portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-10-31 12:04:33 +00:00
Code Releases Activity

Comms optimisation

Browse Source
This commit is contained in:
Peter Boyle
2015-11-06 05:22:18 -06:00
parent f85b9ddd97
commit 880ff88362
1 changed files with 356 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

392
lib/Stencil.h
View File

@@ -48,10 +48,14 @@ namespace Grid {
int _around_the_world;
};
template<class vobj,class cobj, class compressor>
class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal fill in.
public:
typedef uint32_t StencilInteger;
typedef typename cobj::vector_type vector_type;
typedef typename cobj::scalar_type scalar_type;
typedef typename cobj::scalar_object scalar_object;
int _checkerboard;
int _npoints; // Move to template param?
@@ -66,33 +70,334 @@ namespace Grid {
// npoints x Osites() of these
std::vector<std::vector<StencilEntry> > _entries;
// Comms buffers
std::vector<std::vector<scalar_object> > send_buf_extract;
std::vector<std::vector<scalar_object> > recv_buf_extract;
std::vector<scalar_object *> pointers;
std::vector<scalar_object *> rpointers;
Vector<cobj> send_buf;
inline StencilEntry * GetEntry(int &ptype,int point,int osite) { ptype = _permute_type[point]; return & _entries[point][osite]; }
int _unified_buffer_size;
int _request_count;
double buftime;
double gathertime;
double commtime;
double commstime;
double halotime;
double scattertime;
double mergetime;
double gathermtime;
double splicetime;
double nosplicetime;
CartesianStencil(GridBase *grid,
int npoints,
int checkerboard,
const std::vector<int> &directions,
const std::vector<int> &distances);
CartesianStencil(GridBase *grid,
int npoints,
int checkerboard,
const std::vector<int> &directions,
const std::vector<int> &distances)
: _entries(npoints), _permute_type(npoints), _comm_buf_size(npoints)
{
gathertime=0;
commtime=0;
commstime=0;
halotime=0;
scattertime=0;
mergetime=0;
gathermtime=0;
buftime=0;
splicetime=0;
nosplicetime=0;
_npoints = npoints;
_grid = grid;
_directions = directions;
_distances = distances;
_unified_buffer_size=0;
_request_count =0;
int osites = _grid->oSites();
for(int i=0;i<npoints;i++){
int point = i;
_entries[i].resize( osites);
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 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);
// std::cout<<"Comms 0x3"<<std::endl;
} else {
Comms(point,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
Comms(point,dimension,shift,0x2);// both with block stride loop iteration
// std::cout<<"Comms 0x1 ; 0x2"<<std::endl;
}
}
// for(int ss=0;ss<osites;ss++){
// std::cout << "point["<<i<<"] "<<ss<<"-> o"<<_entries[i][ss]._offset<<"; l"<<
// _entries[i][ss]._is_local<<"; p"<<_entries[i][ss]._permute<<std::endl;
// }
}
}
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];
// 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 = 1-wrap;
}
CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
}
}
void Comms (int point,int dimension,int shiftpm,int cbmask)
{
GridBase *grid=_grid;
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(simd_layout==1); // Why?
assert(comm_dim==1);
int shift = (shiftpm + fd) %fd;
assert(shift>=0);
assert(shift<fd);
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 sx = (x+sshift)%rd;
int comm_proc = ((x+sshift)/rd)%pd;
int offnode = (comm_proc!= 0);
// std::cout << "Stencil shift "<<shift<<" sshift "<<sshift<<" fd "<<fd<<" rd " <<rd<<" offnode "<<offnode<<" sx "<<sx<<std::endl;
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;
// GatherPlaneSimple (point,dimension,sx,cbmask);
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++){
_entries[point][lo+o+b]._offset =ro+o+b;
_entries[point][lo+o+b]._is_local=1;
_entries[point][lo+o+b]._permute=permute;
_entries[point][lo+o+b]._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 ) {
_entries[point][lo+o+b]._offset =ro+o+b;
_entries[point][lo+o+b]._is_local=1;
_entries[point][lo+o+b]._permute=permute;
_entries[point][lo+o+b]._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++){
_entries[point][so+o+b]._offset =offset+(bo++);
_entries[point][so+o+b]._is_local=0;
_entries[point][so+o+b]._permute=0;
_entries[point][so+o+b]._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 ) {
_entries[point][so+o+b]._offset =offset+(bo++);
_entries[point][so+o+b]._is_local=0;
_entries[point][so+o+b]._permute =0;
_entries[point][so+o+b]._around_the_world=wrap;
}
}
o +=_grid->_slice_stride[dimension];
}
}
}
// CartesianStencil(GridBase *grid,
// int npoints,
// int checkerboard,
// const std::vector<int> &directions,
// const std::vector<int> &distances);
// Add to tables for various cases; is this mistaken. only local if 1 proc in dim
// Can this be avoided with simpler coding of comms?
