portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Merge GPU support (upstream/develop) into distillation branch.

Browse Source 
This compiles and looks right ... but may need some testing

* develop: (762 commits)
  Tensor ambiguous fix
  Fix for GCC preprocessor/pragma handling bug
  Trips up NVCC for reasons I dont understand on summit
  Fix GCC complaint
  Zero() change
  Force a couple of things to compile on NVCC
  Remove debug code
  nvcc error suppress
  Merge develop
  Reduction finished and hopefully fixes CI regression fail on single precisoin and force
  Double precision variants for summation accuracy
  Update todo list
  Freeze the seed
  Fix compiling of MSource::Gauss for single precision
  Think the reduction is now sorted and cleaned up
  Fix force term
  Printing improvement
  GPU reduction fix and also exit backtrace option
  GPU friendly
  Simplify the comms benchmark
  ...

# Conflicts:
#	Grid/communicator/SharedMemoryMPI.cc
#	Grid/qcd/action/fermion/WilsonKernelsAsm.cc
#	Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h
#	Grid/qcd/smearing/StoutSmearing.h
#	Hadrons/Modules.hpp
#	Hadrons/Utilities/Contractor.cc
#	Hadrons/modules.inc
#	tests/forces/Test_dwf_force_eofa.cc
#	tests/forces/Test_dwf_gpforce_eofa.cc
This commit is contained in:
Michael Marshall
2019-09-13 13:30:00 +01:00
parent 04a661cafe b473405652
commit 61d017d0a5
796 changed files with 41536 additions and 52391 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
Hadrons/Modules/MDistil/BContraction.hpp
View File

@ -76,6 +76,7 @@ template <typename FImpl>
class TBContraction: public Module<BContractionPar>
{
public:
using W = WilsonImplR; // Debug so I can see type info for default FImpl
FERM_TYPE_ALIASES(FImpl,);
public:
// constructor
@ -174,15 +175,25 @@ void TBContraction<FImpl>::execute(void)
int Bindex;
int Nc=3; //Num colours
FermionField tmp1(grid3d);
FermionField tmp2(grid3d);
FermionField tmp3(grid3d);
FermionField ftmp1(grid3d);
FermionField ftmp2(grid3d);
FermionField ftmp3(grid3d);
LatticeView<typename FImpl::SiteSpinor> tmp1{ ftmp1 };
LatticeView<typename FImpl::SiteSpinor> tmp2{ ftmp2 };
LatticeView<typename FImpl::SiteSpinor> tmp3{ ftmp3 };
//std::complex<double> * tmp33 = reinterpret_cast<std::complex<double> *>(&(tmp3[0]()(0)(0)));
#ifdef THIS_IS_NAUGHTY_TODO_FIXME
// The reinterpret_cast gets rid of SIMD attributes
// Plus other badness - e.g. we perhaps shouldn't explicitly say SpinColourVector
// ... but rather use some correct FIMPL types
// REVIEW WITH PETER
#endif
SpinColourVector * tmp11 = reinterpret_cast<SpinColourVector *>(&(tmp1[0]()(0)(0)));
SpinColourVector * tmp22 = reinterpret_cast<SpinColourVector *>(&(tmp2[0]()(0)(0)));
SpinColourVector * tmp33 = reinterpret_cast<SpinColourVector *>(&(tmp3[0]()(0)(0)));
SpinVector tmp11s;
SpinVector tmp22s;
SpinVector tmp33s;
@ -225,10 +236,10 @@ void TBContraction<FImpl>::execute(void)
for (int imom=0 ; imom < Nmom ; imom++){
for (int t=0 ; t < Nt ; t++){
Bindex = i1 + N_1*(i2 + N_2*(i3 + N_3*(imom+Nmom*t)));
ExtractSliceLocal(tmp1,one[i1],0,t,3);
ExtractSliceLocal(tmp2,two[i2],0,t,3);
ExtractSliceLocal(tmp3,three[i3],0,t,3);
parallel_for (unsigned int sU = 0; sU < grid3d->oSites(); ++sU)
ExtractSliceLocal(ftmp1,one[i1],0,t,3);
ExtractSliceLocal(ftmp2,two[i2],0,t,3);
ExtractSliceLocal(ftmp3,three[i3],0,t,3);
accelerator_for(sU, grid3d->oSites(), grid3d->Nsimd(),
{
for (int ie=0 ; ie < 6 ; ie++){
// Why does peekColour not work????
@ -248,7 +259,7 @@ void TBContraction<FImpl>::execute(void)
}
}
}
}
} );
}
}
}

