portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-03 21:44:33 +00:00
Code Releases Activity

merge upstream develop

Browse Source
This commit is contained in:
nmeyer-ur
2020-07-07 20:26:47 +02:00
parent 67db4993c2 64fe5b21b4
commit 8726e94ea7
326 changed files with 10335 additions and 9381 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
Grid/GridCore.h
View File

@@ -47,9 +47,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <Grid/perfmon/PerfCount.h>
#include <Grid/util/Util.h>
#include <Grid/log/Log.h>
#include <Grid/allocator/AlignedAllocator.h>
#include <Grid/allocator/Allocator.h>
#include <Grid/simd/Simd.h>
#include <Grid/threads/Threads.h>
#include <Grid/threads/ThreadReduction.h>
#include <Grid/serialisation/Serialisation.h>
#include <Grid/util/Sha.h>
#include <Grid/communicator/Communicator.h>

1
Grid/GridStd.h
View File

@@ -6,6 +6,7 @@
///////////////////
#include <cassert>
#include <complex>
#include <memory>
#include <vector>
#include <array>
#include <string>

22
Grid/Grid_Eigen_Dense.h
View File

@@ -18,12 +18,23 @@
#pragma push_macro("__CUDA_ARCH__")
#pragma push_macro("__NVCC__")
#pragma push_macro("__CUDACC__")
#undef __CUDA_ARCH__
#undef __NVCC__
#undef __CUDACC__
#undef __CUDA_ARCH__
#define __NVCC__REDEFINE__
#endif
/* SYCL save and restore compile environment*/
#ifdef GRID_SYCL
#pragma push
#pragma push_macro("__SYCL_DEVICE_ONLY__")
#undef __SYCL_DEVICE_ONLY__
#define EIGEN_DONT_VECTORIZE
//#undef EIGEN_USE_SYCL
#define __SYCL__REDEFINE__
#endif
#include <Grid/Eigen/Dense>
#include <Grid/Eigen/unsupported/CXX11/Tensor>
@@ -31,7 +42,13 @@
#ifdef __NVCC__REDEFINE__
#pragma pop_macro("__CUDACC__")
#pragma pop_macro("__NVCC__")
#pragma pop_macro("__CUDA_ARCH__")
#pragma pop_macro("GRID_SIMT")
#pragma pop
#endif
/*SYCL restore*/
#ifdef __SYCL__REDEFINE__
#pragma pop_macro("__SYCL_DEVICE_ONLY__")
#pragma pop
#endif
@@ -39,3 +56,4 @@
#pragma GCC diagnostic pop
#endif

4
Grid/Makefile.am
View File

@@ -21,7 +21,7 @@ if BUILD_HDF5
extra_headers+=serialisation/Hdf5Type.h
endif
all: version-cache
all: version-cache Version.h
version-cache:
@if [ `git status --porcelain | grep -v '??' | wc -l` -gt 0 ]; then\
@@ -42,7 +42,7 @@ version-cache:
fi;\
rm -f vertmp
Version.h:
Version.h: version-cache
cp version-cache Version.h
.PHONY: version-cache

8
Grid/algorithms/Algorithms.h
View File

@@ -29,9 +29,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_ALGORITHMS_H
#define GRID_ALGORITHMS_H
NAMESPACE_CHECK(algorithms);
#include <Grid/algorithms/SparseMatrix.h>
#include <Grid/algorithms/LinearOperator.h>
#include <Grid/algorithms/Preconditioner.h>
NAMESPACE_CHECK(SparseMatrix);
#include <Grid/algorithms/approx/Zolotarev.h>
#include <Grid/algorithms/approx/Chebyshev.h>
@@ -41,10 +43,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/algorithms/approx/Forecast.h>
#include <Grid/algorithms/approx/RemezGeneral.h>
#include <Grid/algorithms/approx/ZMobius.h>
NAMESPACE_CHECK(approx);
#include <Grid/algorithms/iterative/Deflation.h>
#include <Grid/algorithms/iterative/ConjugateGradient.h>
NAMESPACE_CHECK(ConjGrad);
#include <Grid/algorithms/iterative/BiCGSTAB.h>
NAMESPACE_CHECK(BiCGSTAB);
#include <Grid/algorithms/iterative/ConjugateResidual.h>
#include <Grid/algorithms/iterative/NormalEquations.h>
#include <Grid/algorithms/iterative/SchurRedBlack.h>
@@ -62,7 +66,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
#include <Grid/algorithms/iterative/PowerMethod.h>
NAMESPACE_CHECK(PowerMethod);
#include <Grid/algorithms/CoarsenedMatrix.h>
NAMESPACE_CHECK(CoarsendMatrix);
#include <Grid/algorithms/FFT.h>
#endif

422
Grid/algorithms/CoarsenedMatrix.h
View File

@@ -1,14 +1,3 @@
// blockZaxpy in bockPromote - 3s, 5%
// noncoalesced linalg in Preconditionoer ~ 3s 5%
// Lancos tuning or replace 10-20s ~ 25%, open ended
// setup tuning 5s ~ 8%
// -- e.g. ordermin, orderstep tunables.
// MdagM path without norm in LinOp code. few seconds
// Mdir calc blocking kernels
// Fuse kernels in blockMaskedInnerProduct
// preallocate Vectors in Cayley 5D ~ few percent few seconds
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@@ -91,35 +80,8 @@ public:
}
directions [2*_d]=0;
displacements[2*_d]=0;
//// report back
std::cout<<GridLogMessage<<"directions :";
for(int d=0;d<npoint;d++) std::cout<< directions[d]<< " ";
std::cout<<std::endl;
std::cout<<GridLogMessage<<"displacements :";
for(int d=0;d<npoint;d++) std::cout<< displacements[d]<< " ";
std::cout<<std::endl;
}
/*
// Original cleaner code
Geometry(int _d) : dimension(_d), npoint(2*_d+1), directions(npoint), displacements(npoint) {
for(int d=0;d<dimension;d++){
directions[2*d ] = d;
directions[2*d+1] = d;
displacements[2*d ] = +1;
displacements[2*d+1] = -1;
}
directions [2*dimension]=0;
displacements[2*dimension]=0;
}
std::vector<int> GetDelta(int point) {
std::vector<int> delta(dimension,0);
delta[directions[point]] = displacements[point];
return delta;
};
*/
};
template<class Fobj,class CComplex,int nbasis>
@@ -149,24 +111,6 @@ public:
CoarseScalar InnerProd(CoarseGrid);
std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 1"<<std::endl;
blockOrthogonalise(InnerProd,subspace);
// std::cout << GridLogMessage <<" Block Gramm-Schmidt pass 2"<<std::endl; // Really have to do twice? Yuck
// blockOrthogonalise(InnerProd,subspace);
// std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
// CheckOrthogonal();
}
void CheckOrthogonal(void){
CoarseVector iProj(CoarseGrid);
CoarseVector eProj(CoarseGrid);
for(int i=0;i<nbasis;i++){
blockProject(iProj,subspace[i],subspace);
eProj=Zero();
accelerator_for(ss, CoarseGrid->oSites(),1,{
eProj[ss](i)=CComplex(1.0);
});
eProj=eProj - iProj;
std::cout<<GridLogMessage<<"Orthog check error "<<i<<" " << norm2(eProj)<<std::endl;
}
std::cout<<GridLogMessage <<"CheckOrthog done"<<std::endl;
}
void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
blockProject(CoarseVec,FineVec,subspace);
@@ -175,11 +119,6 @@ public:
FineVec.Checkerboard() = subspace[0].Checkerboard();
blockPromote(CoarseVec,FineVec,subspace);
}
void CreateSubspaceRandom(GridParallelRNG &RNG){
for(int i=0;i<nbasis;i++){
random(RNG,subspace[i]);
}
}
virtual void CreateSubspace(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
@@ -218,7 +157,7 @@ public:
// World of possibilities here. But have tried quite a lot of experiments (250+ jobs run on Summit)
// and this is the best I found
////////////////////////////////////////////////////////////////////////////////////////////////
#if 1
virtual void CreateSubspaceChebyshev(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,
int nn,
double hi,
@@ -280,10 +219,10 @@ public:
hermop.HermOp(*Tn,y);
auto y_v = y.View();
auto Tn_v = Tn->View();
auto Tnp_v = Tnp->View();
auto Tnm_v = Tnm->View();
autoView( y_v , y, AcceleratorWrite);
autoView( Tn_v , (*Tn), AcceleratorWrite);
autoView( Tnp_v , (*Tnp), AcceleratorWrite);
autoView( Tnm_v , (*Tnm), AcceleratorWrite);
const int Nsimd = CComplex::Nsimd();
accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
@@ -313,201 +252,6 @@ public:
}
assert(b==nn);
}
#endif
#if 0
virtual void CreateSubspaceChebyshev(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,
int nn,
double hi,
double lo,
int orderfilter,
int ordermin,
int orderstep,
double filterlo
) {
RealD scale;
FineField noise(FineGrid);
FineField Mn(FineGrid);
FineField tmp(FineGrid);
FineField combined(FineGrid);
// New normalised noise
gaussian(RNG,noise);
scale = std::pow(norm2(noise),-0.5);
noise=noise*scale;
// Initial matrix element
hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
int b =0;
#define FILTERb(llo,hhi,oorder) \
{ \
Chebyshev<FineField> Cheb(llo,hhi,oorder); \
Cheb(hermop,noise,Mn); \
scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale; \
subspace[b] = Mn; \
hermop.Op(Mn,tmp); \
std::cout<<GridLogMessage << oorder<< " Cheb filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
b++; \
}
// JacobiPolynomial<FineField> Cheb(0.002,60.0,1500,-0.5,3.5); \
RealD alpha=-0.8;
RealD beta =-0.8;
#define FILTER(llo,hhi,oorder) \
{ \
Chebyshev<FineField> Cheb(llo,hhi,oorder); \
/* JacobiPolynomial<FineField> Cheb(0.0,60.0,oorder,alpha,beta);*/\
Cheb(hermop,noise,Mn); \
scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale; \
subspace[b] = Mn; \
hermop.Op(Mn,tmp); \
std::cout<<GridLogMessage << oorder<< "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
b++; \
}
#define FILTERc(llo,hhi,oorder) \
{ \
Chebyshev<FineField> Cheb(llo,hhi,oorder); \
Cheb(hermop,noise,combined); \
}
double node = 0.000;
FILTERb(lo,hi,orderfilter);// 0
// FILTERc(node,hi,51);// 0
noise = Mn;
int base = 0;
int mult = 100;
FILTER(node,hi,base+1*mult);
FILTER(node,hi,base+2*mult);
FILTER(node,hi,base+3*mult);
FILTER(node,hi,base+4*mult);
FILTER(node,hi,base+5*mult);
FILTER(node,hi,base+6*mult);
FILTER(node,hi,base+7*mult);
FILTER(node,hi,base+8*mult);
FILTER(node,hi,base+9*mult);
FILTER(node,hi,base+10*mult);
FILTER(node,hi,base+11*mult);
FILTER(node,hi,base+12*mult);
FILTER(node,hi,base+13*mult);
FILTER(node,hi,base+14*mult);
FILTER(node,hi,base+15*mult);
assert(b==nn);
}
#endif
#if 0
virtual void CreateSubspaceChebyshev(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,
int nn,
double hi,
double lo,
int orderfilter,
int ordermin,
int orderstep,
double filterlo
) {
RealD scale;
FineField noise(FineGrid);
FineField Mn(FineGrid);
FineField tmp(FineGrid);
FineField combined(FineGrid);
// New normalised noise
gaussian(RNG,noise);
scale = std::pow(norm2(noise),-0.5);
noise=noise*scale;
// Initial matrix element
hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise <n|MdagM|n> "<<norm2(Mn)<<std::endl;
int b =0;
{
Chebyshev<FineField> JacobiPoly(0.005,60.,1500);
// JacobiPolynomial<FineField> JacobiPoly(0.002,60.0,1500,-0.5,3.5);
//JacobiPolynomial<FineField> JacobiPoly(0.03,60.0,500,-0.5,3.5);
// JacobiPolynomial<FineField> JacobiPoly(0.00,60.0,1000,-0.5,3.5);
JacobiPoly(hermop,noise,Mn);
scale = std::pow(norm2(Mn),-0.5); Mn=Mn*scale;
subspace[b] = Mn;
hermop.Op(Mn,tmp);
std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl;
b++;
// scale = std::pow(norm2(tmp),-0.5); tmp=tmp*scale;
// subspace[b] = tmp; b++;
// }
}
#define FILTER(lambda) \
{ \
hermop.HermOp(subspace[0],tmp); \
tmp = tmp - lambda *subspace[0]; \
scale = std::pow(norm2(tmp),-0.5); \
tmp=tmp*scale; \
subspace[b] = tmp; \
hermop.Op(subspace[b],tmp); \
std::cout<<GridLogMessage << "filt ["<<b<<"] <n|MdagM|n> "<<norm2(tmp)<<std::endl; \
b++; \
}
// scale = std::pow(norm2(tmp),-0.5); tmp=tmp*scale;
// subspace[b] = tmp; b++;
// }
FILTER(2.0e-5);
FILTER(2.0e-4);
FILTER(4.0e-4);
FILTER(8.0e-4);
FILTER(8.0e-4);
FILTER(2.0e-3);
FILTER(3.0e-3);
FILTER(4.0e-3);
FILTER(5.0e-3);
FILTER(6.0e-3);
FILTER(2.5e-3);
FILTER(3.5e-3);
FILTER(4.5e-3);
FILTER(5.5e-3);
FILTER(6.5e-3);
// FILTER(6.0e-5);//6
// FILTER(7.0e-5);//8
// FILTER(8.0e-5);//9
// FILTER(9.0e-5);//3
/*
// FILTER(1.0e-4);//10
FILTER(2.0e-4);//11
// FILTER(3.0e-4);//12
// FILTER(4.0e-4);//13
FILTER(5.0e-4);//14
FILTER(6.0e-3);//4
FILTER(7.0e-4);//1
FILTER(8.0e-4);//7
FILTER(9.0e-4);//15
FILTER(1.0e-3);//2
FILTER(2.0e-3);//2
FILTER(3.0e-3);//2
FILTER(4.0e-3);//2
FILTER(5.0e-3);//2
FILTER(6.0e-3);//2
FILTER(7.0e-3);//2
FILTER(8.0e-3);//2
FILTER(1.0e-2);//2
*/
std::cout << GridLogMessage <<"Jacobi filtering done" <<std::endl;
assert(b==nn);
}
#endif
};
@@ -541,36 +285,28 @@ public:
///////////////////////
GridBase * Grid(void) { return _grid; }; // this is all the linalg routines need to know
RealD M (const CoarseVector &in, CoarseVector &out){
void M (const CoarseVector &in, CoarseVector &out)
{
conformable(_grid,in.Grid());
conformable(in.Grid(),out.Grid());
// RealD Nin = norm2(in);
SimpleCompressor<siteVector> compressor;
double comms_usec = -usecond();
Stencil.HaloExchange(in,compressor);
comms_usec += usecond();
auto in_v = in.View();
auto out_v = out.View();
autoView( in_v , in, AcceleratorRead);
autoView( out_v , out, AcceleratorWrite);
typedef LatticeView<Cobj> Aview;
Vector<Aview> AcceleratorViewContainer;
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View());
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
Aview *Aview_p = & AcceleratorViewContainer[0];
const int Nsimd = CComplex::Nsimd();
typedef decltype(coalescedRead(in_v[0])) calcVector;
typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
GridStopWatch ArithmeticTimer;
int osites=Grid()->oSites();
// double flops = osites*Nsimd*nbasis*nbasis*8.0*geom.npoint;
// double bytes = osites*nbasis*nbasis*geom.npoint*sizeof(CComplex);
double usecs =-usecond();
// assert(geom.npoint==9);
accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
int ss = sss/nbasis;
@@ -580,41 +316,28 @@ public:
int ptype;
StencilEntry *SE;
int lane=SIMTlane(Nsimd);
for(int point=0;point<geom.npoint;point++){
SE=Stencil.GetEntry(ptype,point,ss);
if(SE->_is_local) {
nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute,lane);
nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
} else {
nbr = coalescedRead(Stencil.CommBuf()[SE->_offset],lane);
nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
}
synchronise();
acceleratorSynchronise();
for(int bb=0;bb<nbasis;bb++) {
res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
}
}
coalescedWrite(out_v[ss](b),res,lane);
coalescedWrite(out_v[ss](b),res);
});
usecs +=usecond();
double nrm_usec=-usecond();
RealD Nout= norm2(out);
nrm_usec+=usecond();
/*
std::cout << GridLogMessage << "\tNorm " << nrm_usec << " us" <<std::endl;
std::cout << GridLogMessage << "\tHalo " << comms_usec << " us" <<std::endl;
std::cout << GridLogMessage << "\tMatrix " << usecs << " us" <<std::endl;
std::cout << GridLogMessage << "\t mflop/s " << flops/usecs<<std::endl;
std::cout << GridLogMessage << "\t MB/s " << bytes/usecs<<std::endl;
*/
return Nout;
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
};
RealD Mdag (const CoarseVector &in, CoarseVector &out)
void Mdag (const CoarseVector &in, CoarseVector &out)
{
if(hermitian) {
// corresponds to Petrov-Galerkin coarsening
@@ -625,7 +348,6 @@ public:
G5C(tmp, in);
M(tmp, out);
G5C(out, out);
return norm2(out);
}
};
void MdirComms(const CoarseVector &in)
@@ -640,11 +362,11 @@ public:
typedef LatticeView<Cobj> Aview;
Vector<Aview> AcceleratorViewContainer;
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View());
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
Aview *Aview_p = & AcceleratorViewContainer[0];
auto out_v = out.View();
auto in_v = in.View();
autoView( out_v , out, AcceleratorWrite);
autoView( in_v , in, AcceleratorRead);
const int Nsimd = CComplex::Nsimd();
typedef decltype(coalescedRead(in_v[0])) calcVector;
@@ -658,45 +380,21 @@ public:
int ptype;
StencilEntry *SE;
int lane=SIMTlane(Nsimd);
SE=Stencil.GetEntry(ptype,point,ss);
if(SE->_is_local) {
nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute,lane);
nbr = coalescedReadPermute(in_v[SE->_offset],ptype,SE->_permute);
} else {
nbr = coalescedRead(Stencil.CommBuf()[SE->_offset],lane);
nbr = coalescedRead(Stencil.CommBuf()[SE->_offset]);
}
synchronise();
acceleratorSynchronise();
for(int bb=0;bb<nbasis;bb++) {
res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
}
coalescedWrite(out_v[ss](b),res,lane);
coalescedWrite(out_v[ss](b),res);
});
#if 0
accelerator_for(ss,Grid()->oSites(),1,{
siteVector res = Zero();
siteVector nbr;
int ptype;
StencilEntry *SE;
SE=Stencil.GetEntry(ptype,point,ss);
if(SE->_is_local&&SE->_permute) {
permute(nbr,in_v[SE->_offset],ptype);
} else if(SE->_is_local) {
nbr = in_v[SE->_offset];
} else {
nbr = Stencil.CommBuf()[SE->_offset];
}
synchronise();
res = res + Aview_p[point][ss]*nbr;
out_v[ss]=res;
});
#endif
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
}
void MdirAll(const CoarseVector &in,std::vector<CoarseVector> &out)
{
@@ -864,14 +562,12 @@ public:
blockMaskedInnerProduct(oZProj,omask,Subspace.subspace[j],Mphi);
auto iZProj_v = iZProj.View() ;
auto oZProj_v = oZProj.View() ;
auto A_p = A[p].View();
auto A_self = A[self_stencil].View();
autoView( iZProj_v , iZProj, AcceleratorRead) ;
autoView( oZProj_v , oZProj, AcceleratorRead) ;
autoView( A_p , A[p], AcceleratorWrite);
autoView( A_self , A[self_stencil], AcceleratorWrite);
accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });
// if( disp!= 0 ) { accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });}
// accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_self[ss](j,i),A_self(ss)(j,i)+iZProj_v(ss)); });
}
}
@@ -885,11 +581,11 @@ public:
mult(tmp,phi,oddmask ); linop.Op(tmp,Mphio);
{
auto tmp_ = tmp.View();
auto evenmask_ = evenmask.View();
auto oddmask_ = oddmask.View();
auto Mphie_ = Mphie.View();
auto Mphio_ = Mphio.View();
autoView( tmp_ , tmp, AcceleratorWrite);
autoView( evenmask_ , evenmask, AcceleratorRead);
autoView( oddmask_ , oddmask, AcceleratorRead);
autoView( Mphie_ , Mphie, AcceleratorRead);
autoView( Mphio_ , Mphio, AcceleratorRead);
accelerator_for(ss, FineGrid->oSites(), Fobj::Nsimd(),{
coalescedWrite(tmp_[ss],evenmask_(ss)*Mphie_(ss) + oddmask_(ss)*Mphio_(ss));
});
@@ -897,8 +593,8 @@ public:
blockProject(SelfProj,tmp,Subspace.subspace);
auto SelfProj_ = SelfProj.View();
auto A_self = A[self_stencil].View();
autoView( SelfProj_ , SelfProj, AcceleratorRead);
autoView( A_self , A[self_stencil], AcceleratorWrite);
accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{
for(int j=0;j<nbasis;j++){
@@ -912,33 +608,8 @@ public:
std::cout << GridLogMessage << " ForceHermitian, new code "<<std::endl;
ForceHermitian();
}
// AssertHermitian();
// ForceDiagonal();
}
#if 0
///////////////////////////
// test code worth preserving in if block
///////////////////////////
std::cout<<GridLogMessage<< " Computed matrix elements "<< self_stencil <<std::endl;
for(int p=0;p<geom.npoint;p++){
std::cout<<GridLogMessage<< "A["<<p<<"]" << std::endl;
std::cout<<GridLogMessage<< A[p] << std::endl;
}
std::cout<<GridLogMessage<< " picking by block0 "<< self_stencil <<std::endl;
phi=Subspace.subspace[0];
std::vector<int> bc(FineGrid->_ndimension,0);
blockPick(Grid(),phi,tmp,bc); // Pick out a block
linop.Op(tmp,Mphi); // Apply big dop
blockProject(iProj,Mphi,Subspace.subspace); // project it and print it
std::cout<<GridLogMessage<< " Computed matrix elements from block zero only "<<std::endl;
std::cout<<GridLogMessage<< iProj <<std::endl;
std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
#endif
void ForceHermitian(void) {
CoarseMatrix Diff (Grid());
for(int p=0;p<geom.npoint;p++){
@@ -958,27 +629,6 @@ public:
}
}
}
void AssertHermitian(void) {
CoarseMatrix AA (Grid());
CoarseMatrix AAc (Grid());
CoarseMatrix Diff (Grid());
for(int d=0;d<4;d++){
int dd=d+1;
AAc = Cshift(A[2*d+1],dd,1);
AA = A[2*d];
Diff = AA - adj(AAc);
std::cout<<GridLogMessage<<"Norm diff dim "<<d<<" "<< norm2(Diff)<<std::endl;
std::cout<<GridLogMessage<<"Norm dim "<<d<<" "<< norm2(AA)<<std::endl;
}
Diff = A[8] - adj(A[8]);
std::cout<<GridLogMessage<<"Norm diff local "<< norm2(Diff)<<std::endl;
std::cout<<GridLogMessage<<"Norm local "<< norm2(A[8])<<std::endl;
}
};
NAMESPACE_END(Grid);

