portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-15 14:27:06 +01:00
Code Releases Activity

Merge remote-tracking branch 'upstream/develop' into feature/ddalphaamg

Browse Source
This commit is contained in:
Daniel Richtmann
2018-01-08 10:37:10 +01:00
parent 6cf635d61c 67c3fa0f5f
commit c6411f8514
92 changed files with 5093 additions and 1574 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
lib/algorithms/LinearOperator.h
View File

@ -308,32 +308,34 @@ namespace Grid {
public:
SchurStaggeredOperator (Matrix &Mat): _Mat(Mat){};
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
GridLogIterative.TimingMode(1);
std::cout << GridLogIterative << " HermOpAndNorm "<<std::endl;
n2 = Mpc(in,out);
std::cout << GridLogIterative << " HermOpAndNorm.Mpc "<<std::endl;
ComplexD dot= innerProduct(in,out);
std::cout << GridLogIterative << " HermOpAndNorm.innerProduct "<<std::endl;
n1 = real(dot);
}
virtual void HermOp(const Field &in, Field &out){
std::cout << GridLogIterative << " HermOp "<<std::endl;
Mpc(in,out);
}
virtual RealD Mpc (const Field &in, Field &out) {
Field tmp(in._grid);
Field tmp2(in._grid);
std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
_Mat.Mooee(in,out);
_Mat.Mooee(out,tmp);
std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
_Mat.Meooe(in,out);
_Mat.Meooe(out,tmp2);
std::cout << GridLogIterative << " HermOp.MeooeMeooe "<<std::endl;
return axpy_norm(out,-1.0,tmp2,tmp);
#if 0
//... much prefer conventional Schur norm
_Mat.Meooe(in,tmp);
_Mat.MooeeInv(tmp,out);
_Mat.Meooe(out,tmp);
_Mat.Mooee(in,out);
return axpy_norm(out,-1.0,tmp,out);
#endif
RealD nn=axpy_norm(out,-1.0,tmp2,tmp);
std::cout << GridLogIterative << " HermOp.axpy_norm "<<std::endl;
return nn;
}
virtual RealD MpcDag (const Field &in, Field &out){
return Mpc(in,out);

9
lib/algorithms/iterative/SchurRedBlack.h
View File

@ -123,11 +123,14 @@ namespace Grid {
Field tmp(grid);
Field Mtmp(grid);
Field resid(fgrid);
std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve " <<std::endl;
pickCheckerboard(Even,src_e,in);
pickCheckerboard(Odd ,src_o,in);
pickCheckerboard(Even,sol_e,out);
pickCheckerboard(Odd ,sol_o,out);
std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve checkerboards picked" <<std::endl;
/////////////////////////////////////////////////////
// src_o = (source_o - Moe MeeInv source_e)
@ -144,6 +147,7 @@ namespace Grid {
//////////////////////////////////////////////////////////////
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver calling the Mpc solver" <<std::endl;
_HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.checkerboard==Odd);
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver called the Mpc solver" <<std::endl;
///////////////////////////////////////////////////
// sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
@ -152,15 +156,16 @@ namespace Grid {
src_e = src_e-tmp; assert( src_e.checkerboard ==Even);
_Matrix.MooeeInv(src_e,sol_e); assert( sol_e.checkerboard ==Even);
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver reconstructed other CB" <<std::endl;
setCheckerboard(out,sol_e); assert( sol_e.checkerboard ==Even);
setCheckerboard(out,sol_o); assert( sol_o.checkerboard ==Odd );
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver inserted solution" <<std::endl;
// Verify the unprec residual
_Matrix.M(out,resid);
resid = resid-in;
RealD ns = norm2(in);
RealD nr = norm2(resid);
std::cout<<GridLogMessage << "SchurRedBlackStaggered solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
}
};

30
lib/allocator/AlignedAllocator.cc
View File

@ -3,9 +3,12 @@
namespace Grid {
MemoryStats *MemoryProfiler::stats = nullptr;
bool MemoryProfiler::debug = false;
int PointerCache::victim;
PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
void *PointerCache::Insert(void *ptr,size_t bytes) {
@ -94,4 +97,29 @@ void check_huge_pages(void *Buf,uint64_t BYTES)
#endif
}
std::string sizeString(const size_t bytes)
{
constexpr unsigned int bufSize = 256;
const char *suffixes[7] = {"", "K", "M", "G", "T", "P", "E"};
char buf[256];
size_t s = 0;
double count = bytes;
while (count >= 1024 && s < 7)
{
s++;
count /= 1024;
}
if (count - floor(count) == 0.0)
{
snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
}
else
{
snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
}
return std::string(buf);
}
}

90
lib/allocator/AlignedAllocator.h
View File

@ -63,6 +63,64 @@ namespace Grid {
static void *Lookup(size_t bytes) ;
};
std::string sizeString(size_t bytes);
struct MemoryStats
{
size_t totalAllocated{0}, maxAllocated{0},
currentlyAllocated{0}, totalFreed{0};
};
class MemoryProfiler
{
public:
static MemoryStats *stats;
static bool debug;
};
#define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
#define profilerDebugPrint \
if (MemoryProfiler::stats)\
{\
auto s = MemoryProfiler::stats;\
std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl;\
std::cout << GridLogDebug << "[Memory debug] total : " << memString(s->totalAllocated) \
<< std::endl;\
std::cout << GridLogDebug << "[Memory debug] max : " << memString(s->maxAllocated) \
<< std::endl;\
std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
<< std::endl;\
std::cout << GridLogDebug << "[Memory debug] freed : " << memString(s->totalFreed) \
<< std::endl;\
}
#define profilerAllocate(bytes)\
if (MemoryProfiler::stats)\
{\
auto s = MemoryProfiler::stats;\
s->totalAllocated += (bytes);\
s->currentlyAllocated += (bytes);\
s->maxAllocated = std::max(s->maxAllocated, s->currentlyAllocated);\
}\
if (MemoryProfiler::debug)\
{\
std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl;\
profilerDebugPrint;\
}
#define profilerFree(bytes)\
if (MemoryProfiler::stats)\
{\
auto s = MemoryProfiler::stats;\
s->totalFreed += (bytes);\
s->currentlyAllocated -= (bytes);\
}\
if (MemoryProfiler::debug)\
{\
std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl;\
profilerDebugPrint;\
}
void check_huge_pages(void *Buf,uint64_t BYTES);
@ -92,6 +150,7 @@ public:
pointer allocate(size_type __n, const void* _p= 0)
{
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
_Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
// if ( ptr != NULL )
@ -122,6 +181,8 @@ public:
void deallocate(pointer __p, size_type __n) {
size_type bytes = __n * sizeof(_Tp);
profilerFree(bytes);
pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
#ifdef HAVE_MM_MALLOC_H
@ -172,10 +233,13 @@ public:
#ifdef GRID_COMMS_SHMEM
pointer allocate(size_type __n, const void* _p= 0)
{
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
#ifdef CRAY
_Tp *ptr = (_Tp *) shmem_align(__n*sizeof(_Tp),64);
_Tp *ptr = (_Tp *) shmem_align(bytes,64);
#else
_Tp *ptr = (_Tp *) shmem_align(64,__n*sizeof(_Tp));
_Tp *ptr = (_Tp *) shmem_align(64,bytes);
#endif
#ifdef PARANOID_SYMMETRIC_HEAP
static void * bcast;
@ -193,18 +257,23 @@ public:
#endif
return ptr;
}
void deallocate(pointer __p, size_type) {
void deallocate(pointer __p, size_type __n) {
size_type bytes = __n*sizeof(_Tp);
profilerFree(bytes);
shmem_free((void *)__p);
}
#else
pointer allocate(size_type __n, const void* _p= 0)
{
#ifdef HAVE_MM_MALLOC_H
_Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),GRID_ALLOC_ALIGN);
#else
_Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,__n*sizeof(_Tp));
#endif
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
#ifdef HAVE_MM_MALLOC_H
_Tp * ptr = (_Tp *) _mm_malloc(bytes, GRID_ALLOC_ALIGN);
#else
_Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN, bytes);
#endif
uint8_t *cp = (uint8_t *)ptr;
if ( ptr ) {
// One touch per 4k page, static OMP loop to catch same loop order
@ -215,7 +284,10 @@ public:
}
return ptr;
}
void deallocate(pointer __p, size_type) {
void deallocate(pointer __p, size_type __n) {
size_type bytes = __n*sizeof(_Tp);
profilerFree(bytes);
#ifdef HAVE_MM_MALLOC_H
_mm_free((void *)__p);
#else

