portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-09-20 01:05:38 +01:00
Code Releases Activity

Merge branch 'develop' of https://github.com/paboyle/Grid into develop

Browse Source
This commit is contained in:
paboyle 2018-04-25 00:14:03 +01:00
commit 2baf193031
7 changed files with 188 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

52
benchmarks/Benchmark_su3.cc
View File

@ -53,7 +53,7 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage << " L "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
for(int lat=2;lat<=LMAX;lat+=LINC){
for(int lat=4;lat<=LMAX;lat+=LINC){
std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
int64_t vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
@ -148,31 +148,32 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
for(int lat=2;lat<=LMAX;lat+=LINC){
std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
int64_t vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
GridParallelRNG pRNG(&Grid); pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
LatticeColourMatrix z(&Grid); random(pRNG,z);
LatticeColourMatrix x(&Grid); random(pRNG,x);
LatticeColourMatrix y(&Grid); random(pRNG,y);
double start=usecond();
for(int64_t i=0;i<Nloop;i++){
mac(z,x,y);
}
double stop=usecond();
double time = (stop-start)/Nloop*1000.0;
double bytes=3*vol*Nc*Nc*sizeof(Complex);
double flops=Nc*Nc*(8+8+8)*vol;
std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
int64_t vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
GridParallelRNG pRNG(&Grid); pRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
LatticeColourMatrix z(&Grid); random(pRNG,z);
LatticeColourMatrix x(&Grid); random(pRNG,x);
LatticeColourMatrix y(&Grid); random(pRNG,y);
double start=usecond();
for(int64_t i=0;i<Nloop;i++){
mac(z,x,y);
}
double stop=usecond();
double time = (stop-start)/Nloop*1000.0;
double bytes=3*vol*Nc*Nc*sizeof(Complex);
double flops=Nc*Nc*(6+8+8)*vol;
std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
}
#endif
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking SU3xSU3 CovShiftForward(z,x,y)"<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
@ -199,8 +200,9 @@ int main (int argc, char ** argv)
double stop=usecond();
double time = (stop-start)/Nloop*1000.0;
double bytes=3*vol*Nc*Nc*sizeof(Complex);
double flops=Nc*Nc*(8+8+8)*vol;
double flops=Nc*Nc*(6+8+8)*vol;
std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
}
}

13
lib/cshift/Cshift_mpi.h
View File

@ -54,13 +54,13 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
if ( !comm_dim ) {
// std::cout << "Cshift_local" <<std::endl;
//std::cout << "CSHIFT: Cshift_local" <<std::endl;
Cshift_local(ret,rhs,dimension,shift); // Handles checkerboarding
} else if ( splice_dim ) {
// std::cout << "Cshift_comms_simd" <<std::endl;
//std::cout << "CSHIFT: Cshift_comms_simd call - splice_dim = " << splice_dim << " shift " << shift << " dimension = " << dimension << std::endl;
Cshift_comms_simd(ret,rhs,dimension,shift);
} else {
// std::cout << "Cshift_comms" <<std::endl;
//std::cout << "CSHIFT: Cshift_comms" <<std::endl;
Cshift_comms(ret,rhs,dimension,shift);
}
return ret;
@ -91,9 +91,12 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
//std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
if ( sshift[0] == sshift[1] ) {
//std::cout << "Single pass Cshift_comms" <<std::endl;
Cshift_comms_simd(ret,rhs,dimension,shift,0x3);
} else {
//std::cout << "Two pass Cshift_comms" <<std::endl;
Cshift_comms_simd(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
}
@ -175,6 +178,10 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
int simd_layout = grid->_simd_layout[dimension];
int comm_dim = grid->_processors[dimension] >1 ;
//std::cout << "Cshift_comms_simd dim "<< dimension << " fd "<<fd<<" rd "<<rd
// << " ld "<<ld<<" pd " << pd<<" simd_layout "<<simd_layout
// << " comm_dim " << comm_dim << " cbmask " << cbmask <<std::endl;
assert(comm_dim==1);
assert(simd_layout==2);
assert(shift>=0);

