portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-07 04:35:56 +01:00
Code Releases Activity

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

Browse Source
This commit is contained in:
Peter Boyle 2023-03-21 16:04:27 -04:00
commit 7db8dd7a95
19 changed files with 377 additions and 93 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h
View File

@ -81,6 +81,7 @@ public:
using OperatorFunction<FieldD>::operator();
RealD Tolerance;
Integer MaxIterationsMshift;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
std::vector<int> IterationsToCompleteShift; // Iterations for this shift
@ -95,9 +96,9 @@ public:
ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
int _ReliableUpdateFreq
) :
MaxIterations(maxit), shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq)
int _ReliableUpdateFreq) :
MaxIterationsMshift(maxit), shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq),
MaxIterations(20000)
{
verbose=1;
IterationsToCompleteShift.resize(_shifts.order);
@ -244,7 +245,7 @@ public:
// Iteration loop
int k;
for (k=1;k<=MaxIterations;k++){
for (k=1;k<=MaxIterationsMshift;k++){
a = c /cp;
AXPYTimer.Start();
@ -350,12 +351,17 @@ public:
}
}
if ( all_converged ){
if ( all_converged || k == MaxIterationsMshift-1){
SolverTimer.Stop();
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
if ( all_converged ){
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
} else {
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Not all shifts have converged iteration "<<k<<std::endl;
}
// Check answers
for(int s=0; s < nshift; s++) {
Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
@ -396,12 +402,10 @@ public:
return;
}
}
// ugly hack
std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
// assert(0);
assert(0);
}
};

1
Grid/communicator/SharedMemoryMPI.cc
View File

@ -29,6 +29,7 @@ Author: Christoph Lehner <christoph@lhnr.de>
#include <Grid/GridCore.h>
#include <pwd.h>
#include <syscall.h>
#ifdef GRID_CUDA
#include <cuda_runtime_api.h>

34
Grid/qcd/action/fermion/WilsonCompressor.h
View File

@ -484,24 +484,26 @@ public:
int dag = compress.dag;
int face_idx=0;
#define vet_same_node(a,b) \
{ auto tmp = b; }
if ( dag ) {
assert(this->same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
assert(this->same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
assert(this->same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
assert(this->same_node[Tp]==this->HaloGatherDir(source,TpCompress,Tp,face_idx));
assert(this->same_node[Xm]==this->HaloGatherDir(source,XmCompress,Xm,face_idx));
assert(this->same_node[Ym]==this->HaloGatherDir(source,YmCompress,Ym,face_idx));
assert(this->same_node[Zm]==this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
assert(this->same_node[Tm]==this->HaloGatherDir(source,TmCompress,Tm,face_idx));
vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XpCompress,Xp,face_idx));
vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YpCompress,Yp,face_idx));
vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZpCompress,Zp,face_idx));
vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TpCompress,Tp,face_idx));
vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XmCompress,Xm,face_idx));
vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YmCompress,Ym,face_idx));
vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZmCompress,Zm,face_idx));
vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TmCompress,Tm,face_idx));
} else {
assert(this->same_node[Xp]==this->HaloGatherDir(source,XmCompress,Xp,face_idx));
assert(this->same_node[Yp]==this->HaloGatherDir(source,YmCompress,Yp,face_idx));
assert(this->same_node[Zp]==this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
assert(this->same_node[Tp]==this->HaloGatherDir(source,TmCompress,Tp,face_idx));
assert(this->same_node[Xm]==this->HaloGatherDir(source,XpCompress,Xm,face_idx));
assert(this->same_node[Ym]==this->HaloGatherDir(source,YpCompress,Ym,face_idx));
assert(this->same_node[Zm]==this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
assert(this->same_node[Tm]==this->HaloGatherDir(source,TpCompress,Tm,face_idx));
vet_same_node(this->same_node[Xp],this->HaloGatherDir(source,XmCompress,Xp,face_idx));
vet_same_node(this->same_node[Yp],this->HaloGatherDir(source,YmCompress,Yp,face_idx));
vet_same_node(this->same_node[Zp],this->HaloGatherDir(source,ZmCompress,Zp,face_idx));
vet_same_node(this->same_node[Tp],this->HaloGatherDir(source,TmCompress,Tp,face_idx));
vet_same_node(this->same_node[Xm],this->HaloGatherDir(source,XpCompress,Xm,face_idx));
vet_same_node(this->same_node[Ym],this->HaloGatherDir(source,YpCompress,Ym,face_idx));
vet_same_node(this->same_node[Zm],this->HaloGatherDir(source,ZpCompress,Zm,face_idx));
vet_same_node(this->same_node[Tm],this->HaloGatherDir(source,TpCompress,Tm,face_idx));
}
this->face_table_computed=1;
assert(this->u_comm_offset==this->_unified_buffer_size);