void Local (int point, int dimension,int shift,int cbmask);
void Comms (int point, int dimension,int shift,int cbmask);
void CopyPlane(int point, int dimension,int lplane,int rplane,int cbmask,int permute,int wrap);
void ScatterPlane (int point,int dimension,int plane,int cbmask,int offset,int wrap);
// void Local (int point, int dimension,int shift,int cbmask);
// void Comms (int point, int dimension,int shift,int cbmask);
// void CopyPlane(int point, int dimension,int lplane,int rplane,int cbmask,int permute,int wrap);
// void ScatterPlane (int point,int dimension,int plane,int cbmask,int offset,int wrap);
// Could allow a functional munging of the halo to another type during the comms.
// this could implement the 16bit/32bit/64bit compression.
template<class vobj,class cobj, class compressor> void
HaloExchange(const Lattice<vobj> &source,std::vector<cobj,alignedAllocator<cobj> > &u_comm_buf,compressor &compress)
void HaloExchange(const Lattice<vobj> &source,std::vector<cobj,alignedAllocator<cobj> > &u_comm_buf,compressor &compress)
{
// conformable(source._grid,_grid);
assert(source._grid==_grid);
halotime-=usecond();
if (u_comm_buf.size() != _unified_buffer_size ) u_comm_buf.resize(_unified_buffer_size);
int u_comm_offset=0;
@@ -124,27 +429,35 @@ namespace Grid {
if ( comm_dim ) {
sshift[0] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Even);
sshift[1] = _grid->CheckerBoardShiftForCB(_checkerboard,dimension,shift,Odd);
// std::cout << "dim "<<dimension<<"cb "<<_checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
if ( sshift[0] == sshift[1] ) {
if (splice_dim) {
splicetime-=usecond();
GatherStartCommsSimd(source,dimension,shift,0x3,u_comm_buf,u_comm_offset,compress);
splicetime+=usecond();
} else {
nosplicetime-=usecond();
GatherStartComms(source,dimension,shift,0x3,u_comm_buf,u_comm_offset,compress);
nosplicetime+=usecond();
}
} else {
std::cout << "dim "<<dimension<<"cb "<<_checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
if(splice_dim){
splicetime-=usecond();
GatherStartCommsSimd(source,dimension,shift,0x1,u_comm_buf,u_comm_offset,compress);// if checkerboard is unfavourable take two passes
GatherStartCommsSimd(source,dimension,shift,0x2,u_comm_buf,u_comm_offset,compress);// both with block stride loop iteration
splicetime+=usecond();
} else {
nosplicetime-=usecond();
GatherStartComms(source,dimension,shift,0x1,u_comm_buf,u_comm_offset,compress);
GatherStartComms(source,dimension,shift,0x2,u_comm_buf,u_comm_offset,compress);
nosplicetime+=usecond();
}
}
}
}
halotime+=usecond();
}
template<class vobj,class cobj, class compressor>
void GatherStartComms(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,
std::vector<cobj,alignedAllocator<cobj> > &u_comm_buf,
int &u_comm_offset,compressor & compress)
@@ -168,8 +481,7 @@ namespace Grid {
int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
std::vector<cobj,alignedAllocator<cobj> > send_buf(buffer_size); // hmm...
