuration: allocated shared memory region of size ";
std::cout<<std::hex << MAX_MPI_SHM_BYTES <<" ShmCommBuf address = "<<ShmCommBuf << std::dec<<std::endl;
for(int g=0;g<GroupSize;g++){
std::cout<<GridLogMessage<<" Node "<<g<<" led by MPI rank "<<leaders_group[g]<<std::endl;
}
std::cout<<GridLogMessage<<" Boss Node Shm Pointers are {";
for(int g=0;g<ShmSize;g++){
std::cout<<std::hex<<ShmCommBufs[g]<<std::dec;
if(g!=ShmSize-1) std::cout<<",";
else std::cout<<"}"<<std::endl;
}
}
for(int g=0;g<GroupSize;g++){
if ( (ShmRank == 0) && (GroupRank==g) ) std::cout<<GridLogMessage<<"["<<g<<"] Node Group "<<g<<" is ranks {";
for(int r=0;r<ShmSize;r++){
if ( (ShmRank == 0) && (GroupRank==g) ) {
std::cout<<MyGroup[r];
if(r<ShmSize-1) std::cout<<",";
else std::cout<<"}"<<std::endl<<std::flush;
}
MPI_Barrier(communicator_world);
}
}
assert(ShmSetup==0); ShmSetup=1;
GlobalSharedMemory::Init(communicator_world);
GlobalSharedMemory::SharedMemoryAllocate(
GlobalSharedMemory::MAX_MPI_SHM_BYTES,
GlobalSharedMemory::Hugepages);
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Want to implement some magic ... Group sub-cubes into those on same node
////////////////////////////////////////////////////////////////////////////////////////////////////////////
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &dest,int &source)
///////////////////////////////////////////////////////////////////////////
// Use cartesian communicators now even in MPI3
///////////////////////////////////////////////////////////////////////////
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
{
std::vector<int> coor = _processor_coor; // my coord
assert(std::abs(shift) <_processors[dim]);
coor[dim] = (_processor_coor[dim] + shift + _processors[dim])%_processors[dim];
Lexicographic::IndexFromCoor(coor,source,_processors);
source = LexicographicToWorldRank[source];
coor[dim] = (_processor_coor[dim] - shift + _processors[dim])%_processors[dim];
Lexicographic::IndexFromCoor(coor,dest,_processors);
dest = LexicographicToWorldRank[dest];
}// rank is world rank.
int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
assert(ierr==0);
}
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
{
int rank;
Lexicographic::IndexFromCoor(coor,rank,_processors);
rank = LexicographicToWorldRank[rank];
int ierr=MPI_Cart_rank (communicator, &coor[0], &rank);
assert(ierr==0);
return rank;
}// rank is world rank
}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
{
int lr=-1;
for(int r=0;r<WorldSize;r++){// map world Rank to lexico and then to coor
if( LexicographicToWorldRank[r]==rank) lr = r;
}
assert(lr!=-1);
Lexicographic::CoorFromIndex(coor,lr,_processors);
coor.resize(_ndimension);
int ierr=MPI_Cart_coords (communicator, rank, _ndimension,&coor[0]);
assert(ierr==0);
}
////////////////////////////////////////////////////////////////////////////////////////////////////////
// Initialises from communicator_world
////////////////////////////////////////////////////////////////////////////////////////////////////////
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
MPI_Comm optimal_comm;
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm); // Remap using the shared memory optimising routine
InitFromMPICommunicator(processors,optimal_comm);
SetCommunicator(optimal_comm);
}
//////////////////////////////////
// Try to subdivide communicator
//////////////////////////////////
/*
* Use default in MPI compile
*/
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)
: CartesianCommunicator(processors)
{
std::cout << "Attempts to split MPI3 communicators will fail until implemented" <<std::endl;
_ndimension = processors.size();
int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
std::vector<int> parent_processor_coor(_ndimension,0);
std::vector<int> parent_processors (_ndimension,1);
// Can make 5d grid from 4d etc...
int pad = _ndimension-parent_ndimension;
for(int d=0;d<parent_ndimension;d++){
parent_processor_coor[pad+d]=parent._processor_coor[d];
parent_processors [pad+d]=parent._processors[d];
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// split the communicator
//////////////////////////////////////////////////////////////////////////////////////////////////////
// int Nparent = parent._processors ;
// std::cout << " splitting from communicator "<<parent.communicator <<std::endl;
int Nparent;
MPI_Comm_size(parent.communicator,&Nparent);
// std::cout << " Parent size "<<Nparent <<std::endl;
int childsize=1;
for(int d=0;d<processors.size();d++) {
childsize *= processors[d];
}
int Nchild = Nparent/childsize;
assert (childsize * Nchild == Nparent);
// std::cout << " child size "<<childsize <<std::endl;
std::vector<int> ccoor(_ndimension); // coor within subcommunicator
std::vector<int> scoor(_ndimension); // coor of split within parent
std::vector<int> ssize(_ndimension); // coor of split within parent
for(int d=0;d<_ndimension;d++){
ccoor[d] = parent_processor_coor[d] % processors[d];
scoor[d] = parent_processor_coor[d] / processors[d];
ssize[d] = parent_processors[d] / processors[d];
}
// rank within subcomm ; srank is rank of subcomm within blocks of subcomms
int crank;
// Mpi uses the reverse Lexico convention to us; so reversed routines called
Lexicographic::IndexFromCoorReversed(ccoor,crank,processors); // processors is the split grid dimensions
Lexicographic::IndexFromCoorReversed(scoor,srank,ssize); // ssize is the number of split grids
MPI_Comm comm_split;
if ( Nchild > 1 ) {
if(0){
std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"] ";
for(int d=0;d<parent._ndimension;d++) std::cout << parent._processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" child grid["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< parent._ndimension <<"] ";
for(int d=0;d<parent._ndimension;d++) std::cout << parent._processor_coor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new split "<< srank<<" scoor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << scoor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << ccoor[d] << " ";
std::cout<<std::endl;
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Declare victory
//////////////////////////////////////////////////////////////////////////////////////////////////////
std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
<< Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
std::cout << " Split communicator " <<comm_split <<std::endl;
}
////////////////////////////////////////////////////////////////
// Split the communicator
////////////////////////////////////////////////////////////////
int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
assert(ierr==0);
} else {
srank = 0;
comm_split = parent.communicator;
// std::cout << " Inherited communicator " <<comm_split <<std::endl;
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Set up from the new split communicator
//////////////////////////////////////////////////////////////////////////////////////////////////////
InitFromMPICommunicator(processors,comm_split);
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Take the right SHM buffers
//////////////////////////////////////////////////////////////////////////////////////////////////////
SetCommunicator(comm_split);
if(0){
std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
for(int d=0;d<processors.size();d++){
std::cout << d<< " " << _processor_coor[d] <<" " << ccoor[d]<<std::endl;
}
}
for(int d=0;d<processors.size();d++){
assert(_processor_coor[d] == ccoor[d] );
}
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
{
int ierr;
communicator=communicator_world;
_ndimension = processors.size();
_processor_coor.resize(_ndimension);
/////////////////////////////////
// Count the requested nodes
/////////////////////////////////
_Nprocessors=1;
_processors = processors;
for(int i=0;i<_ndimension;i++){
_Nprocessors*=_processors[i];
}
std::vector<int> periodic(_ndimension,1);
MPI_Cart_create(communicator_base, _ndimension,&_processors[0],&periodic[0],0,&communicator);
MPI_Comm_rank(communicator,&_processor);
MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
if ( 0 && (communicator_base != communicator_world) ) {
std::cout << "InitFromMPICommunicator Cartesian communicator created with a non-world communicator"<<std::endl;
std::cout << " new communicator rank "<<_processor<< " coor ["<<_ndimension<<"] ";
for(int d=0;d<_processors.size();d++){
std::cout << _processor_coor[d]<<" ";
}
std::cout << std::endl;
}
int Size;
MPI_Comm_size(communicator,&Size);
communicator_halo.resize (2*_ndimension);
for(int i=0;i<_ndimension*2;i++){
MPI_Comm_dup(communicator,&communicator_halo[i]);
}
assert(Size==_Nprocessors);
}
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
////////////////////////////////////////////////////////////////
int log2size = -1;
for(int i=0;i<=MAXLOG2RANKSPERNODE;i++){
if ( (0x1<<i) == ShmSize ) {
log2size = i;
break;
}
}
assert(log2size != -1);
////////////////////////////////////////////////////////////////
// Identify subblock of ranks on node spreading across dims
// in a maximally symmetrical way
////////////////////////////////////////////////////////////////
std::vector<int> WorldDims = processors;
ShmDims.resize (_ndimension,1);
GroupDims.resize(_ndimension);
ShmCoor.resize (_ndimension);
GroupCoor.resize(_ndimension);
WorldCoor.resize(_ndimension);
int dim = 0;
for(int l2=0;l2<log2size;l2++){
while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%_ndimension;
ShmDims[dim]*=2;
dim=(dim+1)%_ndimension;
}
////////////////////////////////////////////////////////////////
// Establish torus of processes and nodes with sub-blockings
////////////////////////////////////////////////////////////////
for(int d=0;d<_ndimension;d++){
GroupDims[d] = WorldDims[d]/ShmDims[d];
}
////////////////////////////////////////////////////////////////
// Verbose
////////////////////////////////////////////////////////////////
#if 0
std::cout<< GridLogMessage << "MPI-3 usage "<<std::endl;
std::cout<< GridLogMessage << "SHM ";
for(int d=0;d<_ndimension;d++){
std::cout<< ShmDims[d] <<" ";
}
std::cout<< std::endl;
std::cout<< GridLogMessage << "Group ";
for(int d=0;d<_ndimension;d++){
std::cout<< GroupDims[d] <<" ";
}
std::cout<< std::endl;
std::cout<< GridLogMessage<<"World ";
for(int d=0;d<_ndimension;d++){
std::cout<< WorldDims[d] <<" ";
}
std::cout<< std::endl;
#endif
////////////////////////////////////////////////////////////////
// Check processor counts match
////////////////////////////////////////////////////////////////
_Nprocessors=1;
_processors = processors;
_processor_coor.resize(_ndimension);
for(int i=0;i<_ndimension;i++){
_Nprocessors*=_processors[i];
}
assert(WorldSize==_Nprocessors);
////////////////////////////////////////////////////////////////
// Establish mapping between lexico physics coord and WorldRank
////////////////////////////////////////////////////////////////
Lexicographic::CoorFromIndex(GroupCoor,GroupRank,GroupDims);
Lexicographic::CoorFromIndex(ShmCoor,ShmRank,ShmDims);
for(int d=0;d<_ndimension;d++){
WorldCoor[d] = GroupCoor[d]*ShmDims[d]+ShmCoor[d];
}
_processor_coor = WorldCoor;
_processor = WorldRank;
///////////////////////////////////////////////////////////////////
// global sum Lexico to World mapping
///////////////////////////////////////////////////////////////////
int lexico;
LexicographicToWorldRank.resize(WorldSize,0);
Lexicographic::IndexFromCoor(WorldCoor,lexico,WorldDims);
LexicographicToWorldRank[lexico] = WorldRank;
ierr=MPI_Allreduce(MPI_IN_PLACE,&LexicographicToWorldRank[0],WorldSize,MPI_INT,MPI_SUM,communicator);
assert(ierr==0);
for(int i=0;i<WorldSize;i++){
int wr = LexicographicToWorldRank[i];
// int wr = i;
std::vector<int> coor(_ndimension);
ProcessorCoorFromRank(wr,coor); // from world rank
int ck = RankFromProcessorCoor(coor);
assert(ck==wr);
if ( wr == WorldRank ) {
for(int j=0;j<coor.size();j++) {
assert(coor[j] == _processor_coor[j]);
}
}
/*
std::cout << GridLogMessage<< " Lexicographic "<<i;
std::cout << " MPI rank "<<wr;
std::cout << " Coor ";
for(int j=0;j<coor.size();j++) std::cout << coor[j];
std::cout<< std::endl;
*/
/////////////////////////////////////////////////////
// Check everyone agrees on everyone elses coords
/////////////////////////////////////////////////////
std::vector<int> mcoor = coor;
this->Broadcast(0,(void *)&mcoor[0],mcoor.size()*sizeof(int));
for(int d = 0 ; d< _ndimension; d++) {
assert(coor[d] == mcoor[d]);
}
}
};
CartesianCommunicator::~CartesianCommunicator()
{
int MPI_is_finalised;
@ -734,19 +380,15 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
MPI_Request rrq;
int ierr;
int gdest = GroupRanks[dest];
int gfrom = GroupRanks[from];
int gme = GroupRanks[_processor];
int gdest = ShmRanks[dest];
int gfrom = ShmRanks[from];
int gme = ShmRanks[_processor];
assert(dest != _processor);
assert(from != _processor);
assert(gme == ShmRank);
double off_node_bytes=0.0;
#ifdef FORCE_COMMS
gdest = MPI_UNDEFINED;
gfrom = MPI_UNDEFINED;
#endif
if ( gfrom ==MPI_UNDEFINED) {
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
assert(ierr==0);
@ -815,5 +457,38 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
assert(ierr==0);
}
void CartesianCommunicator::AllToAll(int dim,void *in,void *out,uint64_t words,uint64_t bytes)
{
std::vector<int> row(_ndimension,1);
assert(dim>=0 && dim<_ndimension);
// Split the communicator
row[dim] = _processors[dim];
int me;
CartesianCommunicator Comm(row,*this,me);
Comm.AllToAll(in,out,words,bytes);
}
void CartesianCommunicator::AllToAll(void *in,void *out,uint64_t words,uint64_t bytes)
{
// MPI is a pain and uses "int" arguments
// 64*64*64*128*16 == 500Million elements of data.
// When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
// (Turns up on 32^3 x 64 Gparity too)
MPI_Datatype object;
int iwords;
int ibytes;
iwords = words;
ibytes = bytes;
assert(words == iwords); // safe to cast to int ?
assert(bytes == ibytes); // safe to cast to int ?
MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
MPI_Type_commit(&object);
MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
MPI_Type_free(&object);
}
}