7
Grid/qcd/action/fermion/WilsonKernels.h
View File

@ -52,13 +52,6 @@ public:
typedef AcceleratorVector<int,STENCIL_MAX> StencilVector;
public:
#ifdef GRID_SYCL
#define SYCL_HACK
#endif
#ifdef SYCL_HACK
static void HandDhopSiteSycl(StencilVector st_perm,StencilEntry *st_p, SiteDoubledGaugeField *U,SiteHalfSpinor *buf,
int ss,int sU,const SiteSpinor *in, SiteSpinor *out);
#endif
static void DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,

4
Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
View File

@ -63,6 +63,10 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
_tmp(&FiveDimRedBlackGrid),
Dirichlet(0)
{
Stencil.lo = &Lebesgue;
StencilEven.lo = &LebesgueEvenOdd;
StencilOdd.lo = &LebesgueEvenOdd;
// some assertions
assert(FiveDimGrid._ndimension==5);
assert(FourDimGrid._ndimension==4);

3
Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
View File

@ -60,6 +60,9 @@ WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
_tmp(&Hgrid),
anisotropyCoeff(anis)
{
Stencil.lo = &Lebesgue;
StencilEven.lo = &LebesgueEvenOdd;
StencilOdd.lo = &LebesgueEvenOdd;
// Allocate the required comms buffer
ImportGauge(_Umu);
if (anisotropyCoeff.isAnisotropic){

14
Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
View File

@ -433,11 +433,23 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
});
#define ASM_CALL(A) \
thread_for( ss, Nsite, { \
thread_for( sss, Nsite, { \
int ss = st.lo->Reorder(sss); \
int sU = ss; \
int sF = ss*Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v); \
});
#define ASM_CALL_SLICE(A) \
auto grid = in.Grid() ; \
int nt = grid->LocalDimensions()[4]; \
int nxyz = Nsite/nt ; \
for(int t=0;t<nt;t++){ \
thread_for( sss, nxyz, { \
int ss = t*nxyz+sss; \
int sU = ss; \
int sF = ss*Ls; \
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,Ls,1,in_v,out_v); \
});}
template <class Impl>
void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,

2
Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h
View File

@ -127,6 +127,8 @@ NAMESPACE_BEGIN(Grid);
ApproxNegPowerAction.tolerances[i] = action_tolerance[i];
ApproxHalfPowerAction.tolerances[i] = action_tolerance[i];
ApproxNegHalfPowerAction.tolerances[i]= action_tolerance[i];
}
for(int i=0;i<ApproxPowerMD.tolerances.size();i++){
ApproxPowerMD.tolerances[i] = md_tolerance[i];
ApproxNegPowerMD.tolerances[i] = md_tolerance[i];
ApproxHalfPowerMD.tolerances[i] = md_tolerance[i];