83
lib/communicator/Communicator_base.cc
View File

@ -134,8 +134,18 @@ void CartesianCommunicator::AllToAll(void *in,void *out,uint64_t words,uint64_t
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)
{
_ndimension = processors.size();
assert(_ndimension = parent._ndimension);
int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
std::vector<int> parent_processor_coor(_ndimension,0);
std::vector<int> parent_processors (_ndimension,1);
// Can make 5d grid from 4d etc...
int pad = _ndimension-parent_ndimension;
for(int d=0;d<parent_ndimension;d++){
parent_processor_coor[pad+d]=parent._processor_coor[d];
parent_processors [pad+d]=parent._processors[d];
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// split the communicator
//////////////////////////////////////////////////////////////////////////////////////////////////////
@ -154,9 +164,9 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
std::vector<int> ssize(_ndimension); // coor of split within parent
for(int d=0;d<_ndimension;d++){
ccoor[d] = parent._processor_coor[d] % processors[d];
scoor[d] = parent._processor_coor[d] / processors[d];
ssize[d] = parent._processors[d] / processors[d];
ccoor[d] = parent_processor_coor[d] % processors[d];
scoor[d] = parent_processor_coor[d] / processors[d];
ssize[d] = parent_processors[d] / processors[d];
}
int crank; // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
// Mpi uses the reverse Lexico convention to us
@ -166,38 +176,36 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
MPI_Comm comm_split;
if ( Nchild > 1 ) {
/*
std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"] ";
for(int d=0;d<parent._processors.size();d++) std::cout << parent._processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" child grid["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << parent._processor_coor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << ccoor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new coor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << parent._processor_coor[d] << " ";
std::cout<<std::endl;
*/
if(0){
std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"] ";
for(int d=0;d<parent._ndimension;d++) std::cout << parent._processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" child grid["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< parent._ndimension <<"] ";
for(int d=0;d<parent._ndimension;d++) std::cout << parent._processor_coor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new split "<< srank<<" scoor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << scoor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << ccoor[d] << " ";
std::cout<<std::endl;
}
int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
assert(ierr==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Declare victory
//////////////////////////////////////////////////////////////////////////////////////////////////////
/*
std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
<< Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
*/
// std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
// << Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
} else {
comm_split=parent.communicator;
srank = 0;
@ -207,6 +215,17 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
// Set up from the new split communicator
//////////////////////////////////////////////////////////////////////////////////////////////////////
InitFromMPICommunicator(processors,comm_split);
if(0){
std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
for(int d=0;d<processors.size();d++){
std::cout << d<< " " << _processor_coor[d] <<" " << ccoor[d]<<std::endl;
}
}
for(int d=0;d<processors.size();d++){
assert(_processor_coor[d] == ccoor[d] );
}
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
@ -231,7 +250,7 @@ void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &proc
MPI_Comm_rank(communicator,&_processor);
MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
if ( communicator_base != communicator_world ) {
if ( 0 && (communicator_base != communicator_world) ) {
std::cout << "InitFromMPICommunicator Cartesian communicator created with a non-world communicator"<<std::endl;
std::cout << " new communicator rank "<<_processor<< " coor ["<<_ndimension<<"] ";

7
lib/communicator/Communicator_base.h
View File

@ -276,10 +276,11 @@ class CartesianCommunicator {
assert(in.size()==out.size());
uint64_t bytes=sizeof(T);
uint64_t words=in.size()/numnode;
// std:: cout << "AllToAll buffer size "<< in.size()*sizeof(T)<<std::endl;
// std:: cout << "AllToAll datum bytes "<< bytes<<std::endl;
// std:: cout << "AllToAll datum count "<< words<<std::endl;
assert(numnode * words == in.size());
assert(words < (1ULL<<32));
assert(words < (1ULL<<31));
AllToAll(dim,(void *)&in[0],(void *)&out[0],words,bytes);
}
void AllToAll(int dim ,void *in,void *out,uint64_t words,uint64_t bytes);

2
lib/communicator/Communicator_mpi3.cc
View File

@ -606,7 +606,7 @@ CartesianCommunicator::~CartesianCommunicator()
MPI_Finalized(&MPI_is_finalised);
if (communicator && !MPI_is_finalised) {
MPI_Comm_free(&communicator);
for(int i=0;i< communicator_halo.size();i++){
for(int i=0;i<communicator_halo.size();i++){
MPI_Comm_free(&communicator_halo[i]);
}
}