30
Hadrons/Modules/MDistil/DistilVectors.hpp
View File

@ -187,10 +187,10 @@ void TDistilVectors<FImpl>::setup(void)
envCreate(std::vector<FermionField>, SinkName, 1, nnoise*LI*SI*Nt_inv, envGetGrid(FermionField));
grid4d = env().getGrid();
std::vector<int> latt_size = GridDefaultLatt();
std::vector<int> simd_layout = GridDefaultSimd(Nd, vComplex::Nsimd());
std::vector<int> mpi_layout = GridDefaultMpi();
std::vector<int> simd_layout_3 = GridDefaultSimd(Nd-1, vComplex::Nsimd());
Coordinate latt_size = GridDefaultLatt();
Coordinate simd_layout = GridDefaultSimd(Nd, vComplex::Nsimd());
Coordinate mpi_layout = GridDefaultMpi();
Coordinate simd_layout_3 = GridDefaultSimd(Nd-1, vComplex::Nsimd());
latt_size[Nd-1] = 1;
simd_layout_3.push_back( 1 );
mpi_layout[Nd-1] = 1;
@ -233,7 +233,7 @@ void TDistilVectors<FImpl>::execute(void)
const int Ntlocal{ grid4d->LocalDimensions()[3] };
const int Ntfirst{ grid4d->LocalStarts()[3] };
const int Ns{ Grid::QCD::Ns };
const int Ns{ Ns };
const int Nt{ env().getDim(Tdir) };
const int TI{ Hadrons::MDistil::DistilParameters::ParameterDefault( par().TI, Nt, false ) };
const int LI{ static_cast<int>( perambulator.tensor.dimension(2) ) };
@ -254,20 +254,20 @@ void TDistilVectors<FImpl>::execute(void)
for( int dt = 0; dt < Nt_inv; dt++ ) {
for( int ds = 0; ds < SI; ds++ ) {
vecindex = inoise + nnoise * dk + nnoise * LI * ds + nnoise *LI * SI*dt;
rho[vecindex] = zero;
tmp3d_nospin = zero;
rho[vecindex] = 0;
tmp3d_nospin = 0;
for (int it = dt; it < Nt; it += TI){
if (full_tdil) t_inv = tsrc; else t_inv = it;
if( t_inv >= Ntfirst && t_inv < Ntfirst + Ntlocal ) {
for (int ik = dk; ik < nvec; ik += LI){
for (int is = ds; is < Ns; is += SI){
ExtractSliceLocal(evec3d,epack.evec[ik],0,t_inv-Ntfirst,Grid::QCD::Tdir);
ExtractSliceLocal(evec3d,epack.evec[ik],0,t_inv-Ntfirst,Tdir);
//tmp3d_nospin = evec3d * noise[inoise + nnoise*(t_inv + Nt*(ik+nvec*is))];
tmp3d_nospin = evec3d * noise(inoise, t_inv, ik, is);
tmp3d=zero;
tmp3d=0;
pokeSpin(tmp3d,tmp3d_nospin,is);
tmp2=zero;
InsertSliceLocal(tmp3d,tmp2,0,t_inv-Ntfirst,Grid::QCD::Tdir);
tmp2=0;
InsertSliceLocal(tmp3d,tmp2,0,t_inv-Ntfirst,Tdir);
rho[vecindex] += tmp2;
}
}
@ -285,14 +285,14 @@ void TDistilVectors<FImpl>::execute(void)
for( int dt = 0; dt < Nt_inv; dt++ ) {
for( int ds = 0; ds < SI; ds++ ) {
vecindex = inoise + nnoise * dk + nnoise * LI * ds + nnoise *LI * SI*dt;
phi[vecindex] = zero;
phi[vecindex] = 0;
for (int t = Ntfirst; t < Ntfirst + Ntlocal; t++) {
sink_tslice=zero;
sink_tslice=0;
for (int ivec = 0; ivec < nvec; ivec++) {
ExtractSliceLocal(evec3d,epack.evec[ivec],0,t-Ntfirst,Grid::QCD::Tdir);
ExtractSliceLocal(evec3d,epack.evec[ivec],0,t-Ntfirst,Tdir);
sink_tslice += evec3d * perambulator(t, ivec, dk, inoise,dt,ds);
}
InsertSliceLocal(sink_tslice,phi[vecindex],0,t-Ntfirst,Grid::QCD::Tdir);
InsertSliceLocal(sink_tslice,phi[vecindex],0,t-Ntfirst,Tdir);
}
}
}