11
Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h
View File

@ -29,6 +29,8 @@
#ifndef QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
#define QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
#include <Grid/algorithms/iterative/ConjugateGradientMultiShiftCleanup.h>
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -58,7 +60,7 @@ NAMESPACE_BEGIN(Grid);
//Allow derived classes to override the multishift CG
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, FermionFieldD &out){
#if 0
SchurDifferentiableOperator<ImplD> schurOp(numerator ? NumOp : DenOp);
SchurDifferentiableOperator<ImplD> schurOp(numerator ? NumOpD : DenOpD);
ConjugateGradientMultiShift<FermionFieldD> msCG(MaxIter, approx);
msCG(schurOp,in, out);
#else
@ -66,7 +68,8 @@ NAMESPACE_BEGIN(Grid);
SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
FermionFieldD2 inD2(NumOpD2.FermionRedBlackGrid());
FermionFieldD2 outD2(NumOpD2.FermionRedBlackGrid());
// Action better with higher precision?
ConjugateGradientMultiShiftMixedPrec<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
precisionChange(inD2,in);
std::cout << "msCG single solve "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
@ -76,12 +79,12 @@ NAMESPACE_BEGIN(Grid);
}
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, std::vector<FermionFieldD> &out_elems, FermionFieldD &out){
SchurDifferentiableOperator<ImplD2> schurOpD2(numerator ? NumOpD2 : DenOpD2);
SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
SchurDifferentiableOperator<ImplF> schurOpF (numerator ? NumOpF : DenOpF);
FermionFieldD2 inD2(NumOpD2.FermionRedBlackGrid());
FermionFieldD2 outD2(NumOpD2.FermionRedBlackGrid());
std::vector<FermionFieldD2> out_elemsD2(out_elems.size(),NumOpD2.FermionRedBlackGrid());
ConjugateGradientMultiShiftMixedPrec<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
ConjugateGradientMultiShiftMixedPrecCleanup<FermionFieldD2, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
precisionChange(inD2,in);
std::cout << "msCG in "<<norm2(inD2)<<" " <<norm2(in)<<std::endl;
msCG(schurOpD2, inD2, out_elemsD2, outD2);