42
lib/lattice/Lattice_rng.h
View File

@ -77,9 +77,6 @@ namespace Grid {
// merge of April 11 2017
//<<<<<<< HEAD
// this function is necessary for the LS vectorised field
inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
{
@ -91,7 +88,6 @@ namespace Grid {
// all further divisions are local
for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
// then divide the number of local sites
// check that the total number of sims agree, meanse the iSites are the same
@ -102,27 +98,6 @@ namespace Grid {
return fine->lSites() / coarse->lSites();
}
/*
// Wrap seed_seq to give common interface with random_device
class fixedSeed {
public:
typedef std::seed_seq::result_type result_type;
std::seed_seq src;
fixedSeed(const std::vector<int> &seeds) : src(seeds.begin(),seeds.end()) {};
result_type operator () (void){
std::vector<result_type> list(1);
src.generate(list.begin(),list.end());
return list[0];
}
};
=======
>>>>>>> develop
*/
// real scalars are one component
template<class scalar,class distribution,class generator>
@ -171,7 +146,7 @@ namespace Grid {
// support for parallel init
///////////////////////
#ifdef RNG_FAST_DISCARD
static void Skip(RngEngine &eng)
static void Skip(RngEngine &eng,uint64_t site)
{
/////////////////////////////////////////////////////////////////////////////////////
// Skip by 2^40 elements between successive lattice sites
@ -184,8 +159,11 @@ namespace Grid {
// and margin of safety is orders of magnitude.
// We could hack Sitmo to skip in the higher order words of state if necessary
/////////////////////////////////////////////////////////////////////////////////////
uint64_t skip = 0x1; skip = skip<<40;
// uint64_t skip = site+1; // Old init Skipped then drew. Checked compat with faster init
uint64_t skip = site;
skip = skip<<40;
eng.discard(skip);
// std::cout << " Engine " <<site << " state " <<eng<<std::endl;
}
#endif
static RngEngine Reseed(RngEngine &eng)
@ -407,15 +385,14 @@ namespace Grid {
// MT implementation does not implement fast discard even though
// in principle this is possible
////////////////////////////////////////////////
std::vector<int> gcoor;
int rank,o_idx,i_idx;
// Everybody loops over global volume.
for(int gidx=0;gidx<_grid->_gsites;gidx++){
Skip(master_engine); // Skip to next RNG sequence
parallel_for(int gidx=0;gidx<_grid->_gsites;gidx++){
// Where is it?
int rank,o_idx,i_idx;
std::vector<int> gcoor;
_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
@ -423,6 +400,7 @@ namespace Grid {
if( rank == _grid->ThisRank() ){
int l_idx=generator_idx(o_idx,i_idx);
_generators[l_idx] = master_engine;
Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
}
}

