portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 15:55:37 +00:00
Code Releases Activity
b6a4c31b48
Grid/lib/stencil/Stencil.h

1170 lines
37 KiB
C++
Raw Blame History

/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/Stencil.h
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 */
#ifndef GRID_STENCIL_H
#define GRID_STENCIL_H
#include <Grid/stencil/SimpleCompressor.h> // subdir aggregate
#include <Grid/stencil/Lebesgue.h> // subdir aggregate
//////////////////////////////////////////////////////////////////////////////////////////
// Must not lose sight that goal is to be able to construct really efficient
// gather to a point stencil code. CSHIFT is not the best way, so need
// additional stencil support.
//
// Stencil based code will exchange haloes and use a table lookup for neighbours.
// This will be done with generality to allow easier efficient implementations.
// Overlap of comms and compute is enabled by tabulating off-node connected,
//
// Generic services
// 0) Prebuild neighbour tables
// 1) Compute sizes of all haloes/comms buffers; allocate them.
// 2) Gather all faces, and communicate.
// 3) Loop over result sites, giving nbr index/offnode info for each
//
//////////////////////////////////////////////////////////////////////////////////////////
namespace Grid {
///////////////////////////////////////////////////////////////////
// Gather for when there *is* need to SIMD split with compression
///////////////////////////////////////////////////////////////////
void Gather_plane_table_compute (GridBase *grid,int dimension,int plane,int cbmask,
int off,std::vector<std::pair<int,int> > & table);
template<class vobj,class cobj,class compressor>
void Gather_plane_simple_table (std::vector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,cobj *buffer,compressor &compress, int off,int so) __attribute__((noinline));
template<class vobj,class cobj,class compressor>
void Gather_plane_simple_table (std::vector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,cobj *buffer,compressor &compress, int off,int so)
{
int num=table.size();
parallel_for(int i=0;i<num;i++){
compress.Compress(&buffer[off],table[i].first,rhs._odata[so+table[i].second]);
}
}
///////////////////////////////////////////////////////////////////
// Gather for when there *is* need to SIMD split with compression
///////////////////////////////////////////////////////////////////
template<class cobj,class vobj,class compressor>
void Gather_plane_exchange_table(const Lattice<vobj> &rhs,
std::vector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
template<class cobj,class vobj,class compressor>
void Gather_plane_exchange_table(std::vector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
std::vector<cobj *> pointers,int dimension,int plane,int cbmask,
compressor &compress,int type)
{
assert( (table.size()&0x1)==0);
int num=table.size()/2;
int so = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane
parallel_for(int j=0;j<num;j++){
compress.CompressExchange(&pointers[0][0],&pointers[1][0],&rhs._odata[0],
j,so+table[2*j].second,so+table[2*j+1].second,type);
}
}
struct StencilEntry {
uint64_t _offset;
uint64_t _byte_offset;
uint16_t _is_local;
uint16_t _permute;
uint16_t _around_the_world; //256 bits, 32 bytes, 1/2 cacheline
uint16_t _pad;
};
////////////////////////////////////////
// The Stencil Class itself
////////////////////////////////////////
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;
///////////////////////////////////////////
// Helper structs
///////////////////////////////////////////
struct Packet {
void * send_buf;
void * recv_buf;
Integer to_rank;
Integer from_rank;
Integer bytes;
};
struct Merge {
cobj * mpointer;
std::vector<scalar_object *> rpointers;
std::vector<cobj *> vpointers;
Integer buffer_size;
Integer type;
};
struct Decompress {
cobj * kernel_p;
cobj * mpi_p;
Integer buffer_size;
};
////////////////////////////////////////
// Basic Grid and stencil info
////////////////////////////////////////
int face_table_computed;
std::vector<std::vector<std::pair<int,int> > > face_table ;
int _checkerboard;
int _npoints; // Move to template param?