104
Grid/stencil/Stencil.h
View File

@ -300,9 +300,9 @@ public:
protected:
GridBase * _grid;
public:
GridBase *Grid(void) const { return _grid; }
LebesgueOrder *lo;
////////////////////////////////////////////////////////////////////////
// Needed to conveniently communicate gparity parameters into GPU memory
@ -348,6 +348,7 @@ public:
////////////////////////////////////////
// Stencil query
////////////////////////////////////////
#ifdef SHM_FAST_PATH
inline int SameNode(int point) {
int dimension = this->_directions[point];
@ -367,7 +368,40 @@ public:
if ( displacement == 0 ) return 1;
return 0;
}
#else
// fancy calculation for shm code
inline int SameNode(int point) {
int dimension = this->_directions[point];
int displacement = this->_distances[point];
int pd = _grid->_processors[dimension];
int fd = _grid->_fdimensions[dimension];
int ld = _grid->_ldimensions[dimension];
int rd = _grid->_rdimensions[dimension];
int simd_layout = _grid->_simd_layout[dimension];
int comm_dim = _grid->_processors[dimension] >1 ;
int recv_from_rank;
int xmit_to_rank;
if ( ! comm_dim ) return 1;
int nbr_proc;
if (displacement>0) nbr_proc = 1;
else nbr_proc = pd-1;
// FIXME this logic needs to be sorted for three link term
// assert( (displacement==1) || (displacement==-1));
// Present hack only works for >= 4^4 subvol per node
_grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
void *shm = (void *) _grid->ShmBufferTranslate(recv_from_rank,this->u_recv_buf_p);
if ( shm==NULL ) return 0;
return 1;
}
#endif
//////////////////////////////////////////
// Comms packet queue for asynch thread
// Use OpenMP Tasks for cleaner ???
@ -1075,7 +1109,7 @@ public:
int comms_recv = this->_comms_recv[point];
int comms_partial_send = this->_comms_partial_send[point] ;
int comms_partial_recv = this->_comms_partial_recv[point] ;
assert(rhs.Grid()==_grid);
// conformable(_grid,rhs.Grid());
@ -1146,11 +1180,32 @@ public:
recv_buf=this->u_recv_buf_p;
}
// potential SHM fast path for intranode
int shm_send=0;
int shm_recv=0;
#ifdef SHM_FAST_PATH
// Put directly in place if we can
send_buf = (cobj *)_grid->ShmBufferTranslate(xmit_to_rank,recv_buf);
if ( (send_buf==NULL) ) {
shm_send=0;
send_buf = this->u_send_buf_p;
} else {
shm_send=1;
}
void *test_ptr = _grid->ShmBufferTranslate(recv_from_rank,recv_buf);
if ( test_ptr != NULL ) shm_recv = 1;
// static int printed;
// if (!printed){
// std::cout << " GATHER FAST PATH SHM "<<shm_send<< " "<<shm_recv<<std::endl;
// printed = 1;
// }
#else
////////////////////////////////////////////////////////
// Gather locally
////////////////////////////////////////////////////////
send_buf = this->u_send_buf_p; // Gather locally, must send
assert(send_buf!=NULL);
#endif
// std::cout << " GatherPlaneSimple partial send "<< comms_partial_send<<std::endl;
compressor::Gather_plane_simple(face_table[face_idx],rhs,send_buf,compress,comm_off,so,comms_partial_send);
@ -1162,10 +1217,13 @@ public:
// Build a list of things to do after we synchronise GPUs
// Start comms now???
///////////////////////////////////////////////////////////
int do_send = (comms_send|comms_partial_send) && (!shm_send );
int do_recv = (comms_send|comms_partial_send) && (!shm_recv );
AddPacket((void *)&send_buf[comm_off],
(void *)&recv_buf[comm_off],
xmit_to_rank, comms_send|comms_partial_send,
recv_from_rank, comms_recv|comms_partial_recv,
xmit_to_rank, do_send,
recv_from_rank, do_recv,
xbytes,rbytes);
}
@ -1307,19 +1365,47 @@ public:
int recv_from_rank;
int xmit_to_rank;
int shm_send=0;
int shm_recv=0;
_grid->ShiftedRanks(dimension,nbr_proc,xmit_to_rank,recv_from_rank);
#ifdef SHM_FAST_PATH
#warning STENCIL SHM FAST PATH SELECTED
// shm == receive pointer if offnode
// shm == Translate[send pointer] if on node -- my view of his send pointer
cobj *shm = (cobj *) _grid->ShmBufferTranslate(recv_from_rank,sp);
if (shm==NULL) {
shm = rp;
// we found a packet that comes from MPI and contributes to this shift.
// is_same_node is only used in the WilsonStencil, and gets set for this point in the stencil.
// Kernel will add the exterior_terms except if is_same_node.
// leg of stencil
shm_recv=0;
} else {
shm_recv=1;
}
rpointers[i] = shm;
// Test send side
void *test_ptr = (void *) _grid->ShmBufferTranslate(xmit_to_rank,sp);
if ( test_ptr != NULL ) shm_send = 1;
// static int printed;
// if (!printed){
// std::cout << " GATHERSIMD FAST PATH SHM "<<shm_send<< " "<<shm_recv<<std::endl;
// printed = 1;
// }
#else
rpointers[i] = rp;
#endif
int duplicate = CheckForDuplicate(dimension,sx,nbr_proc,(void *)rp,i,xbytes,rbytes,cbmask);
if ( !duplicate ) {
if ( (bytes != rbytes) && (rbytes!=0) ){
acceleratorMemSet(rp,0,bytes); // Zero prefill comms buffer to zero
}
int do_send = (comms_send|comms_partial_send) && (!shm_send );
int do_recv = (comms_send|comms_partial_send) && (!shm_recv );
AddPacket((void *)sp,(void *)rp,
xmit_to_rank,comms_send|comms_partial_send,
recv_from_rank,comms_recv|comms_partial_recv,
xmit_to_rank,do_send,
recv_from_rank,do_send,
xbytes,rbytes);
}
@ -1329,7 +1415,7 @@ public:
}
}
// rpointer may be doing a remote read in the gather over SHM
if ( comms_recv|comms_partial_recv ) {
AddMerge(&this->u_recv_buf_p[comm_off],rpointers,reduced_buffer_size,permute_type,Mergers);
}