991
lib/communicator/Communicator_mpi3_leader.cc
View File

@ -1,991 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/Communicator_mpi.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include "Grid.h"
#include <mpi.h>
//#include <numaif.h>
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/// Workarounds:
/// i) bloody mac os doesn't implement unnamed semaphores since it is "optional" posix.
/// darwin dispatch semaphores don't seem to be multiprocess.
///
/// ii) openmpi under --mca shmem posix works with two squadrons per node;
/// openmpi under default mca settings (I think --mca shmem mmap) on MacOS makes two squadrons map the SAME
/// memory as each other, despite their living on different communicators. This appears to be a bug in OpenMPI.
///
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
#include <semaphore.h>
#include <fcntl.h>
#include <unistd.h>
#include <limits.h>
typedef sem_t *Grid_semaphore;
#error /*THis is deprecated*/
#if 0
#define SEM_INIT(S) S = sem_open(sem_name,0,0600,0); assert ( S != SEM_FAILED );
#define SEM_INIT_EXCL(S) sem_unlink(sem_name); S = sem_open(sem_name,O_CREAT|O_EXCL,0600,0); assert ( S != SEM_FAILED );
#define SEM_POST(S) assert ( sem_post(S) == 0 );
#define SEM_WAIT(S) assert ( sem_wait(S) == 0 );
#else
#define SEM_INIT(S) ;
#define SEM_INIT_EXCL(S) ;
#define SEM_POST(S) ;
#define SEM_WAIT(S) ;
#endif
#include <sys/mman.h>
namespace Grid {
enum { COMMAND_ISEND, COMMAND_IRECV, COMMAND_WAITALL, COMMAND_SENDRECV };
struct Descriptor {
uint64_t buf;
size_t bytes;
int rank;
int tag;
int command;
uint64_t xbuf;
uint64_t rbuf;
int xtag;
int rtag;
int src;
int dest;
MPI_Request request;
};
const int pool = 48;
class SlaveState {
public:
volatile int head;
volatile int start;
volatile int tail;
volatile Descriptor Descrs[pool];
};
class Slave {
public:
Grid_semaphore sem_head;
Grid_semaphore sem_tail;
SlaveState *state;
MPI_Comm squadron;
uint64_t base;
int universe_rank;
int vertical_rank;
char sem_name [NAME_MAX];
////////////////////////////////////////////////////////////
// Descriptor circular pointers
////////////////////////////////////////////////////////////
Slave() {};
void Init(SlaveState * _state,MPI_Comm _squadron,int _universe_rank,int _vertical_rank);
void SemInit(void) {
sprintf(sem_name,"/Grid_mpi3_sem_head_%d",universe_rank);
SEM_INIT(sem_head);
sprintf(sem_name,"/Grid_mpi3_sem_tail_%d",universe_rank);
SEM_INIT(sem_tail);
}
void SemInitExcl(void) {
sprintf(sem_name,"/Grid_mpi3_sem_head_%d",universe_rank);
SEM_INIT_EXCL(sem_head);
sprintf(sem_name,"/Grid_mpi3_sem_tail_%d",universe_rank);
SEM_INIT_EXCL(sem_tail);
}
void WakeUpDMA(void) {
SEM_POST(sem_head);
};
void WakeUpCompute(void) {
SEM_POST(sem_tail);
};
void WaitForCommand(void) {
SEM_WAIT(sem_head);
};
void WaitForComplete(void) {
SEM_WAIT(sem_tail);
};
void EventLoop (void) {
// std::cout<< " Entering event loop "<<std::endl;
while(1){
WaitForCommand();
// std::cout << "Getting command "<<std::endl;
#if 0
_mm_monitor((void *)&state->head,0,0);
int s=state->start;
if ( s != state->head ) {
_mm_mwait(0,0);
}
#endif
Event();
}
}
int Event (void) ;
uint64_t QueueCommand(int command,void *buf, int bytes, int hashtag, MPI_Comm comm,int u_rank) ;
void QueueSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) ;
void WaitAll() {
// std::cout << "Queueing WAIT command "<<std::endl;
QueueCommand(COMMAND_WAITALL,0,0,0,squadron,0);
// std::cout << "Waking up DMA "<<std::endl;
WakeUpDMA();
// std::cout << "Waiting from semaphore "<<std::endl;
WaitForComplete();
// std::cout << "Checking FIFO is empty "<<std::endl;
while ( state->tail != state->head );
}
};
////////////////////////////////////////////////////////////////////////
// One instance of a data mover.
// Master and Slave must agree on location in shared memory
////////////////////////////////////////////////////////////////////////
class MPIoffloadEngine {
public:
static std::vector<Slave> Slaves;
static int ShmSetup;
static int UniverseRank;
static int UniverseSize;
static MPI_Comm communicator_universe;
static MPI_Comm communicator_cached;
static MPI_Comm HorizontalComm;
static int HorizontalRank;
static int HorizontalSize;
static MPI_Comm VerticalComm;
static MPI_Win VerticalWindow;
static int VerticalSize;
static int VerticalRank;
static std::vector<void *> VerticalShmBufs;
static std::vector<std::vector<int> > UniverseRanks;
static std::vector<int> UserCommunicatorToWorldRanks;
static MPI_Group WorldGroup, CachedGroup;
static void CommunicatorInit (MPI_Comm &communicator_world,
MPI_Comm &ShmComm,
void * &ShmCommBuf);
static void MapCommRankToWorldRank(int &hashtag, int & comm_world_peer,int tag, MPI_Comm comm,int commrank);
/////////////////////////////////////////////////////////
// routines for master proc must handle any communicator
/////////////////////////////////////////////////////////
static void QueueSend(int slave,void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
// std::cout<< " Queueing send "<< bytes<< " slave "<< slave << " to comm "<<rank <<std::endl;
Slaves[slave].QueueCommand(COMMAND_ISEND,buf,bytes,tag,comm,rank);
// std::cout << "Queued send command to rank "<< rank<< " via "<<slave <<std::endl;
Slaves[slave].WakeUpDMA();
// std::cout << "Waking up DMA "<< slave<<std::endl;
};
static void QueueSendRecv(int slave,void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src)
{
Slaves[slave].QueueSendRecv(xbuf,rbuf,bytes,xtag,rtag,comm,dest,src);
Slaves[slave].WakeUpDMA();
}
static void QueueRecv(int slave, void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
// std::cout<< " Queueing recv "<< bytes<< " slave "<< slave << " from comm "<<rank <<std::endl;
Slaves[slave].QueueCommand(COMMAND_IRECV,buf,bytes,tag,comm,rank);
// std::cout << "Queued recv command from rank "<< rank<< " via "<<slave <<std::endl;
Slaves[slave].WakeUpDMA();
// std::cout << "Waking up DMA "<< slave<<std::endl;
};
static void WaitAll() {
for(int s=1;s<VerticalSize;s++) {
// std::cout << "Waiting for slave "<< s<<std::endl;
Slaves[s].WaitAll();
}
// std::cout << " Wait all Complete "<<std::endl;
};
static void GetWork(int nwork, int me, int & mywork, int & myoff,int units){
int basework = nwork/units;
int backfill = units-(nwork%units);
if ( me >= units ) {
mywork = myoff = 0;
} else {
mywork = (nwork+me)/units;
myoff = basework * me;
if ( me > backfill )
myoff+= (me-backfill);
}
return;
};
static void QueueRoundRobinSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) {
uint8_t * cxbuf = (uint8_t *) xbuf;
uint8_t * crbuf = (uint8_t *) rbuf;
static int rrp=0;
int procs = VerticalSize-1;
int myoff=0;
int mywork=bytes;
QueueSendRecv(rrp+1,&cxbuf[myoff],&crbuf[myoff],mywork,xtag,rtag,comm,dest,src);
rrp = rrp+1;
if ( rrp == (VerticalSize-1) ) rrp = 0;
}
static void QueueMultiplexedSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) {
uint8_t * cxbuf = (uint8_t *) xbuf;
uint8_t * crbuf = (uint8_t *) rbuf;
int mywork, myoff, procs;
procs = VerticalSize-1;
for(int s=0;s<procs;s++) {
GetWork(bytes,s,mywork,myoff,procs);
QueueSendRecv(s+1,&cxbuf[myoff],&crbuf[myoff],mywork,xtag,rtag,comm,dest,src);
}
};
static void QueueMultiplexedSend(void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
uint8_t * cbuf = (uint8_t *) buf;
int mywork, myoff, procs;
procs = VerticalSize-1;
for(int s=0;s<procs;s++) {
GetWork(bytes,s,mywork,myoff,procs);
QueueSend(s+1,&cbuf[myoff],mywork,tag,comm,rank);
}
};
static void QueueMultiplexedRecv(void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
uint8_t * cbuf = (uint8_t *) buf;
int mywork, myoff, procs;
procs = VerticalSize-1;
for(int s=0;s<procs;s++) {
GetWork(bytes,s,mywork,myoff,procs);
QueueRecv(s+1,&cbuf[myoff],mywork,tag,comm,rank);
}
};
};
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
std::vector<Slave> MPIoffloadEngine::Slaves;
int MPIoffloadEngine::UniverseRank;
int MPIoffloadEngine::UniverseSize;
MPI_Comm MPIoffloadEngine::communicator_universe;
MPI_Comm MPIoffloadEngine::communicator_cached;
MPI_Group MPIoffloadEngine::WorldGroup;
MPI_Group MPIoffloadEngine::CachedGroup;
MPI_Comm MPIoffloadEngine::HorizontalComm;
int MPIoffloadEngine::HorizontalRank;
int MPIoffloadEngine::HorizontalSize;
MPI_Comm MPIoffloadEngine::VerticalComm;
int MPIoffloadEngine::VerticalSize;
int MPIoffloadEngine::VerticalRank;
MPI_Win MPIoffloadEngine::VerticalWindow;
std::vector<void *> MPIoffloadEngine::VerticalShmBufs;
std::vector<std::vector<int> > MPIoffloadEngine::UniverseRanks;
std::vector<int> MPIoffloadEngine::UserCommunicatorToWorldRanks;
int CartesianCommunicator::NodeCount(void) { return HorizontalSize;};
int MPIoffloadEngine::ShmSetup = 0;
void MPIoffloadEngine::CommunicatorInit (MPI_Comm &communicator_world,
MPI_Comm &ShmComm,
void * &ShmCommBuf)
{
int flag;
assert(ShmSetup==0);
//////////////////////////////////////////////////////////////////////
// Universe is all nodes prior to squadron grouping
//////////////////////////////////////////////////////////////////////
MPI_Comm_dup (MPI_COMM_WORLD,&communicator_universe);
MPI_Comm_rank(communicator_universe,&UniverseRank);
MPI_Comm_size(communicator_universe,&UniverseSize);
/////////////////////////////////////////////////////////////////////
// Split into groups that can share memory (Verticals)
/////////////////////////////////////////////////////////////////////
#undef MPI_SHARED_MEM_DEBUG
#ifdef MPI_SHARED_MEM_DEBUG
MPI_Comm_split(communicator_universe,(UniverseRank/4),UniverseRank,&VerticalComm);
#else
MPI_Comm_split_type(communicator_universe, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&VerticalComm);
#endif
MPI_Comm_rank(VerticalComm ,&VerticalRank);
MPI_Comm_size(VerticalComm ,&VerticalSize);
//////////////////////////////////////////////////////////////////////
// Split into horizontal groups by rank in squadron
//////////////////////////////////////////////////////////////////////
MPI_Comm_split(communicator_universe,VerticalRank,UniverseRank,&HorizontalComm);
MPI_Comm_rank(HorizontalComm,&HorizontalRank);
MPI_Comm_size(HorizontalComm,&HorizontalSize);
assert(HorizontalSize*VerticalSize==UniverseSize);
////////////////////////////////////////////////////////////////////////////////
// What is my place in the world
////////////////////////////////////////////////////////////////////////////////
int WorldRank=0;
if(VerticalRank==0) WorldRank = HorizontalRank;
int ierr=MPI_Allreduce(MPI_IN_PLACE,&WorldRank,1,MPI_INT,MPI_SUM,VerticalComm);
assert(ierr==0);
////////////////////////////////////////////////////////////////////////////////
// Where is the world in the universe?
////////////////////////////////////////////////////////////////////////////////
UniverseRanks = std::vector<std::vector<int> >(HorizontalSize,std::vector<int>(VerticalSize,0));
UniverseRanks[WorldRank][VerticalRank] = UniverseRank;
for(int w=0;w<HorizontalSize;w++){
ierr=MPI_Allreduce(MPI_IN_PLACE,&UniverseRanks[w][0],VerticalSize,MPI_INT,MPI_SUM,communicator_universe);
assert(ierr==0);
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// allocate the shared window for our group, pass back Shm info to CartesianCommunicator
//////////////////////////////////////////////////////////////////////////////////////////////////////////
VerticalShmBufs.resize(VerticalSize);
#undef MPI_SHARED_MEM
#ifdef MPI_SHARED_MEM
ierr = MPI_Win_allocate_shared(CartesianCommunicator::MAX_MPI_SHM_BYTES,1,MPI_INFO_NULL,VerticalComm,&ShmCommBuf,&VerticalWindow);
ierr|= MPI_Win_lock_all (MPI_MODE_NOCHECK, VerticalWindow);
assert(ierr==0);
// std::cout<<"SHM "<<ShmCommBuf<<std::endl;
for(int r=0;r<VerticalSize;r++){
MPI_Aint sz;
int dsp_unit;
MPI_Win_shared_query (VerticalWindow, r, &sz, &dsp_unit, &VerticalShmBufs[r]);
// std::cout<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
}
#else
char shm_name [NAME_MAX];
MPI_Barrier(VerticalComm);
if ( VerticalRank == 0 ) {
for(int r=0;r<VerticalSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
if ( r>0 ) size = sizeof(SlaveState);
sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldRank,r);
shm_unlink(shm_name);
int fd=shm_open(shm_name,O_RDWR|O_CREAT,0600);
if ( fd < 0 ) {
perror("failed shm_open");
assert(0);
}
ftruncate(fd, size);
VerticalShmBufs[r] = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if ( VerticalShmBufs[r] == MAP_FAILED ) {
perror("failed mmap");
assert(0);
}
/*
for(uint64_t page=0;page<size;page+=4096){
void *pages = (void *) ( page + (uint64_t)VerticalShmBufs[r] );
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");
}
*/
uint64_t * check = (uint64_t *) VerticalShmBufs[r];
check[0] = WorldRank;
check[1] = r;
// std::cout<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
}
}
MPI_Barrier(VerticalComm);
if ( VerticalRank != 0 ) {
for(int r=0;r<VerticalSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES ;
if ( r>0 ) size = sizeof(SlaveState);
sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldRank,r);
int fd=shm_open(shm_name,O_RDWR|O_CREAT,0600);
if ( fd<0 ) {
perror("failed shm_open");
assert(0);
}
VerticalShmBufs[r] = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
uint64_t * check = (uint64_t *) VerticalShmBufs[r];
assert(check[0]== WorldRank);
assert(check[1]== r);
// std::cerr<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
}
}
#endif
MPI_Barrier(VerticalComm);
//////////////////////////////////////////////////////////////////////