21
Grid/algorithms/FFT.h
View File

@@ -1,4 +1,3 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@@ -37,7 +36,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#endif
#endif
NAMESPACE_BEGIN(Grid);
template<class scalar> struct FFTW { };
@@ -191,7 +189,7 @@ public:
typedef typename sobj::scalar_type scalar;
Lattice<sobj> pgbuf(&pencil_g);
auto pgbuf_v = pgbuf.View();
autoView(pgbuf_v , pgbuf, CpuWrite);
typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
typedef typename FFTW<scalar>::FFTW_plan FFTW_plan;
@@ -232,15 +230,18 @@ public:
result = source;
int pc = processor_coor[dim];
for(int p=0;p<processors[dim];p++) {
{
autoView(r_v,result,CpuRead);
autoView(p_v,pgbuf,CpuWrite);
thread_for(idx, sgrid->lSites(),{
Coordinate cbuf(Nd);
sobj s;
sgrid->LocalIndexToLocalCoor(idx,cbuf);
peekLocalSite(s,result,cbuf);
peekLocalSite(s,r_v,cbuf);
cbuf[dim]+=((pc+p) % processors[dim])*L;
// cbuf[dim]+=p*L;
pokeLocalSite(s,pgbuf,cbuf);
pokeLocalSite(s,p_v,cbuf);
});
}
if (p != processors[dim] - 1) {
result = Cshift(result,dim,L);
}
@@ -269,15 +270,19 @@ public:
flops+= flops_call*NN;
// writing out result
{
autoView(pgbuf_v,pgbuf,CpuRead);
autoView(result_v,result,CpuWrite);
thread_for(idx,sgrid->lSites(),{
Coordinate clbuf(Nd), cgbuf(Nd);
sobj s;
sgrid->LocalIndexToLocalCoor(idx,clbuf);
cgbuf = clbuf;
cgbuf[dim] = clbuf[dim]+L*pc;
peekLocalSite(s,pgbuf,cgbuf);
pokeLocalSite(s,result,clbuf);
peekLocalSite(s,pgbuf_v,cgbuf);
pokeLocalSite(s,result_v,clbuf);
});
}
result = result*div;
// destroying plan

243
Grid/algorithms/LinearOperator.h
View File

@@ -43,7 +43,6 @@ NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////////////////
template<class Field> class LinearOperatorBase {
public:
// Support for coarsening to a multigrid
virtual void OpDiag (const Field &in, Field &out) = 0; // Abstract base
virtual void OpDir (const Field &in, Field &out,int dir,int disp) = 0; // Abstract base
@@ -94,7 +93,10 @@ public:
_Mat.Mdag(in,out);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
_Mat.MdagM(in,out,n1,n2);
_Mat.MdagM(in,out);
ComplexD dot = innerProduct(in,out);
n1=real(dot);
n2=norm2(out);
}
void HermOp(const Field &in, Field &out){
_Mat.MdagM(in,out);
@@ -131,17 +133,14 @@ public:
assert(0);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
_Mat.MdagM(in,out,n1,n2);
out = out + _shift*in;
ComplexD dot;
dot= innerProduct(in,out);
HermOp(in,out);
ComplexD dot = innerProduct(in,out);
n1=real(dot);
n2=norm2(out);
}
void HermOp(const Field &in, Field &out){
RealD n1,n2;
HermOpAndNorm(in,out,n1,n2);
_Mat.MdagM(in,out);
out = out + _shift*in;
}
};
@@ -170,7 +169,7 @@ public:
_Mat.M(in,out);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
_Mat.M(in,out);
HermOp(in,out);
ComplexD dot= innerProduct(in,out); n1=real(dot);
n2=norm2(out);
}
@@ -208,29 +207,32 @@ public:
}
};
//////////////////////////////////////////////////////////
// Even Odd Schur decomp operators; there are several
// ways to introduce the even odd checkerboarding
//////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////
// Even Odd Schur decomp operators; there are several
// ways to introduce the even odd checkerboarding
//////////////////////////////////////////////////////////
template<class Field>
class SchurOperatorBase : public LinearOperatorBase<Field> {
template<class Field>
class SchurOperatorBase : public LinearOperatorBase<Field> {
public:
virtual RealD Mpc (const Field &in, Field &out) =0;
virtual RealD MpcDag (const Field &in, Field &out) =0;
virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
virtual void Mpc (const Field &in, Field &out) =0;
virtual void MpcDag (const Field &in, Field &out) =0;
virtual void MpcDagMpc(const Field &in, Field &out) {
Field tmp(in.Grid());
tmp.Checkerboard() = in.Checkerboard();
ni=Mpc(in,tmp);
no=MpcDag(tmp,out);
Mpc(in,tmp);
MpcDag(tmp,out);
}
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
out.Checkerboard() = in.Checkerboard();
MpcDagMpc(in,out,n1,n2);
MpcDagMpc(in,out);
ComplexD dot= innerProduct(in,out);
n1=real(dot);
n2=norm2(out);
}
virtual void HermOp(const Field &in, Field &out){
RealD n1,n2;
HermOpAndNorm(in,out,n1,n2);
out.Checkerboard() = in.Checkerboard();
MpcDagMpc(in,out);
}
void Op (const Field &in, Field &out){
Mpc(in,out);
@@ -248,72 +250,66 @@ public:
void OpDirAll (const Field &in, std::vector<Field> &out){
assert(0);
};
};
template<class Matrix,class Field>
};
template<class Matrix,class Field>
class SchurDiagMooeeOperator : public SchurOperatorBase<Field> {
public:
Matrix &_Mat;
SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD Mpc (const Field &in, Field &out) {
virtual void Mpc (const Field &in, Field &out) {
Field tmp(in.Grid());
tmp.Checkerboard() = !in.Checkerboard();
//std::cout <<"grid pointers: in._grid="<< in._grid << " out._grid=" << out._grid << " _Mat.Grid=" << _Mat.Grid() << " _Mat.RedBlackGrid=" << _Mat.RedBlackGrid() << std::endl;
_Mat.Meooe(in,tmp);
_Mat.MooeeInv(tmp,out);
_Mat.Meooe(out,tmp);
//std::cout << "cb in " << in.Checkerboard() << " cb out " << out.Checkerboard() << std::endl;
_Mat.Mooee(in,out);
return axpy_norm(out,-1.0,tmp,out);
axpy(out,-1.0,tmp,out);
}
virtual RealD MpcDag (const Field &in, Field &out){
virtual void MpcDag (const Field &in, Field &out){
Field tmp(in.Grid());
_Mat.MeooeDag(in,tmp);
_Mat.MooeeInvDag(tmp,out);
_Mat.MeooeDag(out,tmp);
_Mat.MooeeDag(in,out);
return axpy_norm(out,-1.0,tmp,out);
axpy(out,-1.0,tmp,out);
}
};
template<class Matrix,class Field>
};
template<class Matrix,class Field>
class SchurDiagOneOperator : public SchurOperatorBase<Field> {
protected:
Matrix &_Mat;
public:
SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD Mpc (const Field &in, Field &out) {
virtual void Mpc (const Field &in, Field &out) {
Field tmp(in.Grid());
_Mat.Meooe(in,out);
_Mat.MooeeInv(out,tmp);
_Mat.Meooe(tmp,out);
_Mat.MooeeInv(out,tmp);
return axpy_norm(out,-1.0,tmp,in);
axpy(out,-1.0,tmp,in);
}
virtual RealD MpcDag (const Field &in, Field &out){
virtual void MpcDag (const Field &in, Field &out){
Field tmp(in.Grid());
_Mat.MooeeInvDag(in,out);
_Mat.MeooeDag(out,tmp);
_Mat.MooeeInvDag(tmp,out);
_Mat.MeooeDag(out,tmp);
return axpy_norm(out,-1.0,tmp,in);
axpy(out,-1.0,tmp,in);
}
};
template<class Matrix,class Field>
};
template<class Matrix,class Field>
class SchurDiagTwoOperator : public SchurOperatorBase<Field> {
protected:
Matrix &_Mat;
public:
SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD Mpc (const Field &in, Field &out) {
virtual void Mpc (const Field &in, Field &out) {
Field tmp(in.Grid());
_Mat.MooeeInv(in,out);
@@ -321,9 +317,9 @@ public:
_Mat.MooeeInv(tmp,out);
_Mat.Meooe(out,tmp);
return axpy_norm(out,-1.0,tmp,in);
axpy(out,-1.0,tmp,in);
}
virtual RealD MpcDag (const Field &in, Field &out){
virtual void MpcDag (const Field &in, Field &out){
Field tmp(in.Grid());
_Mat.MeooeDag(in,out);
@@ -331,21 +327,21 @@ public:
_Mat.MeooeDag(tmp,out);
_Mat.MooeeInvDag(out,tmp);
return axpy_norm(out,-1.0,tmp,in);
axpy(out,-1.0,tmp,in);
}
};
};
template<class Field>
class NonHermitianSchurOperatorBase : public LinearOperatorBase<Field>
{
template<class Field>
class NonHermitianSchurOperatorBase : public LinearOperatorBase<Field>
{
public:
virtual RealD Mpc (const Field& in, Field& out) = 0;
virtual RealD MpcDag (const Field& in, Field& out) = 0;
virtual void MpcDagMpc(const Field& in, Field& out, RealD& ni, RealD& no) {
virtual void Mpc (const Field& in, Field& out) = 0;
virtual void MpcDag (const Field& in, Field& out) = 0;
virtual void MpcDagMpc(const Field& in, Field& out) {
Field tmp(in.Grid());
tmp.Checkerboard() = in.Checkerboard();
ni = Mpc(in,tmp);
no = MpcDag(tmp,out);
Mpc(in,tmp);
MpcDag(tmp,out);
}
virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
assert(0);
@@ -366,15 +362,18 @@ public:
void OpDir(const Field& in, Field& out, int dir, int disp) {
assert(0);
}
void OpDirAll(const Field& in, std::vector<Field>& out){
assert(0);
};
};
template<class Matrix, class Field>
class NonHermitianSchurDiagMooeeOperator : public NonHermitianSchurOperatorBase<Field>
{
template<class Matrix, class Field>
class NonHermitianSchurDiagMooeeOperator : public NonHermitianSchurOperatorBase<Field>
{
public:
Matrix& _Mat;
NonHermitianSchurDiagMooeeOperator(Matrix& Mat): _Mat(Mat){};
virtual RealD Mpc(const Field& in, Field& out) {
virtual void Mpc(const Field& in, Field& out) {
Field tmp(in.Grid());
tmp.Checkerboard() = !in.Checkerboard();
@@ -384,9 +383,9 @@ public:
_Mat.Mooee(in, out);
return axpy_norm(out, -1.0, tmp, out);
axpy(out, -1.0, tmp, out);
}
virtual RealD MpcDag(const Field& in, Field& out) {
virtual void MpcDag(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MeooeDag(in, tmp);
@@ -395,19 +394,19 @@ public:
_Mat.MooeeDag(in, out);
return axpy_norm(out, -1.0, tmp, out);
axpy(out, -1.0, tmp, out);
}
};
};
template<class Matrix,class Field>
class NonHermitianSchurDiagOneOperator : public NonHermitianSchurOperatorBase<Field>
{
template<class Matrix,class Field>
class NonHermitianSchurDiagOneOperator : public NonHermitianSchurOperatorBase<Field>
{
protected:
Matrix &_Mat;
public:
NonHermitianSchurDiagOneOperator (Matrix& Mat): _Mat(Mat){};
virtual RealD Mpc(const Field& in, Field& out) {
virtual void Mpc(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.Meooe(in, out);
@@ -415,9 +414,9 @@ public:
_Mat.Meooe(tmp, out);
_Mat.MooeeInv(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
axpy(out, -1.0, tmp, in);
}
virtual RealD MpcDag(const Field& in, Field& out) {
virtual void MpcDag(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MooeeInvDag(in, out);
@@ -425,20 +424,20 @@ public:
_Mat.MooeeInvDag(tmp, out);
_Mat.MeooeDag(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
axpy(out, -1.0, tmp, in);
}
};
};
template<class Matrix, class Field>
class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Field>
{
template<class Matrix, class Field>
class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Field>
{
protected:
Matrix& _Mat;
public:
NonHermitianSchurDiagTwoOperator(Matrix& Mat): _Mat(Mat){};
virtual RealD Mpc(const Field& in, Field& out) {
virtual void Mpc(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MooeeInv(in, out);
@@ -446,9 +445,9 @@ public:
_Mat.MooeeInv(tmp, out);
_Mat.Meooe(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
axpy(out, -1.0, tmp, in);
}
virtual RealD MpcDag(const Field& in, Field& out) {
virtual void MpcDag(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MeooeDag(in, out);
@@ -456,90 +455,57 @@ public:
_Mat.MeooeDag(tmp, out);
_Mat.MooeeInvDag(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
axpy(out, -1.0, tmp, in);
}
};
};
///////////////////////////////////////////////////////////////////////////////////////////////////
// Left handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta --> ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
// Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta --> ( 1 - Moe Mee^-1 Meo Moo^-1) phi=eta ; psi = Moo^-1 phi
///////////////////////////////////////////////////////////////////////////////////////////////////
template<class Matrix,class Field> using SchurDiagOneRH = SchurDiagTwoOperator<Matrix,Field> ;
template<class Matrix,class Field> using SchurDiagOneLH = SchurDiagOneOperator<Matrix,Field> ;
///////////////////////////////////////////////////////////////////////////////////////////////////
// Staggered use
///////////////////////////////////////////////////////////////////////////////////////////////////
template<class Matrix,class Field>
class SchurStaggeredOperator : public SchurOperatorBase<Field> {
///////////////////////////////////////////////////////////////////////////////////////////////////
// Left handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta --> ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
// Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta --> ( 1 - Moe Mee^-1 Meo Moo^-1) phi=eta ; psi = Moo^-1 phi
///////////////////////////////////////////////////////////////////////////////////////////////////
template<class Matrix,class Field> using SchurDiagOneRH = SchurDiagTwoOperator<Matrix,Field> ;
template<class Matrix,class Field> using SchurDiagOneLH = SchurDiagOneOperator<Matrix,Field> ;
///////////////////////////////////////////////////////////////////////////////////////////////////
// Staggered use
///////////////////////////////////////////////////////////////////////////////////////////////////
template<class Matrix,class Field>
class SchurStaggeredOperator : public SchurOperatorBase<Field> {
protected:
Matrix &_Mat;
Field tmp;
RealD mass;
double tMpc;
double tIP;
double tMeo;
double taxpby_norm;
uint64_t ncall;
public:
void Report(void)
{
std::cout << GridLogMessage << " HermOpAndNorm.Mpc "<< tMpc/ncall<<" usec "<<std::endl;
std::cout << GridLogMessage << " HermOpAndNorm.IP "<< tIP /ncall<<" usec "<<std::endl;
std::cout << GridLogMessage << " Mpc.MeoMoe "<< tMeo/ncall<<" usec "<<std::endl;
std::cout << GridLogMessage << " Mpc.axpby_norm "<< taxpby_norm/ncall<<" usec "<<std::endl;
}
public:
SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid())
{
assert( _Mat.isTrivialEE() );
mass = _Mat.Mass();
tMpc=0;
tIP =0;
tMeo=0;
taxpby_norm=0;
ncall=0;
}
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
ncall++;
tMpc-=usecond();
n2 = Mpc(in,out);
tMpc+=usecond();
tIP-=usecond();
Mpc(in,out);
ComplexD dot= innerProduct(in,out);
tIP+=usecond();
n1 = real(dot);
n2 =0.0;
}
virtual void HermOp(const Field &in, Field &out){
ncall++;
tMpc-=usecond();
_Mat.Meooe(in,out);
_Mat.Meooe(out,tmp);
tMpc+=usecond();
taxpby_norm-=usecond();
axpby(out,-1.0,mass*mass,tmp,in);
taxpby_norm+=usecond();
Mpc(in,out);
// _Mat.Meooe(in,out);
// _Mat.Meooe(out,tmp);
// axpby(out,-1.0,mass*mass,tmp,in);
}
virtual RealD Mpc (const Field &in, Field &out)
virtual void 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.Mooee(in,out);
// _Mat.Mooee(out,tmp);
tMeo-=usecond();
_Mat.Meooe(in,out);
_Mat.Meooe(out,tmp);
tMeo+=usecond();
taxpby_norm-=usecond();
RealD nn=axpby_norm(out,-1.0,mass*mass,tmp,in);
taxpby_norm+=usecond();
return nn;
axpby(out,-1.0,mass*mass,tmp,in);
}
virtual RealD MpcDag (const Field &in, Field &out){
return Mpc(in,out);
virtual void MpcDag (const Field &in, Field &out){
Mpc(in,out);
}
virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
assert(0);// Never need with staggered
@@ -547,7 +513,6 @@ public:
};
template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
/////////////////////////////////////////////////////////////
// Base classes for functions of operators
/////////////////////////////////////////////////////////////