4
Hadrons/Modules/MDistil/LapEvec.hpp
View File

@ -236,7 +236,7 @@ void TLapEvec<GImpl>::execute(void)
eig[t].resize(LPar.Nk+LPar.Np,gridLD);
// Construct smearing operator
ExtractSliceLocal(UmuNoTime,Umu_smear,0,t,Grid::QCD::Tdir); // switch to 3d/4d objects
ExtractSliceLocal(UmuNoTime,Umu_smear,0,t,Tdir); // switch to 3d/4d objects
LinOpPeardonNabla<LatticeColourVector> PeardonNabla(UmuNoTime);
LOG(Debug) << "Chebyshev preconditioning to order " << ChebPar.PolyOrder
<< " with parameters (alpha,beta) = (" << ChebPar.alpha << "," << ChebPar.beta << ")" << std::endl;
@ -263,7 +263,7 @@ void TLapEvec<GImpl>::execute(void)
RotateEigen( eig[t].evec ); // Rotate the eigenvectors into our phase convention
for (int i=0;i<LPar.Nvec;i++){
InsertSliceLocal(eig[t].evec[i],eig4d.evec[i],0,t,Grid::QCD::Tdir);
InsertSliceLocal(eig[t].evec[i],eig4d.evec[i],0,t,Tdir);
if(t==0 && Ntfirst==0)
eig4d.eval[i] = eig[t].eval[i]; // TODO: Discuss: is this needed? Is there a better way?
}

4
Hadrons/Modules/MDistil/Noises.hpp
View File

@ -110,7 +110,7 @@ template <typename FImpl>
void TNoises<FImpl>::setup(void)
{
const int Nt{env().getDim(Tdir)};
const int Ns{Grid::QCD::Ns};
//const int Ns{Grid::Ns};
const int nnoise{par().nnoise};
const int nvec{par().nvec};
const int TI{ Hadrons::MDistil::DistilParameters::ParameterDefault( par().TI, Nt, true) };
@ -123,7 +123,7 @@ template <typename FImpl>
void TNoises<FImpl>::execute(void)
{
const int Nt{env().getDim(Tdir)};
const int Ns{Grid::QCD::Ns};
//const int Ns{Grid::Ns};
const int nnoise{par().nnoise};
const int nvec{par().nvec};
const int TI{ Hadrons::MDistil::DistilParameters::ParameterDefault( par().TI, Nt, false) };

