mirror of
https://github.com/paboyle/Grid.git
synced 2024-11-10 07:55:35 +00:00
Coordinate handling GPU ready. No malloc
This commit is contained in:
parent
eed9aa9f0c
commit
34820bec27
@ -28,6 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
#ifndef GRID_COMMUNICATOR_H
|
||||
#define GRID_COMMUNICATOR_H
|
||||
|
||||
#include <Grid/util/Coordinate.h>
|
||||
#include <Grid/communicator/SharedMemory.h>
|
||||
#include <Grid/communicator/Communicator_base.h>
|
||||
|
||||
|
@ -47,8 +47,8 @@ int CartesianCommunicator::Dimensions(void) { return
|
||||
int CartesianCommunicator::IsBoss(void) { return _processor==0; };
|
||||
int CartesianCommunicator::BossRank(void) { return 0; };
|
||||
int CartesianCommunicator::ThisRank(void) { return _processor; };
|
||||
const std::vector<int> & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
|
||||
const std::vector<int> & CartesianCommunicator::ProcessorGrid(void) { return _processors; };
|
||||
const Coordinate & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
|
||||
const Coordinate & CartesianCommunicator::ProcessorGrid(void) { return _processors; };
|
||||
int CartesianCommunicator::ProcessorCount(void) { return _Nprocessors; };
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
@ -52,9 +52,9 @@ public:
|
||||
// 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.
|
||||
Coordinate _processors; // Which dimensions get relayed out over processors lanes.
|
||||
int _processor; // linear processor rank
|
||||
std::vector<int> _processor_coor; // linear processor coordinate
|
||||
Coordinate _processor_coor; // linear processor coordinate
|
||||
unsigned long _ndimension;
|
||||
static Grid_MPI_Comm communicator_world;
|
||||
Grid_MPI_Comm communicator;
|
||||
@ -69,8 +69,8 @@ public:
|
||||
// Constructors to sub-divide a parent communicator
|
||||
// and default to comm world
|
||||
////////////////////////////////////////////////
|
||||
CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank);
|
||||
CartesianCommunicator(const std::vector<int> &pdimensions_in);
|
||||
CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank);
|
||||
CartesianCommunicator(const Coordinate &pdimensions_in);
|
||||
virtual ~CartesianCommunicator();
|
||||
|
||||
private:
|
||||
@ -79,7 +79,7 @@ private:
|
||||
// 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, Grid_MPI_Comm communicator_base);
|
||||
void InitFromMPICommunicator(const Coordinate &processors, Grid_MPI_Comm communicator_base);
|
||||
|
||||
public:
|
||||
|
||||
@ -87,15 +87,15 @@ public:
|
||||
// Wraps MPI_Cart routines, or implements equivalent on other impls
|
||||
////////////////////////////////////////////////////////////////////////////////////////
|
||||
void ShiftedRanks(int dim,int shift,int & source, int & dest);
|
||||
int RankFromProcessorCoor(std::vector<int> &coor);
|
||||
void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
|
||||
int RankFromProcessorCoor(Coordinate &coor);
|
||||
void ProcessorCoorFromRank(int rank,Coordinate &coor);
|
||||
|
||||
int Dimensions(void) ;
|
||||
int IsBoss(void) ;
|
||||
int BossRank(void) ;
|
||||
int ThisRank(void) ;
|
||||
const std::vector<int> & ThisProcessorCoor(void) ;
|
||||
const std::vector<int> & ProcessorGrid(void) ;
|
||||
const Coordinate & ThisProcessorCoor(void) ;
|
||||
const Coordinate & ProcessorGrid(void) ;
|
||||
int ProcessorCount(void) ;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
@ -65,14 +65,14 @@ 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 CartesianCommunicator::RankFromProcessorCoor(Coordinate &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)
|
||||
void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
|
||||
{
|
||||
coor.resize(_ndimension);
|
||||
int ierr=MPI_Cart_coords (communicator, rank, _ndimension,&coor[0]);
|
||||
@ -82,7 +82,7 @@ void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &c
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Initialises from communicator_world
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
|
||||
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
|
||||
{
|
||||
MPI_Comm optimal_comm;
|
||||
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm); // Remap using the shared memory optimising routine
|
||||
@ -93,13 +93,13 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
|
||||
//////////////////////////////////
|
||||
// Try to subdivide communicator
|
||||
//////////////////////////////////
|
||||
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)
|
||||
CartesianCommunicator::CartesianCommunicator(const Coordinate &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);
|
||||
Coordinate parent_processor_coor(_ndimension,0);
|
||||
Coordinate parent_processors (_ndimension,1);
|
||||
|
||||
// Can make 5d grid from 4d etc...