93
lib/lattice/Lattice_transfer.h
View File

@ -599,6 +599,51 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
extract1(in_vobj, out_ptrs, 0);
}
}
template<typename vobj, typename sobj>
typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type
unvectorizeToRevLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
{
typedef typename vobj::vector_type vtype;
GridBase* in_grid = in._grid;
out.resize(in_grid->lSites());
int ndim = in_grid->Nd();
int in_nsimd = vtype::Nsimd();
std::vector<std::vector<int> > in_icoor(in_nsimd);
for(int lane=0; lane < in_nsimd; lane++){
in_icoor[lane].resize(ndim);
in_grid->iCoorFromIindex(in_icoor[lane], lane);
}
parallel_for(int in_oidx = 0; in_oidx < in_grid->oSites(); in_oidx++){ //loop over outer index
//Assemble vector of pointers to output elements
std::vector<sobj*> out_ptrs(in_nsimd);
std::vector<int> in_ocoor(ndim);
in_grid->oCoorFromOindex(in_ocoor, in_oidx);
std::vector<int> lcoor(in_grid->Nd());
for(int lane=0; lane < in_nsimd; lane++){
for(int mu=0;mu<ndim;mu++)
lcoor[mu] = in_ocoor[mu] + in_grid->_rdimensions[mu]*in_icoor[lane][mu];
int lex;
Lexicographic::IndexFromCoorReversed(lcoor, lex, in_grid->_ldimensions);
out_ptrs[lane] = &out[lex];
}
//Unpack into those ptrs
const vobj & in_vobj = in._odata[in_oidx];
extract1(in_vobj, out_ptrs, 0);
}
}
//Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
template<typename vobj, typename sobj>
typename std::enable_if<isSIMDvectorized<vobj>::value
@ -648,6 +693,54 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
}
}
template<typename vobj, typename sobj>
typename std::enable_if<isSIMDvectorized<vobj>::value
&& !isSIMDvectorized<sobj>::value, void>::type
vectorizeFromRevLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
{
typedef typename vobj::vector_type vtype;
GridBase* grid = out._grid;
assert(in.size()==grid->lSites());
int ndim = grid->Nd();
int nsimd = vtype::Nsimd();
std::vector<std::vector<int> > icoor(nsimd);
for(int lane=0; lane < nsimd; lane++){
icoor[lane].resize(ndim);
grid->iCoorFromIindex(icoor[lane],lane);
}
parallel_for(uint64_t oidx = 0; oidx < grid->oSites(); oidx++){ //loop over outer index
//Assemble vector of pointers to output elements
std::vector<sobj*> ptrs(nsimd);
std::vector<int> ocoor(ndim);
grid->oCoorFromOindex(ocoor, oidx);
std::vector<int> lcoor(grid->Nd());
for(int lane=0; lane < nsimd; lane++){
for(int mu=0;mu<ndim;mu++){
lcoor[mu] = ocoor[mu] + grid->_rdimensions[mu]*icoor[lane][mu];
}
int lex;
Lexicographic::IndexFromCoorReversed(lcoor, lex, grid->_ldimensions);
ptrs[lane] = &in[lex];
}
//pack from those ptrs
vobj vecobj;
merge1(vecobj, ptrs, 0);
out._odata[oidx] = vecobj;
}
}
//Convert a Lattice from one precision to another
template<class VobjOut, class VobjIn>
void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){

8
lib/qcd/smearing/WilsonFlow.h
View File

@ -173,8 +173,8 @@ void WilsonFlow<Gimpl>::smear(GaugeField& out, const GaugeField& in) const {
std::cout << "Time to evolve " << diff.count() << " s\n";
#endif
std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "
<< step << " "
<< energyDensityPlaquette(step,out) << std::endl;
<< step << " " << tau(step) << " "
<< energyDensityPlaquette(step,out) << std::endl;
if( step % measure_interval == 0){
std::cout << GridLogMessage << "[WilsonFlow] Top. charge : "
<< step << " "
@ -193,8 +193,8 @@ void WilsonFlow<Gimpl>::smear_adaptive(GaugeField& out, const GaugeField& in, Re
//std::cout << GridLogMessage << "Evolution time :"<< taus << std::endl;
evolve_step_adaptive(out, maxTau);
std::cout << GridLogMessage << "[WilsonFlow] Energy density (plaq) : "
<< step << " "
<< energyDensityPlaquette(out) << std::endl;
<< step << " " << taus << " "
<< energyDensityPlaquette(out) << std::endl;
if( step % measure_interval == 0){
std::cout << GridLogMessage << "[WilsonFlow] Top. charge : "
<< step << " "

10
lib/serialisation/Hdf5IO.cc
View File

@ -55,6 +55,11 @@ void Hdf5Writer::writeDefault(const std::string &s, const char *x)
writeDefault(s, sx);
}
Group & Hdf5Writer::getGroup(void)
{
return group_;
}
// Reader implementation ///////////////////////////////////////////////////////
Hdf5Reader::Hdf5Reader(const std::string &fileName)
: fileName_(fileName)
@ -103,3 +108,8 @@ void Hdf5Reader::readDefault(const std::string &s, std::string &x)
x.resize(strType.getSize());
attribute.read(strType, &(x[0]));
}
Group & Hdf5Reader::getGroup(void)
{
return group_;
}

2
lib/serialisation/Hdf5IO.h
View File

@ -38,6 +38,7 @@ namespace Grid
template <typename U>
typename std::enable_if<!element<std::vector<U>>::is_number, void>::type
writeDefault(const std::string &s, const std::vector<U> &x);
H5NS::Group & getGroup(void);
private:
template <typename U>
void writeSingleAttribute(const U &x, const std::string &name,
@ -65,6 +66,7 @@ namespace Grid
template <typename U>
typename std::enable_if<!element<std::vector<U>>::is_number, void>::type
readDefault(const std::string &s, std::vector<U> &x);
H5NS::Group & getGroup(void);
private:
template <typename U>
void readSingleAttribute(U &x, const std::string &name,

42
lib/serialisation/VectorUtils.h
View File

@ -30,6 +30,48 @@ namespace Grid {
typedef typename std::vector<std::vector<typename TensorToVec<T>::type>> type;
};
template <typename T>
void tensorDim(std::vector<size_t> &dim, const T &t, const bool wipe = true)
{
if (wipe)
{
dim.clear();
}
}
template <typename T>
void tensorDim(std::vector<size_t> &dim, const iScalar<T> &t, const bool wipe = true)
{
if (wipe)
{
dim.clear();
}
tensorDim(dim, t._internal, false);
}
template <typename T, int N>
void tensorDim(std::vector<size_t> &dim, const iVector<T, N> &t, const bool wipe = true)
{
if (wipe)
{
dim.clear();
}
dim.push_back(N);
tensorDim(dim, t._internal[0], false);
}
template <typename T, int N>
void tensorDim(std::vector<size_t> &dim, const iMatrix<T, N> &t, const bool wipe = true)
{
if (wipe)
{
dim.clear();
}
dim.push_back(N);
dim.push_back(N);
tensorDim(dim, t._internal[0][0], false);
}
template <typename T>
typename TensorToVec<T>::type tensorToVec(const T &t)
{