portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-03 05:24:32 +00:00
Code Releases Activity

Bug fixes for cases that physics code couldn't hit but latent

Browse Source 
and discovered on KNL (long vector, y SIMD dir) and checker dir set to y.
Remove the assertions on these code paths now they are tested.
This commit is contained in:
paboyle
2017-02-07 01:01:15 -05:00
parent 0883d6a7ce
commit 85c7bc4321
2 changed files with 56 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
lib/cshift/Cshift_common.h
View File

@@ -1,5 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid Grid physics library, www.github.com/paboyle/Grid
@@ -53,8 +52,7 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<cobj> &buffer,int dimen
cbmask = 0x3; cbmask = 0x3;
} }
int so = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane int so=plane*rhs._grid->_ostride[dimension]; // base offset for start of plane
int e1=rhs._grid->_slice_nblock[dimension]; int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block[dimension]; int e2=rhs._grid->_slice_block[dimension];
@@ -74,7 +72,7 @@ PARALLEL_NESTED_LOOP2
for(int n=0;n<e1;n++){ for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){ for(int b=0;b<e2;b++){
int o = n*stride; int o = n*stride;
int ocb=1<<rhs._grid->CheckerBoardFromOindexTable(o+b); int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
if ( ocb &cbmask ) { if ( ocb &cbmask ) {
table.push_back(std::pair<int,int> (bo++,o+b)); table.push_back(std::pair<int,int> (bo++,o+b));
} }
@@ -105,29 +103,30 @@ Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename cobj::scalar_
int e1=rhs._grid->_slice_nblock[dimension]; int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block[dimension]; int e2=rhs._grid->_slice_block[dimension];
int n1=rhs._grid->_slice_stride[dimension]; int n1=rhs._grid->_slice_stride[dimension];
int n2=rhs._grid->_slice_block[dimension];
if ( cbmask ==0x3){ if ( cbmask ==0x3){
PARALLEL_NESTED_LOOP2 PARALLEL_NESTED_LOOP2
for(int n=0;n<e1;n++){ for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){ for(int b=0;b<e2;b++){
int o = n*n1; int o = n*n1;
int offset = b+n*n2; int offset = b+n*e2;
cobj temp =compress(rhs._odata[so+o+b]); cobj temp =compress(rhs._odata[so+o+b]);
extract<cobj>(temp,pointers,offset); extract<cobj>(temp,pointers,offset);
} }
} }
} else { } else {
assert(0); //Fixme think this is buggy // Case of SIMD split AND checker dim cannot currently be hit, except in
// Test_cshift_red_black code.
std::cout << " Dense packed buffer WARNING " <<std::endl;
PARALLEL_NESTED_LOOP2
for(int n=0;n<e1;n++){ for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){ for(int b=0;b<e2;b++){
int o=n*rhs._grid->_slice_stride[dimension];
int o=n*n1;
int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b); int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
int offset = b+n*rhs._grid->_slice_block[dimension]; int offset = b+n*e2;
if ( ocb & cbmask ) { if ( ocb & cbmask ) {
cobj temp =compress(rhs._odata[so+o+b]); cobj temp =compress(rhs._odata[so+o+b]);
@@ -171,6 +170,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
int e1=rhs._grid->_slice_nblock[dimension]; int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block[dimension]; int e2=rhs._grid->_slice_block[dimension];
int stride=rhs._grid->_slice_stride[dimension];
if ( cbmask ==0x3 ) { if ( cbmask ==0x3 ) {
PARALLEL_NESTED_LOOP2 PARALLEL_NESTED_LOOP2
@@ -182,17 +182,22 @@ PARALLEL_NESTED_LOOP2
} }
} }
} else { } else {
std::vector<std::pair<int,int> > table;
int bo=0; int bo=0;
for(int n=0;n<e1;n++){ for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){ for(int b=0;b<e2;b++){
int o =n*rhs._grid->_slice_stride[dimension]; int o =n*rhs._grid->_slice_stride[dimension];
int bo =n*rhs._grid->_slice_block[dimension];
int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
if ( ocb & cbmask ) { if ( ocb & cbmask ) {
rhs._odata[so+o+b]=buffer[bo++]; table.push_back(std::pair<int,int> (so+o+b,bo++));
} }
} }
} }
PARALLEL_FOR_LOOP
for(int i=0;i<table.size();i++){
// std::cout << "Rcv"<< table[i].first << " " << table[i].second << " " <<buffer[table[i].second]<<std::endl;
rhs._odata[table[i].first]=buffer[table[i].second];
}
} }
} }
@@ -222,7 +227,11 @@ PARALLEL_NESTED_LOOP2
} }
} }
} else { } else {
assert(0); // think this is buggy FIXME
// Case of SIMD split AND checker dim cannot currently be hit, except in
// Test_cshift_red_black code.
// std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
for(int n=0;n<e1;n++){ for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){ for(int b=0;b<e2;b++){
int o = n*rhs._grid->_slice_stride[dimension]; int o = n*rhs._grid->_slice_stride[dimension];
@@ -338,8 +347,8 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
// Map to always positive shift modulo global full dimension. // Map to always positive shift modulo global full dimension.
shift = (shift+fd)%fd; shift = (shift+fd)%fd;
ret.checkerboard = grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
// the permute type // the permute type
ret.checkerboard = grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
int permute_dim =grid->PermuteDim(dimension); int permute_dim =grid->PermuteDim(dimension);
int permute_type=grid->PermuteType(dimension); int permute_type=grid->PermuteType(dimension);
int permute_type_dist; int permute_type_dist;
@@ -348,7 +357,6 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
int o = 0; int o = 0;
int bo = x * grid->_ostride[dimension]; int bo = x * grid->_ostride[dimension];
int cb= (cbmask==0x2)? Odd : Even; int cb= (cbmask==0x2)? Odd : Even;
int sshift = grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb); int sshift = grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
@@ -361,9 +369,23 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
// wrap is whether sshift > rd. // wrap is whether sshift > rd.
// num is sshift mod rd. // num is sshift mod rd.
// //
// shift 7
//
// XoXo YcYc
// oXoX cYcY
// XoXo YcYc
// oXoX cYcY
//
// sshift --
//
// XX YY ; 3
// XX YY ; 0
// XX YY ; 3
// XX YY ; 0
//
int permute_slice=0; int permute_slice=0;
if(permute_dim){ if(permute_dim){
int wrap = sshift/rd; int wrap = sshift/rd; wrap=wrap % ly;
int num = sshift%rd; int num = sshift%rd;
if ( x< rd-num ) permute_slice=wrap; if ( x< rd-num ) permute_slice=wrap;
@@ -375,7 +397,6 @@ template<class vobj> Lattice<vobj> Cshift_local(Lattice<vobj> &ret,const Lattice
} else { } else {
permute_type_dist = permute_type; permute_type_dist = permute_type;
} }
} }
if ( permute_slice ) Copy_plane_permute(ret,rhs,dimension,x,sx,cbmask,permute_type_dist); if ( permute_slice ) Copy_plane_permute(ret,rhs,dimension,x,sx,cbmask,permute_type_dist);