// Map rank of leader on node in their in new world, to the
// rank in this vertical plane's horizontal communicator
//////////////////////////////////////////////////////////////////////
communicator_world = HorizontalComm;
ShmComm = VerticalComm;
ShmCommBuf = VerticalShmBufs[0];
MPI_Comm_group (communicator_world, &WorldGroup);
///////////////////////////////////////////////////////////
// Start the slave data movers
///////////////////////////////////////////////////////////
if ( VerticalRank != 0 ) {
Slave indentured;
indentured.Init( (SlaveState *) VerticalShmBufs[VerticalRank], VerticalComm, UniverseRank,VerticalRank);
indentured.SemInitExcl();// init semaphore in shared memory
MPI_Barrier(VerticalComm);
MPI_Barrier(VerticalComm);
indentured.EventLoop();
assert(0);
} else {
Slaves.resize(VerticalSize);
for(int i=1;i<VerticalSize;i++){
Slaves[i].Init((SlaveState *)VerticalShmBufs[i],VerticalComm, UniverseRanks[HorizontalRank][i],i);
}
MPI_Barrier(VerticalComm);
for(int i=1;i<VerticalSize;i++){
Slaves[i].SemInit();// init semaphore in shared memory
}
MPI_Barrier(VerticalComm);
}
///////////////////////////////////////////////////////////
// Verbose for now
///////////////////////////////////////////////////////////
ShmSetup=1;
if (UniverseRank == 0){
std::cout<<GridLogMessage << "Grid MPI-3 configuration: detected ";
std::cout<<UniverseSize << " Ranks " ;
std::cout<<HorizontalSize << " Nodes " ;
std::cout<<VerticalSize << " with ranks-per-node "<<std::endl;
std::cout<<GridLogMessage << "Grid MPI-3 configuration: using one lead process per node " << std::endl;
std::cout<<GridLogMessage << "Grid MPI-3 configuration: reduced communicator has size " << HorizontalSize << std::endl;
for(int g=0;g<HorizontalSize;g++){
std::cout<<GridLogMessage<<" Node "<<g<<" led by MPI rank "<< UniverseRanks[g][0]<<std::endl;
}
for(int g=0;g<HorizontalSize;g++){
std::cout<<GridLogMessage<<" { ";
for(int s=0;s<VerticalSize;s++){
std::cout<< UniverseRanks[g][s];
if ( s<VerticalSize-1 ) {
std::cout<<",";
}
}
std::cout<<" } "<<std::endl;
}
}
};
///////////////////////////////////////////////////////////////////////////////////////////////
// Map the communicator into communicator_world, and find the neighbour.
// Cache the mappings; cache size is 1.
///////////////////////////////////////////////////////////////////////////////////////////////
void MPIoffloadEngine::MapCommRankToWorldRank(int &hashtag, int & comm_world_peer,int tag, MPI_Comm comm,int rank) {
if ( comm == HorizontalComm ) {
comm_world_peer = rank;
// std::cout << " MapCommRankToWorldRank horiz " <<rank<<"->"<<comm_world_peer<<std::endl;
} else if ( comm == communicator_cached ) {
comm_world_peer = UserCommunicatorToWorldRanks[rank];
// std::cout << " MapCommRankToWorldRank cached " <<rank<<"->"<<comm_world_peer<<std::endl;
} else {
int size;
MPI_Comm_size(comm,&size);
UserCommunicatorToWorldRanks.resize(size);
std::vector<int> cached_ranks(size);
for(int r=0;r<size;r++) {
cached_ranks[r]=r;
}
communicator_cached=comm;
MPI_Comm_group(communicator_cached, &CachedGroup);
MPI_Group_translate_ranks(CachedGroup,size,&cached_ranks[0],WorldGroup, &UserCommunicatorToWorldRanks[0]);
comm_world_peer = UserCommunicatorToWorldRanks[rank];
// std::cout << " MapCommRankToWorldRank cache miss " <<rank<<"->"<<comm_world_peer<<std::endl;
assert(comm_world_peer != MPI_UNDEFINED);
}
assert( (tag & (~0xFFFFL)) ==0);
uint64_t icomm = (uint64_t)comm;
int comm_hash = ((icomm>>0 )&0xFFFF)^((icomm>>16)&0xFFFF)
^ ((icomm>>32)&0xFFFF)^((icomm>>48)&0xFFFF);
// hashtag = (comm_hash<<15) | tag;
hashtag = tag;
};
void Slave::Init(SlaveState * _state,MPI_Comm _squadron,int _universe_rank,int _vertical_rank)
{
squadron=_squadron;
universe_rank=_universe_rank;
vertical_rank=_vertical_rank;
state =_state;
// std::cout << "state "<<_state<<" comm "<<_squadron<<" universe_rank"<<universe_rank <<std::endl;
state->head = state->tail = state->start = 0;
base = (uint64_t)MPIoffloadEngine::VerticalShmBufs[0];
int rank; MPI_Comm_rank(_squadron,&rank);
}
#define PERI_PLUS(A) ( (A+1)%pool )
int Slave::Event (void) {
static int tail_last;
static int head_last;
static int start_last;
int ierr;
MPI_Status stat;
static int i=0;
////////////////////////////////////////////////////
// Try to advance the start pointers
////////////////////////////////////////////////////
int s=state->start;
if ( s != state->head ) {
switch ( state->Descrs[s].command ) {
case COMMAND_ISEND:
ierr = MPI_Isend((void *)(state->Descrs[s].buf+base),
state->Descrs[s].bytes,
MPI_CHAR,
state->Descrs[s].rank,
state->Descrs[s].tag,
MPIoffloadEngine::communicator_universe,
(MPI_Request *)&state->Descrs[s].request);
assert(ierr==0);
state->start = PERI_PLUS(s);
return 1;
break;
case COMMAND_IRECV:
ierr=MPI_Irecv((void *)(state->Descrs[s].buf+base),
state->Descrs[s].bytes,
MPI_CHAR,
state->Descrs[s].rank,
state->Descrs[s].tag,
MPIoffloadEngine::communicator_universe,
(MPI_Request *)&state->Descrs[s].request);
// std::cout<< " Request is "<<state->Descrs[s].request<<std::endl;
// std::cout<< " Request0 is "<<state->Descrs[0].request<<std::endl;
assert(ierr==0);
state->start = PERI_PLUS(s);
return 1;
break;
case COMMAND_SENDRECV:
// fprintf(stderr,"Sendrecv ->%d %d : <-%d %d \n",state->Descrs[s].dest, state->Descrs[s].xtag+i*10,state->Descrs[s].src, state->Descrs[s].rtag+i*10);
ierr=MPI_Sendrecv((void *)(state->Descrs[s].xbuf+base), state->Descrs[s].bytes, MPI_CHAR, state->Descrs[s].dest, state->Descrs[s].xtag+i*10,
(void *)(state->Descrs[s].rbuf+base), state->Descrs[s].bytes, MPI_CHAR, state->Descrs[s].src , state->Descrs[s].rtag+i*10,
MPIoffloadEngine::communicator_universe,MPI_STATUS_IGNORE);
assert(ierr==0);
// fprintf(stderr,"Sendrecv done %d %d\n",ierr,i);
// MPI_Barrier(MPIoffloadEngine::HorizontalComm);
// fprintf(stderr,"Barrier\n");
i++;
state->start = PERI_PLUS(s);
return 1;
break;
case COMMAND_WAITALL:
for(int t=state->tail;t!=s; t=PERI_PLUS(t) ){
if ( state->Descrs[t].command != COMMAND_SENDRECV ) {
MPI_Wait((MPI_Request *)&state->Descrs[t].request,MPI_STATUS_IGNORE);
}
};
s=PERI_PLUS(s);
state->start = s;
state->tail = s;
WakeUpCompute();
return 1;
break;
default:
assert(0);
break;
}
}
return 0;
}
//////////////////////////////////////////////////////////////////////////////
// External interaction with the queue
//////////////////////////////////////////////////////////////////////////////
void Slave::QueueSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src)
{
int head =state->head;
int next = PERI_PLUS(head);
// Set up descriptor
int worldrank;
int hashtag;
MPI_Comm communicator;
MPI_Request request;
uint64_t relative;
relative = (uint64_t)xbuf - base;
state->Descrs[head].xbuf = relative;
relative= (uint64_t)rbuf - base;
state->Descrs[head].rbuf = relative;
state->Descrs[head].bytes = bytes;
MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,xtag,comm,dest);
state->Descrs[head].dest = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
state->Descrs[head].xtag = hashtag;
MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,rtag,comm,src);
state->Descrs[head].src = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
state->Descrs[head].rtag = hashtag;
state->Descrs[head].command= COMMAND_SENDRECV;
// Block until FIFO has space
while( state->tail==next );
// Msync on weak order architectures
// Advance pointer
state->head = next;
};
uint64_t Slave::QueueCommand(int command,void *buf, int bytes, int tag, MPI_Comm comm,int commrank)
{
/////////////////////////////////////////
// Spin; if FIFO is full until not full
/////////////////////////////////////////
int head =state->head;
int next = PERI_PLUS(head);
// Set up descriptor
int worldrank;
int hashtag;
MPI_Comm communicator;
MPI_Request request;
MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,tag,comm,commrank);
uint64_t relative= (uint64_t)buf - base;
state->Descrs[head].buf = relative;
state->Descrs[head].bytes = bytes;
state->Descrs[head].rank = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
state->Descrs[head].tag = hashtag;
state->Descrs[head].command= command;
/*
if ( command == COMMAND_ISEND ) {
std::cout << "QueueSend from "<< universe_rank <<" to commrank " << commrank
<< " to worldrank " << worldrank <<std::endl;
std::cout << " via VerticalRank "<< vertical_rank <<" to universerank " << MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank]<<std::endl;
std::cout << " QueueCommand "<<buf<<"["<<bytes<<"]" << std::endl;
}
if ( command == COMMAND_IRECV ) {
std::cout << "QueueRecv on "<< universe_rank <<" from commrank " << commrank
<< " from worldrank " << worldrank <<std::endl;
std::cout << " via VerticalRank "<< vertical_rank <<" from universerank " << MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank]<<std::endl;
std::cout << " QueueSend "<<buf<<"["<<bytes<<"]" << std::endl;
}
*/
// Block until FIFO has space
while( state->tail==next );
// Msync on weak order architectures
// Advance pointer
state->head = next;
return 0;
}
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Comm CartesianCommunicator::communicator_world;
void CartesianCommunicator::Init(int *argc, char ***argv)
{
int flag;
MPI_Initialized(&flag); // needed to coexist with other libs apparently
if ( !flag ) {
MPI_Init(argc,argv);
}
communicator_world = MPI_COMM_WORLD;
MPI_Comm ShmComm;
MPIoffloadEngine::CommunicatorInit (communicator_world,ShmComm,ShmCommBuf);
}
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
{
int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
assert(ierr==0);
}
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
{
int rank;
int ierr=MPI_Cart_rank (communicator, &coor[0], &rank);
assert(ierr==0);
return rank;
}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
{
coor.resize(_ndimension);
int ierr=MPI_Cart_coords (communicator, rank, _ndimension,&coor[0]);
assert(ierr==0);
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
_ndimension = processors.size();
std::vector<int> periodic(_ndimension,1);
_Nprocessors=1;
_processors = processors;
for(int i=0;i<_ndimension;i++){
_Nprocessors*=_processors[i];
}
int Size;
MPI_Comm_size(communicator_world,&Size);
assert(Size==_Nprocessors);
_processor_coor.resize(_ndimension);
MPI_Cart_create(communicator_world, _ndimension,&_processors[0],&periodic[0],1,&communicator);
MPI_Comm_rank (communicator,&_processor);
MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
};
CartesianCommunicator::~CartesianCommunicator() = default;
void CartesianCommunicator::GlobalSum(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(float &f){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(float *f,int N)
{
int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(double &d)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(double *d,int N)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFrom(void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
std::vector<CommsRequest_t> reqs(0);
SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
SendToRecvFromComplete(reqs);
}
void CartesianCommunicator::SendRecvPacket(void *xmit,
void *recv,
int sender,
int receiver,
int bytes)
{
MPI_Status stat;
assert(sender != receiver);
int tag = sender;
if ( _processor == sender ) {
MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
}
if ( _processor == receiver ) {
MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
}
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
MPI_Request xrq;
MPI_Request rrq;
int rank = _processor;
int ierr;
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
ierr|=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
assert(ierr==0);
list.push_back(xrq);
list.push_back(rrq);
}
void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
uint64_t xmit_i = (uint64_t) xmit;
uint64_t recv_i = (uint64_t) recv;
uint64_t shm = (uint64_t) ShmCommBuf;
// assert xmit and recv lie in shared memory region
assert( (xmit_i >= shm) && (xmit_i+bytes <= shm+MAX_MPI_SHM_BYTES) );
assert( (recv_i >= shm) && (recv_i+bytes <= shm+MAX_MPI_SHM_BYTES) );
assert(from!=_processor);
assert(dest!=_processor);
MPIoffloadEngine::QueueMultiplexedSendRecv(xmit,recv,bytes,_processor,from,communicator,dest,from);
//MPIoffloadEngine::QueueRoundRobinSendRecv(xmit,recv,bytes,_processor,from,communicator,dest,from);
//MPIoffloadEngine::QueueMultiplexedSend(xmit,bytes,_processor,communicator,dest);
//MPIoffloadEngine::QueueMultiplexedRecv(recv,bytes,from,communicator,from);
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list)
{
MPIoffloadEngine::WaitAll();
//this->Barrier();
}
void CartesianCommunicator::StencilBarrier(void) { }
void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
{
int nreq=list.size();
std::vector<MPI_Status> status(nreq);
int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
assert(ierr==0);
}
void CartesianCommunicator::Barrier(void)
{
int ierr = MPI_Barrier(communicator);
assert(ierr==0);
}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
{
int ierr=MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator);
assert(ierr==0);
}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
{
int ierr= MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator_world);
assert(ierr==0);
}
void *CartesianCommunicator::ShmBufferSelf(void) { return ShmCommBuf; }
void *CartesianCommunicator::ShmBuffer(int rank) {
return NULL;
}
void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) {
return NULL;
}
};

