mirror of
https://github.com/paboyle/Grid.git
synced 2025-04-04 19:25:56 +01:00
Fast localCopyRegion, blockProjectFast
This commit is contained in:
parent
42ae36bc28
commit
addc638856
@ -744,7 +744,11 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
|
||||
typedef typename vobj::scalar_type scalar_type;
|
||||
typedef typename vobj::vector_type vector_type;
|
||||
|
||||
static const int words=sizeof(vobj)/sizeof(vector_type);
|
||||
const int words=sizeof(vobj)/sizeof(vector_type);
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////
|
||||
// checks should guarantee that the operations are local
|
||||
//////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
GridBase *Fg = From.Grid();
|
||||
GridBase *Tg = To.Grid();
|
||||
@ -759,52 +763,39 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
|
||||
for(int d=0;d<nd;d++){
|
||||
assert(Fg->_processors[d] == Tg->_processors[d]);
|
||||
}
|
||||
// the above should guarantee that the operations are local
|
||||
|
||||
#if 1
|
||||
|
||||
///////////////////////////////////////////////////////////
|
||||
// do the index calc on the GPU
|
||||
///////////////////////////////////////////////////////////
|
||||
Coordinate f_ostride = Fg->_ostride;
|
||||
Coordinate f_istride = Fg->_istride;
|
||||
Coordinate f_rdimensions = Fg->_rdimensions;
|
||||
Coordinate t_ostride = Tg->_ostride;
|
||||
Coordinate t_istride = Tg->_istride;
|
||||
Coordinate t_rdimensions = Tg->_rdimensions;
|
||||
|
||||
size_t nsite = 1;
|
||||
for(int i=0;i<nd;i++) nsite *= RegionSize[i];
|
||||
|
||||
size_t tbytes = 4*nsite*sizeof(int);
|
||||
int *table = (int*)malloc(tbytes);
|
||||
|
||||
thread_for(idx, nsite, {
|
||||
Coordinate from_coor, to_coor;
|
||||
size_t rem = idx;
|
||||
for(int i=0;i<nd;i++){
|
||||
size_t base_i = rem % RegionSize[i]; rem /= RegionSize[i];
|
||||
from_coor[i] = base_i + FromLowerLeft[i];
|
||||
to_coor[i] = base_i + ToLowerLeft[i];
|
||||
}
|
||||
|
||||
int foidx = Fg->oIndex(from_coor);
|
||||
int fiidx = Fg->iIndex(from_coor);
|
||||
int toidx = Tg->oIndex(to_coor);
|
||||
int tiidx = Tg->iIndex(to_coor);
|
||||
int* tt = table + 4*idx;
|
||||
tt[0] = foidx;
|
||||
tt[1] = fiidx;
|
||||
tt[2] = toidx;
|
||||
tt[3] = tiidx;
|
||||
});
|
||||
|
||||
int* table_d = (int*)acceleratorAllocDevice(tbytes);
|
||||
acceleratorCopyToDevice(table,table_d,tbytes);
|
||||
|
||||
typedef typename vobj::vector_type vector_type;
|
||||
typedef typename vobj::scalar_type scalar_type;
|
||||
|
||||
autoView(from_v,From,AcceleratorRead);
|
||||
autoView(to_v,To,AcceleratorWrite);
|
||||
|
||||
|
||||
|
||||
accelerator_for(idx,nsite,1,{
|
||||
static const int words=sizeof(vobj)/sizeof(vector_type);
|
||||
int* tt = table_d + 4*idx;
|
||||
int from_oidx = *tt++;
|
||||
int from_lane = *tt++;
|
||||
int to_oidx = *tt++;
|
||||
int to_lane = *tt;
|
||||
|
||||
Coordinate from_coor, to_coor, base;
|
||||
Lexicographic::CoorFromIndex(base,idx,RegionSize);
|
||||
for(int i=0;i<nd;i++){
|
||||
from_coor[i] = base[i] + FromLowerLeft[i];
|
||||
to_coor[i] = base[i] + ToLowerLeft[i];
|
||||
}
|
||||
int from_oidx = 0; for(int d=0;d<nd;d++) from_oidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
|
||||
int from_lane = 0; for(int d=0;d<nd;d++) from_lane+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
|
||||