std::vector<cobj,alignedAllocator<cobj> > recv_buf(buffer_size);
if(send_buf.size()<buffer_size) send_buf.resize(buffer_size);
int cb= (cbmask==0x2)? Odd : Even;
int sshift= _grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
@@ -186,7 +498,9 @@ namespace Grid {
int bytes = words * sizeof(cobj);
gathertime-=usecond();
Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask,compress);
gathertime+=usecond();
int rank = _grid->_processor;
int recv_from_rank;
@@ -196,32 +510,27 @@ namespace Grid {
assert (recv_from_rank != _grid->ThisRank());
// FIXME Implement asynchronous send & also avoid buffer copy
commtime-=usecond();
_grid->SendToRecvFrom((void *)&send_buf[0],
xmit_to_rank,
(void *)&recv_buf[0],
(void *)&u_comm_buf[u_comm_offset],
recv_from_rank,
bytes);
commtime+=usecond();
for(int i=0;i<words;i++){
u_comm_buf[u_comm_offset+i]=recv_buf[i];
// std::cout << " Halo["<<i<<"] snd "<<send_buf[i]<< " rcv "<<recv_buf[i]<<" mask 0x"<<cbmask<<std::endl;
}
u_comm_offset+=words;
}
}
}
template<class vobj,class cobj, class compressor>
void GatherStartCommsSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,
std::vector<cobj,alignedAllocator<cobj> > &u_comm_buf,
int &u_comm_offset,compressor &compress)
{
buftime-=usecond();
const int Nsimd = _grid->Nsimd();
typedef typename cobj::vector_type vector_type;
typedef typename cobj::scalar_type scalar_type;
typedef typename cobj::scalar_object scalar_object;
int fd = _grid->_fdimensions[dimension];
int rd = _grid->_rdimensions[dimension];
@@ -244,17 +553,22 @@ namespace Grid {
int words = sizeof(cobj)/sizeof(vector_type);
assert(cbmask==0x3); // Fixme think there is a latent bug if not true
/*
* possibly slow to allocate
* Doesn't matter in this test, but may want to preallocate in the
* dirac operators
*/
std::vector<std::vector<scalar_object> > send_buf_extract(Nsimd,std::vector<scalar_object>(buffer_size) );
std::vector<std::vector<scalar_object> > recv_buf_extract(Nsimd,std::vector<scalar_object>(buffer_size) );
int bytes = buffer_size*sizeof(scalar_object);
std::vector<scalar_object *> pointers(Nsimd); //
std::vector<scalar_object *> rpointers(Nsimd); // received pointers
// Should grow to max size and then cost very little thereafter
send_buf_extract.resize(Nsimd);
recv_buf_extract.resize(Nsimd);
for(int l=0;l<Nsimd;l++){
if( send_buf_extract[l].size() < buffer_size) {
send_buf_extract[l].resize(buffer_size);
recv_buf_extract[l].resize(buffer_size);
}
}
pointers.resize(Nsimd);
rpointers.resize(Nsimd);
int bytes = buffer_size*sizeof(scalar_object);
buftime+=usecond();
///////////////////////////////////////////
// Work out what to send where
@@ -275,7 +589,9 @@ namespace Grid {
}
int sx = (x+sshift)%rd;
gathermtime-=usecond();
Gather_plane_extract<cobj>(rhs,pointers,dimension,sx,cbmask,compress);
gathermtime+=usecond();
for(int i=0;i<Nsimd;i++){
@@ -302,11 +618,13 @@ namespace Grid {
_grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
commstime-=usecond();
_grid->SendToRecvFrom((void *)&send_buf_extract[nbr_lane][0],
xmit_to_rank,
(void *)&recv_buf_extract[i][0],
recv_from_rank,
bytes);
commstime+=usecond();
rpointers[i] = &recv_buf_extract[i][0];
@@ -316,11 +634,13 @@ namespace Grid {
}
// Here we don't want to scatter, just place into a buffer.
mergetime-=usecond();
PARALLEL_FOR_LOOP
for(int i=0;i<buffer_size;i++){
assert(u_comm_offset+i<_unified_buffer_size);
// assert(u_comm_offset+i<_unified_buffer_size);
merge(u_comm_buf[u_comm_offset+i],rpointers,i);
}
mergetime+=usecond();
u_comm_offset+=buffer_size;
}
}