273
lib/communicator/Communicator_mpit.cc
View File

@ -1,273 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/Communicator_mpi.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <Grid/GridQCDcore.h>
#include <Grid/qcd/action/ActionCore.h>
#include <mpi.h>
namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Comm CartesianCommunicator::communicator_world;
// Should error check all MPI calls.
void CartesianCommunicator::Init(int *argc, char ***argv) {
int flag;
int provided;
MPI_Initialized(&flag); // needed to coexist with other libs apparently
if ( !flag ) {
MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
if ( provided != MPI_THREAD_MULTIPLE ) {
QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
}
}
MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
ShmInitGeneric();
}
CartesianCommunicator::~CartesianCommunicator()
{
int MPI_is_finalised;
MPI_Finalized(&MPI_is_finalised);
if (communicator && !MPI_is_finalised){
MPI_Comm_free(&communicator);
for(int i=0;i< communicator_halo.size();i++){
MPI_Comm_free(&communicator_halo[i]);
}
}
}
void CartesianCommunicator::GlobalSum(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(float &f){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(float *f,int N)
{
int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSum(double &d)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(double *d,int N)
{
int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
{
int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
assert(ierr==0);
}
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
{
int rank;
int ierr=MPI_Cart_rank (communicator, &coor[0], &rank);
assert(ierr==0);
return rank;
}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
{
coor.resize(_ndimension);
int ierr=MPI_Cart_coords (communicator, rank, _ndimension,&coor[0]);
assert(ierr==0);
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFrom(void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
std::vector<CommsRequest_t> reqs(0);
SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
SendToRecvFromComplete(reqs);
}
void CartesianCommunicator::SendRecvPacket(void *xmit,
void *recv,
int sender,
int receiver,
int bytes)
{
MPI_Status stat;
assert(sender != receiver);
int tag = sender;
if ( _processor == sender ) {
MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
}
if ( _processor == receiver ) {
MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
}
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
int myrank = _processor;
int ierr;
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
MPI_Request xrq;
MPI_Request rrq;
ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
assert(ierr==0);
list.push_back(xrq);
list.push_back(rrq);
} else {
// Give the CPU to MPI immediately; can use threads to overlap optionally
ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
recv,bytes,MPI_CHAR,from, from,
communicator,MPI_STATUS_IGNORE);
assert(ierr==0);
}
}
void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
{
if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) {
int nreq=list.size();
std::vector<MPI_Status> status(nreq);
int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
assert(ierr==0);
}
}
void CartesianCommunicator::Barrier(void)
{
int ierr = MPI_Barrier(communicator);
assert(ierr==0);
}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
{
int ierr=MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator);
assert(ierr==0);
}
///////////////////////////////////////////////////////
// Should only be used prior to Grid Init finished.
// Check for this?
///////////////////////////////////////////////////////
int CartesianCommunicator::RankWorld(void){
int r;
MPI_Comm_rank(communicator_world,&r);
return r;
}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
{
int ierr= MPI_Bcast(data,
bytes,
MPI_BYTE,
root,
communicator_world);
assert(ierr==0);
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir)
{
int myrank = _processor;
int ierr;
int ncomm =communicator_halo.size();
int commdir=dir%ncomm;
// std::cout << " sending on communicator "<<dir<<" " <<communicator_halo[dir]<<std::endl;
// Give the CPU to MPI immediately; can use threads to overlap optionally
MPI_Request req[2];
MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[commdir],&req[1]);
MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank ,myrank , communicator_halo[commdir],&req[0]);
list.push_back(req[0]);
list.push_back(req[1]);
return 2.0*bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
int nreq=waitall.size();
MPI_Waitall(nreq, &waitall[0], MPI_STATUSES_IGNORE);
}
double CartesianCommunicator::StencilSendToRecvFrom(void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes,int dir)
{
int myrank = _processor;
int ierr;
// std::cout << " sending on communicator "<<dir<<" " <<communicator_halo.size()<< <std::endl;
int ncomm =communicator_halo.size();
int commdir=dir%ncomm;
// Give the CPU to MPI immediately; can use threads to overlap optionally
MPI_Request req[2];
MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[commdir],&req[1]);
MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank ,myrank , communicator_halo[commdir],&req[0]);
MPI_Waitall(2, req, MPI_STATUSES_IGNORE);
return 2.0*bytes;
}
}