4
Hadrons/Modules/MDistil/PerambFromSolve.hpp
View File

@ -169,9 +169,9 @@ void TPerambFromSolve<FImpl>::execute(void)
for (int is = 0; is < Ns; is++) {
result_nospin = peekSpin(solve[inoise+nnoise*(dk+LI*(dt+Nt_inv*ds))],is);
for (int t = Ntfirst; t < Ntfirst + Ntlocal; t++) {
ExtractSliceLocal(result_3d,result_nospin,0,t-Ntfirst,Grid::QCD::Tdir);
ExtractSliceLocal(result_3d,result_nospin,0,t-Ntfirst,Tdir);
for (int ivec = 0; ivec < nvec_reduced; ivec++) {
ExtractSliceLocal(evec3d,epack.evec[ivec],0,t-Ntfirst,Grid::QCD::Tdir);
ExtractSliceLocal(evec3d,epack.evec[ivec],0,t-Ntfirst,Tdir);
pokeSpin(perambulator(t, ivec, dk, inoise,dt,ds),static_cast<Complex>(innerProduct(evec3d, result_3d)),is);
std::cout << "perambulator(t, ivec, dk, inoise,dt,ds)(is) = (" << t << "," << ivec << "," << dk << "," << inoise << "," << dt << "," << ds << ")(" << is << ") = " << perambulator(t, ivec, dk, inoise,dt,ds)()(is)() << std::endl;
}

20
Hadrons/Modules/MDistil/Perambulator.hpp
View File

@ -203,27 +203,27 @@ void TPerambulator<FImpl>::execute(void)
for (int dt = 0; dt < Nt_inv; dt++) {
for (int ds = 0; ds < SI; ds++) {
std::cout << "LapH source vector from noise " << inoise << " and dilution component (d_k,d_t,d_alpha) : (" << dk << ","<< dt << "," << ds << ")" << std::endl;
dist_source = zero;
tmp3d_nospin = zero;
evec3d = zero;
dist_source = 0;
tmp3d_nospin = 0;
evec3d = 0;
for (int it = dt; it < Nt; it += TI){
if (full_tdil) t_inv = tsrc; else t_inv = it;
if( t_inv >= Ntfirst && t_inv < Ntfirst + Ntlocal ) {
for (int ik = dk; ik < nvec; ik += LI){
for (int is = ds; is < Ns; is += SI){
ExtractSliceLocal(evec3d,epack.evec[ik],0,t_inv-Ntfirst,Grid::QCD::Tdir);
ExtractSliceLocal(evec3d,epack.evec[ik],0,t_inv-Ntfirst,Tdir);
//tmp3d_nospin = evec3d * noise[inoise + nnoise*(t_inv + Nt*(ik+nvec*is))];
tmp3d_nospin = evec3d * noise(inoise, t_inv, ik, is);
tmp3d=zero;
tmp3d=0;
pokeSpin(tmp3d,tmp3d_nospin,is);
tmp2=zero;
InsertSliceLocal(tmp3d,tmp2,0,t_inv-Ntfirst,Grid::QCD::Tdir);
tmp2=0;
InsertSliceLocal(tmp3d,tmp2,0,t_inv-Ntfirst,Tdir);
dist_source += tmp2;
}
}
}
}
result=zero;
result=0;
v4dtmp = dist_source;
if (Ls_ == 1){
solver(result, v4dtmp);
@ -238,9 +238,9 @@ void TPerambulator<FImpl>::execute(void)
for (int is = 0; is < Ns; is++) {
result_nospin = peekSpin(result,is);
for (int t = Ntfirst; t < Ntfirst + Ntlocal; t++) {
ExtractSliceLocal(result_3d,result_nospin,0,t-Ntfirst,Grid::QCD::Tdir);
ExtractSliceLocal(result_3d,result_nospin,0,t-Ntfirst,Tdir);
for (int ivec = 0; ivec < nvec; ivec++) {
ExtractSliceLocal(evec3d,epack.evec[ivec],0,t-Ntfirst,Grid::QCD::Tdir);
ExtractSliceLocal(evec3d,epack.evec[ivec],0,t-Ntfirst,Tdir);
pokeSpin(perambulator(t, ivec, dk, inoise,dt,ds),static_cast<Complex>(innerProduct(evec3d, result_3d)),is);
}
}