14
Grid/algorithms/SparseMatrix.h
View File

@@ -38,16 +38,12 @@ template<class Field> class SparseMatrixBase {
public:
virtual GridBase *Grid(void) =0;
// Full checkerboar operations
virtual RealD M (const Field &in, Field &out)=0;
virtual RealD Mdag (const Field &in, Field &out)=0;
virtual void MdagM(const Field &in, Field &out,RealD &ni,RealD &no) {
Field tmp (in.Grid());
ni=M(in,tmp);
no=Mdag(tmp,out);
}
virtual void M (const Field &in, Field &out)=0;
virtual void Mdag (const Field &in, Field &out)=0;
virtual void MdagM(const Field &in, Field &out) {
RealD ni, no;
MdagM(in,out,ni,no);
Field tmp (in.Grid());
M(in,tmp);
Mdag(tmp,out);
}
virtual void Mdiag (const Field &in, Field &out)=0;
virtual void Mdir (const Field &in, Field &out,int dir, int disp)=0;

24
Grid/algorithms/approx/Chebyshev.h
View File

@@ -234,10 +234,8 @@ public:
GridBase *grid=in.Grid();
// std::cout << "Chevyshef(): in.Grid()="<<in.Grid()<<std::endl;
//std::cout <<" Linop.Grid()="<<Linop.Grid()<<"Linop.RedBlackGrid()="<<Linop.RedBlackGrid()<<std::endl;
int vol=grid->gSites();
typedef typename Field::vector_type vector_type;
Field T0(grid); T0 = in;
Field T1(grid);
@@ -260,12 +258,26 @@ public:
for(int n=2;n<order;n++){
Linop.HermOp(*Tn,y);
// y=xscale*y+mscale*(*Tn);
// *Tnp=2.0*y-(*Tnm);
// out=out+Coeffs[n]* (*Tnp);
#if 0
auto y_v = y.View();
auto Tn_v = Tn->View();
auto Tnp_v = Tnp->View();
auto Tnm_v = Tnm->View();
constexpr int Nsimd = vector_type::Nsimd();
accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
});
if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#else
axpby(y,xscale,mscale,y,(*Tn));
axpby(*Tnp,2.0,-1.0,y,(*Tnm));
if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#endif
// Cycle pointers to avoid copies
Field *swizzle = Tnm;
Tnm =Tn;

28
Grid/algorithms/iterative/BiCGSTAB.h
View File

@@ -122,12 +122,14 @@ class BiCGSTAB : public OperatorFunction<Field>
LinearCombTimer.Start();
bo = beta * omega;
auto p_v = p.View();
auto r_v = r.View();
auto v_v = v.View();
{
autoView( p_v , p, AcceleratorWrite);
autoView( r_v , r, AcceleratorRead);
autoView( v_v , v, AcceleratorRead);
accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
});
}
LinearCombTimer.Stop();
LinalgTimer.Stop();
@@ -142,16 +144,20 @@ class BiCGSTAB : public OperatorFunction<Field>
alpha = rho / Calpha.real();
LinearCombTimer.Start();
auto h_v = h.View();
auto psi_v = psi.View();
{
autoView( p_v , p, AcceleratorRead);
autoView( r_v , r, AcceleratorRead);
autoView( v_v , v, AcceleratorRead);
autoView( psi_v,psi, AcceleratorRead);
autoView( h_v , h, AcceleratorWrite);
autoView( s_v , s, AcceleratorWrite);
accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
});
auto s_v = s.View();
accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
});
}
LinearCombTimer.Stop();
LinalgTimer.Stop();
@@ -166,11 +172,17 @@ class BiCGSTAB : public OperatorFunction<Field>
omega = Comega.real() / norm2(t);
LinearCombTimer.Start();
auto t_v = t.View();
{
autoView( psi_v,psi, AcceleratorWrite);
autoView( r_v , r, AcceleratorWrite);
autoView( h_v , h, AcceleratorRead);
autoView( s_v , s, AcceleratorRead);
autoView( t_v , t, AcceleratorRead);
accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));
});
}
LinearCombTimer.Stop();
cp = norm2(r);

8
Grid/algorithms/iterative/ConjugateGradient.h
View File

@@ -140,13 +140,15 @@ public:
b = cp / c;
LinearCombTimer.Start();
auto psi_v = psi.View();
auto p_v = p.View();
auto r_v = r.View();
{
autoView( psi_v , psi, AcceleratorWrite);
autoView( p_v , p, AcceleratorWrite);
autoView( r_v , r, AcceleratorWrite);
accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(psi_v[ss], a * p_v(ss) + psi_v(ss));
coalescedWrite(p_v[ss] , b * p_v(ss) + r_v (ss));
});
}
LinearCombTimer.Stop();
LinalgTimer.Stop();

205
Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@@ -37,211 +37,6 @@ Author: Christoph Lehner <clehner@bnl.gov>
NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////
// Move following 100 LOC to lattice/Lattice_basis.h
////////////////////////////////////////////////////////
template<class Field>
void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k)
{
// If assume basis[j] are already orthonormal,
// can take all inner products in parallel saving 2x bandwidth
// Save 3x bandwidth on the second line of loop.
// perhaps 2.5x speed up.
// 2x overall in Multigrid Lanczos
for(int j=0; j<k; ++j){
auto ip = innerProduct(basis[j],w);
w = w - ip*basis[j];
}
}
template<class Field>
void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm)
{
typedef decltype(basis[0].View()) View;
auto tmp_v = basis[0].View();
Vector<View> basis_v(basis.size(),tmp_v);
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0].Grid();
for(int k=0;k<basis.size();k++){
basis_v[k] = basis[k].View();
}
#if 0
std::vector < vobj , commAllocator<vobj> > Bt(thread_max() * Nm); // Thread private
thread_region
{
vobj* B = Bt.data() + Nm * thread_num();
thread_for_in_region(ss, grid->oSites(),{
for(int j=j0; j<j1; ++j) B[j]=0.;
for(int j=j0; j<j1; ++j){
for(int k=k0; k<k1; ++k){
B[j] +=Qt(j,k) * basis_v[k][ss];
}
}
for(int j=j0; j<j1; ++j){
basis_v[j][ss] = B[j];
}
});
}
#else
int nrot = j1-j0;
uint64_t oSites =grid->oSites();
uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
// printf("BasisRotate %d %d nrot %d siteBlock %d\n",j0,j1,nrot,siteBlock);
Vector <vobj> Bt(siteBlock * nrot);
auto Bp=&Bt[0];
// GPU readable copy of Eigen matrix
Vector<double> Qt_jv(Nm*Nm);
double *Qt_p = & Qt_jv[0];
for(int k=0;k<Nm;++k){
for(int j=0;j<Nm;++j){
Qt_p[j*Nm+k]=Qt(j,k);
}
}
// Block the loop to keep storage footprint down
vobj zz=Zero();
for(uint64_t s=0;s<oSites;s+=siteBlock){
// remaining work in this block
int ssites=MIN(siteBlock,oSites-s);
// zero out the accumulators
accelerator_for(ss,siteBlock*nrot,vobj::Nsimd(),{
auto z=coalescedRead(zz);
coalescedWrite(Bp[ss],z);
});
accelerator_for(sj,ssites*nrot,vobj::Nsimd(),{
int j =sj%nrot;
int jj =j0+j;
int ss =sj/nrot;
int sss=ss+s;
for(int k=k0; k<k1; ++k){
auto tmp = coalescedRead(Bp[ss*nrot+j]);
coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_v[k][sss]));
}
});
accelerator_for(sj,ssites*nrot,vobj::Nsimd(),{
int j =sj%nrot;
int jj =j0+j;
int ss =sj/nrot;
int sss=ss+s;
coalescedWrite(basis_v[jj][sss],coalescedRead(Bp[ss*nrot+j]));
});
}
#endif
}
// Extract a single rotated vector
template<class Field>
void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm)
{
typedef decltype(basis[0].View()) View;
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0].Grid();
result.Checkerboard() = basis[0].Checkerboard();
auto result_v=result.View();
Vector<View> basis_v(basis.size(),result_v);
for(int k=0;k<basis.size();k++){
basis_v[k] = basis[k].View();
}
vobj zz=Zero();
Vector<double> Qt_jv(Nm);
double * Qt_j = & Qt_jv[0];
for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
auto B=coalescedRead(zz);
for(int k=k0; k<k1; ++k){
B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
}
coalescedWrite(result_v[ss], B);
});
}
template<class Field>
void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx)
{
int vlen = idx.size();
assert(vlen>=1);
assert(vlen<=sort_vals.size());
assert(vlen<=_v.size());
for (size_t i=0;i<vlen;i++) {
if (idx[i] != i) {
//////////////////////////////////////
// idx[i] is a table of desired sources giving a permutation.
// Swap v[i] with v[idx[i]].
// Find j>i for which _vnew[j] = _vold[i],
// track the move idx[j] => idx[i]
// track the move idx[i] => i
//////////////////////////////////////
size_t j;
for (j=i;j<idx.size();j++)
if (idx[j]==i)
break;
assert(idx[i] > i); assert(j!=idx.size()); assert(idx[j]==i);
swap(_v[i],_v[idx[i]]); // should use vector move constructor, no data copy
std::swap(sort_vals[i],sort_vals[idx[i]]);
idx[j] = idx[i];
idx[i] = i;
}
}
}
inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals)
{
std::vector<int> idx(sort_vals.size());
std::iota(idx.begin(), idx.end(), 0);
// sort indexes based on comparing values in v
std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
});
return idx;
}
template<class Field>
void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse)
{
std::vector<int> idx = basisSortGetIndex(sort_vals);
if (reverse)
std::reverse(idx.begin(), idx.end());
basisReorderInPlace(_v,sort_vals,idx);
}
// PAB: faster to compute the inner products first then fuse loops.
// If performance critical can improve.
template<class Field>
void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
result = Zero();
assert(_v.size()==eval.size());
int N = (int)_v.size();
for (int i=0;i<N;i++) {
Field& tmp = _v[i];
axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
}
}
/////////////////////////////////////////////////////////////
// Implicitly restarted lanczos
/////////////////////////////////////////////////////////////

241
Grid/algorithms/iterative/PrecGeneralisedConjugateResidualNonHermitian.h Normal file
View File

@@ -0,0 +1,241 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/PrecGeneralisedConjugateResidual.h
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 */
#ifndef GRID_PREC_GCR_NON_HERM_H
#define GRID_PREC_GCR_NON_HERM_H
///////////////////////////////////////////////////////////////////////////////////////////////////////
//VPGCR Abe and Zhang, 2005.
//INTERNATIONAL JOURNAL OF NUMERICAL ANALYSIS AND MODELING
//Computing and Information Volume 2, Number 2, Pages 147-161
//NB. Likely not original reference since they are focussing on a preconditioner variant.
// but VPGCR was nicely written up in their paper
///////////////////////////////////////////////////////////////////////////////////////////////////////
NAMESPACE_BEGIN(Grid);
#define GCRLogLevel std::cout << GridLogMessage <<std::string(level,'\t')<< " Level "<<level<<" "
template<class Field>
class PrecGeneralisedConjugateResidualNonHermitian : public LinearFunction<Field> {
public:
RealD Tolerance;
Integer MaxIterations;
int verbose;
int mmax;
int nstep;
int steps;
int level;
GridStopWatch PrecTimer;
GridStopWatch MatTimer;
GridStopWatch LinalgTimer;
LinearFunction<Field> &Preconditioner;
LinearOperatorBase<Field> &Linop;
void Level(int lv) { level=lv; };
PrecGeneralisedConjugateResidualNonHermitian(RealD tol,Integer maxit,LinearOperatorBase<Field> &_Linop,LinearFunction<Field> &Prec,int _mmax,int _nstep) :
Tolerance(tol),
MaxIterations(maxit),
Linop(_Linop),
Preconditioner(Prec),
mmax(_mmax),
nstep(_nstep)
{
level=1;
verbose=1;
};
void operator() (const Field &src, Field &psi){
psi=Zero();
RealD cp, ssq,rsq;
ssq=norm2(src);
rsq=Tolerance*Tolerance*ssq;
Field r(src.Grid());
PrecTimer.Reset();
MatTimer.Reset();
LinalgTimer.Reset();
GridStopWatch SolverTimer;
SolverTimer.Start();
steps=0;
for(int k=0;k<MaxIterations;k++){
cp=GCRnStep(src,psi,rsq);
GCRLogLevel <<"PGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<" target "<<rsq <<std::endl;
if(cp<rsq) {
SolverTimer.Stop();
Linop.Op(psi,r);
axpy(r,-1.0,src,r);
RealD tr = norm2(r);
GCRLogLevel<<"PGCR: Converged on iteration " <<steps
<< " computed residual "<<sqrt(cp/ssq)
<< " true residual " <<sqrt(tr/ssq)
<< " target " <<Tolerance <<std::endl;
GCRLogLevel<<"PGCR Time elapsed: Total "<< SolverTimer.Elapsed() <<std::endl;
return;
}
}
GCRLogLevel<<"Variable Preconditioned GCR did not converge"<<std::endl;
// assert(0);
}
RealD GCRnStep(const Field &src, Field &psi,RealD rsq){
RealD cp;
ComplexD a, b, zAz;
RealD zAAz;
ComplexD rq;
GridBase *grid = src.Grid();
Field r(grid);
Field z(grid);
Field tmp(grid);
Field ttmp(grid);
Field Az(grid);
////////////////////////////////
// history for flexible orthog
////////////////////////////////
std::vector<Field> q(mmax,grid);
std::vector<Field> p(mmax,grid);
std::vector<RealD> qq(mmax);
GCRLogLevel<< "PGCR nStep("<<nstep<<")"<<std::endl;
//////////////////////////////////
// initial guess x0 is taken as nonzero.
// r0=src-A x0 = src
//////////////////////////////////
MatTimer.Start();
Linop.Op(psi,Az);
zAz = innerProduct(Az,psi);
zAAz= norm2(Az);
MatTimer.Stop();
LinalgTimer.Start();
r=src-Az;
LinalgTimer.Stop();
GCRLogLevel<< "PGCR true residual r = src - A psi "<<norm2(r) <<std::endl;
/////////////////////
// p = Prec(r)
/////////////////////
PrecTimer.Start();
Preconditioner(r,z);
PrecTimer.Stop();
MatTimer.Start();
Linop.Op(z,Az);
MatTimer.Stop();
LinalgTimer.Start();
zAz = innerProduct(Az,psi);
zAAz= norm2(Az);
//p[0],q[0],qq[0]
p[0]= z;
q[0]= Az;
qq[0]= zAAz;
cp =norm2(r);
LinalgTimer.Stop();
for(int k=0;k<nstep;k++){
steps++;
int kp = k+1;
int peri_k = k %mmax;
int peri_kp= kp%mmax;
LinalgTimer.Start();
rq= innerProduct(q[peri_k],r); // what if rAr not real?
a = rq/qq[peri_k];
axpy(psi,a,p[peri_k],psi);
cp = axpy_norm(r,-a,q[peri_k],r);
LinalgTimer.Stop();
GCRLogLevel<< "PGCR step["<<steps<<"] resid " << cp << " target " <<rsq<<std::endl;
if((k==nstep-1)||(cp<rsq)){
return cp;
}
PrecTimer.Start();
Preconditioner(r,z);// solve Az = r
PrecTimer.Stop();
MatTimer.Start();
Linop.Op(z,Az);
MatTimer.Stop();
zAz = innerProduct(Az,psi);
zAAz= norm2(Az);
LinalgTimer.Start();
q[peri_kp]=Az;
p[peri_kp]=z;
int northog = ((kp)>(mmax-1))?(mmax-1):(kp); // if more than mmax done, we orthog all mmax history.
for(int back=0;back<northog;back++){
int peri_back=(k-back)%mmax; assert((k-back)>=0);
b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
p[peri_kp]=p[peri_kp]+b*p[peri_back];
q[peri_kp]=q[peri_kp]+b*q[peri_back];
}
qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
LinalgTimer.Stop();
}
assert(0); // never reached
return cp;
}
};
NAMESPACE_END(Grid);
#endif

66
Grid/allocator/AlignedAllocator.cc
View File

@@ -6,72 +6,6 @@ NAMESPACE_BEGIN(Grid);
MemoryStats *MemoryProfiler::stats = nullptr;
bool MemoryProfiler::debug = false;
#ifdef GRID_NVCC
#define SMALL_LIMIT (0)
#else
#define SMALL_LIMIT (4096)
#endif
#ifdef POINTER_CACHE
int PointerCache::victim;
PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
void *PointerCache::Insert(void *ptr,size_t bytes) {
if (bytes < SMALL_LIMIT ) return ptr;
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif
void * ret = NULL;
int v = -1;
for(int e=0;e<Ncache;e++) {
if ( Entries[e].valid==0 ) {
v=e;
break;
}
}
if ( v==-1 ) {
v=victim;
victim = (victim+1)%Ncache;
}
if ( Entries[v].valid ) {
ret = Entries[v].address;
Entries[v].valid = 0;
Entries[v].address = NULL;
Entries[v].bytes = 0;
}
Entries[v].address=ptr;
Entries[v].bytes =bytes;
Entries[v].valid =1;
return ret;
}
void *PointerCache::Lookup(size_t bytes) {
if (bytes < SMALL_LIMIT ) return NULL;
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif
for(int e=0;e<Ncache;e++){
if ( Entries[e].valid && ( Entries[e].bytes == bytes ) ) {
Entries[e].valid = 0;
return Entries[e].address;
}
}
return NULL;
}
#endif
void check_huge_pages(void *Buf,uint64_t BYTES)
{
#ifdef __linux__