43
lib/communicator/Communicator_none.cc
View File

@ -32,14 +32,22 @@ namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
Grid_MPI_Comm CartesianCommunicator::communicator_world;
void CartesianCommunicator::Init(int *argc, char *** arv)
{
ShmInitGeneric();
GlobalSharedMemory::Init(communicator_world);
GlobalSharedMemory::SharedMemoryAllocate(
GlobalSharedMemory::MAX_MPI_SHM_BYTES,
GlobalSharedMemory::Hugepages);
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)
: CartesianCommunicator(processors) { srank=0;}
: CartesianCommunicator(processors)
{
srank=0;
SetCommunicator(communicator_world);
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
@ -54,6 +62,7 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
assert(_processors[d]==1);
_processor_coor[d] = 0;
}
SetCommunicator(communicator_world);
}
CartesianCommunicator::~CartesianCommunicator(){}
@ -121,6 +130,36 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
dest=0;
}
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes, int dir)
{
std::vector<CommsRequest_t> list;
// Discard the "dir"
SendToRecvFromBegin (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
SendToRecvFromComplete(list);
return 2.0*bytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes, int dir)
{
// Discard the "dir"
SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
return 2.0*bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
SendToRecvFromComplete(waitall);
}
void CartesianCommunicator::StencilBarrier(void){};
}

360
lib/communicator/Communicator_shmem.cc
View File

@ -1,360 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/Communicator_shmem.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid.h>
#include <mpp/shmem.h>
#include <array>
namespace Grid {
// Should error check all MPI calls.
#define SHMEM_VET(addr)
#define SHMEM_VET_DEBUG(addr) { \
if ( ! shmem_addr_accessible(addr,_processor) ) {\
std::fprintf(stderr,"%d Inaccessible shmem address %lx %s %s\n",_processor,addr,__FUNCTION__,#addr); \
BACKTRACEFILE(); \
}\
}
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
///////////////////////////////////////////////////////////////////////////////////////////////////
typedef struct HandShake_t {
uint64_t seq_local;
uint64_t seq_remote;
} HandShake;
std::array<long,_SHMEM_REDUCE_SYNC_SIZE> make_psync_init(void) {
std::array<long,_SHMEM_REDUCE_SYNC_SIZE> ret;
ret.fill(SHMEM_SYNC_VALUE);
return ret;
}
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync_init = make_psync_init();
static Vector< HandShake > XConnections;
static Vector< HandShake > RConnections;
void CartesianCommunicator::Init(int *argc, char ***argv) {
shmem_init();
XConnections.resize(shmem_n_pes());
RConnections.resize(shmem_n_pes());
for(int pe =0 ; pe<shmem_n_pes();pe++){
XConnections[pe].seq_local = 0;
XConnections[pe].seq_remote= 0;
RConnections[pe].seq_local = 0;
RConnections[pe].seq_remote= 0;
}
shmem_barrier_all();
ShmInitGeneric();
}
CartesianCommunicator::~CartesianCommunicator(){}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent)
: CartesianCommunicator(processors)
{
std::cout << "Attempts to split SHMEM communicators will fail " <<std::endl;
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
_ndimension = processors.size();
std::vector<int> periodic(_ndimension,1);
_Nprocessors=1;
_processors = processors;
_processor_coor.resize(_ndimension);
_processor = shmem_my_pe();
Lexicographic::CoorFromIndex(_processor_coor,_processor,_processors);
for(int i=0;i<_ndimension;i++){
_Nprocessors*=_processors[i];
}
int Size = shmem_n_pes();
assert(Size==_Nprocessors);
}
CartesianCommunicator::~CartesianCommunicator() = default;
void CartesianCommunicator::GlobalSum(uint32_t &u){
static long long source ;
static long long dest ;
static long long llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync = psync_init;
// int nreduce=1;
// int pestart=0;
// int logStride=0;
source = u;
dest = 0;
shmem_longlong_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
shmem_barrier_all(); // necessary?
u = dest;
}
void CartesianCommunicator::GlobalSum(uint64_t &u){
static long long source ;
static long long dest ;
static long long llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync = psync_init;
// int nreduce=1;
// int pestart=0;
// int logStride=0;
source = u;
dest = 0;
shmem_longlong_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
shmem_barrier_all(); // necessary?
u = dest;
}
void CartesianCommunicator::GlobalSum(float &f){
static float source ;
static float dest ;
static float llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync = psync_init;
source = f;
dest =0.0;
shmem_float_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
shmem_barrier_all();
f = dest;
}
void CartesianCommunicator::GlobalSumVector(float *f,int N)
{
static float source ;
static float dest = 0 ;
static float llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync = psync_init;
if ( shmem_addr_accessible(f,_processor) ){
shmem_float_sum_to_all(f,f,N,0,0,_Nprocessors,llwrk,psync.data());
shmem_barrier_all();
return;
}
for(int i=0;i<N;i++){
dest =0.0;
source = f[i];
shmem_float_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
shmem_barrier_all();
f[i] = dest;
}
}
void CartesianCommunicator::GlobalSum(double &d)
{
static double source;
static double dest ;
static double llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync = psync_init;
source = d;
dest = 0;
shmem_double_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
shmem_barrier_all();
d = dest;
}
void CartesianCommunicator::GlobalSumVector(double *d,int N)
{
static double source ;
static double dest ;
static double llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync = psync_init;
if ( shmem_addr_accessible(d,_processor) ){
shmem_double_sum_to_all(d,d,N,0,0,_Nprocessors,llwrk,psync.data());
shmem_barrier_all();
return;
}
for(int i=0;i<N;i++){
source = d[i];
dest =0.0;
shmem_double_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
shmem_barrier_all();
d[i] = dest;
}
}
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
{
std::vector<int> coor = _processor_coor;
assert(std::abs(shift) <_processors[dim]);
coor[dim] = (_processor_coor[dim] + shift + _processors[dim])%_processors[dim];
Lexicographic::IndexFromCoor(coor,source,_processors);
coor[dim] = (_processor_coor[dim] - shift + _processors[dim])%_processors[dim];
Lexicographic::IndexFromCoor(coor,dest,_processors);
}
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
{
int rank;
Lexicographic::IndexFromCoor(coor,rank,_processors);
return rank;
}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
{
Lexicographic::CoorFromIndex(coor,rank,_processors);
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFrom(void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
SHMEM_VET(xmit);
SHMEM_VET(recv);
std::vector<CommsRequest_t> reqs(0);
SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
SendToRecvFromComplete(reqs);
}
void CartesianCommunicator::SendRecvPacket(void *xmit,
void *recv,
int sender,
int receiver,
int bytes)
{
static uint64_t seq;
assert(recv!=xmit);
volatile HandShake *RecvSeq = (volatile HandShake *) & RConnections[sender];
volatile HandShake *SendSeq = (volatile HandShake *) & XConnections[receiver];
if ( _processor == sender ) {
// Check he has posted a receive
while(SendSeq->seq_remote == SendSeq->seq_local);
// Advance our send count
seq = ++(SendSeq->seq_local);
// Send this packet
SHMEM_VET(recv);
shmem_putmem(recv,xmit,bytes,receiver);
shmem_fence();
//Notify him we're done
shmem_putmem((void *)&(RecvSeq->seq_remote),&seq,sizeof(seq),receiver);
shmem_fence();
}
if ( _processor == receiver ) {
// Post a receive
seq = ++(RecvSeq->seq_local);
shmem_putmem((void *)&(SendSeq->seq_remote),&seq,sizeof(seq),sender);
// Now wait until he has advanced our reception counter
while(RecvSeq->seq_remote != RecvSeq->seq_local);
}
}
// Basic Halo comms primitive
void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
{
SHMEM_VET(xmit);
SHMEM_VET(recv);
// shmem_putmem_nb(recv,xmit,bytes,dest,NULL);
shmem_putmem(recv,xmit,bytes,dest);
if ( CommunicatorPolicy == CommunicatorPolicySequential ) shmem_barrier_all();
}
void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
{
// shmem_quiet(); // I'm done
if( CommunicatorPolicy == CommunicatorPolicyConcurrent ) shmem_barrier_all();// He's done too
}
void CartesianCommunicator::Barrier(void)
{
shmem_barrier_all();
}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
{
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync = psync_init;
static uint32_t word;
uint32_t *array = (uint32_t *) data;
assert( (bytes % 4)==0);
int words = bytes/4;
if ( shmem_addr_accessible(data,_processor) ){
shmem_broadcast32(data,data,words,root,0,0,shmem_n_pes(),psync.data());
return;
}
for(int w=0;w<words;w++){
word = array[w];
shmem_broadcast32((void *)&word,(void *)&word,1,root,0,0,shmem_n_pes(),psync.data());
if ( shmem_my_pe() != root ) {
array[w] = word;
}
shmem_barrier_all();
}
}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
{
static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync = psync_init;
static uint32_t word;
uint32_t *array = (uint32_t *) data;
assert( (bytes % 4)==0);
int words = bytes/4;
for(int w=0;w<words;w++){
word = array[w];
shmem_broadcast32((void *)&word,(void *)&word,1,root,0,0,shmem_n_pes(),psync.data());
if ( shmem_my_pe() != root ) {
array[w]= word;
}
shmem_barrier_all();
}
}
int CartesianCommunicator::RankWorld(void){
return shmem_my_pe();
}
}