203
lib/lattice/Lattice_transfer.h
View File

@ -50,26 +50,22 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full){
half.checkerboard = cb;
int ssh=0;
//parallel_for
for(int ss=0;ss<full._grid->oSites();ss++){
std::vector<int> coor;
parallel_for(int ss=0;ss<full._grid->oSites();ss++){
int cbos;
std::vector<int> coor;
full._grid->oCoorFromOindex(coor,ss);
cbos=half._grid->CheckerBoard(coor);
if (cbos==cb) {
int ssh=half._grid->oIndex(coor);
half._odata[ssh] = full._odata[ss];
ssh++;
}
}
}
template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
int cb = half.checkerboard;
int ssh=0;
//parallel_for
for(int ss=0;ss<full._grid->oSites();ss++){
parallel_for(int ss=0;ss<full._grid->oSites();ss++){
std::vector<int> coor;
int cbos;
@ -77,8 +73,8 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
cbos=half._grid->CheckerBoard(coor);
if (cbos==cb) {
int ssh=half._grid->oIndex(coor);
full._odata[ss]=half._odata[ssh];
ssh++;
}
}
}
@ -698,30 +694,6 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
////////////////////////////////////////////////////////////////////////////////
// Communicate between grids
////////////////////////////////////////////////////////////////////////////////
//
// All to all plan
//
// Subvolume on fine grid is v. Vectors a,b,c,d
//
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// SIMPLEST CASE:
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Mesh of nodes (2) ; subdivide to 1 subdivisions
//
// Lex ord:
// N0 va0 vb0 N1 va1 vb1
//
// For each dimension do an all to all
//
// full AllToAll(0)
// N0 va0 va1 N1 vb0 vb1
//
// REARRANGE
// N0 va01 N1 vb01
//
// Must also rearrange data to get into the NEW lex order of grid at each stage. Some kind of "insert/extract".
// NB: Easiest to programme if keep in lex order.
//
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// SIMPLE CASE:
///////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -755,9 +727,17 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
//
// Must also rearrange data to get into the NEW lex order of grid at each stage. Some kind of "insert/extract".
// NB: Easiest to programme if keep in lex order.
//
/////////////////////////////////////////////////////////
/*
* Let chunk = (fvol*nvec)/sP be size of a chunk. ( Divide lexico vol * nvec into fP/sP = M chunks )
*
* 2nd A2A (over sP nodes; subdivide the fP into sP chunks of M)
*
* node 0 1st chunk of node 0M..(1M-1); 2nd chunk of node 0M..(1M-1).. data chunk x M x sP = fL / sP * M * sP = fL * M growth
* node 1 1st chunk of node 1M..(2M-1); 2nd chunk of node 1M..(2M-1)..
* node 2 1st chunk of node 2M..(3M-1); 2nd chunk of node 2M..(3M-1)..
* node 3 1st chunk of node 3M..(3M-1); 2nd chunk of node 2M..(3M-1)..
* etc...
*/
template<class Vobj>
void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj> & split)
{
@ -816,57 +796,58 @@ void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj> & split)
int nvec = nvector; // Counts down to 1 as we collapse dims
std::vector<int> ldims = full_grid->_ldimensions;
std::vector<int> lcoor(ndim);
for(int d=ndim-1;d>=0;d--){
if ( ratio[d] != 1 ) {
full_grid ->AllToAll(d,alldata,tmpdata);
// std::cout << GridLogMessage << "Grid_split: dim " <<d<<" ratio "<<ratio[d]<<" nvec "<<nvec<<" procs "<<split_grid->_processors[d]<<std::endl;
// for(int v=0;v<nvec;v++){
// std::cout << "Grid_split: alldata["<<v<<"] " << alldata[v] <<std::endl;
// std::cout << "Grid_split: tmpdata["<<v<<"] " << tmpdata[v] <<std::endl;
// }
//////////////////////////////////////////
//Local volume for this dimension is expanded by ratio of processor extents
// Number of vectors is decreased by same factor
// Rearrange to lexico for bigger volume
//////////////////////////////////////////
nvec /= ratio[d];
if ( split_grid->_processors[d] > 1 ) {
alldata=tmpdata;
split_grid->AllToAll(d,alldata,tmpdata);
}
auto rdims = ldims; rdims[d] *= ratio[d];
auto rsites= lsites*ratio[d];
for(int v=0;v<nvec;v++){
auto rdims = ldims;
auto M = ratio[d];
auto rsites= lsites*M;// increases rsites by M
nvec /= M; // Reduce nvec by subdivision factor
rdims[d] *= M; // increase local dim by same factor
// For loop over each site within old subvol
for(int lsite=0;lsite<lsites;lsite++){
int sP = split_grid->_processors[d];
int fP = full_grid->_processors[d];
Lexicographic::CoorFromIndex(lcoor, lsite, ldims);
int fvol = lsites;
int chunk = (nvec*fvol)/sP; assert(chunk*sP == nvec*fvol);
for(int r=0;r<ratio[d];r++){ // ratio*nvec terms
// Loop over reordered data post A2A
parallel_for(int c=0;c<chunk;c++){
std::vector<int> coor(ndim);
for(int m=0;m<M;m++){
for(int s=0;s<sP;s++){
// addressing; use lexico
int lex_r;
uint64_t lex_c = c+chunk*m+chunk*M*s;
uint64_t lex_fvol_vec = c+chunk*s;
uint64_t lex_fvol = lex_fvol_vec%fvol;
uint64_t lex_vec = lex_fvol_vec/fvol;
auto rcoor = lcoor; rcoor[d] += r*ldims[d];
// which node sets an adder to the coordinate
Lexicographic::CoorFromIndex(coor, lex_fvol, ldims);
coor[d] += m*ldims[d];
Lexicographic::IndexFromCoor(coor, lex_r, rdims);
lex_r += lex_vec * rsites;
int rsite; Lexicographic::IndexFromCoor(rcoor, rsite, rdims);
rsite += v * rsites;
// LexicoFind coordinate & vector number within split lattice
alldata[lex_r] = tmpdata[lex_c];
int rmul=nvec*lsites;
int vmul= lsites;
alldata[rsite] = tmpdata[lsite+r*rmul+v*vmul];
// if ( lsite==0 ) {
// std::cout << "Grid_split: grow alldata["<<rsite<<"] " << alldata[rsite] << " <- tmpdata["<< lsite+r*rmul+v*vmul<<"] "<<tmpdata[lsite+r*rmul+v*vmul] <<std::endl;
// }
}
}
}
ldims[d]*= ratio[d];
lsites *= ratio[d];
if ( split_grid->_processors[d] > 1 ) {
tmpdata = alldata;
split_grid->AllToAll(d,tmpdata,alldata);
}
}
}
vectorizeFromLexOrdArray(alldata,split);
@ -937,72 +918,74 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj> & split)
/////////////////////////////////////////////////////////////////
// Start from split grid and work towards full grid
/////////////////////////////////////////////////////////////////
std::vector<int> lcoor(ndim);
std::vector<int> rcoor(ndim);
int nvec = 1;
lsites = split_grid->lSites();
std::vector<int> ldims = split_grid->_ldimensions;
uint64_t rsites = split_grid->lSites();
std::vector<int> rdims = split_grid->_ldimensions;
// for(int d=ndim-1;d>=0;d--){
for(int d=0;d<ndim;d++){
if ( ratio[d] != 1 ) {
auto M = ratio[d];
if ( split_grid->_processors[d] > 1 ) {
tmpdata = alldata;
split_grid->AllToAll(d,tmpdata,alldata);
}
//////////////////////////////////////////
//Local volume for this dimension is expanded by ratio of processor extents
// Number of vectors is decreased by same factor
// Rearrange to lexico for bigger volume
//////////////////////////////////////////
auto rsites= lsites/ratio[d];
auto rdims = ldims; rdims[d]/=ratio[d];
for(int v=0;v<nvec;v++){
// rsite, rcoor --> smaller local volume
// lsite, lcoor --> bigger original (single node?) volume
// For loop over each site within smaller subvol
for(int rsite=0;rsite<rsites;rsite++){
Lexicographic::CoorFromIndex(rcoor, rsite, rdims);
int lsite;
for(int r=0;r<ratio[d];r++){
lcoor = rcoor; lcoor[d] += r*rdims[d];
Lexicographic::IndexFromCoor(lcoor, lsite, ldims); lsite += v * lsites;
int rmul=nvec*rsites;
int vmul= rsites;
tmpdata[rsite+r*rmul+v*vmul]=alldata[lsite];
int sP = split_grid->_processors[d];
int fP = full_grid->_processors[d];
auto ldims = rdims; ldims[d] /= M; // Decrease local dims by same factor
auto lsites= rsites/M; // Decreases rsites by M
int fvol = lsites;
int chunk = (nvec*fvol)/sP; assert(chunk*sP == nvec*fvol);
{
// Loop over reordered data post A2A
parallel_for(int c=0;c<chunk;c++){
std::vector<int> coor(ndim);
for(int m=0;m<M;m++){
for(int s=0;s<sP;s++){
// addressing; use lexico
int lex_r;
uint64_t lex_c = c+chunk*m+chunk*M*s;
uint64_t lex_fvol_vec = c+chunk*s;
uint64_t lex_fvol = lex_fvol_vec%fvol;
uint64_t lex_vec = lex_fvol_vec/fvol;
// which node sets an adder to the coordinate
Lexicographic::CoorFromIndex(coor, lex_fvol, ldims);
coor[d] += m*ldims[d];
Lexicographic::IndexFromCoor(coor, lex_r, rdims);
lex_r += lex_vec * rsites;
// LexicoFind coordinate & vector number within split lattice
tmpdata[lex_c] = alldata[lex_r];
}
}
}
}
nvec *= ratio[d];
ldims[d]=rdims[d];
lsites =rsites;
if ( split_grid->_processors[d] > 1 ) {
split_grid->AllToAll(d,tmpdata,alldata);
tmpdata=alldata;
}
full_grid ->AllToAll(d,tmpdata,alldata);
rdims[d]/= M;
rsites /= M;
nvec *= M; // Increase nvec by subdivision factor
}
}
lsites = full_grid->lSites();
for(int v=0;v<nvector;v++){
assert(v<full.size());
// assert(v<full.size());
parallel_for(int site=0;site<lsites;site++){
// assert(v*lsites+site < alldata.size());
scalardata[site] = alldata[v*lsites+site];
}
vectorizeFromLexOrdArray(scalardata,full[v]);
}
}
}
#endif

8
lib/qcd/QCD.h
View File

@ -492,6 +492,14 @@ namespace QCD {
return traceIndex<ColourIndex>(lhs);
}
//////////////////////////////////////////
// Current types
//////////////////////////////////////////
GRID_SERIALIZABLE_ENUM(Current, undef,
Vector, 0,
Axial, 1,
Tadpole, 2);
} //namespace QCD
} // Grid