16
Grid/threads/Accelerator.h
View File

@ -248,17 +248,23 @@ inline int acceleratorIsCommunicable(void *ptr)
//////////////////////////////////////////////
// SyCL acceleration
//////////////////////////////////////////////
#ifdef GRID_SYCL
NAMESPACE_END(Grid);
#include <CL/sycl.hpp>
#include <CL/sycl/usm.hpp>
#ifdef GRID_SYCL
#define GRID_SYCL_LEVEL_ZERO_IPC
#ifdef GRID_SYCL_LEVEL_ZERO_IPC
NAMESPACE_END(Grid);
#if 0
#include <CL/sycl.hpp>
#include <CL/sycl/usm.hpp>
#include <level_zero/ze_api.h>
#include <CL/sycl/backend/level_zero.hpp>
#else
#include <sycl/CL/sycl.hpp>
#include <sycl/usm.hpp>
#include <level_zero/ze_api.h>
#include <sycl/ext/oneapi/backend/level_zero.hpp>
#endif
NAMESPACE_BEGIN(Grid);
extern cl::sycl::queue *theGridAccelerator;

38
HMC/Mobius2p1f_DD_EOFA_96I_mshift.cc
View File

@ -232,31 +232,34 @@ int main(int argc, char **argv) {
// std::vector<Real> hasenbusch({ light_mass, 0.005, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass }); // Updated
// std::vector<Real> hasenbusch({ light_mass, 0.0145, 0.045, 0.108, 0.25, 0.51 , 0.75 , pv_mass });
OneFlavourRationalParams OFRp; // Up/down
OFRp.lo = 4.0e-5;
int SP_iters=10000;
RationalActionParams OFRp; // Up/down
OFRp.lo = 6.0e-5;
OFRp.hi = 90.0;
OFRp.MaxIter = 60000;
OFRp.tolerance= 1.0e-5;
OFRp.mdtolerance= 1.0e-3;
OFRp.inv_pow = 2;
OFRp.MaxIter = SP_iters; // get most shifts by 2000, stop sharing space
OFRp.action_tolerance= 1.0e-8;
OFRp.action_degree = 18;
OFRp.md_tolerance= 1.0e-5;
OFRp.md_degree = 14;
// OFRp.degree = 20; converges
// OFRp.degree = 16;
OFRp.degree = 18;
OFRp.precision= 80;
OFRp.BoundsCheckFreq=0;
std::vector<RealD> ActionTolByPole({
1.0e-8,1.0e-8,1.0e-8,1.0e-8,
1.0e-7,1.0e-8,1.0e-8,1.0e-8,
1.0e-8,1.0e-8,1.0e-8,1.0e-8,
1.0e-8,1.0e-8,1.0e-8,1.0e-8,
1.0e-8,1.0e-8,1.0e-8,1.0e-8,
1.0e-8,1.0e-8
});
std::vector<RealD> MDTolByPole({
1.0e-5,5.0e-6,1.0e-6,1.0e-7, // soften convergence more more
1.6e-5,5.0e-6,1.0e-6,3.0e-7, // soften convergence more more
// 1.0e-6,3.0e-7,1.0e-7,1.0e-7,
// 3.0e-6,1.0e-6,1.0e-7,1.0e-7, // soften convergence
1.0e-8,1.0e-8,1.0e-8,1.0e-8,
1.0e-8,1.0e-8,1.0e-8,1.0e-8,
1.0e-8,1.0e-8,1.0e-8,1.0e-8,
1.0e-8,1.0e-8
});
@ -340,6 +343,7 @@ int main(int argc, char **argv) {
ParamsDirF.dirichlet=Dirichlet;
ParamsDir.partialDirichlet=1;
ParamsDirF.partialDirichlet=1;
std::cout << GridLogMessage<< "Partial Dirichlet depth is "<<dwf_compressor_depth<<std::endl;
// double StoppingCondition = 1e-14;
// double MDStoppingCondition = 1e-9;
@ -457,9 +461,9 @@ int main(int argc, char **argv) {
#define MIXED_PRECISION
#ifdef MIXED_PRECISION
std::vector<OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2> *> Bdys;
std::vector<GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2> *> Bdys;
#else
std::vector<OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
std::vector<GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy> *> Bdys;
#endif
typedef SchurDiagMooeeOperator<FermionActionF,FermionFieldF> LinearOperatorF;
@ -544,19 +548,19 @@ int main(int argc, char **argv) {
ParamsNumD2.partialDirichlet = ParamsNum.partialDirichlet;
NumeratorsD2.push_back (new FermionActionD2(UD2,*FGridF,*FrbGridF,*GridPtrF,*GridRBPtrF,light_num[h],M5,b,c, ParamsNumD2));
Bdys.push_back( new OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2>(
Bdys.push_back( new GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2>(
*Numerators[h],*Denominators[h],
*NumeratorsF[h],*DenominatorsF[h],
*NumeratorsD2[h],*DenominatorsD2[h],
OFRp, 400) );
Bdys.push_back( new OneFlavourEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2>(
OFRp, SP_iters) );
Bdys.push_back( new GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction<FermionImplPolicy,FermionImplPolicyF,FermionImplPolicyD2>(
*Numerators[h],*Denominators[h],
*NumeratorsF[h],*DenominatorsF[h],
*NumeratorsD2[h],*DenominatorsD2[h],
OFRp, 400) );
OFRp, SP_iters) );
#else
Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
Bdys.push_back( new OneFlavourEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
Bdys.push_back( new GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
Bdys.push_back( new GeneralEvenOddRatioRationalPseudoFermionAction<FermionImplPolicy>(*Numerators[h],*Denominators[h],OFRp));
#endif
}
}