92
lib/communicator/SharedMemory.cc Normal file
View File

@ -0,0 +1,92 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/SharedMemory.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
namespace Grid {
// static data
uint64_t GlobalSharedMemory::MAX_MPI_SHM_BYTES = 1024LL*1024LL*1024LL;
int GlobalSharedMemory::Hugepages = 0;
int GlobalSharedMemory::_ShmSetup;
int GlobalSharedMemory::_ShmAlloc;
uint64_t GlobalSharedMemory::_ShmAllocBytes;
std::vector<void *> GlobalSharedMemory::WorldShmCommBufs;
Grid_MPI_Comm GlobalSharedMemory::WorldShmComm;
int GlobalSharedMemory::WorldShmRank;
int GlobalSharedMemory::WorldShmSize;
std::vector<int> GlobalSharedMemory::WorldShmRanks;
Grid_MPI_Comm GlobalSharedMemory::WorldComm;
int GlobalSharedMemory::WorldSize;
int GlobalSharedMemory::WorldRank;
int GlobalSharedMemory::WorldNodes;
int GlobalSharedMemory::WorldNode;
void GlobalSharedMemory::SharedMemoryFree(void)
{
assert(_ShmAlloc);
assert(_ShmAllocBytes>0);
for(int r=0;r<WorldShmSize;r++){
munmap(WorldShmCommBufs[r],_ShmAllocBytes);
}
_ShmAlloc = 0;
_ShmAllocBytes = 0;
}
/////////////////////////////////
// Alloc, free shmem region
/////////////////////////////////
void *SharedMemory::ShmBufferMalloc(size_t bytes){
// bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
void *ptr = (void *)heap_top;
heap_top += bytes;
heap_bytes+= bytes;
if (heap_bytes >= heap_size) {
std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
std::cout<< " Current value is " << (heap_size/(1024*1024)) <<std::endl;
assert(heap_bytes<heap_size);
}
return ptr;
}
void SharedMemory::ShmBufferFreeAll(void) {
heap_top =(size_t)ShmBufferSelf();
heap_bytes=0;
}
void *SharedMemory::ShmBufferSelf(void)
{
return ShmCommBufs[ShmRank];
}
}

164
lib/communicator/SharedMemory.h Normal file
View File

@ -0,0 +1,164 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/SharedMemory.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
// TODO
// 1) move includes into SharedMemory.cc
//
// 2) split shared memory into a) optimal communicator creation from comm world
//
// b) shared memory buffers container
// -- static globally shared; init once
// -- per instance set of buffers.
//
#pragma once
#include <Grid/GridCore.h>
#if defined (GRID_COMMS_MPI3)
#include <mpi.h>
#endif
#include <semaphore.h>
#include <fcntl.h>
#include <unistd.h>
#include <limits.h>
#include <sys/types.h>
#include <sys/ipc.h>
#include <sys/shm.h>
#include <sys/mman.h>
#include <zlib.h>
#ifdef HAVE_NUMAIF_H
#include <numaif.h>
#endif
namespace Grid {
#if defined (GRID_COMMS_MPI3)
typedef MPI_Comm Grid_MPI_Comm;
typedef MPI_Request CommsRequest_t;
#else
typedef int CommsRequest_t;
typedef int Grid_MPI_Comm;
#endif
class GlobalSharedMemory {
private:
static const int MAXLOG2RANKSPERNODE = 16;
// Init once lock on the buffer allocation
static int _ShmSetup;
static int _ShmAlloc;
static uint64_t _ShmAllocBytes;
public:
static int ShmSetup(void) { return _ShmSetup; }
static int ShmAlloc(void) { return _ShmAlloc; }
static uint64_t ShmAllocBytes(void) { return _ShmAllocBytes; }
static uint64_t MAX_MPI_SHM_BYTES;
static int Hugepages;
static std::vector<void *> WorldShmCommBufs;
static Grid_MPI_Comm WorldComm;
static int WorldRank;
static int WorldSize;
static Grid_MPI_Comm WorldShmComm;
static int WorldShmRank;
static int WorldShmSize;
static int WorldNodes;
static int WorldNode;
static std::vector<int> WorldShmRanks;
//////////////////////////////////////////////////////////////////////////////////////
// Create an optimal reordered communicator that makes MPI_Cart_create get it right
//////////////////////////////////////////////////////////////////////////////////////
static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
static void OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
///////////////////////////////////////////////////
// Provide shared memory facilities off comm world
///////////////////////////////////////////////////
static void SharedMemoryAllocate(uint64_t bytes, int flags);
static void SharedMemoryFree(void);
};
//////////////////////////////
// one per communicator
//////////////////////////////
class SharedMemory
{
private:
static const int MAXLOG2RANKSPERNODE = 16;
size_t heap_top;
size_t heap_bytes;
size_t heap_size;
protected:
Grid_MPI_Comm ShmComm; // for barriers
int ShmRank;
int ShmSize;
std::vector<void *> ShmCommBufs;
std::vector<int> ShmRanks;// Mapping comm ranks to Shm ranks
public:
SharedMemory() {};
///////////////////////////////////////////////////////////////////////////////////////
// set the buffers & sizes
///////////////////////////////////////////////////////////////////////////////////////
void SetCommunicator(Grid_MPI_Comm comm);
////////////////////////////////////////////////////////////////////////
// For this instance ; disjoint buffer sets between splits if split grid
////////////////////////////////////////////////////////////////////////
void ShmBarrier(void);
///////////////////////////////////////////////////
// Call on any instance
///////////////////////////////////////////////////
void SharedMemoryTest(void);
void *ShmBufferSelf(void);
void *ShmBuffer (int rank);
void *ShmBufferTranslate(int rank,void * local_p);
void *ShmBufferMalloc(size_t bytes);
void ShmBufferFreeAll(void) ;
//////////////////////////////////////////////////////////////////////////
// Make info on Nodes & ranks and Shared memory available
//////////////////////////////////////////////////////////////////////////
int NodeCount(void) { return GlobalSharedMemory::WorldNodes;};
int RankCount(void) { return GlobalSharedMemory::WorldSize;};
};
}

395
lib/communicator/SharedMemoryMPI.cc Normal file
View File

@ -0,0 +1,395 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/SharedMemory.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
namespace Grid {
/*Construct from an MPI communicator*/
void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
{
assert(_ShmSetup==0);
WorldComm = comm;
MPI_Comm_rank(WorldComm,&WorldRank);
MPI_Comm_size(WorldComm,&WorldSize);
// WorldComm, WorldSize, WorldRank
/////////////////////////////////////////////////////////////////////
// Split into groups that can share memory
/////////////////////////////////////////////////////////////////////
MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&WorldShmComm);
MPI_Comm_rank(WorldShmComm ,&WorldShmRank);
MPI_Comm_size(WorldShmComm ,&WorldShmSize);
// WorldShmComm, WorldShmSize, WorldShmRank
// WorldNodes
WorldNodes = WorldSize/WorldShmSize;
assert( (WorldNodes * WorldShmSize) == WorldSize );
// FIXME: Check all WorldShmSize are the same ?
/////////////////////////////////////////////////////////////////////
// find world ranks in our SHM group (i.e. which ranks are on our node)
/////////////////////////////////////////////////////////////////////
MPI_Group WorldGroup, ShmGroup;
MPI_Comm_group (WorldComm, &WorldGroup);
MPI_Comm_group (WorldShmComm, &ShmGroup);
std::vector<int> world_ranks(WorldSize); for(int r=0;r<WorldSize;r++) world_ranks[r]=r;
WorldShmRanks.resize(WorldSize);
MPI_Group_translate_ranks (WorldGroup,WorldSize,&world_ranks[0],ShmGroup, &WorldShmRanks[0]);
///////////////////////////////////////////////////////////////////
// Identify who is in my group and nominate the leader
///////////////////////////////////////////////////////////////////
int g=0;
std::vector<int> MyGroup;
MyGroup.resize(WorldShmSize);
for(int rank=0;rank<WorldSize;rank++){
if(WorldShmRanks[rank]!=MPI_UNDEFINED){
assert(g<WorldShmSize);
MyGroup[g++] = rank;
}
}
std::sort(MyGroup.begin(),MyGroup.end(),std::less<int>());
int myleader = MyGroup[0];
std::vector<int> leaders_1hot(WorldSize,0);
std::vector<int> leaders_group(WorldNodes,0);
leaders_1hot [ myleader ] = 1;
///////////////////////////////////////////////////////////////////
// global sum leaders over comm world
///////////////////////////////////////////////////////////////////
int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,WorldComm);
assert(ierr==0);
///////////////////////////////////////////////////////////////////
// find the group leaders world rank
///////////////////////////////////////////////////////////////////
int group=0;
for(int l=0;l<WorldSize;l++){
if(leaders_1hot[l]){
leaders_group[group++] = l;
}
}
///////////////////////////////////////////////////////////////////
// Identify the node of the group in which I (and my leader) live
///////////////////////////////////////////////////////////////////
WorldNode=-1;
for(int g=0;g<WorldNodes;g++){
if (myleader == leaders_group[g]){
WorldNode=g;
}
}
assert(WorldNode!=-1);
_ShmSetup=1;
}
void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
{
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
////////////////////////////////////////////////////////////////
int log2size = -1;
for(int i=0;i<=MAXLOG2RANKSPERNODE;i++){
if ( (0x1<<i) == WorldShmSize ) {
log2size = i;
break;
}
}
assert(log2size != -1);
////////////////////////////////////////////////////////////////
// Identify subblock of ranks on node spreading across dims
// in a maximally symmetrical way
////////////////////////////////////////////////////////////////
int ndimension = processors.size();
std::vector<int> processor_coor(ndimension);
std::vector<int> WorldDims = processors; std::vector<int> ShmDims (ndimension,1); std::vector<int> NodeDims (ndimension);
std::vector<int> ShmCoor (ndimension); std::vector<int> NodeCoor (ndimension); std::vector<int> WorldCoor(ndimension);
int dim = 0;
for(int l2=0;l2<log2size;l2++){
while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
ShmDims[dim]*=2;
dim=(dim+1)%ndimension;
}
////////////////////////////////////////////////////////////////
// Establish torus of processes and nodes with sub-blockings
////////////////////////////////////////////////////////////////
for(int d=0;d<ndimension;d++){
NodeDims[d] = WorldDims[d]/ShmDims[d];
}
////////////////////////////////////////////////////////////////
// Check processor counts match
////////////////////////////////////////////////////////////////
int Nprocessors=1;
for(int i=0;i<ndimension;i++){
Nprocessors*=processors[i];
}
assert(WorldSize==Nprocessors);
////////////////////////////////////////////////////////////////
// Establish mapping between lexico physics coord and WorldRank
////////////////////////////////////////////////////////////////
int rank;
Lexicographic::CoorFromIndexReversed(NodeCoor,WorldNode ,NodeDims);
Lexicographic::CoorFromIndexReversed(ShmCoor ,WorldShmRank,ShmDims);
for(int d=0;d<ndimension;d++) WorldCoor[d] = NodeCoor[d]*ShmDims[d]+ShmCoor[d];
Lexicographic::IndexFromCoorReversed(WorldCoor,rank,WorldDims);
/////////////////////////////////////////////////////////////////
// Build the new communicator
/////////////////////////////////////////////////////////////////
int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
assert(ierr==0);
}
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbfs mapping intended
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMMMAP
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// allocate the shared windows for our group
//////////////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Barrier(WorldShmComm);
WorldShmCommBufs.resize(WorldShmSize);
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbf and others map filesystems as mappable huge pages
////////////////////////////////////////////////////////////////////////////////////////////
char shm_name [NAME_MAX];
for(int r=0;r<WorldShmSize;r++){
sprintf(shm_name,GRID_SHM_PATH "/Grid_mpi3_shm_%d_%d",WorldNode,r);
int fd=open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd == -1) {
printf("open %s failed\n",shm_name);
perror("open hugetlbfs");
exit(0);
}
int mmap_flag = MAP_SHARED ;
#ifdef MAP_POPULATE
mmap_flag|=MAP_POPULATE;
#endif
#ifdef MAP_HUGETLB
if ( flags ) mmap_flag |= MAP_HUGETLB;
#endif
void *ptr = (void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag,fd, 0);
if ( ptr == (void *)MAP_FAILED ) {
printf("mmap %s failed\n",shm_name);
perror("failed mmap"); assert(0);
}
assert(((uint64_t)ptr&0x3F)==0);
close(fd);
WorldShmCommBufs[r] =ptr;
}
_ShmAlloc=1;
_ShmAllocBytes = bytes;
};
#endif // MMAP
#ifdef GRID_MPI3_SHMOPEN
////////////////////////////////////////////////////////////////////////////////////////////
// POSIX SHMOPEN ; as far as I know Linux does not allow EXPLICIT HugePages with this case
// tmpfs (Larry Meadows says) does not support explicit huge page, and this is used for
// the posix shm virtual file system
////////////////////////////////////////////////////////////////////////////////////////////
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
MPI_Barrier(WorldShmComm);
WorldShmCommBufs.resize(WorldShmSize);
char shm_name [NAME_MAX];
if ( WorldShmRank == 0 ) {
for(int r=0;r<WorldShmSize;r++){
size_t size = bytes;
sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldNode,r);
shm_unlink(shm_name);
int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd < 0 ) { perror("failed shm_open"); assert(0); }
ftruncate(fd, size);
int mmap_flag = MAP_SHARED;
#ifdef MAP_POPULATE
mmap_flag |= MAP_POPULATE;
#endif
#ifdef MAP_HUGETLB
if (flags) mmap_flag |= MAP_HUGETLB;
#endif
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
if ( ptr == (void * )MAP_FAILED ) { perror("failed mmap"); assert(0); }
assert(((uint64_t)ptr&0x3F)==0);
WorldShmCommBufs[r] =ptr;
close(fd);
}
}
MPI_Barrier(WorldShmComm);
if ( WorldShmRank != 0 ) {
for(int r=0;r<WorldShmSize;r++){
size_t size = bytes ;
sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldNode,r);
int fd=shm_open(shm_name,O_RDWR,0666);
if ( fd<0 ) { perror("failed shm_open"); assert(0); }
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if ( ptr == MAP_FAILED ) { perror("failed mmap"); assert(0); }
assert(((uint64_t)ptr&0x3F)==0);
WorldShmCommBufs[r] =ptr;
close(fd);
}
}
_ShmAlloc=1;
_ShmAllocBytes = bytes;
}
#endif
////////////////////////////////////////////////////////
// Global shared functionality finished
// Now move to per communicator functionality
////////////////////////////////////////////////////////
void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
{
int rank, size;
MPI_Comm_rank(comm,&rank);
MPI_Comm_size(comm,&size);
ShmRanks.resize(size);
/////////////////////////////////////////////////////////////////////
// Split into groups that can share memory
/////////////////////////////////////////////////////////////////////
MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
MPI_Comm_rank(ShmComm ,&ShmRank);
MPI_Comm_size(ShmComm ,&ShmSize);
ShmCommBufs.resize(ShmSize);
//////////////////////////////////////////////////////////////////////
// Map ShmRank to WorldShmRank and use the right buffer
//////////////////////////////////////////////////////////////////////
assert (GlobalSharedMemory::ShmAlloc()==1);
heap_size = GlobalSharedMemory::ShmAllocBytes();
for(int r=0;r<ShmSize;r++){
uint32_t sr = (r==ShmRank) ? GlobalSharedMemory::WorldRank : 0 ;
MPI_Allreduce(MPI_IN_PLACE,&sr,1,MPI_UINT32_T,MPI_SUM,comm);
ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[sr];
}
ShmBufferFreeAll();
/////////////////////////////////////////////////////////////////////
// find comm ranks in our SHM group (i.e. which ranks are on our node)
/////////////////////////////////////////////////////////////////////
MPI_Group FullGroup, ShmGroup;
MPI_Comm_group (comm , &FullGroup);
MPI_Comm_group (ShmComm, &ShmGroup);
std::vector<int> ranks(size); for(int r=0;r<size;r++) ranks[r]=r;
MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]);
}
//////////////////////////////////////////////////////////////////
// On node barrier
//////////////////////////////////////////////////////////////////
void SharedMemory::ShmBarrier(void)
{
MPI_Barrier (ShmComm);
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// Test the shared memory is working
//////////////////////////////////////////////////////////////////////////////////////////////////////////
void SharedMemory::SharedMemoryTest(void)
{
ShmBarrier();
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
uint64_t * check = (uint64_t *) ShmCommBufs[r];
check[0] = GlobalSharedMemory::WorldNode;
check[1] = r;
check[2] = 0x5A5A5A;
}
}
ShmBarrier();
for(int r=0;r<ShmSize;r++){
uint64_t * check = (uint64_t *) ShmCommBufs[r];
assert(check[0]==GlobalSharedMemory::WorldNode);
assert(check[1]==r);
assert(check[2]==0x5A5A5A);
}
ShmBarrier();
}
void *SharedMemory::ShmBuffer(int rank)
{
int gpeer = ShmRanks[rank];
if (gpeer == MPI_UNDEFINED){
return NULL;
} else {
return ShmCommBufs[gpeer];
}
}
void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
{
static int count =0;
int gpeer = ShmRanks[rank];
assert(gpeer!=ShmRank); // never send to self
if (gpeer == MPI_UNDEFINED){
return NULL;
} else {
uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
return (void *) remote;
}
}
}