int to_oidx = 0; for(int d=0;d<nd;d++) to_oidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
|
||||
int to_lane = 0; for(int d=0;d<nd;d++) to_lane+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
|
||||
|
||||
const vector_type* from = (const vector_type *)&from_v[from_oidx];
|
||||
vector_type* to = (vector_type *)&to_v[to_oidx];
|
||||
@ -814,54 +805,7 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
|
||||
stmp = getlane(from[w], from_lane);
|
||||
putlane(to[w], stmp, to_lane);
|
||||
}
|
||||
});
|
||||
|
||||
acceleratorFreeDevice(table_d);
|
||||
free(table);
|
||||
|
||||
|
||||
#else
|
||||
Coordinate ldf = Fg->_ldimensions;
|
||||
Coordinate rdf = Fg->_rdimensions;
|
||||
Coordinate isf = Fg->_istride;
|
||||
Coordinate osf = Fg->_ostride;
|
||||
Coordinate rdt = Tg->_rdimensions;
|
||||
Coordinate ist = Tg->_istride;
|
||||
Coordinate ost = Tg->_ostride;
|
||||
|
||||
autoView( t_v , To, CpuWrite);
|
||||
autoView( f_v , From, CpuRead);
|
||||
thread_for(idx,Fg->lSites(),{
|
||||
sobj s;
|
||||
Coordinate Fcoor(nd);
|
||||
Coordinate Tcoor(nd);
|
||||
Lexicographic::CoorFromIndex(Fcoor,idx,ldf);
|
||||
int in_region=1;
|
||||
for(int d=0;d<nd;d++){
|
||||
if ( (Fcoor[d] < FromLowerLeft[d]) || (Fcoor[d]>=FromLowerLeft[d]+RegionSize[d]) ){
|
||||
in_region=0;
|
||||
}
|
||||
Tcoor[d] = ToLowerLeft[d]+ Fcoor[d]-FromLowerLeft[d];
|
||||
}
|
||||
if (in_region) {
|
||||
#if 0
|
||||
Integer idx_f = 0; for(int d=0;d<nd;d++) idx_f+=isf[d]*(Fcoor[d]/rdf[d]); // inner index from
|
||||
Integer idx_t = 0; for(int d=0;d<nd;d++) idx_t+=ist[d]*(Tcoor[d]/rdt[d]); // inner index to
|
||||
Integer odx_f = 0; for(int d=0;d<nd;d++) odx_f+=osf[d]*(Fcoor[d]%rdf[d]); // outer index from
|
||||
Integer odx_t = 0; for(int d=0;d<nd;d++) odx_t+=ost[d]*(Tcoor[d]%rdt[d]); // outer index to
|
||||
scalar_type * fp = (scalar_type *)&f_v[odx_f];
|
||||
scalar_type * tp = (scalar_type *)&t_v[odx_t];
|
||||
for(int w=0;w<words;w++){
|
||||
tp[w].putlane(fp[w].getlane(idx_f),idx_t);
|
||||
}
|
||||
#else
|
||||
peekLocalSite(s,f_v,Fcoor);
|
||||
pokeLocalSite(s,t_v,Tcoor);
|
||||
#endif
|
||||
}
|
||||
});
|
||||
|
||||
#endif
|
||||
}
|
||||
|
||||
|
||||
@ -1802,5 +1746,35 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj> & split)
|
||||
}
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
// Faster but less accurate blockProject
|
||||
//////////////////////////////////////////////////////
|
||||
template<class vobj,class CComplex,int nbasis,class VLattice>
|
||||
inline void blockProjectFast(Lattice<iVector<CComplex,nbasis > > &coarseData,
|
||||
const Lattice<vobj> &fineData,
|
||||
const VLattice &Basis)
|
||||
{
|
||||
GridBase * fine = fineData.Grid();
|
||||
GridBase * coarse= coarseData.Grid();
|
||||
|
||||
Lattice<iScalar<CComplex> > ip(coarse);
|
||||
|
||||
autoView( coarseData_ , coarseData, AcceleratorWrite);
|
||||
autoView( ip_ , ip, AcceleratorWrite);
|
||||
RealD t_IP=0;
|
||||
RealD t_co=0;
|
||||
for(int v=0;v<nbasis;v++) {
|
||||
t_IP-=usecond();
|
||||
blockInnerProductD(ip,Basis[v],fineData);
|
||||
t_IP+=usecond();
|
||||
t_co-=usecond();
|
||||
accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
|
||||
convertType(coarseData_[sc](v),ip_[sc]);
|
||||
});
|
||||
t_co+=usecond();
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
|
Loading…
x
Reference in New Issue
Block a user