16
lib/qcd/action/fermion/FermionOperator.h
View File

@ -47,6 +47,7 @@ namespace Grid {
INHERIT_IMPL_TYPES(Impl);
FermionOperator(const ImplParams &p= ImplParams()) : Impl(p) {};
virtual ~FermionOperator(void) = default;
virtual FermionField &tmp(void) = 0;
@ -112,6 +113,21 @@ namespace Grid {
///////////////////////////////////////////////
virtual void ImportGauge(const GaugeField & _U)=0;
//////////////////////////////////////////////////////////////////////
// Conserved currents, either contract at sink or insert sequentially.
//////////////////////////////////////////////////////////////////////
virtual void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu)=0;
virtual void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax)=0;
};
}

29
lib/qcd/action/fermion/FermionOperatorImpl.h
View File

@ -212,6 +212,13 @@ namespace QCD {
StencilImpl &St) {
mult(&phi(), &U(mu), &chi());
}
inline void multLinkProp(SitePropagator &phi,
const SiteDoubledGaugeField &U,
const SitePropagator &chi,
int mu) {
mult(&phi(), &U(mu), &chi());
}
template <class ref>
inline void loadLinkElement(Simd &reg, ref &memory) {
@ -340,7 +347,20 @@ class DomainWallVec5dImpl : public PeriodicGaugeImpl< GaugeImplTypes< S,Nrepres
}
mult(&phi(), &UU(), &chi());
}
inline void multLinkProp(SitePropagator &phi,
const SiteDoubledGaugeField &U,
const SitePropagator &chi,
int mu) {
SiteGaugeLink UU;
for (int i = 0; i < Nrepresentation; i++) {
for (int j = 0; j < Nrepresentation; j++) {
vsplat(UU()()(i, j), U(mu)()(i, j));
}
}
mult(&phi(), &UU(), &chi());
}
inline void DoubleStore(GridBase *GaugeGrid, DoubledGaugeField &Uds,const GaugeField &Umu)
{
SiteScalarGaugeField ScalarUmu;
@ -537,7 +557,12 @@ class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Nrepresent
}
}
// Fixme: Gparity prop * link
inline void multLinkProp(SitePropagator &phi, const SiteDoubledGaugeField &U,
const SitePropagator &chi, int mu)
{
assert(0);
}
template <class ref>
inline void loadLinkElement(Simd &reg, ref &memory) {

25
lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
View File

@ -393,6 +393,31 @@ void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder
}
};
////////////////////////////////////////////////////////
// Conserved current - not yet implemented.
////////////////////////////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu)
{
assert(0);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax)
{
assert(0);
}
FermOpStaggeredTemplateInstantiate(ImprovedStaggeredFermion);
//AdjointFermOpTemplateInstantiate(ImprovedStaggeredFermion);

16
lib/qcd/action/fermion/ImprovedStaggeredFermion.h
View File

@ -157,6 +157,22 @@ class ImprovedStaggeredFermion : public StaggeredKernels<Impl>, public ImprovedS
LebesgueOrder Lebesgue;
LebesgueOrder LebesgueEvenOdd;
///////////////////////////////////////////////////////////////
// Conserved current utilities
///////////////////////////////////////////////////////////////
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax);
};
typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;

24
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc
View File

@ -405,6 +405,30 @@ void ImprovedStaggeredFermion5D<Impl>::MooeeInvDag(const FermionField &in,
MooeeInv(in, out);
}
////////////////////////////////////////////////////////
// Conserved current - not yet implemented.
////////////////////////////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu)
{
assert(0);
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax)
{
assert(0);
}
FermOpStaggeredTemplateInstantiate(ImprovedStaggeredFermion5D);
FermOpStaggeredVec5dTemplateInstantiate(ImprovedStaggeredFermion5D);

15
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.h
View File

@ -170,6 +170,21 @@ namespace QCD {
// Comms buffer
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > comm_buf;
///////////////////////////////////////////////////////////////
// Conserved current utilities
///////////////////////////////////////////////////////////////
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax);
};
}}

106
lib/qcd/action/fermion/WilsonFermion.cc
View File

@ -345,6 +345,112 @@ void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
}
};
/*******************************************************************************
* Conserved current utilities for Wilson fermions, for contracting propagators
* to make a conserved current sink or inserting the conserved current
* sequentially.
******************************************************************************/
template <class Impl>
void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu)
{
Gamma g5(Gamma::Algebra::Gamma5);
conformable(_grid, q_in_1._grid);
conformable(_grid, q_in_2._grid);
conformable(_grid, q_out._grid);
PropagatorField tmp1(_grid), tmp2(_grid);
q_out = zero;
// Forward, need q1(x + mu), q2(x). Backward, need q1(x), q2(x + mu).
// Inefficient comms method but not performance critical.
tmp1 = Cshift(q_in_1, mu, 1);
tmp2 = Cshift(q_in_2, mu, 1);
parallel_for (unsigned int sU = 0; sU < Umu._grid->oSites(); ++sU)
{
Kernels::ContractConservedCurrentSiteFwd(tmp1._odata[sU],
q_in_2._odata[sU],
q_out._odata[sU],
Umu, sU, mu);
Kernels::ContractConservedCurrentSiteBwd(q_in_1._odata[sU],
tmp2._odata[sU],
q_out._odata[sU],
Umu, sU, mu);
}
}
template <class Impl>
void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax)
{
conformable(_grid, q_in._grid);
conformable(_grid, q_out._grid);
Lattice<iSinglet<Simd>> ph(_grid), coor(_grid);
Complex i(0.0,1.0);
PropagatorField tmpFwd(_grid), tmpBwd(_grid), tmp(_grid);
unsigned int tshift = (mu == Tp) ? 1 : 0;
unsigned int LLt = GridDefaultLatt()[Tp];
// Momentum projection
ph = zero;
for(unsigned int mu = 0; mu < Nd - 1; mu++)
{
LatticeCoordinate(coor, mu);
ph = ph + mom[mu]*coor*((1./(_grid->_fdimensions[mu])));
}
ph = exp((Real)(2*M_PI)*i*ph);
q_out = zero;
LatticeInteger coords(_grid);
LatticeCoordinate(coords, Tp);
// Need q(x + mu) and q(x - mu).
tmp = Cshift(q_in, mu, 1);
tmpFwd = tmp*ph;
tmp = ph*q_in;
tmpBwd = Cshift(tmp, mu, -1);
parallel_for (unsigned int sU = 0; sU < Umu._grid->oSites(); ++sU)
{
// Compute the sequential conserved current insertion only if our simd
// object contains a timeslice we need.
vInteger t_mask = ((coords._odata[sU] >= tmin) &&
(coords._odata[sU] <= tmax));
Integer timeSlices = Reduce(t_mask);
if (timeSlices > 0)
{
Kernels::SeqConservedCurrentSiteFwd(tmpFwd._odata[sU],
q_out._odata[sU],
Umu, sU, mu, t_mask);
}
// Repeat for backward direction.
t_mask = ((coords._odata[sU] >= (tmin + tshift)) &&
(coords._odata[sU] <= (tmax + tshift)));
//if tmax = LLt-1 (last timeslice) include timeslice 0 if the time is shifted (mu=3)
unsigned int t0 = 0;
if((tmax==LLt-1) && (tshift==1)) t_mask = (t_mask || (coords._odata[sU] == t0 ));
timeSlices = Reduce(t_mask);
if (timeSlices > 0)
{
Kernels::SeqConservedCurrentSiteBwd(tmpBwd._odata[sU],
q_out._odata[sU],
Umu, sU, mu, t_mask);
}
}
}
FermOpTemplateInstantiate(WilsonFermion);
AdjointFermOpTemplateInstantiate(WilsonFermion);
TwoIndexFermOpTemplateInstantiate(WilsonFermion);