126
lib/communicator/SharedMemoryNone.cc Normal file
View File

@ -0,0 +1,126 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/communicator/SharedMemory.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
namespace Grid {
/*Construct from an MPI communicator*/
void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
{
assert(_ShmSetup==0);
WorldComm = 0;
WorldRank = 0;
WorldSize = 1;
WorldShmComm = 0 ;
WorldShmRank = 0 ;
WorldShmSize = 1 ;
WorldNodes = 1 ;
WorldNode = 0 ;
WorldShmRanks.resize(WorldSize); WorldShmRanks[0] = 0;
WorldShmCommBufs.resize(1);
_ShmSetup=1;
}
void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
{
optimal_comm = WorldComm;
}
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbfs mapping intended, use anonymous mmap
////////////////////////////////////////////////////////////////////////////////////////////
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
void * ShmCommBuf ;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
int mmap_flag =0;
#ifdef MAP_ANONYMOUS
mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
#endif
#ifdef MAP_ANON
mmap_flag = mmap_flag| MAP_SHARED | MAP_ANON;
#endif
#ifdef MAP_HUGETLB
if ( flags ) mmap_flag |= MAP_HUGETLB;
#endif
ShmCommBuf =(void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag, -1, 0);
if (ShmCommBuf == (void *)MAP_FAILED) {
perror("mmap failed ");
exit(EXIT_FAILURE);
}
#ifdef MADV_HUGEPAGE
if (!Hugepages ) madvise(ShmCommBuf,bytes,MADV_HUGEPAGE);
#endif
bzero(ShmCommBuf,bytes);
WorldShmCommBufs[0] = ShmCommBuf;
_ShmAllocBytes=bytes;
_ShmAlloc=1;
};
////////////////////////////////////////////////////////
// Global shared functionality finished
// Now move to per communicator functionality
////////////////////////////////////////////////////////
void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
{
assert(GlobalSharedMemory::ShmAlloc()==1);
ShmRanks.resize(1);
ShmCommBufs.resize(1);
ShmRanks[0] = 0;
ShmRank = 0;
ShmSize = 1;
//////////////////////////////////////////////////////////////////////
// Map ShmRank to WorldShmRank and use the right buffer
//////////////////////////////////////////////////////////////////////
ShmCommBufs[0] = GlobalSharedMemory::WorldShmCommBufs[0];
heap_size = GlobalSharedMemory::ShmAllocBytes();
ShmBufferFreeAll();
return;
}
//////////////////////////////////////////////////////////////////
// On node barrier
//////////////////////////////////////////////////////////////////
void SharedMemory::ShmBarrier(void){ return ; }
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// Test the shared memory is working
//////////////////////////////////////////////////////////////////////////////////////////////////////////
void SharedMemory::SharedMemoryTest(void) { return; }
void *SharedMemory::ShmBuffer(int rank)
{
return NULL;
}
void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
{
return NULL;
}
}

1
lib/qcd/action/fermion/WilsonCompressor.h
View File

@ -265,7 +265,6 @@ public:
if ( timer3 ) std::cout << GridLogMessage << " timer3 (commsMergeShm) " <<timer3/calls <<std::endl;
if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
}
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
std::vector<int> same_node;
std::vector<int> surface_list;

1
lib/stencil/Stencil.h
View File

@ -105,7 +105,6 @@ template<class vobj,class cobj>
class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal fill in.
public:
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
typedef typename cobj::vector_type vector_type;
typedef typename cobj::scalar_type scalar_type;
typedef typename cobj::scalar_object scalar_object;

8
lib/util/Init.cc
View File

@ -220,11 +220,11 @@ void Grid_init(int *argc,char ***argv)
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--shm");
GridCmdOptionInt(arg,MB);
uint64_t MB64 = MB;
CartesianCommunicator::MAX_MPI_SHM_BYTES = MB64*1024LL*1024LL;
GlobalSharedMemory::MAX_MPI_SHM_BYTES = MB64*1024LL*1024LL;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--shm-hugepages") ){
CartesianCommunicator::Hugepages = 1;
GlobalSharedMemory::Hugepages = 1;
}
@ -398,8 +398,8 @@ void Grid_init(int *argc,char ***argv)
Grid_default_latt,
Grid_default_mpi);
std::cout << GridLogMessage << "Requesting "<< CartesianCommunicator::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
if ( CartesianCommunicator::Hugepages) {
std::cout << GridLogMessage << "Requesting "<< GlobalSharedMemory::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
if ( GlobalSharedMemory::Hugepages) {
std::cout << GridLogMessage << "Mapped stencil comms buffers as MAP_HUGETLB "<<std::endl;
}

2
scripts/filelist
View File

@ -6,7 +6,7 @@ home=`pwd`
cd $home/lib
HFILES=`find . -type f -name '*.h' -not -name '*Hdf5*' -not -path '*/gamma-gen/*' -not -path '*/Old/*' -not -path '*/Eigen/*'`
HFILES="$HFILES"
CCFILES=`find . -type f -name '*.cc' -not -path '*/gamma-gen/*' -not -name '*Communicator*.cc' -not -name '*Hdf5*'`
CCFILES=`find . -type f -name '*.cc' -not -path '*/gamma-gen/*' -not -name '*Communicator*.cc' -not -name '*SharedMemory*.cc' -not -name '*Hdf5*'`
HPPFILES=`find . -type f -name '*.hpp'`
echo HFILES=$HFILES $HPPFILES > Make.inc
echo >> Make.inc

330
tests/lanczos/Test_dwf_compressed_lanczos_reorg_synthetic.cc Normal file
View File

