mirror of
https://github.com/paboyle/Grid.git
synced 2025-08-12 09:17:05 +01:00
Update to use shared memory to contain the stencil comms buffers
Tested on 2.1.1.1 1.2.1.1 4.1.1.1 1.4.1.1 2.2.1.1 subnode decompositions
This commit is contained in:
azusayamaguchi
@@ -40,169 +40,188 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
#ifdef GRID_COMMS_SHMEM
|
||||
#include <mpp/shmem.h>
|
||||
#endif
|
||||
|
||||
namespace Grid {
|
||||
|
||||
class CartesianCommunicator {
|
||||
public:
|
||||
|
||||
// 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 const uint64_t MAX_MPI_SHM_BYTES = 128*1024*1024;
|
||||
|
||||
// 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;
|
||||
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;
|
||||
|
||||
#ifdef GRID_COMMS_MPI
|
||||
MPI_Comm communicator;
|
||||
typedef MPI_Request CommsRequest_t;
|
||||
#elif GRID_COMMS_MPI3
|
||||
int shm_mode;
|
||||
|
||||
MPI_Comm communicator;
|
||||
typedef MPI_Request CommsRequest_t;
|
||||
|
||||
const int MAXLOG2RANKSPERNODE = 16; // 65536 ranks per node adequate for now
|
||||
const uint64_t MAX_MPI_SHM_BYTES = 256*1024*1024; // 256MB shared memory for comms enought for 48^4 local vol comms
|
||||
|
||||
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;
|
||||
|
||||
void * ShmCommBuf;
|
||||
std::vector<void *> ShmCommBufs;
|
||||
|
||||
int WorldRank;
|
||||
int WorldSize;
|
||||
|
||||
static int ShmRank;
|
||||
static int ShmSize;
|
||||
static int GroupSize;
|
||||
static int GroupRank;
|
||||
|
||||
std::vector<int> LexicographicToWorldRank;
|
||||
#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3)
|
||||
MPI_Comm communicator;
|
||||
static MPI_Comm communicator_world;
|
||||
typedef MPI_Request CommsRequest_t;
|
||||
#else
|
||||
typedef int CommsRequest_t;
|
||||
typedef int CommsRequest_t;
|
||||
#endif
|
||||
|
||||
static void Init(int *argc, char ***argv);
|
||||
////////////////////////////////////////////////////////////////////
|
||||
// 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
|
||||
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
|
||||
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) ;
|
||||
|
||||
////////////////////////////////////////////////
|
||||
// Must call in Grid startup
|
||||
////////////////////////////////////////////////
|
||||
static void Init(int *argc, char ***argv);
|
||||
|
||||
////////////////////////////////////////////////
|
||||
// Constructor of any given grid
|
||||
////////////////////////////////////////////////
|
||||
CartesianCommunicator(const std::vector<int> &pdimensions_in);
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////
|
||||
// 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);
|
||||
|
||||
/////////////////////////////////
|
||||
// Grid information and queries
|
||||
/////////////////////////////////
|
||||
static int ShmRank;
|
||||
static int ShmSize;
|
||||
static int GroupSize;
|
||||
static int GroupRank;
|
||||
static int WorldRank;
|
||||
static int WorldSize;
|
||||
static int Slave;
|
||||
|
||||
int IsBoss(void) ;
|
||||
int BossRank(void) ;
|
||||
int ThisRank(void) ;
|
||||
const std::vector<int> & ThisProcessorCoor(void) ;
|
||||
const std::vector<int> & ProcessorGrid(void) ;
|
||||
int ProcessorCount(void) ;
|
||||
static int Ranks (void);
|
||||
static int Nodes (void);
|
||||
static int Cores (void);
|
||||
static int NodeRank (void);
|
||||
static int CoreRank (void);
|
||||
|
||||
// Constructor
|
||||
CartesianCommunicator(const std::vector<int> &pdimensions_in);
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
// very VERY rarely (Log, serial RNG) we need world without a grid
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
static int RankWorld(void) ;
|
||||
static void BroadcastWorld(int root,void* data, int bytes);
|
||||
|
||||
////////////////////////////////////////////////////////////
|
||||
// Reduction
|
||||
////////////////////////////////////////////////////////////
|
||||
void GlobalSum(RealF &);
|
||||
void GlobalSumVector(RealF *,int N);
|
||||
void GlobalSum(RealD &);
|
||||
void GlobalSumVector(RealD *,int N);
|
||||
void GlobalSum(uint32_t &);
|
||||
void GlobalSum(uint64_t &);
|
||||
void GlobalSum(ComplexF &c);
|
||||
void GlobalSumVector(ComplexF *c,int N);
|
||||
void GlobalSum(ComplexD &c);
|
||||
void GlobalSumVector(ComplexD *c,int N);
|
||||
|
||||
template<class obj> void GlobalSum(obj &o){
|
||||
typedef typename obj::scalar_type scalar_type;
|
||||
int words = sizeof(obj)/sizeof(scalar_type);
|
||||
scalar_type * ptr = (scalar_type *)& o;
|
||||
GlobalSumVector(ptr,words);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////
|
||||
// Face exchange, buffer swap in translational invariant way
|
||||
////////////////////////////////////////////////////////////
|