16
HMC/Mobius2p1f_EOFA_96I_hmc.cc
View File

@ -183,7 +183,7 @@ int main(int argc, char **argv) {
// 4/2 => 0.6 dH
// 3/3 => 0.8 dH .. depth 3, slower
//MD.MDsteps = 4;
MD.MDsteps = 3;
MD.MDsteps = 12;
MD.trajL = 0.5;
HMCparameters HMCparams;
@ -200,8 +200,8 @@ int main(int argc, char **argv) {
TheHMC.Resources.AddFourDimGrid("gauge"); // use default simd lanes decomposition
CheckpointerParameters CPparams;
CPparams.config_prefix = "ckpoint_DDHMC_lat";
CPparams.rng_prefix = "ckpoint_DDHMC_rng";
CPparams.config_prefix = "ckpoint_HMC_lat";
CPparams.rng_prefix = "ckpoint_HMC_rng";
CPparams.saveInterval = 1;
CPparams.format = "IEEE64BIG";
TheHMC.Resources.LoadNerscCheckpointer(CPparams);
@ -228,7 +228,7 @@ int main(int argc, char **argv) {
Real pv_mass = 1.0;
// std::vector<Real> hasenbusch({ 0.01, 0.045, 0.108, 0.25, 0.51 , pv_mass });
// std::vector<Real> hasenbusch({ light_mass, 0.01, 0.045, 0.108, 0.25, 0.51 , pv_mass });
std::vector<Real> hasenbusch({ light_mass, 0.005, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass }); // Updated
std::vector<Real> hasenbusch({ 0.005, 0.0145, 0.045, 0.108, 0.25, 0.51 , pv_mass }); // Updated
// std::vector<Real> hasenbusch({ light_mass, 0.0145, 0.045, 0.108, 0.25, 0.51 , 0.75 , pv_mass });
auto GridPtr = TheHMC.Resources.GetCartesian();
@ -299,8 +299,8 @@ int main(int argc, char **argv) {
////////////////////////////////////
// Collect actions
////////////////////////////////////
ActionLevel<HMCWrapper::Field> Level1(1);
ActionLevel<HMCWrapper::Field> Level2(3);
// ActionLevel<HMCWrapper::Field> Level1(1);
ActionLevel<HMCWrapper::Field> Level2(1);
ActionLevel<HMCWrapper::Field> Level3(15);
////////////////////////////////////
@ -369,7 +369,7 @@ int main(int argc, char **argv) {
ActionCGL, ActionCGR,
DerivativeCGL, DerivativeCGR,
SFRp, true);
// Level2.push_back(&EOFA);
Level2.push_back(&EOFA);
////////////////////////////////////
// up down action
@ -477,7 +477,7 @@ int main(int argc, char **argv) {
// Gauge action
/////////////////////////////////////////////////////////////
Level3.push_back(&GaugeAction);
TheHMC.TheAction.push_back(Level1);
// TheHMC.TheAction.push_back(Level1);
TheHMC.TheAction.push_back(Level2);
TheHMC.TheAction.push_back(Level3);
std::cout << GridLogMessage << " Action complete "<< std::endl;