228
Grid/allocator/AlignedAllocator.h
View File

@@ -26,118 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_ALIGNED_ALLOCATOR_H
#define GRID_ALIGNED_ALLOCATOR_H
#ifdef HAVE_MALLOC_MALLOC_H
#include <malloc/malloc.h>
#endif
#ifdef HAVE_MALLOC_H
#include <malloc.h>
#endif
#ifdef HAVE_MM_MALLOC_H
#include <mm_malloc.h>
#endif
#define POINTER_CACHE
#define GRID_ALLOC_ALIGN (2*1024*1024)
#pragma once
NAMESPACE_BEGIN(Grid);
// Move control to configure.ac and Config.h?
#ifdef POINTER_CACHE
class PointerCache {
private:
/*Pinning pages is costly*/
/*Could maintain separate large and small allocation caches*/
#ifdef GRID_NVCC
static const int Ncache=128;
#else
static const int Ncache=8;
#endif
static int victim;
typedef struct {
void *address;
size_t bytes;
int valid;
} PointerCacheEntry;
static PointerCacheEntry Entries[Ncache];
public:
static void *Insert(void *ptr,size_t bytes) ;
static void *Lookup(size_t bytes) ;
};
#endif
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);
////////////////////////////////////////////////////////////////////
// A lattice of something, but assume the something is SIMDized.
////////////////////////////////////////////////////////////////////
template<typename _Tp>
class alignedAllocator {
public:
@@ -161,70 +53,60 @@ public:
{
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
#ifdef POINTER_CACHE
_Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
#else
pointer ptr = nullptr;
#endif
#ifdef GRID_NVCC
////////////////////////////////////
// Unified (managed) memory
////////////////////////////////////
if ( ptr == (_Tp *) NULL ) {
// printf(" alignedAllocater cache miss %ld bytes ",bytes); BACKTRACEFP(stdout);
auto err = cudaMallocManaged((void **)&ptr,bytes);
if( err != cudaSuccess ) {
ptr = (_Tp *) NULL;
std::cerr << " cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
assert(0);
}
}
assert( ptr != (_Tp *)NULL);
#else
//////////////////////////////////////////////////////////////////////////////////////////
// 2MB align; could make option probably doesn't need configurability
//////////////////////////////////////////////////////////////////////////////////////////
#ifdef HAVE_MM_MALLOC_H
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
#else
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
#endif
assert( ptr != (_Tp *)NULL);
//////////////////////////////////////////////////
// First touch optimise in threaded loop
//////////////////////////////////////////////////
uint64_t *cp = (uint64_t *)ptr;
thread_for(n,bytes/sizeof(uint64_t), { // need only one touch per page
cp[n]=0;
});
#endif
_Tp *ptr = (_Tp*) MemoryManager::CpuAllocate(bytes);
assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
return ptr;
}
void deallocate(pointer __p, size_type __n) {
void deallocate(pointer __p, size_type __n)
{
size_type bytes = __n * sizeof(_Tp);
profilerFree(bytes);
MemoryManager::CpuFree((void *)__p,bytes);
}
#ifdef POINTER_CACHE
pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
#else
pointer __freeme = __p;
#endif
// FIXME: hack for the copy constructor, eventually it must be avoided
//void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
void construct(pointer __p, const _Tp& __val) { assert(0);};
void construct(pointer __p) { };
void destroy(pointer __p) { };
};
template<typename _Tp> inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
template<typename _Tp> inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
#ifdef GRID_NVCC
if ( __freeme ) cudaFree((void *)__freeme);
#else
#ifdef HAVE_MM_MALLOC_H
if ( __freeme ) _mm_free((void *)__freeme);
#else
if ( __freeme ) free((void *)__freeme);
#endif
#endif
template<typename _Tp>
class uvmAllocator {
public:
typedef std::size_t size_type;
typedef std::ptrdiff_t difference_type;
typedef _Tp* pointer;
typedef const _Tp* const_pointer;
typedef _Tp& reference;
typedef const _Tp& const_reference;
typedef _Tp value_type;
template<typename _Tp1> struct rebind { typedef uvmAllocator<_Tp1> other; };
uvmAllocator() throw() { }
uvmAllocator(const uvmAllocator&) throw() { }
template<typename _Tp1> uvmAllocator(const uvmAllocator<_Tp1>&) throw() { }
~uvmAllocator() throw() { }
pointer address(reference __x) const { return &__x; }
size_type max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
pointer allocate(size_type __n, const void* _p= 0)
{
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
_Tp *ptr = (_Tp*) MemoryManager::SharedAllocate(bytes);
assert( ( (_Tp*)ptr != (_Tp *)NULL ) );
return ptr;
}
void deallocate(pointer __p, size_type __n)
{
size_type bytes = __n * sizeof(_Tp);
profilerFree(bytes);
MemoryManager::SharedFree((void *)__p,bytes);
}
// FIXME: hack for the copy constructor, eventually it must be avoided
@@ -233,17 +115,17 @@ public:
void construct(pointer __p) { };
void destroy(pointer __p) { };
};
template<typename _Tp> inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
template<typename _Tp> inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
template<typename _Tp> inline bool operator==(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return true; }
template<typename _Tp> inline bool operator!=(const uvmAllocator<_Tp>&, const uvmAllocator<_Tp>&){ return false; }
////////////////////////////////////////////////////////////////////////////////
// Template typedefs
////////////////////////////////////////////////////////////////////////////////
template<class T> using commAllocator = alignedAllocator<T>;
template<class T> using Vector = std::vector<T,alignedAllocator<T> >;
template<class T> using commVector = std::vector<T,alignedAllocator<T> >;
template<class T> using Matrix = std::vector<std::vector<T,alignedAllocator<T> > >;
template<class T> using commAllocator = uvmAllocator<T>;
template<class T> using Vector = std::vector<T,uvmAllocator<T> >;
template<class T> using commVector = std::vector<T,uvmAllocator<T> >;
//template<class T> using Matrix = std::vector<std::vector<T,alignedAllocator<T> > >;
NAMESPACE_END(Grid);
#endif

4
Grid/allocator/Allocator.h Normal file
View File

@@ -0,0 +1,4 @@
#pragma once
#include <Grid/allocator/MemoryStats.h>
#include <Grid/allocator/MemoryManager.h>
#include <Grid/allocator/AlignedAllocator.h>

244
Grid/allocator/MemoryManager.cc Normal file
View File

@@ -0,0 +1,244 @@
#include <Grid/GridCore.h>
NAMESPACE_BEGIN(Grid);
/*Allocation types, saying which pointer cache should be used*/
#define Cpu (0)
#define CpuSmall (1)
#define Acc (2)
#define AccSmall (3)
#define Shared (4)
#define SharedSmall (5)
uint64_t total_shared;
uint64_t total_device;
uint64_t total_host;;
void MemoryManager::PrintBytes(void)
{
std::cout << " MemoryManager : "<<total_shared<<" shared bytes "<<std::endl;
std::cout << " MemoryManager : "<<total_device<<" accelerator bytes "<<std::endl;
std::cout << " MemoryManager : "<<total_host <<" cpu bytes "<<std::endl;
}
//////////////////////////////////////////////////////////////////////
// Data tables for recently freed pooiniter caches
//////////////////////////////////////////////////////////////////////
MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
int MemoryManager::Victim[MemoryManager::NallocType];
int MemoryManager::Ncache[MemoryManager::NallocType] = { 8, 32, 8, 32, 8, 32 };
//////////////////////////////////////////////////////////////////////
// Actual allocation and deallocation utils
//////////////////////////////////////////////////////////////////////
void *MemoryManager::AcceleratorAllocate(size_t bytes)
{
void *ptr = (void *) Lookup(bytes,Acc);
if ( ptr == (void *) NULL ) {
ptr = (void *) acceleratorAllocDevice(bytes);
total_device+=bytes;
}
return ptr;
}
void MemoryManager::AcceleratorFree (void *ptr,size_t bytes)
{
void *__freeme = Insert(ptr,bytes,Acc);
if ( __freeme ) {
acceleratorFreeDevice(__freeme);
total_device-=bytes;
// PrintBytes();
}
}
void *MemoryManager::SharedAllocate(size_t bytes)
{
void *ptr = (void *) Lookup(bytes,Shared);
if ( ptr == (void *) NULL ) {
ptr = (void *) acceleratorAllocShared(bytes);
total_shared+=bytes;
// std::cout <<"AcceleratorAllocate: allocated Shared pointer "<<std::hex<<ptr<<std::dec<<std::endl;
// PrintBytes();
}
return ptr;
}
void MemoryManager::SharedFree (void *ptr,size_t bytes)
{
void *__freeme = Insert(ptr,bytes,Shared);
if ( __freeme ) {
acceleratorFreeShared(__freeme);
total_shared-=bytes;
// PrintBytes();
}
}
#ifdef GRID_UVM
void *MemoryManager::CpuAllocate(size_t bytes)
{
void *ptr = (void *) Lookup(bytes,Cpu);
if ( ptr == (void *) NULL ) {
ptr = (void *) acceleratorAllocShared(bytes);
total_host+=bytes;
}
return ptr;
}
void MemoryManager::CpuFree (void *_ptr,size_t bytes)
{
NotifyDeletion(_ptr);
void *__freeme = Insert(_ptr,bytes,Cpu);
if ( __freeme ) {
acceleratorFreeShared(__freeme);
total_host-=bytes;
}
}
#else
void *MemoryManager::CpuAllocate(size_t bytes)
{
void *ptr = (void *) Lookup(bytes,Cpu);
if ( ptr == (void *) NULL ) {
ptr = (void *) acceleratorAllocCpu(bytes);
total_host+=bytes;
}
return ptr;
}
void MemoryManager::CpuFree (void *_ptr,size_t bytes)
{
NotifyDeletion(_ptr);
void *__freeme = Insert(_ptr,bytes,Cpu);
if ( __freeme ) {
acceleratorFreeCpu(__freeme);
total_host-=bytes;
}
}
#endif
//////////////////////////////////////////
// call only once
//////////////////////////////////////////
void MemoryManager::Init(void)
{
char * str;
int Nc;
int NcS;
str= getenv("GRID_ALLOC_NCACHE_LARGE");
if ( str ) {
Nc = atoi(str);
if ( (Nc>=0) && (Nc < NallocCacheMax)) {
Ncache[Cpu]=Nc;
Ncache[Acc]=Nc;
Ncache[Shared]=Nc;
}
}
str= getenv("GRID_ALLOC_NCACHE_SMALL");
if ( str ) {
Nc = atoi(str);
if ( (Nc>=0) && (Nc < NallocCacheMax)) {
Ncache[CpuSmall]=Nc;
Ncache[AccSmall]=Nc;
Ncache[SharedSmall]=Nc;
}
}
std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
#ifdef ALLOCATION_CACHE
std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
#endif
#ifdef GRID_UVM
std::cout << GridLogMessage<< "MemoryManager::Init() Unified memory space"<<std::endl;
#ifdef GRID_CUDA
std::cout << GridLogMessage<< "MemoryManager::Init() Using cudaMallocManaged"<<std::endl;
#endif
#ifdef GRID_HIP
std::cout << GridLogMessage<< "MemoryManager::Init() Using hipMallocManaged"<<std::endl;
#endif
#ifdef GRID_SYCL
std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_shared"<<std::endl;
#endif
#else
std::cout << GridLogMessage<< "MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory"<<std::endl;
#ifdef GRID_CUDA
std::cout << GridLogMessage<< "MemoryManager::Init() Using cudaMalloc"<<std::endl;
#endif
#ifdef GRID_HIP
std::cout << GridLogMessage<< "MemoryManager::Init() Using hipMalloc"<<std::endl;
#endif
#ifdef GRID_SYCL
std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_device"<<std::endl;
#endif
#endif
}
void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
{
#ifdef ALLOCATION_CACHE
bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
int cache = type + small;
return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache]);
#else
return ptr;
#endif
}
void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim)
{
assert(ncache>0);
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif
void * ret = NULL;
int v = -1;
for(int e=0;e<ncache;e++) {
if ( entries[e].valid==0 ) {
v=e;
break;
}
}
if ( v==-1 ) {
v=victim;
victim = (victim+1)%ncache;
}
if ( entries[v].valid ) {
ret = entries[v].address;
entries[v].valid = 0;
entries[v].address = NULL;
entries[v].bytes = 0;
}
entries[v].address=ptr;
entries[v].bytes =bytes;
entries[v].valid =1;
return ret;
}
void *MemoryManager::Lookup(size_t bytes,int type)
{
#ifdef ALLOCATION_CACHE
bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
int cache = type+small;
return Lookup(bytes,Entries[cache],Ncache[cache]);
#else
return NULL;
#endif
}
void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache)
{
assert(ncache>0);
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif
for(int e=0;e<ncache;e++){
if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
entries[e].valid = 0;
return entries[e].address;
}
}
return NULL;
}
NAMESPACE_END(Grid);

181
Grid/allocator/MemoryManager.h Normal file
View File

@@ -0,0 +1,181 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/MemoryManager.h
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 */
#pragma once
#include <list>
#include <unordered_map>
NAMESPACE_BEGIN(Grid);
// Move control to configure.ac and Config.h?
#define ALLOCATION_CACHE
#define GRID_ALLOC_ALIGN (2*1024*1024)
#define GRID_ALLOC_SMALL_LIMIT (4096)
/*Pinning pages is costly*/
////////////////////////////////////////////////////////////////////////////
// Advise the LatticeAccelerator class
////////////////////////////////////////////////////////////////////////////
enum ViewAdvise {
AdviseDefault = 0x0, // Regular data
AdviseInfrequentUse = 0x1 // Advise that the data is used infrequently. This can
// significantly influence performance of bulk storage.
// AdviseTransient = 0x2, // Data will mostly be read. On some architectures
// enables read-only copies of memory to be kept on
// host and device.
// AdviseAcceleratorWriteDiscard = 0x4 // Field will be written in entirety on device
};
////////////////////////////////////////////////////////////////////////////
// View Access Mode
////////////////////////////////////////////////////////////////////////////
enum ViewMode {
AcceleratorRead = 0x01,
AcceleratorWrite = 0x02,
AcceleratorWriteDiscard = 0x04,
CpuRead = 0x08,
CpuWrite = 0x10,
CpuWriteDiscard = 0x10 // same for now
};
class MemoryManager {
private:
////////////////////////////////////////////////////////////
// For caching recently freed allocations
////////////////////////////////////////////////////////////
typedef struct {
void *address;
size_t bytes;
int valid;
} AllocationCacheEntry;
static const int NallocCacheMax=128;
static const int NallocType=6;
static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
static int Victim[NallocType];
static int Ncache[NallocType];
/////////////////////////////////////////////////
// Free pool
/////////////////////////////////////////////////
static void *Insert(void *ptr,size_t bytes,int type) ;
static void *Lookup(size_t bytes,int type) ;
static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim) ;
static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache) ;
static void *AcceleratorAllocate(size_t bytes);
static void AcceleratorFree (void *ptr,size_t bytes);
static void PrintBytes(void);
public:
static void Init(void);
static void *SharedAllocate(size_t bytes);
static void SharedFree (void *ptr,size_t bytes);
static void *CpuAllocate(size_t bytes);
static void CpuFree (void *ptr,size_t bytes);
////////////////////////////////////////////////////////
// Footprint tracking
////////////////////////////////////////////////////////
static uint64_t DeviceBytes;
static uint64_t DeviceLRUBytes;
static uint64_t DeviceMaxBytes;
static uint64_t HostToDeviceBytes;
static uint64_t DeviceToHostBytes;
static uint64_t HostToDeviceXfer;
static uint64_t DeviceToHostXfer;
private:
#ifndef GRID_UVM
//////////////////////////////////////////////////////////////////////
// Data tables for ViewCache
//////////////////////////////////////////////////////////////////////
typedef std::list<uint64_t> LRU_t;
typedef typename LRU_t::iterator LRUiterator;
typedef struct {
int LRU_valid;
LRUiterator LRU_entry;
uint64_t CpuPtr;
uint64_t AccPtr;
size_t bytes;
uint32_t transient;
uint32_t state;
uint32_t accLock;
uint32_t cpuLock;
} AcceleratorViewEntry;
typedef std::unordered_map<uint64_t,AcceleratorViewEntry> AccViewTable_t;
typedef typename AccViewTable_t::iterator AccViewTableIterator ;
static AccViewTable_t AccViewTable;
static LRU_t LRU;
/////////////////////////////////////////////////
// Device motion
/////////////////////////////////////////////////
static void Create(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
static void EvictVictims(uint64_t bytes); // Frees up <bytes>
static void Evict(AcceleratorViewEntry &AccCache);
static void Flush(AcceleratorViewEntry &AccCache);
static void Clone(AcceleratorViewEntry &AccCache);
static void AccDiscard(AcceleratorViewEntry &AccCache);
static void CpuDiscard(AcceleratorViewEntry &AccCache);
// static void LRUupdate(AcceleratorViewEntry &AccCache);
static void LRUinsert(AcceleratorViewEntry &AccCache);
static void LRUremove(AcceleratorViewEntry &AccCache);
// manage entries in the table
static int EntryPresent(uint64_t CpuPtr);
static void EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
static void EntryErase (uint64_t CpuPtr);
static AccViewTableIterator EntryLookup(uint64_t CpuPtr);
static void EntrySet (uint64_t CpuPtr,AcceleratorViewEntry &entry);
static void AcceleratorViewClose(uint64_t AccPtr);
static uint64_t AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
static void CpuViewClose(uint64_t Ptr);
static uint64_t CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
#endif
static void NotifyDeletion(void * CpuPtr);
public:
static void Print(void);
static int isOpen (void* CpuPtr);
static void ViewClose(void* CpuPtr,ViewMode mode);
static void *ViewOpen (void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint);
};
NAMESPACE_END(Grid);