@ -0,0 +1,330 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./tests/Test_dwf_compressed_lanczos_reorg.cc
Copyright (C) 2017
Author: Leans heavily on Christoph Lehner's code
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
/*
* Reimplement the badly named "multigrid" lanczos as compressed Lanczos using the features
* in Grid that were intended to be used to support blocked Aggregates, from
*/
#include <Grid/Grid.h>
#include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
using namespace std;
using namespace Grid;
using namespace Grid::QCD;
template<class Fobj,class CComplex,int nbasis>
class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
public:
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj> FineField;
LinearOperatorBase<FineField> &_Linop;
Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
ProjectedHermOp(LinearOperatorBase<FineField>& linop, Aggregation<Fobj,CComplex,nbasis> &aggregate) :
_Linop(linop),
_Aggregate(aggregate) { };
void operator()(const CoarseField& in, CoarseField& out) {
GridBase *FineGrid = _Aggregate.FineGrid;
FineField fin(FineGrid);
FineField fout(FineGrid);
_Aggregate.PromoteFromSubspace(in,fin);
_Linop.HermOp(fin,fout);
_Aggregate.ProjectToSubspace(out,fout);
}
};
template<class Fobj,class CComplex,int nbasis>
class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
public:
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj> FineField;
OperatorFunction<FineField> & _poly;
LinearOperatorBase<FineField> &_Linop;
Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,LinearOperatorBase<FineField>& linop,
Aggregation<Fobj,CComplex,nbasis> &aggregate) :
_poly(poly),
_Linop(linop),
_Aggregate(aggregate) { };
void operator()(const CoarseField& in, CoarseField& out) {
GridBase *FineGrid = _Aggregate.FineGrid;
FineField fin(FineGrid) ;fin.checkerboard =_Aggregate.checkerboard;
FineField fout(FineGrid);fout.checkerboard =_Aggregate.checkerboard;
_Aggregate.PromoteFromSubspace(in,fin);
_poly(_Linop,fin,fout);
_Aggregate.ProjectToSubspace(out,fout);
}
};
// Make serializable Lanczos params
template<class Fobj,class CComplex,int nbasis>
class CoarseFineIRL
{
public:
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<Fobj> FineField;
private:
GridBase *_CoarseGrid;
GridBase *_FineGrid;
int _checkerboard;
LinearOperatorBase<FineField> & _FineOp;
Aggregation<Fobj,CComplex,nbasis> _Aggregate;
public:
CoarseFineIRL(GridBase *FineGrid,
GridBase *CoarseGrid,
LinearOperatorBase<FineField> &FineOp,
int checkerboard) :
_CoarseGrid(CoarseGrid),
_FineGrid(FineGrid),
_Aggregate(CoarseGrid,FineGrid,checkerboard),
_FineOp(FineOp),
_checkerboard(checkerboard)
{};
template<typename T> static RealD normalise(T& v)
{
RealD nn = norm2(v);
nn = ::sqrt(nn);
v = v * (1.0/nn);
return nn;
}
void testFine(void)
{
int Nk = nbasis;
_Aggregate.subspace.resize(Nk,_FineGrid);
_Aggregate.subspace[0]=1.0;
_Aggregate.subspace[0].checkerboard=_checkerboard;
normalise(_Aggregate.subspace[0]);
PlainHermOp<FineField> Op(_FineOp);
for(int k=1;k<Nk;k++){
Op(_Aggregate.subspace[k-1],_Aggregate.subspace[k]);
normalise(_Aggregate.subspace[k]);
std::cout << GridLogMessage << "testFine subspace "<<k<<" " <<norm2(_Aggregate.subspace[k])<<std::endl;
}
for(int k=0;k<Nk;k++){
std::cout << GridLogMessage << "testFine subspace "<<k<<" cb " <<_Aggregate.subspace[k].checkerboard<<std::endl;
}
_Aggregate.Orthogonalise();
}
void calcFine(RealD alpha, RealD beta,int Npoly,int Nm,RealD resid,
RealD MaxIt, RealD betastp, int MinRes)
{
assert(nbasis<=Nm);
Chebyshev<FineField> Cheby(alpha,beta,Npoly);
FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
PlainHermOp<FineField> Op(_FineOp);
int Nk = nbasis;
std::vector<RealD> eval(Nm);
FineField src(_FineGrid); src=1.0; src.checkerboard = _checkerboard;
ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nk,Nk,Nm,resid,MaxIt,betastp,MinRes);
_Aggregate.subspace.resize(Nm,_FineGrid);
IRL.calc(eval,_Aggregate.subspace,src,Nk,false);
_Aggregate.subspace.resize(Nk,_FineGrid);
for(int k=0;k<Nk;k++){
std::cout << GridLogMessage << "testFine subspace "<<k<<" cb " <<_Aggregate.subspace[k].checkerboard<<std::endl;
}
_Aggregate.Orthogonalise();
}
void calcCoarse(RealD alpha, RealD beta,int Npoly,
int Nk, int Nm,RealD resid,
RealD MaxIt, RealD betastp, int MinRes)
{
Chebyshev<FineField> Cheby(alpha,beta,Npoly);
ProjectedHermOp<Fobj,CComplex,nbasis> Op(_FineOp,_Aggregate);
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp(Cheby,_FineOp,_Aggregate);
std::vector<RealD> eval(Nm);
std::vector<CoarseField> evec(Nm,_CoarseGrid);
CoarseField src(_CoarseGrid); src=1.0;
ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,Nk,Nk,Nm,resid,MaxIt,betastp,MinRes);
IRL.calc(eval,evec,src,Nk,false);
// We got the evalues of the Cheby operator;
// Reconstruct eigenvalues of original operator via Chebyshev inverse
for (int i=0;i<Nk;i++){
RealD eval_guess;
if (i==0) eval_guess = 0;
else eval[i-1] = 0;
RealD eval_poly = eval[i];
RealD eval_op = Cheby.approxInv(eval_poly,eval_guess,100,1e-10);
std::cout << i << " Reconstructed eval = " << eval_op << " from quess " << eval_op << std::endl;
eval[i] = eval_op;
}
}
};
struct LanczosParams : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParams,
RealD, alpha,
RealD, beta,
int, Npoly,
int, Nk,
int, Nm,
RealD, resid,
int, MaxIt,
RealD, betastp,
int, MinRes);
};
struct CompressedLanczosParams : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(CompressedLanczosParams,
LanczosParams, FineParams,
LanczosParams, CoarseParams,
std::vector<int>, blockSize,
std::string, config,
std::vector < std::complex<double> >, omega,
RealD, mass,
RealD, M5
);
};
int main (int argc, char ** argv) {
Grid_init(&argc,&argv);
CompressedLanczosParams Params;
{
Params.omega.resize(10);
Params.blockSize.resize(5);
XmlWriter writer("Params_template.xml");
write(writer,"Params",Params);
std::cout << GridLogMessage << " Written Params_template.xml" <<std::endl;
}
{
XmlReader reader("./Params.xml");
read(reader, "Params", Params);
}
int Ls = (int)Params.omega.size();
RealD mass = Params.mass;
RealD M5 = Params.M5;
std::vector<int> blockSize = Params.blockSize;
// Grids
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
std::vector<int> fineLatt = GridDefaultLatt();
int dims=fineLatt.size();
assert(blockSize.size()==dims+1);
std::vector<int> coarseLatt(dims);
std::vector<int> coarseLatt5d ;
for (int d=0;d<coarseLatt.size();d++){
coarseLatt[d] = fineLatt[d]/blockSize[d]; assert(coarseLatt[d]*blockSize[d]==fineLatt[d]);
}
std::cout << GridLogMessage<< " 5d coarse lattice is ";
for (int i=0;i<coarseLatt.size();i++){
std::cout << coarseLatt[i]<<"x";
}
int cLs = Ls/blockSize[dims]; assert(cLs*blockSize[dims]==Ls);
std::cout << cLs<<std::endl;
GridCartesian * CoarseGrid4 = SpaceTimeGrid::makeFourDimGrid(coarseLatt, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * CoarseGrid4rb = SpaceTimeGrid::makeFourDimRedBlackGrid(CoarseGrid4);
GridCartesian * CoarseGrid5 = SpaceTimeGrid::makeFiveDimGrid(cLs,CoarseGrid4);
GridRedBlackCartesian * CoarseGrid5rb = SpaceTimeGrid::makeFourDimRedBlackGrid(CoarseGrid5);
// Gauge field
LatticeGaugeField Umu(UGrid);
// FieldMetaData header;
// NerscIO::readConfiguration(Umu,header,Params.config);
{
std::vector<int> seeds4({1,2,3,4});
GridParallelRNG RNG4(UGrid); RNG4.SeedFixedIntegers(seeds4);
SU3::HotConfiguration(RNG4, Umu);
}
std::cout << GridLogMessage << "Lattice dimensions: " << GridDefaultLatt() << " Ls: " << Ls << std::endl;
// ZMobius EO Operator
ZMobiusFermionR Ddwf(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5, Params.omega,1.,0.);
SchurDiagTwoOperator<ZMobiusFermionR,LatticeFermion> HermOp(Ddwf);
// Eigenvector storage
LanczosParams fine =Params.FineParams;
LanczosParams coarse=Params.CoarseParams;
const int Nm1 = fine.Nm;
const int Nm2 = coarse.Nm;
std::cout << GridLogMessage << "Keep " << fine.Nk << " full vectors" << std::endl;
std::cout << GridLogMessage << "Keep " << coarse.Nk << " total vectors" << std::endl;
assert(Nm2 >= Nm1);
const int nbasis= 70;
CoarseFineIRL<vSpinColourVector,vTComplex,nbasis> IRL(FrbGrid,CoarseGrid5rb,HermOp,Odd);
std::cout << GridLogMessage << "Constructed CoarseFine IRL" << std::endl;
std::cout << GridLogMessage << "Performing fine grid IRL Nk "<< nbasis<<" Nm "<<Nm1<< std::endl;
IRL.testFine();
// IRL.calcFine(fine.alpha,fine.beta,fine.Npoly,fine.Nm,fine.resid, fine.MaxIt, fine.betastp, fine.MinRes);
std::cout << GridLogMessage << "Performing coarse grid (poly) IRL " << nbasis<<" Nm "<<Nm2<< std::endl;
IRL.calcCoarse(coarse.alpha,coarse.beta,coarse.Npoly,coarse.Nk,coarse.Nm,coarse.resid, coarse.MaxIt, coarse.betastp, coarse.MinRes);
// IRL.smoothedCoarseEigenvalues();
Grid_finalize();
}

9
tests/solver/Test_dwf_mrhs_cg_mpi.cc
View File

@ -72,14 +72,17 @@ int main (int argc, char ** argv)
int nrhs = 1;
int me;
for(int i=0;i<mpi_layout.size();i++) nrhs *= (mpi_layout[i]/mpi_split[i]);
std::cout << GridLogMessage << "Creating split grids " <<std::endl;
GridCartesian * SGrid = new GridCartesian(GridDefaultLatt(),
GridDefaultSimd(Nd,vComplex::Nsimd()),
mpi_split,
*UGrid,me);
std::cout << GridLogMessage <<"Creating split ferm grids " <<std::endl;
GridCartesian * SFGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,SGrid);
std::cout << GridLogMessage <<"Creating split rb grids " <<std::endl;
GridRedBlackCartesian * SrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(SGrid);
std::cout << GridLogMessage <<"Creating split ferm rb grids " <<std::endl;
GridRedBlackCartesian * SFrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,SGrid);
std::cout << GridLogMessage << "Made the grids"<<std::endl;
///////////////////////////////////////////////
@ -119,7 +122,7 @@ int main (int argc, char ** argv)
GridParallelRNG pRNG5(FGrid); pRNG5.SeedFixedIntegers(seeds);
for(int s=0;s<nrhs;s++) {
random(pRNG5,src[s]);
tmp = 100.0*s;
tmp = 10.0*s;
src[s] = (src[s] * 0.1) + tmp;
std::cout << GridLogMessage << " src ["<<s<<"] "<<norm2(src[s])<<std::endl;
}
@ -133,7 +136,7 @@ int main (int argc, char ** argv)
pRNG.SeedFixedIntegers(seeds);
std::cout << GridLogMessage << "Intialised 4D RNG "<<std::endl;
SU3::HotConfiguration(pRNG,Umu);
std::cout << "Intialised the HOT Gauge Field"<<std::endl;
std::cout << GridLogMessage << "Intialised the HOT Gauge Field"<<std::endl;
// std::cout << " Site zero "<< Umu._odata[0] <<std::endl;
} else {
SU3::ColdConfiguration(Umu);