GridBase * _grid;
// npoints of these
std::vector<int> _directions;
std::vector<int> _distances;
std::vector<int> _comm_buf_size;
std::vector<int> _permute_type;
Vector<StencilEntry> _entries;
std::vector<Packet> Packets;
std::vector<Merge> Mergers;
std::vector<Merge> MergersSHM;
std::vector<Decompress> Decompressions;
std::vector<Decompress> DecompressionsSHM;
///////////////////////////////////////////////////////////
// Unified Comms buffers for all directions
///////////////////////////////////////////////////////////
// Vectors that live on the symmetric heap in case of SHMEM
// These are used; either SHM objects or refs to the above symmetric heap vectors
// depending on comms target
cobj* u_recv_buf_p;
cobj* u_send_buf_p;
std::vector<cobj *> u_simd_send_buf;
std::vector<cobj *> u_simd_recv_buf;
int u_comm_offset;
int _unified_buffer_size;
cobj *CommBuf(void) { return u_recv_buf_p; }
/////////////////////////////////////////
// Timing info; ugly; possibly temporary
/////////////////////////////////////////
double commtime;
double mpi3synctime;
double mpi3synctime_g;
double shmmergetime;
double gathertime;
double gathermtime;
double halogtime;
double mergetime;
double decompresstime;
double comms_bytes;
double splicetime;
double nosplicetime;
double calls;
std::vector<double> comm_bytes_thr;
std::vector<double> comm_time_thr;
std::vector<double> comm_enter_thr;
std::vector<double> comm_leave_thr;
////////////////////////////////////////
// Stencil query
////////////////////////////////////////
inline int SameNode(int point) {
int dimension = _directions[point];
int displacement = _distances[point];
assert( (displacement==1) || (displacement==-1));
int pd = _grid->_processors[dimension];
int fd = _grid->_fdimensions[dimension];
int ld = _grid->_ldimensions[dimension];
int rd = _grid->_rdimensions[dimension];
int simd_layout = _grid->_simd_layout[dimension];
int comm_dim = _grid->_processors[dimension] >1 ;
int recv_from_rank;
int xmit_to_rank;
if ( ! comm_dim ) return 1;
int nbr_proc;
if (displacement==1) nbr_proc = 1;
else nbr_proc = pd-1;
_grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
void *shm = (void *) _grid->ShmBufferTranslate(recv_from_rank,u_recv_buf_p);
if ( shm==NULL ) return 0;
return 1;
}
inline int GetNodeLocal(int osite,int point) {
return _entries[point+_npoints*osite]._is_local;
}
inline StencilEntry * GetEntry(int &ptype,int point,int osite) {
ptype = _permute_type[point]; return & _entries[point+_npoints*osite];
}
inline uint64_t GetInfo(int &ptype,int &local,int &perm,int point,int ent,uint64_t base) {
uint64_t cbase = (uint64_t)&u_recv_buf_p[0];
local = _entries[ent]._is_local;
perm = _entries[ent]._permute;
if (perm) ptype = _permute_type[point];
if (local) {
return base + _entries[ent]._byte_offset;
} else {
return cbase + _entries[ent]._byte_offset;
}
}
inline uint64_t GetPFInfo(int ent,uint64_t base) {
uint64_t cbase = (uint64_t)&u_recv_buf_p[0];
int local = _entries[ent]._is_local;
if (local) return base + _entries[ent]._byte_offset;
else return cbase + _entries[ent]._byte_offset;
}
//////////////////////////////////////////
// Comms packet queue for asynch thread
//////////////////////////////////////////
void CommunicateThreaded()
{
#ifdef GRID_OMP
// must be called in parallel region
int mythread = omp_get_thread_num();
int nthreads = CartesianCommunicator::nCommThreads;
#else
int mythread = 0;
int nthreads = 1;
#endif
if (nthreads == -1) nthreads = 1;
if (mythread < nthreads) {
comm_enter_thr[mythread] = usecond();
for (int i = mythread; i < Packets.size(); i += nthreads) {
uint64_t bytes = _grid->StencilSendToRecvFrom(Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
comm_bytes_thr[mythread] += bytes;
}
comm_leave_thr[mythread]= usecond();
comm_time_thr[mythread] += comm_leave_thr[mythread] - comm_enter_thr[mythread];
}
}
void CollateThreads(void)
{