468
Grid/allocator/MemoryManagerCache.cc Normal file
View File

@@ -0,0 +1,468 @@
#include <Grid/GridCore.h>
#ifndef GRID_UVM
#warning "Using explicit device memory copies"
NAMESPACE_BEGIN(Grid);
#define dprintf(...)
////////////////////////////////////////////////////////////
// For caching copies of data on device
////////////////////////////////////////////////////////////
MemoryManager::AccViewTable_t MemoryManager::AccViewTable;
MemoryManager::LRU_t MemoryManager::LRU;
////////////////////////////////////////////////////////
// Footprint tracking
////////////////////////////////////////////////////////
uint64_t MemoryManager::DeviceBytes;
uint64_t MemoryManager::DeviceLRUBytes;
uint64_t MemoryManager::DeviceMaxBytes = 1024*1024*128;
uint64_t MemoryManager::HostToDeviceBytes;
uint64_t MemoryManager::DeviceToHostBytes;
uint64_t MemoryManager::HostToDeviceXfer;
uint64_t MemoryManager::DeviceToHostXfer;
////////////////////////////////////
// Priority ordering for unlocked entries
// Empty
// CpuDirty
// Consistent
// AccDirty
////////////////////////////////////
#define Empty (0x0) /*Entry unoccupied */
#define CpuDirty (0x1) /*CPU copy is golden, Acc buffer MAY not be allocated*/
#define Consistent (0x2) /*ACC copy AND CPU copy are valid */
#define AccDirty (0x4) /*ACC copy is golden */
#define EvictNext (0x8) /*Priority for eviction*/
/////////////////////////////////////////////////
// Mechanics of data table maintenance
/////////////////////////////////////////////////
int MemoryManager::EntryPresent(uint64_t CpuPtr)
{
if(AccViewTable.empty()) return 0;
auto count = AccViewTable.count(CpuPtr); assert((count==0)||(count==1));
return count;
}
void MemoryManager::EntryCreate(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
{
assert(!EntryPresent(CpuPtr));
AcceleratorViewEntry AccCache;
AccCache.CpuPtr = CpuPtr;
AccCache.AccPtr = (uint64_t)NULL;
AccCache.bytes = bytes;
AccCache.state = CpuDirty;
AccCache.LRU_valid=0;
AccCache.transient=0;
AccCache.accLock=0;
AccCache.cpuLock=0;
AccViewTable[CpuPtr] = AccCache;
}
MemoryManager::AccViewTableIterator MemoryManager::EntryLookup(uint64_t CpuPtr)
{
assert(EntryPresent(CpuPtr));
auto AccCacheIterator = AccViewTable.find(CpuPtr);
assert(AccCacheIterator!=AccViewTable.end());
return AccCacheIterator;
}
void MemoryManager::EntryErase(uint64_t CpuPtr)
{
auto AccCache = EntryLookup(CpuPtr);
AccViewTable.erase(CpuPtr);
}
void MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
{
assert(AccCache.LRU_valid==0);
if (AccCache.transient) {
LRU.push_back(AccCache.CpuPtr);
AccCache.LRU_entry = --LRU.end();
} else {
LRU.push_front(AccCache.CpuPtr);
AccCache.LRU_entry = LRU.begin();
}
AccCache.LRU_valid = 1;
DeviceLRUBytes+=AccCache.bytes;
}
void MemoryManager::LRUremove(AcceleratorViewEntry &AccCache)
{
assert(AccCache.LRU_valid==1);
LRU.erase(AccCache.LRU_entry);
AccCache.LRU_valid = 0;
DeviceLRUBytes-=AccCache.bytes;
}
/////////////////////////////////////////////////
// Accelerator cache motion & consistency logic
/////////////////////////////////////////////////
void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
{
///////////////////////////////////////////////////////////
// Remove from Accelerator, remove entry, without flush
// Cannot be locked. If allocated Must be in LRU pool.
///////////////////////////////////////////////////////////
assert(AccCache.state!=Empty);
// dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);
assert(AccCache.accLock==0);
assert(AccCache.cpuLock==0);
assert(AccCache.CpuPtr!=(uint64_t)NULL);
if(AccCache.AccPtr) {
AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
DeviceBytes -=AccCache.bytes;
LRUremove(AccCache);
// dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);
}
uint64_t CpuPtr = AccCache.CpuPtr;
EntryErase(CpuPtr);
}
void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
{
///////////////////////////////////////////////////////////////////////////
// Make CPU consistent, remove from Accelerator, remove entry
// Cannot be locked. If allocated must be in LRU pool.
///////////////////////////////////////////////////////////////////////////
assert(AccCache.state!=Empty);
// dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);
assert(AccCache.accLock==0);
assert(AccCache.cpuLock==0);
if(AccCache.state==AccDirty) {
Flush(AccCache);
}
assert(AccCache.CpuPtr!=(uint64_t)NULL);
if(AccCache.AccPtr) {
AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
DeviceBytes -=AccCache.bytes;
LRUremove(AccCache);
// dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);
}
uint64_t CpuPtr = AccCache.CpuPtr;
EntryErase(CpuPtr);
}
void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
{
assert(AccCache.state==AccDirty);
assert(AccCache.cpuLock==0);
assert(AccCache.accLock==0);
assert(AccCache.AccPtr!=(uint64_t)NULL);
assert(AccCache.CpuPtr!=(uint64_t)NULL);
acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
// dprintf("MemoryManager: Flush %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
DeviceToHostBytes+=AccCache.bytes;
DeviceToHostXfer++;
AccCache.state=Consistent;
}
void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
{
assert(AccCache.state==CpuDirty);
assert(AccCache.cpuLock==0);
assert(AccCache.accLock==0);
assert(AccCache.CpuPtr!=(uint64_t)NULL);
if(AccCache.AccPtr==(uint64_t)NULL){
AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
DeviceBytes+=AccCache.bytes;
}
// dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
HostToDeviceBytes+=AccCache.bytes;
HostToDeviceXfer++;
AccCache.state=Consistent;
}
void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
{
assert(AccCache.state!=Empty);
assert(AccCache.cpuLock==0);
assert(AccCache.accLock==0);
assert(AccCache.CpuPtr!=(uint64_t)NULL);
if(AccCache.AccPtr==(uint64_t)NULL){
AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
DeviceBytes+=AccCache.bytes;
}
AccCache.state=AccDirty;
}
/////////////////////////////////////////////////////////////////////////////////
// View management
/////////////////////////////////////////////////////////////////////////////////
void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
{
if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
AcceleratorViewClose((uint64_t)Ptr);
} else if( (mode==CpuRead)||(mode==CpuWrite)){
CpuViewClose((uint64_t)Ptr);
} else {
assert(0);
}
}
void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
{
uint64_t CpuPtr = (uint64_t)_CpuPtr;
if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
} else if( (mode==CpuRead)||(mode==CpuWrite)){
return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
} else {
assert(0);
return NULL;
}
}
void MemoryManager::EvictVictims(uint64_t bytes)
{
while(bytes+DeviceLRUBytes > DeviceMaxBytes){
if ( DeviceLRUBytes > 0){
assert(LRU.size()>0);
uint64_t victim = LRU.back();
auto AccCacheIterator = EntryLookup(victim);
auto & AccCache = AccCacheIterator->second;
Evict(AccCache);
}
}
}
uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint)
{
////////////////////////////////////////////////////////////////////////////
// Find if present, otherwise get or force an empty
////////////////////////////////////////////////////////////////////////////
if ( EntryPresent(CpuPtr)==0 ){
EvictVictims(bytes);
EntryCreate(CpuPtr,bytes,mode,hint);
}
auto AccCacheIterator = EntryLookup(CpuPtr);
auto & AccCache = AccCacheIterator->second;
assert((mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard));
assert(AccCache.cpuLock==0); // Programming error
if(AccCache.state!=Empty) {
assert(AccCache.CpuPtr == CpuPtr);
assert(AccCache.bytes ==bytes);
}
/*
* State transitions and actions
*
* Action State StateNext Flush Clone
*
* AccRead Empty Consistent - Y
* AccWrite Empty AccDirty - Y
* AccRead CpuDirty Consistent - Y
* AccWrite CpuDirty AccDirty - Y
* AccRead Consistent Consistent - -
* AccWrite Consistent AccDirty - -
* AccRead AccDirty AccDirty - -
* AccWrite AccDirty AccDirty - -
*/
if(AccCache.state==Empty) {
assert(AccCache.LRU_valid==0);
AccCache.CpuPtr = CpuPtr;
AccCache.AccPtr = (uint64_t)NULL;
AccCache.bytes = bytes;
AccCache.state = CpuDirty; // Cpu starts primary
if(mode==AcceleratorWriteDiscard){
CpuDiscard(AccCache);
AccCache.state = AccDirty; // Empty + AcceleratorWrite=> AccDirty
} else if(mode==AcceleratorWrite){
Clone(AccCache);
AccCache.state = AccDirty; // Empty + AcceleratorWrite=> AccDirty
} else {
Clone(AccCache);
AccCache.state = Consistent; // Empty + AccRead => Consistent
}
AccCache.accLock= 1;
} else if(AccCache.state==CpuDirty ){
if(mode==AcceleratorWriteDiscard) {
CpuDiscard(AccCache);
AccCache.state = AccDirty; // CpuDirty + AcceleratorWrite=> AccDirty
} else if(mode==AcceleratorWrite) {
Clone(AccCache);
AccCache.state = AccDirty; // CpuDirty + AcceleratorWrite=> AccDirty
} else {
Clone(AccCache);
AccCache.state = Consistent; // CpuDirty + AccRead => Consistent
}
AccCache.accLock++;
// printf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
} else if(AccCache.state==Consistent) {
if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
AccCache.state = AccDirty; // Consistent + AcceleratorWrite=> AccDirty
else
AccCache.state = Consistent; // Consistent + AccRead => Consistent
AccCache.accLock++;
// printf("Consistent entry into device accLock %d\n",AccCache.accLock);
} else if(AccCache.state==AccDirty) {
if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
AccCache.state = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
else
AccCache.state = AccDirty; // AccDirty + AccRead => AccDirty
AccCache.accLock++;
// printf("AccDirty entry into device accLock %d\n",AccCache.accLock);
} else {
assert(0);
}
// If view is opened on device remove from LRU
if(AccCache.LRU_valid==1){
// must possibly remove from LRU as now locked on GPU
LRUremove(AccCache);
}
int transient =hint;
AccCache.transient= transient? EvictNext : 0;
return AccCache.AccPtr;
}
////////////////////////////////////
// look up & decrement lock count
////////////////////////////////////
void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
{
auto AccCacheIterator = EntryLookup(CpuPtr);
auto & AccCache = AccCacheIterator->second;
assert(AccCache.cpuLock==0);
assert(AccCache.accLock>0);
AccCache.accLock--;
// Move to LRU queue if not locked and close on device
if(AccCache.accLock==0) {
LRUinsert(AccCache);
}
}
void MemoryManager::CpuViewClose(uint64_t CpuPtr)
{
auto AccCacheIterator = EntryLookup(CpuPtr);
auto & AccCache = AccCacheIterator->second;
assert(AccCache.cpuLock>0);
assert(AccCache.accLock==0);
AccCache.cpuLock--;
}
/*
* Action State StateNext Flush Clone
*
* CpuRead Empty CpuDirty - -
* CpuWrite Empty CpuDirty - -
* CpuRead CpuDirty CpuDirty - -
* CpuWrite CpuDirty CpuDirty - -
* CpuRead Consistent Consistent - -
* CpuWrite Consistent CpuDirty - -
* CpuRead AccDirty Consistent Y -
* CpuWrite AccDirty CpuDirty Y -
*/
uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,ViewAdvise transient)
{
////////////////////////////////////////////////////////////////////////////
// Find if present, otherwise get or force an empty
////////////////////////////////////////////////////////////////////////////
if ( EntryPresent(CpuPtr)==0 ){
EvictVictims(bytes);
EntryCreate(CpuPtr,bytes,mode,transient);
}
auto AccCacheIterator = EntryLookup(CpuPtr);
auto & AccCache = AccCacheIterator->second;
assert((mode==CpuRead)||(mode==CpuWrite));
assert(AccCache.accLock==0); // Programming error
if(AccCache.state!=Empty) {
assert(AccCache.CpuPtr == CpuPtr);
assert(AccCache.bytes==bytes);
}
if(AccCache.state==Empty) {
AccCache.CpuPtr = CpuPtr;
AccCache.AccPtr = (uint64_t)NULL;
AccCache.bytes = bytes;
AccCache.state = CpuDirty; // Empty + CpuRead/CpuWrite => CpuDirty
AccCache.accLock= 0;
AccCache.cpuLock= 1;
} else if(AccCache.state==CpuDirty ){
// AccPtr dont care, deferred allocate
AccCache.state = CpuDirty; // CpuDirty +CpuRead/CpuWrite => CpuDirty
AccCache.cpuLock++;
} else if(AccCache.state==Consistent) {
assert(AccCache.AccPtr != (uint64_t)NULL);
if(mode==CpuWrite)
AccCache.state = CpuDirty; // Consistent +CpuWrite => CpuDirty
else
AccCache.state = Consistent; // Consistent +CpuRead => Consistent
AccCache.cpuLock++;
} else if(AccCache.state==AccDirty) {
assert(AccCache.AccPtr != (uint64_t)NULL);
Flush(AccCache);
if(mode==CpuWrite) AccCache.state = CpuDirty; // AccDirty +CpuWrite => CpuDirty, Flush
else AccCache.state = Consistent; // AccDirty +CpuRead => Consistent, Flush
AccCache.cpuLock++;
} else {
assert(0); // should be unreachable
}
AccCache.transient= transient? EvictNext : 0;
return AccCache.CpuPtr;
}
void MemoryManager::NotifyDeletion(void *_ptr)
{
// Look up in ViewCache
uint64_t ptr = (uint64_t)_ptr;
if(EntryPresent(ptr)) {
auto e = EntryLookup(ptr);
AccDiscard(e->second);
}
}
void MemoryManager::Print(void)
{
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogDebug << "Memory Manager " << std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogDebug << DeviceBytes << " bytes allocated on device " << std::endl;
std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device " << std::endl;
std::cout << GridLogDebug << HostToDeviceXfer << " transfers to device " << std::endl;
std::cout << GridLogDebug << DeviceToHostXfer << " transfers from device " << std::endl;
std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to device " << std::endl;
std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
auto &AccCache = it->second;
std::string str;
if ( AccCache.state==Empty ) str = std::string("Empty");
if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
if ( AccCache.state==Consistent)str = std::string("Consistent");
std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
<< "\t" << AccCache.cpuLock
<< "\t" << AccCache.accLock
<< "\t" << AccCache.LRU_valid<<std::endl;
}
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
};
int MemoryManager::isOpen (void* _CpuPtr)
{
uint64_t CpuPtr = (uint64_t)_CpuPtr;
if ( EntryPresent(CpuPtr) ){
auto AccCacheIterator = EntryLookup(CpuPtr);
auto & AccCache = AccCacheIterator->second;
return AccCache.cpuLock+AccCache.accLock;
} else {
return 0;
}
}
NAMESPACE_END(Grid);
#endif

24
Grid/allocator/MemoryManagerShared.cc Normal file
View File

@@ -0,0 +1,24 @@
#include <Grid/GridCore.h>
#ifdef GRID_UVM
#warning "Grid is assuming unified virtual memory address space"
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////
// View management is 1:1 address space mapping
/////////////////////////////////////////////////////////////////////////////////
uint64_t MemoryManager::DeviceBytes;
uint64_t MemoryManager::DeviceLRUBytes;
uint64_t MemoryManager::DeviceMaxBytes = 1024*1024*128;
uint64_t MemoryManager::HostToDeviceBytes;
uint64_t MemoryManager::DeviceToHostBytes;
uint64_t MemoryManager::HostToDeviceXfer;
uint64_t MemoryManager::DeviceToHostXfer;
void MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
int MemoryManager::isOpen (void* CpuPtr) { return 0;}
void MemoryManager::Print(void){};
void MemoryManager::NotifyDeletion(void *ptr){};
NAMESPACE_END(Grid);
#endif

67
Grid/allocator/MemoryStats.cc Normal file
View File

@@ -0,0 +1,67 @@
#include <Grid/GridCore.h>
#include <fcntl.h>
NAMESPACE_BEGIN(Grid);
MemoryStats *MemoryProfiler::stats = nullptr;
bool MemoryProfiler::debug = false;
void check_huge_pages(void *Buf,uint64_t BYTES)
{
#ifdef __linux__
int fd = open("/proc/self/pagemap", O_RDONLY);
assert(fd >= 0);
const int page_size = 4096;
uint64_t virt_pfn = (uint64_t)Buf / page_size;
off_t offset = sizeof(uint64_t) * virt_pfn;
uint64_t npages = (BYTES + page_size-1) / page_size;
uint64_t pagedata[npages];
uint64_t ret = lseek(fd, offset, SEEK_SET);
assert(ret == offset);
ret = ::read(fd, pagedata, sizeof(uint64_t)*npages);
assert(ret == sizeof(uint64_t) * npages);
int nhugepages = npages / 512;
int n4ktotal, nnothuge;
n4ktotal = 0;
nnothuge = 0;
for (int i = 0; i < nhugepages; ++i) {
uint64_t baseaddr = (pagedata[i*512] & 0x7fffffffffffffULL) * page_size;
for (int j = 0; j < 512; ++j) {
uint64_t pageaddr = (pagedata[i*512+j] & 0x7fffffffffffffULL) * page_size;
++n4ktotal;
if (pageaddr != baseaddr + j * page_size)
++nnothuge;
}
}
int rank = CartesianCommunicator::RankWorld();
printf("rank %d Allocated %d 4k pages, %d not in huge pages\n", rank, n4ktotal, nnothuge);
#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);
}
NAMESPACE_END(Grid);

95
Grid/allocator/MemoryStats.h Normal file
View File

@@ -0,0 +1,95 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/MemoryStats.h
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 */
#pragma once
NAMESPACE_BEGIN(Grid);
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);
NAMESPACE_END(Grid);

1
Grid/cartesian/Cartesian_base.h
View File

@@ -81,6 +81,7 @@ public:
bool _isCheckerBoarded;
int LocallyPeriodic;
Coordinate _checker_dim_mask;
public:

4
Grid/cartesian/Cartesian_full.h
View File

@@ -38,6 +38,7 @@ class GridCartesian: public GridBase {
public:
int dummy;
Coordinate _checker_dim_mask;
virtual int CheckerBoardFromOindexTable (int Oindex) {
return 0;
}
@@ -104,6 +105,7 @@ public:
_ldimensions.resize(_ndimension);
_rdimensions.resize(_ndimension);
_simd_layout.resize(_ndimension);
_checker_dim_mask.resize(_ndimension);;
_lstart.resize(_ndimension);
_lend.resize(_ndimension);
@@ -114,6 +116,8 @@ public:
for (int d = 0; d < _ndimension; d++)
{
_checker_dim_mask[d]=0;
_fdimensions[d] = dimensions[d]; // Global dimensions
_gdimensions[d] = _fdimensions[d]; // Global dimensions
_simd_layout[d] = simd_layout[d];

18
Grid/cartesian/Cartesian_red_black.h
View File

@@ -36,11 +36,27 @@ static const int CbBlack=1;
static const int Even =CbRed;
static const int Odd =CbBlack;
accelerator_inline int RedBlackCheckerBoardFromOindex (int oindex, Coordinate &rdim, Coordinate &chk_dim_msk)
{
int nd=rdim.size();
Coordinate coor(nd);
Lexicographic::CoorFromIndex(coor,oindex,rdim);
int linear=0;
for(int d=0;d<nd;d++){
if(chk_dim_msk[d])
linear=linear+coor[d];
}
return (linear&0x1);
}
// Specialise this for red black grids storing half the data like a chess board.
class GridRedBlackCartesian : public GridBase
{
public:
Coordinate _checker_dim_mask;
// Coordinate _checker_dim_mask;
int _checker_dim;
std::vector<int> _checker_board;

1
Grid/communicator/Communicator_base.h
View File

@@ -114,6 +114,7 @@ public:
void GlobalSumVector(RealD *,int N);
void GlobalSum(uint32_t &);
void GlobalSum(uint64_t &);
void GlobalSumVector(uint64_t*,int N);
void GlobalSum(ComplexF &c);
void GlobalSumVector(ComplexF *c,int N);
void GlobalSum(ComplexD &c);

4
Grid/communicator/Communicator_mpi3.cc
View File

@@ -275,6 +275,10 @@ void CartesianCommunicator::GlobalSum(uint64_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalSumVector(uint64_t* u,int N){
int ierr=MPI_Allreduce(MPI_IN_PLACE,u,N,MPI_UINT64_T,MPI_SUM,communicator);
assert(ierr==0);
}
void CartesianCommunicator::GlobalXOR(uint32_t &u){
int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
assert(ierr==0);

3
Grid/communicator/Communicator_none.cc
View File

@@ -70,9 +70,10 @@ CartesianCommunicator::~CartesianCommunicator(){}
void CartesianCommunicator::GlobalSum(float &){}
void CartesianCommunicator::GlobalSumVector(float *,int N){}
void CartesianCommunicator::GlobalSum(double &){}
void CartesianCommunicator::GlobalSumVector(double *,int N){}
void CartesianCommunicator::GlobalSum(uint32_t &){}
void CartesianCommunicator::GlobalSum(uint64_t &){}
void CartesianCommunicator::GlobalSumVector(double *,int N){}
void CartesianCommunicator::GlobalSumVector(uint64_t *,int N){}
void CartesianCommunicator::GlobalXOR(uint32_t &){}
void CartesianCommunicator::GlobalXOR(uint64_t &){}

4
Grid/communicator/SharedMemory.cc
View File

@@ -74,7 +74,9 @@ void *SharedMemory::ShmBufferMalloc(size_t bytes){
if (heap_bytes >= heap_size) {
std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
std::cout<< " Current value is " << (heap_size/(1024*1024)) <<std::endl;
std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
std::cout<< " Current bytes is " << (heap_bytes/(1024*1024)) <<"MB"<<std::endl;
std::cout<< " Current heap is " << (heap_size/(1024*1024)) <<"MB"<<std::endl;
assert(heap_bytes<heap_size);
}
//std::cerr << "ShmBufferMalloc "<<std::hex<< ptr<<" - "<<((uint64_t)ptr+bytes)<<std::dec<<std::endl;

24
Grid/communicator/SharedMemoryMPI.cc
View File

@@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
#include <pwd.h>
#ifdef GRID_NVCC
#ifdef GRID_CUDA
#include <cuda_runtime_api.h>
#endif
@@ -170,17 +170,24 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
std::vector<int> primes({2,3,5});
int dim = 0;
int last_dim = ndimension - 1;
int AutoShmSize = 1;
while(AutoShmSize != WorldShmSize) {
for(int p=0;p<primes.size();p++) {
int p;
for(p=0;p<primes.size();p++) {
int prime=primes[p];
if ( divides(prime,WorldDims[dim]/ShmDims[dim])
&& divides(prime,WorldShmSize/AutoShmSize) ) {
AutoShmSize*=prime;
ShmDims[dim]*=prime;
last_dim = dim;
break;
}
}
if (p == primes.size() && last_dim == dim) {
std::cerr << "GlobalSharedMemory::GetShmDims failed" << std::endl;
exit(EXIT_FAILURE);
}
dim=(dim+1) %ndimension;
}
}
@@ -413,7 +420,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbfs mapping intended
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_NVCC
#ifdef GRID_CUDA
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
void * ShmCommBuf ;
@@ -433,13 +440,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
#ifdef GRID_IBM_SUMMIT
// IBM Jsrun makes cuda Device numbering screwy and not match rank
std::cout << "IBM Summit or similar - NOT setting device to WorldShmRank"<<std::endl;
#else
std::cout << "setting device to WorldShmRank"<<std::endl;
cudaSetDevice(WorldShmRank);
#endif
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -677,7 +677,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
/////////////////////////////////////////////////////////////////////////
void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
{
#ifdef GRID_NVCC
#ifdef GRID_CUDA
cudaMemset(dest,0,bytes);
#else
bzero(dest,bytes);
@@ -685,7 +685,7 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
}
void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
{
#ifdef GRID_NVCC
#ifdef GRID_CUDA
cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
#else
bcopy(src,dest,bytes);