8
configure.ac
View File

@ -646,6 +646,14 @@ case ${ac_SHM_FORCE_MPI} in
;;
*) ;;
esac
############### force MPI in SMP
AC_ARG_ENABLE([shm-fast-path],[AS_HELP_STRING([--enable-shm-fast-path],[Allow kernels to remote copy over intranode])],[ac_SHM_FAST_PATH=${enable_shm_fast_path}],[ac_SHM_FAST_PATH=no])
case ${ac_SHM_FAST_PATH} in
yes)
AC_DEFINE([SHM_FAST_PATH],[1],[SHM_FAST_PATH] )
;;
*) ;;
esac
############### communication type selection
AC_ARG_ENABLE([comms-threads],[AS_HELP_STRING([--enable-comms-threads | --disable-comms-threads],[Use multiple threads in MPI calls])],[ac_COMMS_THREADS=${enable_comms_threads}],[ac_COMMS_THREADS=yes])

9
systems/PVC/benchmarks/run-2tile-mpi.sh
View File

@ -23,12 +23,7 @@ export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0
for i in 0
do
mpiexec -launcher ssh -n 2 -host localhost ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT --shm-mpi 1 --device-mem 32768
mpiexec -launcher ssh -n 2 -host localhost ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --shm-mpi 1 --device-mem 32768
done
#mpiexec -launcher ssh -n 2 -host localhost ./wrap4gpu.sh ./Benchmark_halo --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT --shm-mpi 1 > halo.2tile.1x2.log
#mpiexec -launcher ssh -n 2 -host localhost ./wrap4gpu.sh ./Benchmark_halo --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --shm-mpi 1 > halo.2tile.2x1.log
#mpiexec -launcher ssh -n 1 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 32.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 0 > 1tile.log
mpiexec -launcher ssh -n 2 -host localhost ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT --comms-sequential --shm-mpi 0

12
systems/PVC/benchmarks/wrap.sh
View File

@ -5,10 +5,10 @@ export ZE_AFFINITY_MASK=0.$MPI_LOCALRANKID
echo Ranke $MPI_LOCALRANKID ZE_AFFINITY_MASK is $ZE_AFFINITY_MASK
if [ $MPI_LOCALRANKID = "0" ]
then
# ~psteinbr/build_pti/ze_tracer -h $@
onetrace --chrome-device-timeline $@
else
#if [ $MPI_LOCALRANKID = "0" ]
#then
# ~psteinbr/build_pti/ze_tracer -c $@
# onetrace --chrome-kernel-timeline $@
#else
$@
fi
#fi

2
systems/PVC/config-command
View File

@ -1,4 +1,4 @@
INSTALL=/nfs/site/home/azusayax/install
INSTALL=/nfs/site/home/paboylx/prereqs/
../../configure \
--enable-simd=GPU \
--enable-gen-simd-width=64 \

3
systems/PVC/setup.sh
View File

@ -1,5 +1,6 @@
export https_proxy=http://proxy-chain.intel.com:911
export LD_LIBRARY_PATH=/nfs/site/home/azusayax/install/lib:$LD_LIBRARY_PATH
#export LD_LIBRARY_PATH=/nfs/site/home/azusayax/install/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=$HOME/prereqs/lib/:$LD_LIBRARY_PATH
module load intel-release
source /opt/intel/oneapi/PVC_setup.sh