int nthreads = CartesianCommunicator::nCommThreads;
double first=0.0;
double last =0.0;
for(int t=0;t<nthreads;t++) {
double t0 = comm_enter_thr[t];
double t1 = comm_leave_thr[t];
comms_bytes+=comm_bytes_thr[t];
comm_enter_thr[t] = 0.0;
comm_leave_thr[t] = 0.0;
comm_time_thr[t] = 0.0;
comm_bytes_thr[t]=0;
if ( first == 0.0 ) first = t0; // first is t0
if ( (t0 > 0.0) && ( t0 < first ) ) first = t0; // min time seen
if ( t1 > last ) last = t1; // max time seen
}
commtime+= last-first;
}
void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
{
reqs.resize(Packets.size());
commtime-=usecond();
for(int i=0;i<Packets.size();i++){
comms_bytes+=_grid->StencilSendToRecvFromBegin(reqs[i],
Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
}
}
void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
{
for(int i=0;i<Packets.size();i++){
_grid->StencilSendToRecvFromComplete(reqs[i],i);
}
commtime+=usecond();
}
void Communicate(void)
{
#ifdef GRID_OMP
#pragma omp parallel
{
// must be called in parallel region
int mythread = omp_get_thread_num();
int maxthreads= omp_get_max_threads();
int nthreads = CartesianCommunicator::nCommThreads;
assert(nthreads <= maxthreads);
if (nthreads == -1) nthreads = 1;
#else
int mythread = 0;
int nthreads = 1;
#endif
if (mythread < nthreads) {
for (int i = mythread; i < Packets.size(); i += nthreads) {
double start = usecond();
comm_bytes_thr[mythread] += _grid->StencilSendToRecvFrom(Packets[i].send_buf,
Packets[i].to_rank,
Packets[i].recv_buf,
Packets[i].from_rank,
Packets[i].bytes,i);
comm_time_thr[mythread] += usecond() - start;
}
}
#ifdef GRID_OMP
}
#endif
}
template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress)
{
std::vector<std::vector<CommsRequest_t> > reqs;
Prepare();
HaloGather(source,compress);
// Concurrent
//CommunicateBegin(reqs);
//CommunicateComplete(reqs);
// Sequential, possibly threaded
Communicate();
CommsMergeSHM(compress);
CommsMerge(compress);
}
template<class compressor> int HaloGatherDir(const Lattice<vobj> &source,compressor &compress,int point,int & face_idx)
{
int dimension = _directions[point];
int displacement = _distances[point];
int fd = _grid->_fdimensions[dimension];
int rd = _grid->_rdimensions[dimension];
// Map to always positive shift modulo global full dimension.
int shift = (displacement+fd)%fd;
assert (source.checkerboard== _checkerboard);
// the permute type
int simd_layout = _grid->_simd_layout[dimension];
int comm_dim = _grid->_processors[dimension] >1 ;
int splice_dim = _grid->_simd_layout[dimension]>1 && (comm_dim);
int same_node = 1;
// Gather phase
int sshift [2];
if ( comm_dim ) {
sshift[0] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Even);
sshift[1] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Odd);
if ( sshift[0] == sshift[1] ) {
if (splice_dim) {
splicetime-=usecond();
auto tmp = GatherSimd(source,dimension,shift,0x3,compress,face_idx);
same_node = same_node && tmp;
splicetime+=usecond();
} else {
nosplicetime-=usecond();
auto tmp = Gather(source,dimension,shift,0x3,compress,face_idx);
same_node = same_node && tmp;
nosplicetime+=usecond();
}
} else {
if(splice_dim){
splicetime-=usecond();
// if checkerboard is unfavourable take two passes
// both with block stride loop iteration
auto tmp1 = GatherSimd(source,dimension,shift,0x1,compress,face_idx);
auto tmp2 = GatherSimd(source,dimension,shift,0x2,compress,face_idx);
same_node = same_node && tmp1 && tmp2;
splicetime+=usecond();
} else {
nosplicetime-=usecond();
auto tmp1 = Gather(source,dimension,shift,0x1,compress,face_idx);
auto tmp2 = Gather(source,dimension,shift,0x2,compress,face_idx);
same_node = same_node && tmp1 && tmp2;
nosplicetime+=usecond();
}
}
}
return same_node;
}
template<class compressor>
void HaloGather(const Lattice<vobj> &source,compressor &compress)
{
mpi3synctime_g-=usecond();
_grid->StencilBarrier();// Synch shared memory on a single nodes
mpi3synctime_g+=usecond();