25
Grid/cshift/Cshift.h
View File

@@ -49,4 +49,29 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#ifdef GRID_COMMS_SHMEM
#include <Grid/cshift/Cshift_mpi.h> // uses same implementation of communicator
#endif
NAMESPACE_BEGIN(Grid);
template<typename Op, typename T1>
auto Cshift(const LatticeUnaryExpression<Op,T1> &expr,int dim,int shift)
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))>
{
return Cshift(closure(expr),dim,shift);
}
template <class Op, class T1, class T2>
auto Cshift(const LatticeBinaryExpression<Op,T1,T2> &expr,int dim,int shift)
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))>
{
return Cshift(closure(expr),dim,shift);
}
template <class Op, class T1, class T2, class T3>
auto Cshift(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr,int dim,int shift)
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1),
eval(0, expr.arg2),
eval(0, expr.arg3)))>
{
return Cshift(closure(expr),dim,shift);
}
NAMESPACE_END(Grid);
#endif

106
Grid/cshift/Cshift_common.h
View File

@@ -29,6 +29,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
NAMESPACE_BEGIN(Grid);
extern Vector<std::pair<int,int> > Cshift_table;
///////////////////////////////////////////////////////////////////
// Gather for when there is no need to SIMD split
///////////////////////////////////////////////////////////////////
@@ -46,16 +48,16 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
int e2=rhs.Grid()->_slice_block[dimension];
int ent = 0;
static Vector<std::pair<int,int> > table; table.resize(e1*e2);
if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
int stride=rhs.Grid()->_slice_stride[dimension];
auto rhs_v = rhs.View();
if ( cbmask == 0x3 ) {
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o = n*stride;
int bo = n*e2;
table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
Cshift_table[ent++] = std::pair<int,int>(off+bo+b,so+o+b);
}
}
} else {
@@ -65,14 +67,19 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
int o = n*stride;
int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
if ( ocb &cbmask ) {
table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
Cshift_table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
}
}
}
}
thread_for(i,ent,{
buffer[table[i].first]=rhs_v[table[i].second];
{
autoView(rhs_v , rhs, AcceleratorRead);
auto buffer_p = & buffer[0];
auto table = &Cshift_table[0];
accelerator_for(i,ent,1,{
buffer_p[table[i].first]=rhs_v[table[i].second];
});
}
}
///////////////////////////////////////////////////////////////////
@@ -95,35 +102,37 @@ Gather_plane_extract(const Lattice<vobj> &rhs,
int e2=rhs.Grid()->_slice_block[dimension];
int n1=rhs.Grid()->_slice_stride[dimension];
auto rhs_v = rhs.View();
if ( cbmask ==0x3){
thread_for_collapse(2,n,e1,{
for(int b=0;b<e2;b++){
autoView(rhs_v , rhs, AcceleratorRead);
accelerator_for2d(n,e1,b,e2,1,{
int o = n*n1;
int offset = b+n*e2;
vobj temp =rhs_v[so+o+b];
extract<vobj>(temp,pointers,offset);
}
});
} else {
autoView(rhs_v , rhs, AcceleratorRead);
// Case of SIMD split AND checker dim cannot currently be hit, except in
// Test_cshift_red_black code.
std::cout << " Dense packed buffer WARNING " <<std::endl;
thread_for_collapse(2,n,e1,{
for(int b=0;b<e2;b++){
Coordinate rdim=rhs.Grid()->_rdimensions;
Coordinate cdm =rhs.Grid()->_checker_dim_mask;
std::cout << " Dense packed buffer WARNING " <<std::endl; // Does this get called twice once for each cb?
accelerator_for2d(n,e1,b,e2,1,{
Coordinate coor;
int o=n*n1;
int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
int oindex = o+b;
int cb = RedBlackCheckerBoardFromOindex(oindex, rdim, cdm);
int ocb=1<<cb;
int offset = b+n*e2;
if ( ocb & cbmask ) {
vobj temp =rhs_v[so+o+b];
extract<vobj>(temp,pointers,offset);
}
}
});
}
}
@@ -145,7 +154,8 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
int e2=rhs.Grid()->_slice_block[dimension];
int stride=rhs.Grid()->_slice_stride[dimension];
static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
int ent =0;
if ( cbmask ==0x3 ) {
@@ -154,7 +164,7 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
for(int b=0;b<e2;b++){
int o =n*rhs.Grid()->_slice_stride[dimension];
int bo =n*rhs.Grid()->_slice_block[dimension];
table[ent++] = std::pair<int,int>(so+o+b,bo+b);
Cshift_table[ent++] = std::pair<int,int>(so+o+b,bo+b);
}
}
@@ -165,16 +175,20 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
int o =n*rhs.Grid()->_slice_stride[dimension];
int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
if ( ocb & cbmask ) {
table[ent++]=std::pair<int,int> (so+o+b,bo++);
Cshift_table[ent++]=std::pair<int,int> (so+o+b,bo++);
}
}
}
}
auto rhs_v = rhs.View();
thread_for(i,ent,{
rhs_v[table[i].first]=buffer[table[i].second];
{
autoView( rhs_v, rhs, AcceleratorWrite);
auto buffer_p = & buffer[0];
auto table = &Cshift_table[0];
accelerator_for(i,ent,1,{
rhs_v[table[i].first]=buffer_p[table[i].second];
});
}
}
//////////////////////////////////////////////////////
@@ -194,13 +208,11 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
int e2=rhs.Grid()->_slice_block[dimension];
if(cbmask ==0x3 ) {
auto rhs_v = rhs.View();
thread_for_collapse(2,n,e1,{
for(int b=0;b<e2;b++){
autoView( rhs_v , rhs, AcceleratorWrite);
accelerator_for2d(n,e1,b,e2,1,{
int o = n*rhs.Grid()->_slice_stride[dimension];
int offset = b+n*rhs.Grid()->_slice_block[dimension];
merge(rhs_v[so+o+b],pointers,offset);
}
});
} else {
@@ -208,7 +220,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
// Test_cshift_red_black code.
// std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
auto rhs_v = rhs.View();
autoView( rhs_v, rhs, CpuWrite);
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o = n*rhs.Grid()->_slice_stride[dimension];
@@ -225,6 +237,7 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
//////////////////////////////////////////////////////
// local to node block strided copies
//////////////////////////////////////////////////////
template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask)
{
int rd = rhs.Grid()->_rdimensions[dimension];
@@ -239,14 +252,16 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
int e1=rhs.Grid()->_slice_nblock[dimension]; // clearly loop invariant for icpc
int e2=rhs.Grid()->_slice_block[dimension];
int stride = rhs.Grid()->_slice_stride[dimension];
static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
int ent=0;
if(cbmask == 0x3 ){
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*stride+b;
table[ent++] = std::pair<int,int>(lo+o,ro+o);
Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
}
}
} else {
@@ -255,23 +270,24 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
int o =n*stride+b;
int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o);
if ( ocb&cbmask ) {
table[ent++] = std::pair<int,int>(lo+o,ro+o);
Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
}
}
}
}
auto rhs_v = rhs.View();
auto lhs_v = lhs.View();
thread_for(i,ent,{
{
autoView(rhs_v , rhs, AcceleratorRead);
autoView(lhs_v , lhs, AcceleratorWrite);
auto table = &Cshift_table[0];
accelerator_for(i,ent,1,{
lhs_v[table[i].first]=rhs_v[table[i].second];
});
}
}
template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask,int permute_type)
{
int rd = rhs.Grid()->_rdimensions[dimension];
if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
@@ -285,29 +301,33 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
int e2=rhs.Grid()->_slice_block [dimension];
int stride = rhs.Grid()->_slice_stride[dimension];
static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
if(Cshift_table.size()<e1*e2) Cshift_table.resize(e1*e2); // Let it grow to biggest
int ent=0;
if ( cbmask == 0x3 ) {
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*stride;
table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
}}
} else {
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*stride;
int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o+b);
if ( ocb&cbmask ) table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
if ( ocb&cbmask ) Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
}}
}
auto rhs_v = rhs.View();
auto lhs_v = lhs.View();
thread_for(i,ent,{
{
autoView( rhs_v, rhs, AcceleratorRead);
autoView( lhs_v, lhs, AcceleratorWrite);
auto table = &Cshift_table[0];
accelerator_for(i,ent,1,{
permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
});
}
}
//////////////////////////////////////////////////////

4
Grid/cshift/Cshift_table.cc Normal file
View File

@@ -0,0 +1,4 @@
#include <Grid/GridCore.h>
NAMESPACE_BEGIN(Grid);
Vector<std::pair<int,int> > Cshift_table;
NAMESPACE_END(Grid);

5
Grid/lattice/Lattice.h
View File

@@ -26,6 +26,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
*************************************************************************************/
/* END LEGAL */
#pragma once
#include <Grid/lattice/Lattice_view.h>
#include <Grid/lattice/Lattice_base.h>
#include <Grid/lattice/Lattice_conformable.h>
#include <Grid/lattice/Lattice_ET.h>
@@ -35,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/lattice/Lattice_local.h>
#include <Grid/lattice/Lattice_reduction.h>
#include <Grid/lattice/Lattice_peekpoke.h>
#include <Grid/lattice/Lattice_reality.h>
//#include <Grid/lattice/Lattice_reality.h>
#include <Grid/lattice/Lattice_comparison_utils.h>
#include <Grid/lattice/Lattice_comparison.h>
#include <Grid/lattice/Lattice_coordinate.h>
@@ -43,4 +44,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/lattice/Lattice_rng.h>
#include <Grid/lattice/Lattice_unary.h>
#include <Grid/lattice/Lattice_transfer.h>
#include <Grid/lattice/Lattice_basis.h>

77
Grid/lattice/Lattice_ET.h
View File

@@ -9,6 +9,7 @@ Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: neo <cossu@post.kek.jp>
Author: Christoph Lehner <christoph@lhnr.de
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
@@ -91,12 +92,18 @@ const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg)
{
return arg[ss];
}
// What needs this?
// Cannot be legal on accelerator
// Comparison must convert
#if 1
template <class lobj> accelerator_inline
const lobj & eval(const uint64_t ss, const Lattice<lobj> &arg)
{
auto view = arg.View();
auto view = arg.View(AcceleratorRead);
return view[ss];
}
#endif
///////////////////////////////////////////////////
// handle nodes in syntax tree- eval one operand
@@ -179,16 +186,12 @@ inline void CBFromExpression(int &cb, const T1 &lat) // Lattice leaf
cb = lat.Checkerboard();
}
template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
inline void CBFromExpression(int &cb, const T1 &notlat) // non-lattice leaf
{
}
inline void CBFromExpression(int &cb, const T1 &notlat) {} // non-lattice leaf
template <typename Op, typename T1> inline
void CBFromExpression(int &cb,const LatticeUnaryExpression<Op, T1> &expr)
{
CBFromExpression(cb, expr.arg1); // recurse AST
}
template <typename Op, typename T1, typename T2> inline
void CBFromExpression(int &cb,const LatticeBinaryExpression<Op, T1, T2> &expr)
{
@@ -203,6 +206,68 @@ inline void CBFromExpression(int &cb, const LatticeTrinaryExpression<Op, T1, T2,
CBFromExpression(cb, expr.arg3); // recurse AST
}
//////////////////////////////////////////////////////////////////////////
// ViewOpen
//////////////////////////////////////////////////////////////////////////
template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
inline void ExpressionViewOpen(T1 &lat) // Lattice leaf
{
lat.ViewOpen(AcceleratorRead);
}
template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
inline void ExpressionViewOpen(T1 &notlat) {}
template <typename Op, typename T1> inline
void ExpressionViewOpen(LatticeUnaryExpression<Op, T1> &expr)
{
ExpressionViewOpen(expr.arg1); // recurse AST
}
template <typename Op, typename T1, typename T2> inline
void ExpressionViewOpen(LatticeBinaryExpression<Op, T1, T2> &expr)
{
ExpressionViewOpen(expr.arg1); // recurse AST
ExpressionViewOpen(expr.arg2); // recurse AST
}
template <typename Op, typename T1, typename T2, typename T3>
inline void ExpressionViewOpen(LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
{
ExpressionViewOpen(expr.arg1); // recurse AST
ExpressionViewOpen(expr.arg2); // recurse AST
ExpressionViewOpen(expr.arg3); // recurse AST
}
//////////////////////////////////////////////////////////////////////////
// ViewClose
//////////////////////////////////////////////////////////////////////////
template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
inline void ExpressionViewClose( T1 &lat) // Lattice leaf
{
lat.ViewClose();
}
template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
inline void ExpressionViewClose(T1 &notlat) {}
template <typename Op, typename T1> inline
void ExpressionViewClose(LatticeUnaryExpression<Op, T1> &expr)
{
ExpressionViewClose(expr.arg1); // recurse AST
}
template <typename Op, typename T1, typename T2> inline
void ExpressionViewClose(LatticeBinaryExpression<Op, T1, T2> &expr)
{
ExpressionViewClose(expr.arg1); // recurse AST
ExpressionViewClose(expr.arg2); // recurse AST
}
template <typename Op, typename T1, typename T2, typename T3>
inline void ExpressionViewClose(LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
{
ExpressionViewClose(expr.arg1); // recurse AST
ExpressionViewClose(expr.arg2); // recurse AST
ExpressionViewClose(expr.arg3); // recurse AST
}
////////////////////////////////////////////
// Unary operators and funcs
////////////////////////////////////////////

69
Grid/lattice/Lattice_arith.h
View File

@@ -7,6 +7,7 @@
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Christoph Lehner <christoph@lhnr.de>
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
@@ -36,9 +37,9 @@ NAMESPACE_BEGIN(Grid);
template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = lhs.Checkerboard();
auto ret_v = ret.View();
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
autoView( rhs_v , rhs, AcceleratorRead);
conformable(ret,rhs);
conformable(lhs,rhs);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
@@ -55,9 +56,9 @@ void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
autoView( rhs_v , rhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
@@ -72,9 +73,9 @@ void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
autoView( rhs_v , rhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
@@ -88,9 +89,9 @@ void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
autoView( rhs_v , rhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
@@ -107,8 +108,8 @@ template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(lhs,ret);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
mult(&tmp,&lhs_v(ss),&rhs);
@@ -120,8 +121,8 @@ template<class obj1,class obj2,class obj3> inline
void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,lhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
@@ -134,8 +135,8 @@ template<class obj1,class obj2,class obj3> inline
void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,lhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
@@ -147,8 +148,8 @@ template<class obj1,class obj2,class obj3> inline
void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(lhs,ret);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
@@ -164,8 +165,8 @@ template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
auto ret_v = ret.View();
auto rhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( rhs_v , lhs, AcceleratorRead);
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto rhs_t=rhs_v(ss);
@@ -178,8 +179,8 @@ template<class obj1,class obj2,class obj3> inline
void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
auto ret_v = ret.View();
auto rhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( rhs_v , lhs, AcceleratorRead);
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto rhs_t=rhs_v(ss);
@@ -192,8 +193,8 @@ template<class obj1,class obj2,class obj3> inline
void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
auto ret_v = ret.View();
auto rhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( rhs_v , lhs, AcceleratorRead);
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto rhs_t=rhs_v(ss);
@@ -205,8 +206,8 @@ template<class obj1,class obj2,class obj3> inline
void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
auto ret_v = ret.View();
auto rhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( rhs_v , lhs, AcceleratorRead);
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto rhs_t=rhs_v(ss);
@@ -220,9 +221,9 @@ void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &
ret.Checkerboard() = x.Checkerboard();
conformable(ret,x);
conformable(x,y);
auto ret_v = ret.View();
auto x_v = x.View();
auto y_v = y.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( x_v , x, AcceleratorRead);
autoView( y_v , y, AcceleratorRead);
accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
auto tmp = a*x_v(ss)+y_v(ss);
coalescedWrite(ret_v[ss],tmp);
@@ -233,9 +234,9 @@ void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice
ret.Checkerboard() = x.Checkerboard();
conformable(ret,x);
conformable(x,y);
auto ret_v = ret.View();
auto x_v = x.View();
auto y_v = y.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( x_v , x, AcceleratorRead);
autoView( y_v , y, AcceleratorRead);
accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
auto tmp = a*x_v(ss)+b*y_v(ss);
coalescedWrite(ret_v[ss],tmp);