16
lib/qcd/action/fermion/WilsonFermion.h
View File

@ -146,6 +146,22 @@ class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic {
LebesgueOrder Lebesgue;
LebesgueOrder LebesgueEvenOdd;
///////////////////////////////////////////////////////////////
// Conserved current utilities
///////////////////////////////////////////////////////////////
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax);
};
typedef WilsonFermion<WilsonImplF> WilsonFermionF;

163
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -12,6 +12,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
Author: Andrew Lawson <andrew.lawson1991@gmail.com>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -702,6 +703,168 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHw(FermionField &out,const Fe
}
/*******************************************************************************
* Conserved current utilities for Wilson fermions, for contracting propagators
* to make a conserved current sink or inserting the conserved current
* sequentially.
******************************************************************************/
// Helper macro to reverse Simd vector. Fixme: slow, generic implementation.
#define REVERSE_LS(qSite, qSiteRev, Nsimd) \
{ \
std::vector<typename SitePropagator::scalar_object> qSiteVec(Nsimd); \
extract(qSite, qSiteVec); \
for (int i = 0; i < Nsimd / 2; ++i) \
{ \
typename SitePropagator::scalar_object tmp = qSiteVec[i]; \
qSiteVec[i] = qSiteVec[Nsimd - i - 1]; \
qSiteVec[Nsimd - i - 1] = tmp; \
} \
merge(qSiteRev, qSiteVec); \
}
template <class Impl>
void WilsonFermion5D<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu)
{
conformable(q_in_1._grid, FermionGrid());
conformable(q_in_1._grid, q_in_2._grid);
conformable(_FourDimGrid, q_out._grid);
PropagatorField tmp1(FermionGrid()), tmp2(FermionGrid());
unsigned int LLs = q_in_1._grid->_rdimensions[0];
q_out = zero;
// Forward, need q1(x + mu, s), q2(x, Ls - 1 - s). Backward, need q1(x, s),
// q2(x + mu, Ls - 1 - s). 5D lattice so shift 4D coordinate mu by one.
tmp1 = Cshift(q_in_1, mu + 1, 1);
tmp2 = Cshift(q_in_2, mu + 1, 1);
parallel_for (unsigned int sU = 0; sU < Umu._grid->oSites(); ++sU)
{
unsigned int sF1 = sU * LLs;
unsigned int sF2 = (sU + 1) * LLs - 1;
for (unsigned int s = 0; s < LLs; ++s)
{
bool axial_sign = ((curr_type == Current::Axial) && \
(s < (LLs / 2)));
SitePropagator qSite2, qmuSite2;
// If vectorised in 5th dimension, reverse q2 vector to match up
// sites correctly.
if (Impl::LsVectorised)
{
REVERSE_LS(q_in_2._odata[sF2], qSite2, Ls / LLs);
REVERSE_LS(tmp2._odata[sF2], qmuSite2, Ls / LLs);
}
else
{
qSite2 = q_in_2._odata[sF2];
qmuSite2 = tmp2._odata[sF2];
}
Kernels::ContractConservedCurrentSiteFwd(tmp1._odata[sF1],
qSite2,
q_out._odata[sU],
Umu, sU, mu, axial_sign);
Kernels::ContractConservedCurrentSiteBwd(q_in_1._odata[sF1],
qmuSite2,
q_out._odata[sU],
Umu, sU, mu, axial_sign);
sF1++;
sF2--;
}
}
}
template <class Impl>
void WilsonFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax)
{
conformable(q_in._grid, FermionGrid());
conformable(q_in._grid, q_out._grid);
Lattice<iSinglet<Simd>> ph(FermionGrid()), coor(FermionGrid());
PropagatorField tmpFwd(FermionGrid()), tmpBwd(FermionGrid()),
tmp(FermionGrid());
Complex i(0.0, 1.0);
unsigned int tshift = (mu == Tp) ? 1 : 0;
unsigned int LLs = q_in._grid->_rdimensions[0];
unsigned int LLt = GridDefaultLatt()[Tp];
// Momentum projection.
ph = zero;
for(unsigned int nu = 0; nu < Nd - 1; nu++)
{
// Shift coordinate lattice index by 1 to account for 5th dimension.
LatticeCoordinate(coor, nu + 1);
ph = ph + mom[nu]*coor*((1./(_FourDimGrid->_fdimensions[nu])));
}
ph = exp((Real)(2*M_PI)*i*ph);
q_out = zero;
LatticeInteger coords(_FourDimGrid);
LatticeCoordinate(coords, Tp);
// Need q(x + mu, s) and q(x - mu, s). 5D lattice so shift 4D coordinate mu
// by one.
tmp = Cshift(q_in, mu + 1, 1);
tmpFwd = tmp*ph;
tmp = ph*q_in;
tmpBwd = Cshift(tmp, mu + 1, -1);
parallel_for (unsigned int sU = 0; sU < Umu._grid->oSites(); ++sU)
{
// Compute the sequential conserved current insertion only if our simd
// object contains a timeslice we need.
vInteger t_mask = ((coords._odata[sU] >= tmin) &&
(coords._odata[sU] <= tmax));
Integer timeSlices = Reduce(t_mask);
if (timeSlices > 0)
{
unsigned int sF = sU * LLs;
for (unsigned int s = 0; s < LLs; ++s)
{
bool axial_sign = ((curr_type == Current::Axial) && (s < (LLs / 2)));
Kernels::SeqConservedCurrentSiteFwd(tmpFwd._odata[sF],
q_out._odata[sF], Umu, sU,
mu, t_mask, axial_sign);
++sF;
}
}
// Repeat for backward direction.
t_mask = ((coords._odata[sU] >= (tmin + tshift)) &&
(coords._odata[sU] <= (tmax + tshift)));
//if tmax = LLt-1 (last timeslice) include timeslice 0 if the time is shifted (mu=3)
unsigned int t0 = 0;
if((tmax==LLt-1) && (tshift==1)) t_mask = (t_mask || (coords._odata[sU] == t0 ));
timeSlices = Reduce(t_mask);
if (timeSlices > 0)
{
unsigned int sF = sU * LLs;
for (unsigned int s = 0; s < LLs; ++s)
{
bool axial_sign = ((curr_type == Current::Axial) && (s < (LLs / 2)));
Kernels::SeqConservedCurrentSiteBwd(tmpBwd._odata[sF],
q_out._odata[sF], Umu, sU,
mu, t_mask, axial_sign);
++sF;
}
}
}
}
FermOpTemplateInstantiate(WilsonFermion5D);
GparityFermOpTemplateInstantiate(WilsonFermion5D);