// conformable(source._grid,_grid);
assert(source._grid==_grid);
halogtime-=usecond();
u_comm_offset=0;
// Gather all comms buffers
int face_idx=0;
for(int point = 0 ; point < _npoints; point++) {
compress.Point(point);
HaloGatherDir(source,compress,point,face_idx);
}
face_table_computed=1;
assert(u_comm_offset==_unified_buffer_size);
halogtime+=usecond();
}
/////////////////////////
// Implementation
/////////////////////////
void Prepare(void)
{
Decompressions.resize(0);
DecompressionsSHM.resize(0);
Mergers.resize(0);
MergersSHM.resize(0);
Packets.resize(0);
calls++;
}
void AddPacket(void *xmit,void * rcv, Integer to,Integer from,Integer bytes){
Packet p;
p.send_buf = xmit;
p.recv_buf = rcv;
p.to_rank = to;
p.from_rank= from;
p.bytes = bytes;
Packets.push_back(p);
}
void AddDecompress(cobj *k_p,cobj *m_p,Integer buffer_size,std::vector<Decompress> &dv) {
Decompress d;
d.kernel_p = k_p;
d.mpi_p = m_p;
d.buffer_size = buffer_size;
dv.push_back(d);
}
void AddMerge(cobj *merge_p,std::vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
Merge m;
m.type = type;
m.mpointer = merge_p;
m.vpointers= rpointers;
m.buffer_size = buffer_size;
mv.push_back(m);
}
template<class decompressor> void CommsMerge(decompressor decompress) {
CommsMerge(decompress,Mergers,Decompressions);
}
template<class decompressor> void CommsMergeSHM(decompressor decompress) {
mpi3synctime-=usecond();
_grid->StencilBarrier();// Synch shared memory on a single nodes
mpi3synctime+=usecond();
shmmergetime-=usecond();
CommsMerge(decompress,MergersSHM,DecompressionsSHM);
shmmergetime+=usecond();
}
template<class decompressor>
void CommsMerge(decompressor decompress,std::vector<Merge> &mm,std::vector<Decompress> &dd) {
for(int i=0;i<mm.size();i++){
mergetime-=usecond();
parallel_for(int o=0;o<mm[i].buffer_size/2;o++){
decompress.Exchange(mm[i].mpointer,
mm[i].vpointers[0],
mm[i].vpointers[1],
mm[i].type,o);
}
mergetime+=usecond();
}
for(int i=0;i<dd.size();i++){
decompresstime-=usecond();
parallel_for(int o=0;o<dd[i].buffer_size;o++){
decompress.Decompress(dd[i].kernel_p,dd[i].mpi_p,o);
}
decompresstime+=usecond();
}
}
////////////////////////////////////////
// Set up routines
////////////////////////////////////////
void PrecomputeByteOffsets(void){
for(int i=0;i<_entries.size();i++){
if( _entries[i]._is_local ) {
_entries[i]._byte_offset = _entries[i]._offset*sizeof(vobj);
} else {
_entries[i]._byte_offset = _entries[i]._offset*sizeof(cobj);
}
}
};
CartesianStencil(GridBase *grid,
int npoints,
int checkerboard,
const std::vector<int> &directions,
const std::vector<int> &distances)
: _permute_type(npoints),
_comm_buf_size(npoints),
comm_bytes_thr(npoints),
comm_enter_thr(npoints),
comm_leave_thr(npoints),
comm_time_thr(npoints)
{
face_table_computed=0;
_npoints = npoints;
_grid = grid;
_directions = directions;
_distances = distances;
_unified_buffer_size=0;
int osites = _grid->oSites();
_entries.resize(_npoints* osites);
for(int ii=0;ii<npoints;ii++){
int i = ii; // reverse direction to get SIMD comms done first
int point = i;
int dimension = directions[i];
int displacement = distances[i];
int shift = displacement;
int fd = _grid->_fdimensions[dimension];
int rd = _grid->_rdimensions[dimension];
_permute_type[point]=_grid->PermuteType(dimension);
_checkerboard = checkerboard;
//////////////////////////
// the permute type
//////////////////////////
int simd_layout = _grid->_simd_layout[dimension];
int comm_dim = _grid->_processors[dimension] >1 ;
int splice_dim = _grid->_simd_layout[dimension]>1 && (comm_dim);
int rotate_dim = _grid->_simd_layout[dimension]>2;
assert ( (rotate_dim && comm_dim) == false) ; // Do not think spread out is supported
int sshift[2];
//////////////////////////
// Underlying approach. For each local site build
// up a table containing the npoint "neighbours" and whether they
// live in lattice or a comms buffer.