22
lib/cshift/Cshift_mpi.h
View File

@@ -74,7 +74,6 @@ template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &r
sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd); sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
// std::cout << "Cshift_comms dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl; // std::cout << "Cshift_comms dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
if ( sshift[0] == sshift[1] ) { if ( sshift[0] == sshift[1] ) {
// std::cout << "Single pass Cshift_comms" <<std::endl; // std::cout << "Single pass Cshift_comms" <<std::endl;
Cshift_comms(ret,rhs,dimension,shift,0x3); Cshift_comms(ret,rhs,dimension,shift,0x3);
@@ -154,10 +153,14 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
(void *)&recv_buf[0], (void *)&recv_buf[0],
recv_from_rank, recv_from_rank,
bytes); bytes);
grid->Barrier();
// for(int i=0;i<words;i++){ /*
// std::cout << "SendRecv ["<<i<<"] snd "<<send_buf[i]<<" rcv " << recv_buf[i] << " 0x" << cbmask<<std::endl; for(int i=0;i<send_buf.size();i++){
// } assert(recv_buf.size()==buffer_size);
assert(send_buf.size()==buffer_size);
std::cout << "SendRecv_Cshift_comms ["<<i<<" "<< dimension<<"] snd "<<send_buf[i]<<" rcv " << recv_buf[i] << " 0x" << cbmask<<std::endl;
}
*/
Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask); Scatter_plane_simple (ret,recv_buf,dimension,x,cbmask);
} }
} }
@@ -243,7 +246,14 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
(void *)&recv_buf_extract[i][0], (void *)&recv_buf_extract[i][0],
recv_from_rank, recv_from_rank,
bytes); bytes);
/*
for(int w=0;w<recv_buf_extract[i].size();w++){
assert(recv_buf_extract[i].size()==buffer_size);
assert(send_buf_extract[i].size()==buffer_size);
std::cout << "SendRecv_Cshift_comms ["<<w<<" "<< dimension<<"] recv "<<recv_buf_extract[i][w]<<" send " << send_buf_extract[nbr_lane][w] << cbmask<<std::endl;
}
*/
grid->Barrier();
rpointers[i] = &recv_buf_extract[i][0]; rpointers[i] = &recv_buf_extract[i][0];
} else { } else {
rpointers[i] = &send_buf_extract[nbr_lane][0]; rpointers[i] = &send_buf_extract[nbr_lane][0];