181
Grid/lattice/Lattice_base.h
View File

@@ -9,6 +9,7 @@ Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Christoph Lehner <christoph@lhnr.de>
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
@@ -28,6 +29,7 @@ See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#pragma once
#define STREAMING_STORES
@@ -36,129 +38,6 @@ NAMESPACE_BEGIN(Grid);
extern int GridCshiftPermuteMap[4][16];
///////////////////////////////////////////////////////////////////
// Base class which can be used by traits to pick up behaviour
///////////////////////////////////////////////////////////////////
class LatticeBase {};
/////////////////////////////////////////////////////////////////////////////////////////
// Conformable checks; same instance of Grid required
/////////////////////////////////////////////////////////////////////////////////////////
void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
{
assert(lhs == rhs);
}
////////////////////////////////////////////////////////////////////////////
// Minimal base class containing only data valid to access from accelerator
// _odata will be a managed pointer in CUDA
////////////////////////////////////////////////////////////////////////////
// Force access to lattice through a view object.
// prevents writing of code that will not offload to GPU, but perhaps annoyingly
// strict since host could could in principle direct access through the lattice object
// Need to decide programming model.
#define LATTICE_VIEW_STRICT
template<class vobj> class LatticeAccelerator : public LatticeBase
{
protected:
GridBase *_grid;
int checkerboard;
vobj *_odata; // A managed pointer
uint64_t _odata_size;
public:
accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr) { };
accelerator_inline uint64_t oSites(void) const { return _odata_size; };
accelerator_inline int Checkerboard(void) const { return checkerboard; };
accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
accelerator_inline void Conformable(GridBase * &grid) const
{
if (grid) conformable(grid, _grid);
else grid = _grid;
};
};
/////////////////////////////////////////////////////////////////////////////////////////
// A View class which provides accessor to the data.
// This will be safe to call from accelerator_for and is trivially copy constructible
// The copy constructor for this will need to be used by device lambda functions
/////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
class LatticeView : public LatticeAccelerator<vobj>
{
public:
// Rvalue
#ifdef __CUDA_ARCH__
accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { return coalescedRead(this->_odata[i]); }
#else
accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
#endif
accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
accelerator_inline vobj & operator[](size_t i) { return this->_odata[i]; };
accelerator_inline uint64_t begin(void) const { return 0;};
accelerator_inline uint64_t end(void) const { return this->_odata_size; };
accelerator_inline uint64_t size(void) const { return this->_odata_size; };
LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me)
{
}
};
/////////////////////////////////////////////////////////////////////////////////////////
// Lattice expression types used by ET to assemble the AST
//
// Need to be able to detect code paths according to the whether a lattice object or not
// so introduce some trait type things
/////////////////////////////////////////////////////////////////////////////////////////
class LatticeExpressionBase {};
template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
template<class T> struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
template <typename Op, typename _T1>
class LatticeUnaryExpression : public LatticeExpressionBase
{
public:
typedef typename ViewMap<_T1>::Type T1;
Op op;
T1 arg1;
LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
};
template <typename Op, typename _T1, typename _T2>
class LatticeBinaryExpression : public LatticeExpressionBase
{
public:
typedef typename ViewMap<_T1>::Type T1;
typedef typename ViewMap<_T2>::Type T2;
Op op;
T1 arg1;
T2 arg2;
LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
};
template <typename Op, typename _T1, typename _T2, typename _T3>
class LatticeTrinaryExpression : public LatticeExpressionBase
{
public:
typedef typename ViewMap<_T1>::Type T1;
typedef typename ViewMap<_T2>::Type T2;
typedef typename ViewMap<_T3>::Type T3;
Op op;
T1 arg1;
T2 arg2;
T3 arg3;
LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
};
/////////////////////////////////////////////////////////////////////////////////////////
// The real lattice class, with normal copy and assignment semantics.
// This contains extra (host resident) grid pointer data that may be accessed by host code
@@ -201,14 +80,23 @@ private:
}
}
public:
/////////////////////////////////////////////////////////////////////////////////
// Can use to make accelerator dirty without copy from host ; useful for temporaries "dont care" prev contents
/////////////////////////////////////////////////////////////////////////////////
void SetViewMode(ViewMode mode) {
LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
accessor.ViewClose();
}
/////////////////////////////////////////////////////////////////////////////////
// Return a view object that may be dereferenced in site loops.
// The view is trivially copy constructible and may be copied to an accelerator device
// in device lambdas
/////////////////////////////////////////////////////////////////////////////////
LatticeView<vobj> View (void) const
LatticeView<vobj> View (ViewMode mode) const
{
LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);
return accessor;
}
@@ -232,11 +120,15 @@ public:
assert( (cb==Odd) || (cb==Even));
this->checkerboard=cb;
auto me = View();
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWriteDiscard);
accelerator_for(ss,me.size(),1,{
auto tmp = eval(ss,expr);
auto tmp = eval(ss,exprCopy);
vstream(me[ss],tmp);
});
me.ViewClose();
ExpressionViewClose(exprCopy);
return *this;
}
template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
@@ -251,11 +143,15 @@ public:
assert( (cb==Odd) || (cb==Even));
this->checkerboard=cb;
auto me = View();
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWriteDiscard);
accelerator_for(ss,me.size(),1,{
auto tmp = eval(ss,expr);
auto tmp = eval(ss,exprCopy);
vstream(me[ss],tmp);
});
me.ViewClose();
ExpressionViewClose(exprCopy);
return *this;
}
template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
@@ -269,11 +165,15 @@ public:
CBFromExpression(cb,expr);
assert( (cb==Odd) || (cb==Even));
this->checkerboard=cb;
auto me = View();
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWriteDiscard);
accelerator_for(ss,me.size(),1,{
auto tmp = eval(ss,expr);
auto tmp = eval(ss,exprCopy);
vstream(me[ss],tmp);
});
me.ViewClose();
ExpressionViewClose(exprCopy);
return *this;
}
//GridFromExpression is tricky to do
@@ -324,10 +224,11 @@ public:
}
template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
auto me = View();
auto me = View(CpuWrite);
thread_for(ss,me.size(),{
me[ss] = r;
me[ss]= r;
});
me.ViewClose();
return *this;
}
@@ -337,11 +238,12 @@ public:
///////////////////////////////////////////
// user defined constructor
///////////////////////////////////////////
Lattice(GridBase *grid) {
Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) {
this->_grid = grid;
resize(this->_grid->oSites());
assert((((uint64_t)&this->_odata[0])&0xF) ==0);
this->checkerboard=0;
SetViewMode(mode);
}
// virtual ~Lattice(void) = default;
@@ -357,7 +259,6 @@ public:
// copy constructor
///////////////////////////////////////////
Lattice(const Lattice& r){
// std::cout << "Lattice constructor(const Lattice &) "<<this<<std::endl;
this->_grid = r.Grid();
resize(this->_grid->oSites());
*this = r;
@@ -380,11 +281,12 @@ public:
typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
conformable(*this,r);
this->checkerboard = r.Checkerboard();
auto me = View();
auto him= r.View();
auto me = View(AcceleratorWriteDiscard);
auto him= r.View(AcceleratorRead);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
coalescedWrite(me[ss],him(ss));
});
me.ViewClose(); him.ViewClose();
return *this;
}
@@ -394,11 +296,12 @@ public:
inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
this->checkerboard = r.Checkerboard();
conformable(*this,r);
auto me = View();
auto him= r.View();
auto me = View(AcceleratorWriteDiscard);
auto him= r.View(AcceleratorRead);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
coalescedWrite(me[ss],him(ss));
});
me.ViewClose(); him.ViewClose();
return *this;
}
///////////////////////////////////////////

226
Grid/lattice/Lattice_basis.h Normal file
View File

@@ -0,0 +1,226 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/lattice/Lattice_basis.h
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Christoph Lehner <christoph@lhnr.de>
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 */
#pragma once
NAMESPACE_BEGIN(Grid);
template<class Field>
void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k)
{
// If assume basis[j] are already orthonormal,
// can take all inner products in parallel saving 2x bandwidth
// Save 3x bandwidth on the second line of loop.
// perhaps 2.5x speed up.
// 2x overall in Multigrid Lanczos
for(int j=0; j<k; ++j){
auto ip = innerProduct(basis[j],w);
w = w - ip*basis[j];
}
}
template<class VField, class Matrix>
void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
{
typedef decltype(basis[0]) Field;
typedef decltype(basis[0].View(AcceleratorRead)) View;
Vector<View> basis_v; basis_v.reserve(basis.size());
GridBase* grid = basis[0].Grid();
for(int k=0;k<basis.size();k++){
basis_v.push_back(basis[k].View(AcceleratorWrite));
}
View *basis_vp = &basis_v[0];
int nrot = j1-j0;
if (!nrot) // edge case not handled gracefully by Cuda
return;
uint64_t oSites =grid->oSites();
uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
Vector <vobj> Bt(siteBlock * nrot);
auto Bp=&Bt[0];
// GPU readable copy of matrix
Vector<double> Qt_jv(Nm*Nm);
double *Qt_p = & Qt_jv[0];
thread_for(i,Nm*Nm,{
int j = i/Nm;
int k = i%Nm;
Qt_p[i]=Qt(j,k);
});
// Block the loop to keep storage footprint down
for(uint64_t s=0;s<oSites;s+=siteBlock){
// remaining work in this block
int ssites=MIN(siteBlock,oSites-s);
// zero out the accumulators
accelerator_for(ss,siteBlock*nrot,vobj::Nsimd(),{
decltype(coalescedRead(Bp[ss])) z;
z=Zero();
coalescedWrite(Bp[ss],z);
});
accelerator_for(sj,ssites*nrot,vobj::Nsimd(),{
int j =sj%nrot;
int jj =j0+j;
int ss =sj/nrot;
int sss=ss+s;
for(int k=k0; k<k1; ++k){
auto tmp = coalescedRead(Bp[ss*nrot+j]);
coalescedWrite(Bp[ss*nrot+j],tmp+ Qt_p[jj*Nm+k] * coalescedRead(basis_v[k][sss]));
}
});
accelerator_for(sj,ssites*nrot,vobj::Nsimd(),{
int j =sj%nrot;
int jj =j0+j;
int ss =sj/nrot;
int sss=ss+s;
coalescedWrite(basis_v[jj][sss],coalescedRead(Bp[ss*nrot+j]));
});
}
for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
}
// Extract a single rotated vector
template<class Field>
void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm)
{
typedef decltype(basis[0].View(AcceleratorRead)) View;
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0].Grid();
result.Checkerboard() = basis[0].Checkerboard();
Vector<View> basis_v; basis_v.reserve(basis.size());
for(int k=0;k<basis.size();k++){
basis_v.push_back(basis[k].View(AcceleratorRead));
}
vobj zz=Zero();
Vector<double> Qt_jv(Nm);
double * Qt_j = & Qt_jv[0];
for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
autoView(result_v,result,AcceleratorWrite);
accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
auto B=coalescedRead(zz);
for(int k=k0; k<k1; ++k){
B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
}
coalescedWrite(result_v[ss], B);
});
for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
}
template<class Field>
void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx)
{
int vlen = idx.size();
assert(vlen>=1);
assert(vlen<=sort_vals.size());
assert(vlen<=_v.size());
for (size_t i=0;i<vlen;i++) {
if (idx[i] != i) {
//////////////////////////////////////
// idx[i] is a table of desired sources giving a permutation.
// Swap v[i] with v[idx[i]].
// Find j>i for which _vnew[j] = _vold[i],
// track the move idx[j] => idx[i]
// track the move idx[i] => i
//////////////////////////////////////
size_t j;
for (j=i;j<idx.size();j++)
if (idx[j]==i)
break;
assert(idx[i] > i); assert(j!=idx.size()); assert(idx[j]==i);
swap(_v[i],_v[idx[i]]); // should use vector move constructor, no data copy
std::swap(sort_vals[i],sort_vals[idx[i]]);
idx[j] = idx[i];
idx[i] = i;
}
}
}
inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals)
{
std::vector<int> idx(sort_vals.size());
std::iota(idx.begin(), idx.end(), 0);
// sort indexes based on comparing values in v
std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
});
return idx;
}
template<class Field>
void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse)
{
std::vector<int> idx = basisSortGetIndex(sort_vals);
if (reverse)
std::reverse(idx.begin(), idx.end());
basisReorderInPlace(_v,sort_vals,idx);
}
// PAB: faster to compute the inner products first then fuse loops.
// If performance critical can improve.
template<class Field>
void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
result = Zero();
assert(_v.size()==eval.size());
int N = (int)_v.size();
for (int i=0;i<N;i++) {
Field& tmp = _v[i];
axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
}
}
NAMESPACE_END(Grid);

14
Grid/lattice/Lattice_comparison.h
View File

@@ -78,9 +78,9 @@ template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
{
Lattice<vPredicate> ret(rhs.Grid());
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
auto ret_v = ret.View();
autoView( lhs_v, lhs, CpuRead);
autoView( rhs_v, rhs, CpuRead);
autoView( ret_v, ret, CpuWrite);
thread_for( ss, rhs_v.size(), {
ret_v[ss]=op(lhs_v[ss],rhs_v[ss]);
});
@@ -93,8 +93,8 @@ template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
{
Lattice<vPredicate> ret(lhs.Grid());
auto lhs_v = lhs.View();
auto ret_v = ret.View();
autoView( lhs_v, lhs, CpuRead);
autoView( ret_v, ret, CpuWrite);
thread_for( ss, lhs_v.size(), {
ret_v[ss]=op(lhs_v[ss],rhs);
});
@@ -107,8 +107,8 @@ template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
{
Lattice<vPredicate> ret(rhs.Grid());
auto rhs_v = rhs.View();
auto ret_v = ret.View();
autoView( rhs_v, rhs, CpuRead);
autoView( ret_v, ret, CpuWrite);
thread_for( ss, rhs_v.size(), {
ret_v[ss]=op(lhs,rhs_v[ss]);
});

20
Grid/lattice/Lattice_coordinate.h
View File

@@ -37,7 +37,7 @@ template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
GridBase *grid = l.Grid();
int Nsimd = grid->iSites();
auto l_v = l.View();
autoView(l_v, l, CpuWrite);
thread_for( o, grid->oSites(), {
vector_type vI;
Coordinate gcoor;
@@ -51,23 +51,5 @@ template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
});
};
// LatticeCoordinate();
// FIXME for debug; deprecate this; made obscelete by
template<class vobj> void lex_sites(Lattice<vobj> &l){
auto l_v = l.View();
Real *v_ptr = (Real *)&l_v[0];
size_t o_len = l.Grid()->oSites();
size_t v_len = sizeof(vobj)/sizeof(vRealF);
size_t vec_len = vRealF::Nsimd();
for(int i=0;i<o_len;i++){
for(int j=0;j<v_len;j++){
for(int vv=0;vv<vec_len;vv+=2){
v_ptr[i*v_len*vec_len+j*vec_len+vv ]= i+vv*500;
v_ptr[i*v_len*vec_len+j*vec_len+vv+1]= i+vv*500;
}
}}
}
NAMESPACE_END(Grid);

16
Grid/lattice/Lattice_local.h
View File

@@ -43,8 +43,8 @@ template<class vobj>
inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
{
Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
auto rhs_v = rhs.View();
auto ret_v = ret.View();
autoView( rhs_v , rhs, AcceleratorRead);
autoView( ret_v , ret, AcceleratorWrite);
accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss],innerProduct(rhs_v(ss),rhs_v(ss)));
});
@@ -56,9 +56,9 @@ template<class vobj>
inline auto localInnerProduct (const Lattice<vobj> &lhs,const Lattice<vobj> &rhs) -> Lattice<typename vobj::tensor_reduced>
{
Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
auto ret_v = ret.View();
autoView( lhs_v , lhs, AcceleratorRead);
autoView( rhs_v , rhs, AcceleratorRead);
autoView( ret_v , ret, AcceleratorWrite);
accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss],innerProduct(lhs_v(ss),rhs_v(ss)));
});
@@ -73,9 +73,9 @@ inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Latt
typedef decltype(coalescedRead(ll())) sll;
typedef decltype(coalescedRead(rr())) srr;
Lattice<decltype(outerProduct(ll(),rr()))> ret(rhs.Grid());
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
auto ret_v = ret.View();
autoView( lhs_v , lhs, AcceleratorRead);
autoView( rhs_v , rhs, AcceleratorRead);
autoView( ret_v , ret, AcceleratorWrite);
accelerator_for(ss,rhs_v.size(),1,{
// FIXME had issues with scalar version of outer
// Use vector [] operator and don't read coalesce this loop

14
Grid/lattice/Lattice_matrix_reduction.h
View File

@@ -51,9 +51,9 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
auto X_v = X.View();
auto Y_v = Y.View();
auto R_v = R.View();
autoView( X_v , X, CpuRead);
autoView( Y_v , Y, CpuRead);
autoView( R_v , R, CpuWrite);
thread_region
{
std::vector<vobj> s_x(Nblock);
@@ -97,8 +97,8 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
auto X_v = X.View();
auto R_v = R.View();
autoView( X_v , X, CpuRead);
autoView( R_v , R, CpuWrite);
thread_region
{
@@ -156,8 +156,8 @@ static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj>
int ostride=FullGrid->_ostride[Orthog];
typedef typename vobj::vector_typeD vector_typeD;
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
autoView( lhs_v , lhs, CpuRead);
autoView( rhs_v , rhs, CpuRead);
thread_region {
std::vector<vobj> Left(Nblock);
std::vector<vobj> Right(Nblock);

58
Grid/lattice/Lattice_peekpoke.h
View File

@@ -46,9 +46,9 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Ind
{
Lattice<decltype(peekIndex<Index>(vobj(),i))> ret(lhs.Grid());
ret.Checkerboard()=lhs.Checkerboard();
auto ret_v = ret.View();
auto lhs_v = lhs.View();
thread_for( ss, lhs_v.size(), {
autoView( ret_v, ret, AcceleratorWrite);
autoView( lhs_v, lhs, AcceleratorRead);
accelerator_for( ss, lhs_v.size(), 1, {
ret_v[ss] = peekIndex<Index>(lhs_v[ss],i);
});
return ret;
@@ -58,9 +58,9 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekInd
{
Lattice<decltype(peekIndex<Index>(vobj(),i,j))> ret(lhs.Grid());
ret.Checkerboard()=lhs.Checkerboard();
auto ret_v = ret.View();
auto lhs_v = lhs.View();
thread_for( ss, lhs_v.size(), {
autoView( ret_v, ret, AcceleratorWrite);
autoView( lhs_v, lhs, AcceleratorRead);
accelerator_for( ss, lhs_v.size(), 1, {
ret_v[ss] = peekIndex<Index>(lhs_v[ss],i,j);
});
return ret;
@@ -72,18 +72,18 @@ auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekInd
template<int Index,class vobj>
void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0))> & rhs,int i)
{
auto rhs_v = rhs.View();
auto lhs_v = lhs.View();
thread_for( ss, lhs_v.size(), {
autoView( rhs_v, rhs, AcceleratorRead);
autoView( lhs_v, lhs, AcceleratorWrite);
accelerator_for( ss, lhs_v.size(), 1, {
pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i);
});
}
template<int Index,class vobj>
void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0,0))> & rhs,int i,int j)
{
auto rhs_v = rhs.View();
auto lhs_v = lhs.View();
thread_for( ss, lhs_v.size(), {
autoView( rhs_v, rhs, AcceleratorRead);
autoView( lhs_v, lhs, AcceleratorWrite);
accelerator_for( ss, lhs_v.size(), 1, {
pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i,j);
});
}
@@ -111,7 +111,7 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
// extract-modify-merge cycle is easiest way and this is not perf critical
ExtractBuffer<sobj> buf(Nsimd);
auto l_v = l.View();
autoView( l_v , l, CpuWrite);
if ( rank == grid->ThisRank() ) {
extract(l_v[odx],buf);
buf[idx] = s;
@@ -141,7 +141,7 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
grid->GlobalCoorToRankIndex(rank,odx,idx,site);
ExtractBuffer<sobj> buf(Nsimd);
auto l_v = l.View();
autoView( l_v , l, CpuWrite);
extract(l_v[odx],buf);
s = buf[idx];
@@ -151,21 +151,21 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
return;
};
//////////////////////////////////////////////////////////
// Peek a scalar object from the SIMD array
//////////////////////////////////////////////////////////
// Must be CPU read view
template<class vobj,class sobj>
accelerator_inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
GridBase *grid = l.Grid();
inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
{
GridBase *grid = l.getGrid();
assert(l.mode==CpuRead);
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
assert( l.Checkerboard()== grid->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -173,8 +173,7 @@ accelerator_inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate
idx= grid->iIndex(site);
odx= grid->oIndex(site);
auto l_v = l.View();
scalar_type * vp = (scalar_type *)&l_v[odx];
scalar_type * vp = (scalar_type *)&l[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
@@ -183,18 +182,19 @@ accelerator_inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate
return;
};
// Must be CPU write view
template<class vobj,class sobj>
accelerator_inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
GridBase *grid=l.Grid();
inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
{
GridBase *grid=l.getGrid();
assert(l.mode==CpuWrite);
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
assert( l.Checkerboard()== grid->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
static const int words=sizeof(vobj)/sizeof(vector_type);
@@ -202,13 +202,11 @@ accelerator_inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate
idx= grid->iIndex(site);
odx= grid->oIndex(site);
auto l_v = l.View();
scalar_type * vp = (scalar_type *)&l_v[odx];
scalar_type * vp = (scalar_type *)&l[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
vp[idx+w*Nsimd] = pt[w];
}
return;
};

14
Grid/lattice/Lattice_reality.h
View File

@@ -40,8 +40,11 @@ NAMESPACE_BEGIN(Grid);
template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
auto lhs_v = lhs.View();
auto ret_v = ret.View();
autoView( lhs_v, lhs, AcceleratorRead);
autoView( ret_v, ret, AcceleratorWrite);
ret.Checkerboard()=lhs.Checkerboard();
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
coalescedWrite(ret_v[ss], adj(lhs_v(ss)));
});
@@ -50,8 +53,11 @@ template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
auto lhs_v = lhs.View();
auto ret_v = ret.View();
autoView( lhs_v, lhs, AcceleratorRead);
autoView( ret_v, ret, AcceleratorWrite);
ret.Checkerboard() = lhs.Checkerboard();
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
coalescedWrite( ret_v[ss] , conjugate(lhs_v(ss)));
});