//////////////////////////
if ( !comm_dim ) {
sshift[0] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Even);
sshift[1] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Odd);
if ( sshift[0] == sshift[1] ) {
Local(point,dimension,shift,0x3);
} else {
Local(point,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
Local(point,dimension,shift,0x2);// both with block stride loop iteration
}
} else {
// All permute extract done in comms phase prior to Stencil application
// So tables are the same whether comm_dim or splice_dim
sshift[0] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Even);
sshift[1] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Odd);
if ( sshift[0] == sshift[1] ) {
Comms(point,dimension,shift,0x3);
} else {
Comms(point,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
Comms(point,dimension,shift,0x2);// both with block stride loop iteration
}
}
}
/////////////////////////////////////////////////////////////////////////////////
// Try to allocate for receiving in a shared memory region, fall back to buffer
/////////////////////////////////////////////////////////////////////////////////
const int Nsimd = grid->Nsimd();
_grid->ShmBufferFreeAll();
u_simd_send_buf.resize(Nsimd);
u_simd_recv_buf.resize(Nsimd);
u_send_buf_p=(cobj *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(cobj));
u_recv_buf_p=(cobj *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(cobj));
for(int l=0;l<2;l++){
u_simd_recv_buf[l] = (cobj *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(cobj));
u_simd_send_buf[l] = (cobj *)_grid->ShmBufferMalloc(_unified_buffer_size*sizeof(cobj));
}
PrecomputeByteOffsets();
}
void Local (int point, int dimension,int shiftpm,int cbmask)
{
int fd = _grid->_fdimensions[dimension];
int rd = _grid->_rdimensions[dimension];
int ld = _grid->_ldimensions[dimension];
int gd = _grid->_gdimensions[dimension];
int ly = _grid->_simd_layout[dimension];
// Map to always positive shift modulo global full dimension.
int shift = (shiftpm+fd)%fd;
// the permute type
int permute_dim =_grid->PermuteDim(dimension);
for(int x=0;x<rd;x++){
int o = 0;
int bo = x * _grid->_ostride[dimension];
int cb= (cbmask==0x2)? Odd : Even;
int sshift = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,cb);
int sx = (x+sshift)%rd;
int wraparound=0;
if ( (shiftpm==-1) && (sx>x) ) {
wraparound = 1;
}
if ( (shiftpm== 1) && (sx<x) ) {
wraparound = 1;
}
int permute_slice=0;
if(permute_dim){
int wrap = sshift/rd;
int num = sshift%rd;
if ( x< rd-num ) permute_slice=wrap;
else permute_slice = (wrap+1)%ly;
}
CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
}
}
void Comms (int point,int dimension,int shiftpm,int cbmask)
{
GridBase *grid=_grid;
const int Nsimd = grid->Nsimd();
int fd = _grid->_fdimensions[dimension];
int ld = _grid->_ldimensions[dimension];
int rd = _grid->_rdimensions[dimension];
int pd = _grid->_processors[dimension];
int simd_layout = _grid->_simd_layout[dimension];
int comm_dim = _grid->_processors[dimension] >1 ;
assert(comm_dim==1);
int shift = (shiftpm + fd) %fd;
assert(shift>=0);
assert(shift<fd);
// done in reduced dims, so SIMD factored
int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
_comm_buf_size[point] = buffer_size; // Size of _one_ plane. Multiple planes may be gathered and
// send to one or more remote nodes.
