portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

Fixed the stencil sector and Wilson now agrees between stencil based implementation

Browse Source 
and the cshift based implementation. Managed to reduce the volume of code in this
sector a little, but consolidation would be good, perhaps taking common
logic out into simple helper functions
This commit is contained in:
Peter Boyle 2015-04-29 06:23:56 +01:00
parent 25d523c0f4
commit c72db6c6f6
4 changed files with 17 additions and 51 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
lib/Grid_stencil.h
View File

@ -39,13 +39,6 @@
namespace Grid {
struct CommsRequest {
int words;
int unified_buffer_offset;
int tag;
int to_rank;
int from_rank;
} ;
class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal fill in.
@ -69,7 +62,6 @@ namespace Grid {
int _unified_buffer_size;
int _request_count;
std::vector<CommsRequest> CommsRequests;
CartesianStencil(GridBase *grid,
int npoints,
@ -90,7 +82,6 @@ namespace Grid {
template<class vobj,class cobj, class compressor> void
HaloExchange(const Lattice<vobj> &source,std::vector<cobj,alignedAllocator<cobj> > &u_comm_buf,compressor &compress)
{
std::cout<< "HaloExchange comm_buf.size()="<< u_comm_buf.size()<<" unified_buffer_size"<< _unified_buffer_size<< std::endl;
// conformable(source._grid,_grid);
assert(source._grid==_grid);
if (u_comm_buf.size() != _unified_buffer_size ) u_comm_buf.resize(_unified_buffer_size);
@ -141,7 +132,6 @@ namespace Grid {
}
}
}
std::cout<< "HaloExchange complete"<< std::endl;
}
template<class vobj,class cobj, class compressor>
@ -194,24 +184,18 @@ namespace Grid {
_grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
assert (xmit_to_rank != _grid->ThisRank());
assert (recv_from_rank != _grid->ThisRank());
// FIXME Implement asynchronous send & also avoid buffer copy
_grid->SendToRecvFrom((void *)&send_buf[0],
xmit_to_rank,
(void *)&recv_buf[0],
recv_from_rank,
bytes);
printf("GatherStartComms communicated offnode x %d\n",x);fflush(stdout);
printf("GatherStartComms inserting %le to u_comm_offset %d buf size %d for dim %d shift %d\n",
*( (RealF *) &recv_buf[0]),
u_comm_offset,buffer_size,
dimension,shift
); fflush(stdout);
for(int i=0;i<buffer_size;i++){
u_comm_buf[u_comm_offset+i]=recv_buf[i];
}
u_comm_offset+=buffer_size;
printf("GatherStartComms inserted x %d\n",x);fflush(stdout);
}
}
}
@ -248,7 +232,7 @@ namespace Grid {
int buffer_size = _grid->_slice_nblock[dimension]*_grid->_slice_block[dimension];
int words = sizeof(cobj)/sizeof(vector_type);
/* FIXME ALTERNATE BUFFER DETERMINATION */
/* FIXME ALTERNATE BUFFER DETERMINATION ; possibly slow to allocate*/
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);
@ -267,25 +251,21 @@ namespace Grid {
for(int x=0;x<rd;x++){
int any_offnode = ( ((x+sshift)%fd) >= rd );
std::cout<<"any_offnode ="<<any_offnode<<std::endl;
if ( any_offnode ) {
// FIXME call local permute copy if none are offnode.
for(int i=0;i<Nsimd;i++){
pointers[i] = &send_buf_extract[i][0];
}
int sx = (x+sshift)%rd;
std::cout<< "Gathering "<< x <<std::endl;
Gather_plane_extract<cobj>(rhs,pointers,dimension,sx,cbmask,compress);
std::cout<< "Gathered "<<std::endl;
for(int i=0;i<Nsimd;i++){
std::vector<int> icoor;
_grid->iCoorFromIindex(icoor,i);
int inner_bit = (Nsimd>>(permute_type+1));
int ic= (i&inner_bit)? 1:0;
assert(ic==icoor[dimension]);
int my_coor = rd*ic + x;
int nbr_coor = my_coor+sshift;
@ -301,12 +281,9 @@ namespace Grid {
if (nbr_ic) nbr_lane|=inner_bit;
assert (sx == nbr_ox);
std::cout<<"nbr_proc "<<nbr_proc<< " x "<<x<<" nbr_x "<<nbr_ox << " lane "<<i << " nbr_lane "<<nbr_lane
<< " nbr_ic "<<nbr_ic << " mycoor "<< my_coor<< " nbr_coor "<<nbr_coor<<std::endl;
if(nbr_proc){
std::cout<< "MPI sending "<<std::endl;
_grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
_grid->SendToRecvFrom((void *)&send_buf_extract[nbr_lane][0],
@ -314,23 +291,20 @@ namespace Grid {
(void *)&recv_buf_extract[i][0],
recv_from_rank,
bytes);
std::cout<< "MPI complete "<<std::endl;
rpointers[i] = &recv_buf_extract[i][0];
std::cout<<"lane "<<i<<" data "<<*( (Real *) rpointers[i])<<std::endl;
} else {
rpointers[i] = &send_buf_extract[nbr_lane][0];
std::cout<<"lane "<<i<<" data "<<*( (Real *) rpointers[i])<<std::endl;
}
}
// Here we don't want to scatter, just place into a buffer.
std::cout<< "merging u_comm_offset "<< u_comm_offset<<" comm_buf_size" << u_comm_buf.size() <<std::endl;
for(int i=0;i<buffer_size;i++){
assert(u_comm_offset+i<_unified_buffer_size);
merge(u_comm_buf[u_comm_offset+i],rpointers,i);
}
u_comm_offset+=buffer_size;
}
}