15
lib/qcd/action/fermion/WilsonFermion5D.h
View File

@ -214,6 +214,21 @@ namespace QCD {
// Comms buffer
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > comm_buf;
///////////////////////////////////////////////////////////////
// Conserved current utilities
///////////////////////////////////////////////////////////////
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
std::vector<Real> mom,
unsigned int tmin,
unsigned int tmax);
};
}}

166
lib/qcd/action/fermion/WilsonKernels.cc
View File

@ -281,6 +281,172 @@ void WilsonKernels<Impl>::DhopDir( StencilImpl &st, DoubledGaugeField &U,SiteHal
vstream(out._odata[sF], result);
}
/*******************************************************************************
* Conserved current utilities for Wilson fermions, for contracting propagators
* to make a conserved current sink or inserting the conserved current
* sequentially. Common to both 4D and 5D.
******************************************************************************/
// N.B. Functions below assume a -1/2 factor within U.
#define WilsonCurrentFwd(expr, mu) ((expr - Gamma::gmu[mu]*expr))
#define WilsonCurrentBwd(expr, mu) ((expr + Gamma::gmu[mu]*expr))
/*******************************************************************************
* Name: ContractConservedCurrentSiteFwd
* Operation: (1/2) * q2[x] * U(x) * (g[mu] - 1) * q1[x + mu]
* Notes: - DoubledGaugeField U assumed to contain -1/2 factor.
* - Pass in q_in_1 shifted in +ve mu direction.
******************************************************************************/
template<class Impl>
void WilsonKernels<Impl>::ContractConservedCurrentSiteFwd(
const SitePropagator &q_in_1,
const SitePropagator &q_in_2,
SitePropagator &q_out,
DoubledGaugeField &U,
unsigned int sU,
unsigned int mu,
bool switch_sign)
{
SitePropagator result, tmp;
Gamma g5(Gamma::Algebra::Gamma5);
Impl::multLinkProp(tmp, U._odata[sU], q_in_1, mu);
result = g5 * adj(q_in_2) * g5 * WilsonCurrentFwd(tmp, mu);
if (switch_sign)
{
q_out -= result;
}
else
{
q_out += result;
}
}
/*******************************************************************************
* Name: ContractConservedCurrentSiteBwd
* Operation: (1/2) * q2[x + mu] * U^dag(x) * (g[mu] + 1) * q1[x]
* Notes: - DoubledGaugeField U assumed to contain -1/2 factor.
* - Pass in q_in_2 shifted in +ve mu direction.
******************************************************************************/
template<class Impl>
void WilsonKernels<Impl>::ContractConservedCurrentSiteBwd(
const SitePropagator &q_in_1,
const SitePropagator &q_in_2,
SitePropagator &q_out,
DoubledGaugeField &U,
unsigned int sU,
unsigned int mu,
bool switch_sign)
{
SitePropagator result, tmp;
Gamma g5(Gamma::Algebra::Gamma5);
Impl::multLinkProp(tmp, U._odata[sU], q_in_1, mu + Nd);
result = g5 * adj(q_in_2) * g5 * WilsonCurrentBwd(tmp, mu);
if (switch_sign)
{
q_out += result;
}
else
{
q_out -= result;
}
}
// G-parity requires more specialised implementation.
#define NO_CURR_SITE(Impl) \
template <> \
void WilsonKernels<Impl>::ContractConservedCurrentSiteFwd( \
const SitePropagator &q_in_1, \
const SitePropagator &q_in_2, \
SitePropagator &q_out, \
DoubledGaugeField &U, \
unsigned int sU, \
unsigned int mu, \
bool switch_sign) \
{ \
assert(0); \
} \
template <> \
void WilsonKernels<Impl>::ContractConservedCurrentSiteBwd( \
const SitePropagator &q_in_1, \
const SitePropagator &q_in_2, \
SitePropagator &q_out, \
DoubledGaugeField &U, \
unsigned int mu, \
unsigned int sU, \
bool switch_sign) \
{ \
assert(0); \
}
NO_CURR_SITE(GparityWilsonImplF);
NO_CURR_SITE(GparityWilsonImplD);
NO_CURR_SITE(GparityWilsonImplFH);
NO_CURR_SITE(GparityWilsonImplDF);
/*******************************************************************************
* Name: SeqConservedCurrentSiteFwd
* Operation: (1/2) * U(x) * (g[mu] - 1) * q[x + mu]
* Notes: - DoubledGaugeField U assumed to contain -1/2 factor.
* - Pass in q_in shifted in +ve mu direction.
******************************************************************************/
template<class Impl>
void WilsonKernels<Impl>::SeqConservedCurrentSiteFwd(const SitePropagator &q_in,
SitePropagator &q_out,
DoubledGaugeField &U,
unsigned int sU,
unsigned int mu,
vInteger t_mask,
bool switch_sign)
{
SitePropagator result;
Impl::multLinkProp(result, U._odata[sU], q_in, mu);
result = WilsonCurrentFwd(result, mu);
// Zero any unwanted timeslice entries.
result = predicatedWhere(t_mask, result, 0.*result);
if (switch_sign)
{
q_out -= result;
}
else
{
q_out += result;
}
}
/*******************************************************************************
* Name: SeqConservedCurrentSiteFwd
* Operation: (1/2) * U^dag(x) * (g[mu] + 1) * q[x - mu]
* Notes: - DoubledGaugeField U assumed to contain -1/2 factor.
* - Pass in q_in shifted in -ve mu direction.
******************************************************************************/
template<class Impl>
void WilsonKernels<Impl>::SeqConservedCurrentSiteBwd(const SitePropagator &q_in,
SitePropagator &q_out,
DoubledGaugeField &U,
unsigned int sU,
unsigned int mu,
vInteger t_mask,
bool switch_sign)
{
SitePropagator result;
Impl::multLinkProp(result, U._odata[sU], q_in, mu + Nd);
result = WilsonCurrentBwd(result, mu);
// Zero any unwanted timeslice entries.
result = predicatedWhere(t_mask, result, 0.*result);
if (switch_sign)
{
q_out += result;
}
else
{
q_out -= result;
}
}
FermOpTemplateInstantiate(WilsonKernels);
AdjointFermOpTemplateInstantiate(WilsonKernels);
TwoIndexFermOpTemplateInstantiate(WilsonKernels);