int cb= (cbmask==0x2)? Odd : Even;
int sshift= _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,cb);
for(int x=0;x<rd;x++){
int permute_type=grid->PermuteType(dimension);
int sx = (x+sshift)%rd;
int offnode = 0;
if ( simd_layout > 1 ) {
for(int i=0;i<Nsimd;i++){
int inner_bit = (Nsimd>>(permute_type+1));
int ic= (i&inner_bit)? 1:0;
int my_coor = rd*ic + x;
int nbr_coor = my_coor+sshift;
int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
if ( nbr_proc ) {
offnode =1;
}
}
} else {
int comm_proc = ((x+sshift)/rd)%pd;
offnode = (comm_proc!= 0);
}
int wraparound=0;
if ( (shiftpm==-1) && (sx>x) && (grid->_processor_coor[dimension]==0) ) {
wraparound = 1;
}
if ( (shiftpm== 1) && (sx<x) && (grid->_processor_coor[dimension]==grid->_processors[dimension]-1) ) {
wraparound = 1;
}
if (!offnode) {
int permute_slice=0;
CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
} else {
int words = buffer_size;
if (cbmask != 0x3) words=words>>1;
int rank = grid->_processor;
int recv_from_rank;
int xmit_to_rank;
int unified_buffer_offset = _unified_buffer_size;
_unified_buffer_size += words;
ScatterPlane(point,dimension,x,cbmask,unified_buffer_offset,wraparound); // permute/extract/merge is done in comms phase
}
}
}
// Routine builds up integer table for each site in _offsets, _is_local, _permute
void CopyPlane(int point, int dimension,int lplane,int rplane,int cbmask,int permute,int wrap)
{
int rd = _grid->_rdimensions[dimension];
if ( !_grid->CheckerBoarded(dimension) ) {
int o = 0; // relative offset to base within plane
int ro = rplane*_grid->_ostride[dimension]; // base offset for start of plane
int lo = lplane*_grid->_ostride[dimension]; // offset in buffer
// Simple block stride gather of SIMD objects
for(int n=0;n<_grid->_slice_nblock[dimension];n++){
for(int b=0;b<_grid->_slice_block[dimension];b++){
int idx=point+(lo+o+b)*_npoints;
_entries[idx]._offset =ro+o+b;
_entries[idx]._permute=permute;
_entries[idx]._is_local=1;
_entries[idx]._around_the_world=wrap;
}
o +=_grid->_slice_stride[dimension];
}
} else {
int ro = rplane*_grid->_ostride[dimension]; // base offset for start of plane
int lo = lplane*_grid->_ostride[dimension]; // base offset for start of plane
int o = 0; // relative offset to base within plane
for(int n=0;n<_grid->_slice_nblock[dimension];n++){
for(int b=0;b<_grid->_slice_block[dimension];b++){
int ocb=1<<_grid->CheckerBoardFromOindex(o+b);
if ( ocb&cbmask ) {
int idx = point+(lo+o+b)*_npoints;
_entries[idx]._offset =ro+o+b;
_entries[idx]._is_local=1;
_entries[idx]._permute=permute;
_entries[idx]._around_the_world=wrap;
}
}
o +=_grid->_slice_stride[dimension];
}
}
}
// Routine builds up integer table for each site in _offsets, _is_local, _permute
void ScatterPlane (int point,int dimension,int plane,int cbmask,int offset, int wrap)
{
int rd = _grid->_rdimensions[dimension];
if ( !_grid->CheckerBoarded(dimension) ) {
int so = plane*_grid->_ostride[dimension]; // base offset for start of plane
int o = 0; // relative offset to base within plane
int bo = 0; // offset in buffer
// Simple block stride gather of SIMD objects
for(int n=0;n<_grid->_slice_nblock[dimension];n++){
for(int b=0;b<_grid->_slice_block[dimension];b++){
int idx=point+(so+o+b)*_npoints;
_entries[idx]._offset =offset+(bo++);
_entries[idx]._is_local=0;
_entries[idx]._permute=0;
_entries[idx]._around_the_world=wrap;
}
o +=_grid->_slice_stride[dimension];
}
} else {
int so = plane*_grid->_ostride[dimension]; // base offset for start of plane
int o = 0; // relative offset to base within plane
int bo = 0; // offset in buffer
for(int n=0;n<_grid->_slice_nblock[dimension];n++){
for(int b=0;b<_grid->_slice_block[dimension];b++){
int ocb=1<<_grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
if ( ocb & cbmask ) {
int idx = point+(so+o+b)*_npoints;
_entries[idx]._offset =offset+(bo++);
_entries[idx]._is_local=0;
_entries[idx]._permute =0;
_entries[idx]._around_the_world=wrap;
}
}
o +=_grid->_slice_stride[dimension];
}
}
}
template<class compressor>
int Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress,int &face_idx)
{
typedef typename cobj::vector_type vector_type;
typedef typename cobj::scalar_type scalar_type;
GridBase *grid=_grid;
assert(rhs._grid==_grid);
// conformable(_grid,rhs._grid);
int fd = _grid->_fdimensions[dimension];
int rd = _grid->_rdimensions[dimension];
int pd = _grid->_processors[dimension];
int simd_layout = _grid->_simd_layout[dimension];
int comm_dim = _grid->_processors[dimension] >1 ;
assert(simd_layout==1);
assert(comm_dim==1);
assert(shift>=0);
assert(shift<fd);
int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
int cb= (cbmask==0x2)? Odd : Even;
int sshift= _grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
int shm_receive_only = 1;
for(int x=0;x<rd;x++){
int sx = (x+sshift)%rd;
int comm_proc = ((x+sshift)/rd)%pd;
if (comm_proc) {
int words = buffer_size;
if (cbmask != 0x3) words=words>>1;
int bytes = words * compress.CommDatumSize();
int so = sx*rhs._grid->_ostride[dimension]; // base offset for start of plane
if ( !face_table_computed ) {
face_table.resize(face_idx+1);
Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table[face_idx]);
}
int rank = _grid->_processor;
int recv_from_rank;
int xmit_to_rank;
_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
assert (xmit_to_rank != _grid->ThisRank());
assert (recv_from_rank != _grid->ThisRank());
/////////////////////////////////////////////////////////
// try the direct copy if possible
/////////////////////////////////////////////////////////
cobj *send_buf;
cobj *recv_buf;
if ( compress.DecompressionStep() ) {
recv_buf=u_simd_recv_buf[0];
} else {
recv_buf=u_recv_buf_p;
}
send_buf = (cobj *)_grid->ShmBufferTranslate(xmit_to_rank,recv_buf);
if ( send_buf==NULL ) {
send_buf = u_send_buf_p;
}
// Find out if we get the direct copy.