2
lib/qcd/Grid_qcd_wilson_dop.cc
View File

@ -24,7 +24,6 @@ const int WilsonMatrix::Tm = 7;
void Point(int p) {
mu=p;
std::cout << "WilsonCompressor.Point " << mu<<std::endl;
};
vHalfSpinColourVector operator () (const vSpinColourVector &in)
@ -193,7 +192,6 @@ void WilsonMatrix::Dhop(const LatticeFermion &in, LatticeFermion &out)
chi_p = &tmp;
}
}
std::cout<<"Xm for site "<<ss<<" l "<<local<<" p "<<perm<<" chi "<<Reduce(TensorRemove(innerProduct(*chi_p,*chi_p)))<<std::endl;
mult(&(Uchi()),&(Umu._odata[ss](Xm)),&(*chi_p)());
accumReconXm(result,Uchi);

21
lib/stencil/Grid_stencil_common.cc
View File

@ -19,7 +19,6 @@ namespace Grid {
_distances = distances;
_unified_buffer_size=0;
_request_count =0;
CommsRequests.resize(0);
int osites = _grid->oSites();
@ -117,6 +116,7 @@ namespace Grid {
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];
@ -137,9 +137,10 @@ namespace Grid {
for(int x=0;x<rd;x++){
int comm_proc = ((x+sshift)/rd)%pd;
int offnode = (comm_proc!=0);
int sx = (x+sshift)%rd;
int offnode = (((x+sshift)%fd) >= rd );
// int comm_proc = ((x+sshift)/ld)%pd;
// int offnode = (comm_proc!=0);
int sx = (x+sshift)%rd;
if (!offnode) {
@ -157,17 +158,9 @@ namespace Grid {
int recv_from_rank;
int xmit_to_rank;
CommsRequest cr;
cr.tag = _request_count++;
cr.words = words;
cr.unified_buffer_offset = _unified_buffer_size;
int unified_buffer_offset = _unified_buffer_size;
_unified_buffer_size += words;
grid->ShiftedRanks(dimension,comm_proc,cr.to_rank,cr.from_rank);
CommsRequests.push_back(cr);
ScatterPlane(point,dimension,x,cbmask,cr.unified_buffer_offset); // permute/extract/merge is done in comms phase
ScatterPlane(point,dimension,x,cbmask,unified_buffer_offset); // permute/extract/merge is done in comms phase
}
}

5
tests/Grid_wilson.cc
View File

@ -21,7 +21,7 @@ int main (int argc, char ** argv)
Grid_init(&argc,&argv);
std::vector<int> simd_layout({1,1,2,2});
std::vector<int> mpi_layout ({2,1,1,2});
std::vector<int> mpi_layout ({2,2,2,2});
std::vector<int> latt_size ({8,8,8,8});
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
@ -76,7 +76,8 @@ int main (int argc, char ** argv)
std::cout << "norm result "<< norm2(result)<<std::endl;
std::cout << "norm ref "<< norm2(ref)<<std::endl;
for(int ss=0;ss<10;ss++ ){
// for(int ss=0;ss<10;ss++ ){
for(int ss=0;ss<0;ss++ ){
for(int i=0;i<Ns;i++){
for(int j=0;j<Nc;j++){
ComplexF * ref_p = (ComplexF *)&ref._odata[ss]()(i)(j);