|
||||
int pad = _ndimension-parent_ndimension;
|
||||
@ -126,9 +126,9 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
|
||||
|
||||
// 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
|
||||
Coordinate ccoor(_ndimension); // coor within subcommunicator
|
||||
Coordinate scoor(_ndimension); // coor of split within parent
|
||||
Coordinate ssize(_ndimension); // coor of split within parent
|
||||
|
||||
for(int d=0;d<_ndimension;d++){
|
||||
ccoor[d] = parent_processor_coor[d] % processors[d];
|
||||
@ -208,7 +208,7 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
|
||||
}
|
||||
}
|
||||
|
||||
void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
|
||||
void CartesianCommunicator::InitFromMPICommunicator(const Coordinate &processors, MPI_Comm communicator_base)
|
||||
{
|
||||
_ndimension = processors.size();
|
||||
_processor_coor.resize(_ndimension);
|
||||
@ -222,7 +222,7 @@ void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &proc
|
||||
_Nprocessors*=_processors[i];
|
||||
}
|
||||
|
||||
std::vector<int> periodic(_ndimension,1);
|
||||
Coordinate 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]);
|
||||
@ -459,7 +459,7 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
|
||||
|
||||
void CartesianCommunicator::AllToAll(int dim,void *in,void *out,uint64_t words,uint64_t bytes)
|
||||
{
|
||||
std::vector<int> row(_ndimension,1);
|
||||
Coordinate row(_ndimension,1);
|
||||
assert(dim>=0 && dim<_ndimension);
|
||||
|
||||
// Split the communicator
|
||||
|
@ -42,14 +42,14 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
|
||||
GlobalSharedMemory::Hugepages);
|
||||
}
|
||||
|
||||
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)
|
||||
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
|
||||
: CartesianCommunicator(processors)
|
||||
{
|
||||
srank=0;
|
||||
SetCommunicator(communicator_world);
|
||||
}
|
||||
|
||||
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
|
||||
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
|
||||
{
|
||||
_processors = processors;
|
||||
_ndimension = processors.size();
|
||||
@ -122,8 +122,8 @@ int CartesianCommunicator::RankWorld(void){return 0;}
|
||||
void CartesianCommunicator::Barrier(void){}
|
||||
void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
|
||||
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
|
||||
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor) { return 0;}
|
||||
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor){ coor = _processor_coor; }
|
||||
int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) { return 0;}
|
||||
void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){ coor = _processor_coor; }
|
||||
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
|
||||
{
|
||||
source =0;
|
||||
|
@ -90,7 +90,7 @@ public:
|
||||
// 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
|
||||
static void OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
|
||||
///////////////////////////////////////////////////
|
||||
// Provide shared memory facilities off comm world
|
||||
///////////////////////////////////////////////////
|
||||
|
@ -114,7 +114,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
|
||||
_ShmSetup=1;
|
||||
}
|
||||
|
||||
void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
|
||||
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
|
||||
{
|
||||
////////////////////////////////////////////////////////////////
|
||||
// Assert power of two shm_size.
|
||||
@ -133,9 +133,9 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
|
||||
// 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);
|
||||
Coordinate processor_coor(ndimension);
|
||||
Coordinate WorldDims = processors; Coordinate ShmDims(ndimension,1); Coordinate NodeDims (ndimension);
|
||||
Coordinate ShmCoor(ndimension); Coordinate NodeCoor(ndimension); Coordinate WorldCoor(ndimension);
|
||||
int dim = 0;
|
||||
for(int l2=0;l2<log2size;l2++){
|
||||
while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
|
||||
|
@ -47,7 +47,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
|
||||
_ShmSetup=1;
|
||||
}
|
||||
|
||||
void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
|
||||
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
|
||||
{
|
||||
optimal_comm = WorldComm;
|
||||
}
|
||||
|
Loading…
Reference in New Issue
Block a user