void *success = (void *) _grid->ShmBufferTranslate(recv_from_rank,u_send_buf_p);
if (success==NULL) {
// we found a packet that comes from MPI and contributes to this leg of stencil
shm_receive_only = 0;
}
gathertime-=usecond();
assert(send_buf!=NULL);
Gather_plane_simple_table(face_table[face_idx],rhs,send_buf,compress,u_comm_offset,so); face_idx++;
gathertime+=usecond();
if ( compress.DecompressionStep() ) {
if ( shm_receive_only ) { // Early decompress before MPI is finished is possible
AddDecompress(&u_recv_buf_p[u_comm_offset],
&recv_buf[u_comm_offset],
words,DecompressionsSHM);
} else { // Decompress after MPI is finished
AddDecompress(&u_recv_buf_p[u_comm_offset],
&recv_buf[u_comm_offset],
words,Decompressions);
}
AddPacket((void *)&send_buf[u_comm_offset],
(void *)&recv_buf[u_comm_offset],
xmit_to_rank,
recv_from_rank,
bytes);
} else {
AddPacket((void *)&send_buf[u_comm_offset],
(void *)&u_recv_buf_p[u_comm_offset],
xmit_to_rank,
recv_from_rank,
bytes);
}
u_comm_offset+=words;
}
}
return shm_receive_only;
}
template<class compressor>
int GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress,int & face_idx)
{
const int Nsimd = _grid->Nsimd();
const int maxl =2;// max layout in a direction
int fd = _grid->_fdimensions[dimension];
int rd = _grid->_rdimensions[dimension];
int ld = _grid->_ldimensions[dimension];
int pd = _grid->_processors[dimension];
int simd_layout = _grid->_simd_layout[dimension];
int comm_dim = _grid->_processors[dimension] >1 ;
assert(comm_dim==1);
// This will not work with a rotate dim
assert(simd_layout==maxl);
assert(shift>=0);
assert(shift<fd);
int permute_type=_grid->PermuteType(dimension);
// std::cout << "SimdNew permute type "<<permute_type<<std::endl;
///////////////////////////////////////////////
// Simd direction uses an extract/merge pair
///////////////////////////////////////////////
int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
int words = sizeof(cobj)/sizeof(vector_type);
assert(cbmask==0x3); // Fixme think there is a latent bug if not true
int reduced_buffer_size = buffer_size;
if (cbmask != 0x3) reduced_buffer_size=buffer_size>>1;
int datum_bytes = compress.CommDatumSize();
int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
std::vector<cobj *> rpointers(maxl);
std::vector<cobj *> spointers(maxl);
///////////////////////////////////////////
// Work out what to send where
///////////////////////////////////////////
int cb = (cbmask==0x2)? Odd : Even;
int sshift= _grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
// loop over outer coord planes orthog to dim
int shm_receive_only = 1;
for(int x=0;x<rd;x++){
int any_offnode = ( ((x+sshift)%fd) >= rd );
if ( any_offnode ) {
for(int i=0;i<maxl;i++){
spointers[i] = (cobj *) &u_simd_send_buf[i][u_comm_offset];
}
int sx = (x+sshift)%rd;
if ( !face_table_computed ) {
face_table.resize(face_idx+1);
Gather_plane_table_compute ((GridBase *)_grid,dimension,sx,cbmask,u_comm_offset,face_table[face_idx]);
}
gathermtime-=usecond();
Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type); face_idx++;
gathermtime+=usecond();
//spointers[0] -- low
//spointers[1] -- high
for(int i=0;i<maxl;i++){