160
tests/core/Test_fft_matt.cc Normal file
View File

@ -0,0 +1,160 @@
/*************************************************************************************
grid` physics library, www.github.com/paboyle/Grid
Source file: ./tests/Test_cshift.cc
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid.h>
using namespace Grid;
;
int main (int argc, char ** argv)
{
Grid_init(&argc,&argv);
int threads = GridThread::GetThreads();
std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
Coordinate latt_size = GridDefaultLatt();
Coordinate simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
Coordinate mpi_layout = GridDefaultMpi();
int vol = 1;
for(int d=0;d<latt_size.size();d++){
vol = vol * latt_size[d];
}
GridCartesian GRID(latt_size,simd_layout,mpi_layout);
GridRedBlackCartesian RBGRID(&GRID);
LatticeComplexD one(&GRID);
LatticeComplexD zz(&GRID);
LatticeComplexD C(&GRID);
LatticeComplexD Ctilde(&GRID);
LatticeComplexD Cref (&GRID);
LatticeComplexD Csav (&GRID);
LatticeComplexD coor(&GRID);
LatticeSpinMatrixD S(&GRID);
LatticeSpinMatrixD Stilde(&GRID);
Coordinate p({1,3,2,3});
one = ComplexD(1.0,0.0);
zz = ComplexD(0.0,0.0);
ComplexD ci(0.0,1.0);
std::vector<int> seeds({1,2,3,4});
GridSerialRNG sRNG; sRNG.SeedFixedIntegers(seeds); // naughty seeding
GridParallelRNG pRNG(&GRID);
pRNG.SeedFixedIntegers(seeds);
LatticeGaugeFieldD Umu(&GRID);
SU<Nc>::ColdConfiguration(pRNG,Umu); // Unit gauge
////////////////////////////////////////////////////
// Wilson test
////////////////////////////////////////////////////
{
LatticeFermionD src(&GRID); gaussian(pRNG,src);
LatticeFermionD tmp(&GRID);
LatticeFermionD ref(&GRID);
RealD mass=0.01;
WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
Dw.M(src,tmp);
std::cout << "Dw src = " <<norm2(src)<<std::endl;
std::cout << "Dw tmp = " <<norm2(tmp)<<std::endl;
Dw.FreePropagator(tmp,ref,mass);
std::cout << "Dw ref = " <<norm2(ref)<<std::endl;
ref = ref - src;
std::cout << "Dw ref-src = " <<norm2(ref)<<std::endl;
}
////////////////////////////////////////////////////
// Wilson prop
////////////////////////////////////////////////////
{
std::cout<<"****************************************"<<std::endl;
std::cout << "Wilson Mom space 4d propagator \n";
std::cout<<"****************************************"<<std::endl;
LatticeFermionD src(&GRID); gaussian(pRNG,src);
LatticeFermionD tmp(&GRID);
LatticeFermionD ref(&GRID);
LatticeFermionD diff(&GRID);
src=Zero();
Coordinate point(4,0); // 0,0,0,0
SpinColourVectorD ferm;
ferm=Zero();
ferm()(0)(0) = ComplexD(1.0);
pokeSite(ferm,src,point);
RealD mass=0.01;
WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
// Momentum space prop
std::cout << " Solving by FFT and Feynman rules" <<std::endl;
Dw.FreePropagator(src,ref,mass) ;
Gamma G5(Gamma::Algebra::Gamma5);
LatticeFermionD result(&GRID);
const int sdir=0;
////////////////////////////////////////////////////////////////////////
// Conjugate gradient on normal equations system
////////////////////////////////////////////////////////////////////////
std::cout << " Solving by Conjugate Gradient (CGNE)" <<std::endl;
Dw.Mdag(src,tmp);
src=tmp;
MdagMLinearOperator<WilsonFermionD,LatticeFermionD> HermOp(Dw);
ConjugateGradient<LatticeFermionD> CG(1.0e-10,10000);
CG(HermOp,src,result);
////////////////////////////////////////////////////////////////////////
std::cout << " Taking difference" <<std::endl;
std::cout << "Dw result "<<norm2(result)<<std::endl;
std::cout << "Dw ref "<<norm2(ref)<<std::endl;
diff = ref - result;
std::cout << "result - ref "<<norm2(diff)<<std::endl;
DumpSliceNorm("Slice Norm Solution ",result,Nd-1);
}
Grid_finalize();
}