32
lib/qcd/action/fermion/WilsonKernels.h
View File

@ -180,6 +180,38 @@ public:
void DhopDir(StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor * buf,
int sF, int sU, const FermionField &in, FermionField &out, int dirdisp, int gamma);
//////////////////////////////////////////////////////////////////////////////
// Utilities for inserting Wilson conserved current.
//////////////////////////////////////////////////////////////////////////////
void ContractConservedCurrentSiteFwd(const SitePropagator &q_in_1,
const SitePropagator &q_in_2,
SitePropagator &q_out,
DoubledGaugeField &U,
unsigned int sU,
unsigned int mu,
bool switch_sign = false);
void ContractConservedCurrentSiteBwd(const SitePropagator &q_in_1,
const SitePropagator &q_in_2,
SitePropagator &q_out,
DoubledGaugeField &U,
unsigned int sU,
unsigned int mu,
bool switch_sign = false);
void SeqConservedCurrentSiteFwd(const SitePropagator &q_in,
SitePropagator &q_out,
DoubledGaugeField &U,
unsigned int sU,
unsigned int mu,
vInteger t_mask,
bool switch_sign = false);
void SeqConservedCurrentSiteBwd(const SitePropagator &q_in,
SitePropagator &q_out,
DoubledGaugeField &U,
unsigned int sU,
unsigned int mu,
vInteger t_mask,
bool switch_sign = false);
private:
// Specialised variants
void GenericDhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteHalfSpinor * buf,

6
lib/qcd/utils/SUn.h
View File

@ -746,7 +746,7 @@ template<typename GaugeField,typename GaugeMat>
}
}
template<typename GaugeField>
static void ColdConfiguration(GridParallelRNG &pRNG,GaugeField &out){
static void ColdConfiguration(GaugeField &out){
typedef typename GaugeField::vector_type vector_type;
typedef iSUnMatrix<vector_type> vMatrixType;
typedef Lattice<vMatrixType> LatticeMatrixType;
@ -757,6 +757,10 @@ template<typename GaugeField,typename GaugeMat>
PokeIndex<LorentzIndex>(out,Umu,mu);
}
}
template<typename GaugeField>
static void ColdConfiguration(GridParallelRNG &pRNG,GaugeField &out){
ColdConfiguration(out);
}
template<typename LatticeMatrixType>
static void taProj( const LatticeMatrixType &in, LatticeMatrixType &out){

2
lib/serialisation/JSON_IO.cc
View File

@ -25,7 +25,7 @@
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid.h>
#include <Grid/Grid.h>
using namespace Grid;
using namespace std;

6
lib/serialisation/MacroMagic.h
View File

@ -125,7 +125,11 @@ static inline void write(Writer<T> &WR,const std::string &s, const cname &obj){
}\
template <typename T>\
static inline void read(Reader<T> &RD,const std::string &s, cname &obj){ \
push(RD,s);\
if (!push(RD,s))\
{\
std::cout << Grid::GridLogWarning << "IO: Cannot open node '" << s << "'" << std::endl;\
return;\
};\
GRID_MACRO_EVAL(GRID_MACRO_MAP(GRID_MACRO_READ_MEMBER,__VA_ARGS__)) \
pop(RD);\
}\

41
lib/serialisation/XmlIO.cc
View File

@ -100,13 +100,16 @@ XmlReader::XmlReader(const string &fileName,string toplev) : fileName_(fileName)
bool XmlReader::push(const string &s)
{
if (node_.child(s.c_str()))
{
node_ = node_.child(s.c_str());
if (node_.child(s.c_str()) == NULL )
return true;
}
else
{
return false;
node_ = node_.child(s.c_str());
return true;
}
}
void XmlReader::pop(void)
@ -117,20 +120,30 @@ void XmlReader::pop(void)
bool XmlReader::nextElement(const std::string &s)
{
if (node_.next_sibling(s.c_str()))
{
node_ = node_.next_sibling(s.c_str());
return true;
}
{
node_ = node_.next_sibling(s.c_str());
return true;
}
else
{
return false;
}
{
return false;
}
}
template <>
void XmlReader::readDefault(const string &s, string &output)
{
output = node_.child(s.c_str()).first_child().value();
if (node_.child(s.c_str()))
{
output = node_.child(s.c_str()).first_child().value();
}
else
{
std::cout << GridLogWarning << "XML: cannot open node '" << s << "'";
std::cout << std::endl;
output = "";
}
}

10
lib/serialisation/XmlIO.h
View File

@ -39,6 +39,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <cassert>
#include <Grid/pugixml/pugixml.h>
#include <Grid/GridCore.h>
namespace Grid
{
@ -119,7 +120,6 @@ namespace Grid
std::string buf;
readDefault(s, buf);
// std::cout << s << " " << buf << std::endl;
fromString(output, buf);
}
@ -132,7 +132,13 @@ namespace Grid
std::string buf;
unsigned int i = 0;
push(s);
if (!push(s))
{
std::cout << GridLogWarning << "XML: cannot open node '" << s << "'";
std::cout << std::endl;
return;
}
while (node_.child("elem"))
{
output.resize(i + 1);

8
lib/util/Init.cc
View File

@ -204,7 +204,7 @@ std::string GridCmdVectorIntToString(const std::vector<int> & vec){
// Reinit guard
/////////////////////////////////////////////////////////
static int Grid_is_initialised = 0;
static MemoryStats dbgMemStats;
void Grid_init(int *argc,char ***argv)
{
@ -251,6 +251,11 @@ void Grid_init(int *argc,char ***argv)
assert(fp!=(FILE *)NULL);
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--debug-mem") ){
MemoryProfiler::debug = true;
MemoryProfiler::stats = &dbgMemStats;
}
////////////////////////////////////
// Banner
////////////////////////////////////
@ -324,6 +329,7 @@ void Grid_init(int *argc,char ***argv)
std::cout<<GridLogMessage<<" --decomposition : report on default omp,mpi and simd decomposition"<<std::endl;
std::cout<<GridLogMessage<<" --debug-signals : catch sigsegv and print a blame report"<<std::endl;
std::cout<<GridLogMessage<<" --debug-stdout : print stdout from EVERY node"<<std::endl;
std::cout<<GridLogMessage<<" --debug-mem : print Grid allocator activity"<<std::endl;
std::cout<<GridLogMessage<<" --notimestamp : suppress millisecond resolution stamps"<<std::endl;
std::cout<<GridLogMessage<<std::endl;
std::cout<<GridLogMessage<<"Performance:"<<std::endl;