int my_coor = rd*i + x; // self explanatory
int nbr_coor = my_coor+sshift; // self explanatory
int nbr_proc = ((nbr_coor)/ld) % pd;// relative shift in processors
int nbr_lcoor= (nbr_coor%ld); // local plane coor on neighbour node
int nbr_ic = (nbr_lcoor)/rd; // inner coord of peer simd lane "i"
int nbr_ox = (nbr_lcoor%rd); // outer coord of peer "x"
int nbr_plane = nbr_ic;
assert (sx == nbr_ox);
auto rp = &u_simd_recv_buf[i ][u_comm_offset];
auto sp = &u_simd_send_buf[nbr_plane][u_comm_offset];
if(nbr_proc){
int recv_from_rank;
int xmit_to_rank;
_grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
// shm == receive pointer if offnode
// shm == Translate[send pointer] if on node -- my view of his send pointer
cobj *shm = (cobj *) _grid->ShmBufferTranslate(recv_from_rank,sp);
if (shm==NULL) {
shm = rp;
// we found a packet that comes from MPI and contributes to this shift.
// same_node is only used in the WilsonStencil, and gets set for this point in the stencil.
// Kernel will add the exterior_terms except if same_node.
shm_receive_only = 0;
// leg of stencil
}
// if Direct, StencilSendToRecvFrom will suppress copy to a peer on node
// assuming above pointer flip
rpointers[i] = shm;
AddPacket((void *)sp,(void *)rp,xmit_to_rank,recv_from_rank,bytes);
} else {
rpointers[i] = sp;
}
}
if ( shm_receive_only ) {
AddMerge(&u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,MergersSHM);
} else {
AddMerge(&u_recv_buf_p[u_comm_offset],rpointers,reduced_buffer_size,permute_type,Mergers);
}
u_comm_offset +=buffer_size;
}
}
return shm_receive_only;
}
void ZeroCounters(void) {
gathertime = 0.;
commtime = 0.;
mpi3synctime=0.;
mpi3synctime_g=0.;
shmmergetime=0.;
for(int i=0;i<_npoints;i++){
comm_time_thr[i]=0;
comm_bytes_thr[i]=0;
comm_enter_thr[i]=0;
comm_leave_thr[i]=0;
}
halogtime = 0.;
mergetime = 0.;
decompresstime = 0.;
gathermtime = 0.;
splicetime = 0.;
nosplicetime = 0.;
comms_bytes = 0.;
calls = 0.;
};
void Report(void) {
#define AVERAGE(A) _grid->GlobalSum(A);A/=NP;
#define PRINTIT(A) AVERAGE(A); std::cout << GridLogMessage << " Stencil " << #A << " "<< A/calls<<std::endl;
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
double t = 0;
// if comm_time_thr is set they were all done in parallel so take the max
// but add up the bytes
int threaded = 0 ;
for (int i = 0; i < 8; ++i) {
if ( comm_time_thr[i]>0.0 ) {
threaded = 1;
comms_bytes += comm_bytes_thr[i];
if (t < comm_time_thr[i]) t = comm_time_thr[i];
}
}
if (threaded) commtime += t;
_grid->GlobalSum(commtime); commtime/=NP;
if ( calls > 0. ) {
std::cout << GridLogMessage << " Stencil calls "<<calls<<std::endl;
PRINTIT(halogtime);
PRINTIT(gathertime);
PRINTIT(gathermtime);
PRINTIT(mergetime);
PRINTIT(decompresstime);
if(comms_bytes>1.0){
PRINTIT(comms_bytes);
PRINTIT(commtime);
std::cout << GridLogMessage << " Stencil " << comms_bytes/commtime/1000. << " GB/s per rank"<<std::endl;
std::cout << GridLogMessage << " Stencil " << comms_bytes/commtime/1000.*NP/NN << " GB/s per node"<<std::endl;
}
PRINTIT(mpi3synctime);
PRINTIT(mpi3synctime_g);
PRINTIT(shmmergetime);
PRINTIT(splicetime);
PRINTIT(nosplicetime);
}
#undef PRINTIT
#undef AVERAGE
};
};
}
#endif
View Git Blame Copy Permalink