portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-19 08:17:05 +01:00
Code Releases Activity

MultiRHS work

Browse Source
This commit is contained in:
Peter Boyle
2023-11-28 07:43:37 -05:00
parent 59abaeb5cd
commit 0a3682ad0b
7 changed files with 136 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
Grid/lattice/Lattice_transfer.h
View File

@ -745,8 +745,6 @@ 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);
GridBase *Fg = From.Grid();
GridBase *Tg = To.Grid();
assert(!Fg->_isCheckerBoarded);
@ -763,13 +761,14 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
// the above should guarantee that the operations are local
#if 1
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);
RealD t_cpu=-usecond();
#if 0
thread_for(idx, nsite, {
Coordinate from_coor, to_coor;
size_t rem = idx;
@ -792,15 +791,44 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
int* table_d = (int*)acceleratorAllocDevice(tbytes);
acceleratorCopyToDevice(table,table_d,tbytes);
#else
int* table_d = (int*)acceleratorAllocDevice(tbytes);
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;
accelerator_for(idx, nsite, 1, {
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 = 0; for(int d=0;d<nd;d++) foidx+=f_ostride[d]*(from_coor[d]%f_rdimensions[d]);
int fiidx = 0; for(int d=0;d<nd;d++) fiidx+=f_istride[d]*(from_coor[d]/f_rdimensions[d]);
int toidx = 0; for(int d=0;d<nd;d++) toidx+=t_ostride[d]*(to_coor[d]%t_rdimensions[d]);
int tiidx = 0; for(int d=0;d<nd;d++) tiidx+=t_istride[d]*(to_coor[d]/t_rdimensions[d]);
int* tt = table_d + 4*idx;
tt[0] = foidx;
tt[1] = fiidx;
tt[2] = toidx;
tt[3] = tiidx;
});
#endif
t_cpu+=usecond();
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);
RealD t_acc=-usecond();
const int words=sizeof(vobj)/sizeof(vector_type);
accelerator_for(idx,nsite,words,{
int* tt = table_d + 4*idx;
int from_oidx = *tt++;
int from_lane = *tt++;
@ -811,12 +839,20 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
vector_type* to = (vector_type *)&to_v[to_oidx];
scalar_type stmp;
#ifdef GRID_SIMT
int w = acceleratorSIMTlane(words);
stmp = getlane(from[w], from_lane);
putlane(to[w], stmp, to_lane);
#else
for(int w=0;w<words;w++){
stmp = getlane(from[w], from_lane);
putlane(to[w], stmp, to_lane);
}
#endif
});
t_acc+=usecond();
std::cout << " localCopyRegion cpu " <<t_cpu/1000<<" ms"<<std::endl;
std::cout << " localCopyRegion acc " <<t_acc/1000<<" ms"<<std::endl;
acceleratorFreeDevice(table_d);
free(table);

15
Grid/lattice/PaddedCell.h
View File

@ -403,18 +403,8 @@ public:
double t = usecond();
padded = in;
tins += usecond() - t;
// return in; ?
} else {
//////////////////////////////////////////////
// Replace sequence with
// ---------------------
// (i) Gather high face(s); start comms
// (ii) Gather low face(s); start comms
// (iii) Copy middle bit with localCopyRegion
// (iv) Complete high face(s), insert slice(s)
// (iv) Complete low face(s), insert slice(s)
//////////////////////////////////////////////
Face_exchange(in,padded,dim,depth);
}
return padded;
@ -482,6 +472,7 @@ public:
// Gather all surface terms up to depth "d"
////////////////////////////////////////////////////////////////////////////
RealD t;
RealD t_tot=-usecond();
int plane=0;
for ( int d=0;d < depth ; d ++ ) {
int tag = d*1024 + dimension*2+0;
@ -549,6 +540,7 @@ public:
}
t_scatter+= usecond() - t;
// DumpSliceNorm(std::string("Face_exchange to scatter 1st "),to,dimension);
t_tot+=usecond();
//DumpSliceNorm(std::string("Face_exchange to done"),to,dimension);
std::cout << GridLogPerformance << "PaddedCell::Expand new timings: gather :" << t_gather/1000 << "ms"<<std::endl;
@ -557,6 +549,7 @@ public:
// std::cout << GridLogPerformance << "PaddedCell::Expand new timings: scatter:" << 2.0*bytes/t_scatter<< "MB/s"<<std::endl;
std::cout << GridLogPerformance << "PaddedCell::Expand new timings: copy :" << t_copy/1000 << "ms"<<std::endl;
std::cout << GridLogPerformance << "PaddedCell::Expand new timings: comms :" << t_comms/1000 << "ms"<<std::endl;
std::cout << GridLogPerformance << "PaddedCell::Expand new timings: total :" << t_tot/1000 << "ms"<<std::endl;
// std::cout << GridLogPerformance << "PaddedCell::Expand new timings: comms :" << (RealD)4.0*bytes/t_comms << "MB/s"<<std::endl;
}