||||
void SendToRecvFrom(void *xmit,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,
|
||||
int bytes);
|
||||
|
||||
void SendRecvPacket(void *xmit,
|
||||
void *recv,
|
||||
int xmit_to_rank,
|
||||
int recv_from_rank,
|
||||
int bytes);
|
||||
|
||||
void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
|
||||
void *xmit,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,
|
||||
int bytes);
|
||||
|
||||
void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
|
||||
|
||||
// Wraps MPI_Cart routines
|
||||
void ShiftedRanks(int dim,int shift,int & source, int & dest);
|
||||
int RankFromProcessorCoor(std::vector<int> &coor);
|
||||
void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
|
||||
void StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
|
||||
void *xmit,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,
|
||||
int bytes);
|
||||
|
||||
void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
|
||||
void StencilBarrier(void);
|
||||
|
||||
// Helper function for SHM Windows in MPI3
|
||||
void *ShmBufferSelf(void);
|
||||
void *ShmBuffer(int rank);
|
||||
|
||||
/////////////////////////////////
|
||||
// Grid information queries
|
||||
/////////////////////////////////
|
||||
int IsBoss(void) { return _processor==0; };
|
||||
int BossRank(void) { return 0; };
|
||||
int ThisRank(void) { return _processor; };
|
||||
const std::vector<int> & ThisProcessorCoor(void) { return _processor_coor; };
|
||||
const std::vector<int> & ProcessorGrid(void) { return _processors; };
|
||||
int ProcessorCount(void) { return _Nprocessors; };
|
||||
|
||||
////////////////////////////////////////////////////////////
|
||||
// Reduction
|
||||
////////////////////////////////////////////////////////////
|
||||
void GlobalSum(RealF &);
|
||||
void GlobalSumVector(RealF *,int N);
|
||||
|
||||
void GlobalSum(RealD &);
|
||||
void GlobalSumVector(RealD *,int N);
|
||||
|
||||
void GlobalSum(uint32_t &);
|
||||
void GlobalSum(uint64_t &);
|
||||
|
||||
void GlobalSum(ComplexF &c)
|
||||
{
|
||||
GlobalSumVector((float *)&c,2);
|
||||
}
|
||||
void GlobalSumVector(ComplexF *c,int N)
|
||||
{
|
||||
GlobalSumVector((float *)c,2*N);
|
||||
}
|
||||
|
||||
void GlobalSum(ComplexD &c)
|
||||
{
|
||||
GlobalSumVector((double *)&c,2);
|
||||
}
|
||||
void GlobalSumVector(ComplexD *c,int N)
|
||||
{
|
||||
GlobalSumVector((double *)c,2*N);
|
||||
}
|
||||
|
||||
template<class obj> void GlobalSum(obj &o){
|
||||
typedef typename obj::scalar_type scalar_type;
|
||||
int words = sizeof(obj)/sizeof(scalar_type);
|
||||
scalar_type * ptr = (scalar_type *)& o;
|
||||
GlobalSumVector(ptr,words);
|
||||
}
|
||||
////////////////////////////////////////////////////////////
|
||||
// Face exchange, buffer swap in translational invariant way
|
||||
////////////////////////////////////////////////////////////
|
||||
void SendToRecvFrom(void *xmit,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,
|
||||
int bytes);
|
||||
|
||||
void SendRecvPacket(void *xmit,
|
||||
void *recv,
|
||||
int xmit_to_rank,
|
||||
int recv_from_rank,
|
||||
int bytes);
|
||||
|
||||
void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
|
||||
void *xmit,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,
|
||||
int bytes);
|
||||
void SendToRecvFromComplete(std::vector<CommsRequest_t> &waitall);
|
||||
void StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
|
||||
void *xmit,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,
|
||||
int bytes);
|
||||
void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
|
||||
{
|
||||
SendToRecvFromComplete(waitall);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////
|
||||
// Barrier
|
||||
////////////////////////////////////////////////////////////
|
||||
void Barrier(void);
|
||||
|
||||
////////////////////////////////////////////////////////////
|
||||
// Broadcast a buffer and composite larger
|
||||
////////////////////////////////////////////////////////////
|
||||
void Broadcast(int root,void* data, int bytes);
|
||||
template<class obj> void Broadcast(int root,obj &data)
|
||||
////////////////////////////////////////////////////////////
|
||||
// Barrier
|
||||
////////////////////////////////////////////////////////////
|
||||
void Barrier(void);
|
||||
|
||||
////////////////////////////////////////////////////////////
|
||||
// Broadcast a buffer and composite larger
|
||||
////////////////////////////////////////////////////////////
|
||||
void Broadcast(int root,void* data, int bytes);
|
||||
|
||||
template<class obj> void Broadcast(int root,obj &data)
|
||||
{
|
||||
Broadcast(root,(void *)&data,sizeof(data));
|
||||
};
|
||||
|
||||
static void BroadcastWorld(int root,void* data, int bytes);
|
||||
|
||||
};
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user