213
Grid/lattice/Lattice_reduction.h
View File

@@ -5,6 +5,7 @@
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Christoph Lehner <christoph@lhnr.de>
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
@@ -24,7 +25,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <Grid/Grid_Eigen_Dense.h>
#ifdef GRID_NVCC
#if defined(GRID_CUDA)||defined(GRID_HIP)
#include <Grid/lattice/Lattice_reduction_gpu.h>
#endif
@@ -38,7 +39,36 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
{
typedef typename vobj::scalar_object sobj;
const int Nsimd = vobj::Nsimd();
// const int Nsimd = vobj::Nsimd();
const int nthread = GridThread::GetThreads();
Vector<sobj> sumarray(nthread);
for(int i=0;i<nthread;i++){
sumarray[i]=Zero();
}
thread_for(thr,nthread, {
int nwork, mywork, myoff;
nwork = osites;
GridThread::GetWork(nwork,thr,mywork,myoff);
vobj vvsum=Zero();
for(int ss=myoff;ss<mywork+myoff; ss++){
vvsum = vvsum + arg[ss];
}
sumarray[thr]=Reduce(vvsum);
});
sobj ssum=Zero(); // sum across threads
for(int i=0;i<nthread;i++){
ssum = ssum+sumarray[i];
}
return ssum;
}
template<class vobj>
inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
{
typedef typename vobj::scalar_objectD sobj;
const int nthread = GridThread::GetThreads();
Vector<sobj> sumarray(nthread);
@@ -62,23 +92,43 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
ssum = ssum+sumarray[i];
}
return ssum;
typedef typename vobj::scalar_object ssobj;
ssobj ret = ssum;
return ret;
}
template<class vobj>
inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
{
#ifdef GRID_NVCC
#if defined(GRID_CUDA)||defined(GRID_HIP)
return sum_gpu(arg,osites);
#else
return sum_cpu(arg,osites);
#endif
}
template<class vobj>
inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
{
#if defined(GRID_CUDA)||defined(GRID_HIP)
return sumD_gpu(arg,osites);
#else
return sumD_cpu(arg,osites);
#endif
}
template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
{
auto arg_v = arg.View();
#if defined(GRID_CUDA)||defined(GRID_HIP)
autoView( arg_v, arg, AcceleratorRead);
Integer osites = arg.Grid()->oSites();
auto ssum= sum(&arg_v[0],osites);
auto ssum= sum_gpu(&arg_v[0],osites);
#else
autoView(arg_v, arg, CpuRead);
Integer osites = arg.Grid()->oSites();
auto ssum= sum_cpu(&arg_v[0],osites);
#endif
arg.Grid()->GlobalSum(ssum);
return ssum;
}
@@ -93,7 +143,7 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
// Double inner product
template<class vobj>
inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
{
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_typeD vector_type;
@@ -101,47 +151,41 @@ inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &righ
GridBase *grid = left.Grid();
// Might make all code paths go this way.
auto left_v = left.View();
auto right_v=right.View();
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites = grid->oSites();
#ifdef GRID_NVCC
// Might make all code paths go this way.
typedef decltype(innerProductD(vobj(),vobj())) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
{
autoView( left_v , left, AcceleratorRead);
autoView( right_v,right, AcceleratorRead);
// GPU - SIMT lane compliance...
typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
accelerator_for( ss, sites, nsimd,{
auto x_l = left_v(ss);
auto y_l = right_v(ss);
coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
})
// This is in single precision and fails some tests
// Need a sumD that sums in double
nrm = TensorRemove(sumD_gpu(inner_tmp_v,sites));
#else
// CPU
typedef decltype(innerProductD(left_v[0],right_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
accelerator_for( ss, sites, nsimd,{
accelerator_for( ss, sites, 1,{
auto x_l = left_v[ss];
auto y_l = right_v[ss];
inner_tmp_v[ss]=innerProductD(x_l,y_l);
})
nrm = TensorRemove(sum(inner_tmp_v,sites));
#endif
grid->GlobalSum(nrm);
});
}
// This is in single precision and fails some tests
auto anrm = sum(inner_tmp_v,sites);
nrm = anrm;
return nrm;
}
template<class vobj>
inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right) {
GridBase *grid = left.Grid();
ComplexD nrm = rankInnerProduct(left,right);
grid->GlobalSum(nrm);
return nrm;
}
/////////////////////////
// Fast axpby_norm
// z = a x + b y
@@ -167,44 +211,66 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
GridBase *grid = x.Grid();
auto x_v=x.View();
auto y_v=y.View();
auto z_v=z.View();
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites = grid->oSites();
#ifdef GRID_NVCC
// GPU
typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
autoView( x_v, x, AcceleratorRead);
autoView( y_v, y, AcceleratorRead);
autoView( z_v, z, AcceleratorWrite);
accelerator_for( ss, sites, nsimd,{
auto tmp = a*x_v(ss)+b*y_v(ss);
coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
coalescedWrite(z_v[ss],tmp);
});
nrm = real(TensorRemove(sumD_gpu(inner_tmp_v,sites)));
#else
// CPU
typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
accelerator_for( ss, sites, nsimd,{
auto tmp = a*x_v(ss)+b*y_v(ss);
accelerator_for( ss, sites, 1,{
auto tmp = a*x_v[ss]+b*y_v[ss];
inner_tmp_v[ss]=innerProductD(tmp,tmp);
z_v[ss]=tmp;
});
// Already promoted to double
nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
#endif
grid->GlobalSum(nrm);
return nrm;
}
template<class vobj> strong_inline void
innerProductNorm(ComplexD& ip, RealD &nrm, const Lattice<vobj> &left,const Lattice<vobj> &right)
{
conformable(left,right);
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_typeD vector_type;
Vector<ComplexD> tmp(2);
GridBase *grid = left.Grid();
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites = grid->oSites();
// GPU
typedef decltype(innerProductD(vobj(),vobj())) inner_t;
typedef decltype(innerProductD(vobj(),vobj())) norm_t;
Vector<inner_t> inner_tmp(sites);
Vector<norm_t> norm_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
auto norm_tmp_v = &norm_tmp[0];
{
autoView(left_v,left, AcceleratorRead);
autoView(right_v,right,AcceleratorRead);
accelerator_for( ss, sites, 1,{
auto left_tmp = left_v[ss];
inner_tmp_v[ss]=innerProductD(left_tmp,right_v[ss]);
norm_tmp_v [ss]=innerProductD(left_tmp,left_tmp);
});
}
tmp[0] = TensorRemove(sum(inner_tmp_v,sites));
tmp[1] = TensorRemove(sum(norm_tmp_v,sites));
grid->GlobalSumVector(&tmp[0],2); // keep norm Complex -> can use GlobalSumVector
ip = tmp[0];
nrm = real(tmp[1]);
}
template<class Op,class T1>
inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
@@ -271,7 +337,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
// sum over reduced dimension planes, breaking out orthog dir
// Parallel over orthog direction
auto Data_v=Data.View();
autoView( Data_v, Data, CpuRead);
thread_for( r,rd, {
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
for(int n=0;n<e1;n++){
@@ -349,8 +415,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
int e2= grid->_slice_block [orthogdim];
int stride=grid->_slice_stride[orthogdim];
auto lhv=lhs.View();
auto rhv=rhs.View();
autoView( lhv, lhs, CpuRead);
autoView( rhv, rhs, CpuRead);
thread_for( r,rd,{
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
@@ -457,14 +523,12 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
tensor_reduced at; at=av;
auto Rv=R.View();
auto Xv=X.View();
auto Yv=Y.View();
thread_for_collapse(2, n, e1, {
for(int b=0;b<e2;b++){
autoView( Rv, R, CpuWrite);
autoView( Xv, X, CpuRead);
autoView( Yv, Y, CpuRead);
thread_for2d( n, e1, b,e2, {
int ss= so+n*stride+b;
Rv[ss] = at*Xv[ss]+Yv[ss];
}
});
}
};
@@ -517,9 +581,9 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
auto X_v=X.View();
auto Y_v=Y.View();
auto R_v=R.View();
autoView( X_v, X, CpuRead);
autoView( Y_v, Y, CpuRead);
autoView( R_v, R, CpuWrite);
thread_region
{
Vector<vobj> s_x(Nblock);
@@ -564,13 +628,14 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
// int nl=1;
//FIXME package in a convenient iterator
// thread_for2d_in_region
//Should loop over a plane orthogonal to direction "Orthog"
int stride=FullGrid->_slice_stride[Orthog];
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
auto R_v = R.View();
auto X_v = X.View();
autoView( R_v, R, CpuWrite);
autoView( X_v, X, CpuRead);
thread_region
{
std::vector<vobj> s_x(Nblock);
@@ -628,8 +693,8 @@ static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj>
typedef typename vobj::vector_typeD vector_typeD;
auto lhs_v=lhs.View();
auto rhs_v=rhs.View();
autoView( lhs_v, lhs, CpuRead);
autoView( rhs_v, rhs, CpuRead);
thread_region
{
std::vector<vobj> Left(Nblock);

25
Grid/lattice/Lattice_reduction_gpu.h
View File

@@ -1,7 +1,13 @@
NAMESPACE_BEGIN(Grid);
#define WARP_SIZE 32
#ifdef GRID_HIP
extern hipDeviceProp_t *gpu_props;
#endif
#ifdef GRID_CUDA
extern cudaDeviceProp *gpu_props;
#endif
#define WARP_SIZE 32
__device__ unsigned int retirementCount = 0;
template <class Iterator>
@@ -19,7 +25,12 @@ template <class Iterator>
void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
int device;
#ifdef GRID_CUDA
cudaGetDevice(&device);
#endif
#ifdef GRID_HIP
hipGetDevice(&device);
#endif
Iterator warpSize = gpu_props[device].warpSize;
Iterator sharedMemPerBlock = gpu_props[device].sharedMemPerBlock;
@@ -147,7 +158,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
sobj *smem = (sobj *)shmem_pointer;
// wait until all outstanding memory instructions in this thread are finished
__threadfence();
acceleratorFence();
if (tid==0) {
unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
@@ -156,7 +167,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
}
// each thread must read the correct value of amLast
__syncthreads();
acceleratorSynchroniseAll();
if (amLast) {
// reduce buffer[0], ..., buffer[gridDim.x-1]
@@ -199,13 +210,7 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
sobj *buffer_v = &buffer[0];
reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
cudaDeviceSynchronize();
cudaError err = cudaGetLastError();
if ( cudaSuccess != err ) {
printf("Cuda error %s\n",cudaGetErrorString( err ));
exit(0);
}
accelerator_barrier();
auto result = buffer_v[0];
return result;
}

4
Grid/lattice/Lattice_rng.h
View File

@@ -375,7 +375,7 @@ public:
int osites = _grid->oSites(); // guaranteed to be <= l.Grid()->oSites() by a factor multiplicity
int words = sizeof(scalar_object) / sizeof(scalar_type);
auto l_v = l.View();
autoView(l_v, l, CpuWrite);
thread_for( ss, osites, {
ExtractBuffer<scalar_object> buf(Nsimd);
for (int m = 0; m < multiplicity; m++) { // Draw from same generator multiplicity times
@@ -462,7 +462,7 @@ public:
{
// Obtain one reseeded generator per thread
int Nthread = GridThread::GetThreads();
int Nthread = 32; // Hardwire a good level or parallelism
std::vector<RngEngine> seeders(Nthread);
for(int t=0;t<Nthread;t++){
seeders[t] = Reseed(master_engine);

10
Grid/lattice/Lattice_trace.h
View File

@@ -37,17 +37,19 @@ NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////////////////////////////////////////////////
// Trace
////////////////////////////////////////////////////////////////////////////////////////////////////
/*
template<class vobj>
inline auto trace(const Lattice<vobj> &lhs) -> Lattice<decltype(trace(vobj()))>
{
Lattice<decltype(trace(vobj()))> ret(lhs.Grid());
auto ret_v = ret.View();
auto lhs_v = lhs.View();
autoView(ret_v , ret, AcceleratorWrite);
autoView(lhs_v , lhs, AcceleratorRead);
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
coalescedWrite(ret_v[ss], trace(lhs_v(ss)));
});
return ret;
};
*/
////////////////////////////////////////////////////////////////////////////////////////////////////
// Trace Index level dependent operation
@@ -56,8 +58,8 @@ template<int Index,class vobj>
inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(vobj()))>
{
Lattice<decltype(traceIndex<Index>(vobj()))> ret(lhs.Grid());
auto ret_v = ret.View();
auto lhs_v = lhs.View();
autoView( ret_v , ret, AcceleratorWrite);
autoView( lhs_v , lhs, AcceleratorRead);
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
});

348
Grid/lattice/Lattice_transfer.h
View File

@@ -6,6 +6,7 @@
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Christoph Lehner <christoph@lhnr.de>
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
@@ -46,11 +47,12 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
////////////////////////////////////////////////////////////////////////////////////////////
// remove and insert a half checkerboard
////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full){
template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full)
{
half.Checkerboard() = cb;
auto half_v = half.View();
auto full_v = full.View();
autoView( half_v, half, CpuWrite);
autoView( full_v, full, CpuRead);
thread_for(ss, full.Grid()->oSites(),{
int cbos;
Coordinate coor;
@@ -63,10 +65,11 @@ template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,con
}
});
}
template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half)
{
int cb = half.Checkerboard();
auto half_v = half.View();
auto full_v = full.View();
autoView( half_v , half, CpuRead);
autoView( full_v , full, CpuWrite);
thread_for(ss,full.Grid()->oSites(),{
Coordinate coor;
@@ -82,94 +85,136 @@ template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Latti
});
}
template<class vobj,class CComplex,int nbasis>
////////////////////////////////////////////////////////////////////////////////////////////
// Flexible Type Conversion for internal promotion to double as well as graceful
// treatment of scalar-compatible types
////////////////////////////////////////////////////////////////////////////////////////////
accelerator_inline void convertType(ComplexD & out, const std::complex<double> & in) {
out = in;
}
accelerator_inline void convertType(ComplexF & out, const std::complex<float> & in) {
out = in;
}
#ifdef GRID_SIMT
accelerator_inline void convertType(vComplexF & out, const ComplexF & in) {
((ComplexF*)&out)[acceleratorSIMTlane(vComplexF::Nsimd())] = in;
}
accelerator_inline void convertType(vComplexD & out, const ComplexD & in) {
((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd())] = in;
}
accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
((ComplexD*)&out)[acceleratorSIMTlane(vComplexD::Nsimd()*2)] = in;
}
#endif
accelerator_inline void convertType(vComplexF & out, const vComplexD2 & in) {
out.v = Optimization::PrecisionChange::DtoS(in._internal[0].v,in._internal[1].v);
}
accelerator_inline void convertType(vComplexD2 & out, const vComplexF & in) {
Optimization::PrecisionChange::StoD(in.v,out._internal[0].v,out._internal[1].v);
}
template<typename T1,typename T2,int N>
accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in);
template<typename T1,typename T2,int N>
accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in);
template<typename T1,typename T2, typename std::enable_if<!isGridScalar<T1>::value, T1>::type* = nullptr>
accelerator_inline void convertType(T1 & out, const iScalar<T2> & in) {
convertType(out,in._internal);
}
template<typename T1,typename T2>
accelerator_inline void convertType(iScalar<T1> & out, const T2 & in) {
convertType(out._internal,in);
}
template<typename T1,typename T2,int N>
accelerator_inline void convertType(iMatrix<T1,N> & out, const iMatrix<T2,N> & in) {
for (int i=0;i<N;i++)
for (int j=0;j<N;j++)
convertType(out._internal[i][j],in._internal[i][j]);
}
template<typename T1,typename T2,int N>
accelerator_inline void convertType(iVector<T1,N> & out, const iVector<T2,N> & in) {
for (int i=0;i<N;i++)
convertType(out._internal[i],in._internal[i]);
}
template<typename T, typename std::enable_if<isGridFundamental<T>::value, T>::type* = nullptr>
accelerator_inline void convertType(T & out, const T & in) {
out = in;
}
template<typename T1,typename T2>
accelerator_inline void convertType(Lattice<T1> & out, const Lattice<T2> & in) {
autoView( out_v , out,AcceleratorWrite);
autoView( in_v , in ,AcceleratorRead);
accelerator_for(ss,out_v.size(),T1::Nsimd(),{
convertType(out_v[ss],in_v(ss));
});
}
////////////////////////////////////////////////////////////////////////////////////////////
// precision-promoted local inner product
////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
inline auto localInnerProductD(const Lattice<vobj> &lhs,const Lattice<vobj> &rhs)
-> Lattice<iScalar<decltype(TensorRemove(innerProductD2(lhs.View(CpuRead)[0],rhs.View(CpuRead)[0])))>>
{
autoView( lhs_v , lhs, AcceleratorRead);
autoView( rhs_v , rhs, AcceleratorRead);
typedef decltype(TensorRemove(innerProductD2(lhs_v[0],rhs_v[0]))) t_inner;
Lattice<iScalar<t_inner>> ret(lhs.Grid());
{
autoView(ret_v, ret,AcceleratorWrite);
accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
convertType(ret_v[ss],innerProductD2(lhs_v(ss),rhs_v(ss)));
});
}
return ret;
}
////////////////////////////////////////////////////////////////////////////////////////////
// block routines
////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
const Lattice<vobj> &fineData,
const std::vector<Lattice<vobj> > &Basis)
const VLattice &Basis)
{
GridBase * fine = fineData.Grid();
GridBase * coarse= coarseData.Grid();
Lattice<CComplex> ip(coarse);
Lattice<iScalar<CComplex>> ip(coarse);
Lattice<vobj> fineDataRed = fineData;
// auto fineData_ = fineData.View();
auto coarseData_ = coarseData.View();
auto ip_ = ip.View();
autoView( coarseData_ , coarseData, AcceleratorWrite);
autoView( ip_ , ip, AcceleratorWrite);
for(int v=0;v<nbasis;v++) {
blockInnerProduct(ip,Basis[v],fineData);
blockInnerProductD(ip,Basis[v],fineDataRed); // ip = <basis|fine>
accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
coalescedWrite(coarseData_[sc](v),ip_(sc));
convertType(coarseData_[sc](v),ip_[sc]);
});
// improve numerical stability of projection
// |fine> = |fine> - <basis|fine> |basis>
ip=-ip;
blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed);
}
}
template<class vobj,class CComplex,int nbasis>
inline void blockProject1(Lattice<iVector<CComplex,nbasis > > &coarseData,
const Lattice<vobj> &fineData,
const std::vector<Lattice<vobj> > &Basis)
{
typedef iVector<CComplex,nbasis > coarseSiteData;
coarseSiteData elide;
typedef decltype(coalescedRead(elide)) ScalarComplex;
GridBase * fine = fineData.Grid();
GridBase * coarse= coarseData.Grid();
int _ndimension = coarse->_ndimension;
// checks
assert( nbasis == Basis.size() );
subdivides(coarse,fine);
for(int i=0;i<nbasis;i++){
conformable(Basis[i],fineData);
}
Coordinate block_r (_ndimension);
for(int d=0 ; d<_ndimension;d++){
block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
assert(block_r[d]*coarse->_rdimensions[d] == fine->_rdimensions[d]);
}
int blockVol = fine->oSites()/coarse->oSites();
coarseData=Zero();
auto fineData_ = fineData.View();
auto coarseData_ = coarseData.View();
////////////////////////////////////////////////////////////////////////////////////////////////////////
// To make this lock free, loop over coars parallel, and then loop over fine associated with coarse.
// Otherwise do fine inner product per site, and make the update atomic
////////////////////////////////////////////////////////////////////////////////////////////////////////
accelerator_for( sci, nbasis*coarse->oSites(), vobj::Nsimd(), {
auto sc=sci/nbasis;
auto i=sci%nbasis;
auto Basis_ = Basis[i].View();
Coordinate coor_c(_ndimension);
Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions); // Block coordinate
int sf;
decltype(innerProduct(Basis_(sf),fineData_(sf))) reduce=Zero();
for(int sb=0;sb<blockVol;sb++){
Coordinate coor_b(_ndimension);
Coordinate coor_f(_ndimension);
Lexicographic::CoorFromIndex(coor_b,sb,block_r);
for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d]+coor_b[d];
Lexicographic::IndexFromCoor(coor_f,sf,fine->_rdimensions);
reduce=reduce+innerProduct(Basis_(sf),fineData_(sf));
}
coalescedWrite(coarseData_[sc](i),reduce);
});
return;
}
template<class vobj,class CComplex>
inline void blockZAXPY(Lattice<vobj> &fineZ,
template<class vobj,class vobj2,class CComplex>
inline void blockZAXPY(Lattice<vobj> &fineZ,
const Lattice<CComplex> &coarseA,
const Lattice<vobj> &fineX,
const Lattice<vobj2> &fineX,
const Lattice<vobj> &fineY)
{
GridBase * fine = fineZ.Grid();
@@ -191,10 +236,10 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
assert(block_r[d]*coarse->_rdimensions[d]==fine->_rdimensions[d]);
}
auto fineZ_ = fineZ.View();
auto fineX_ = fineX.View();
auto fineY_ = fineY.View();
auto coarseA_= coarseA.View();
autoView( fineZ_ , fineZ, AcceleratorWrite);
autoView( fineX_ , fineX, AcceleratorRead);
autoView( fineY_ , fineY, AcceleratorRead);
autoView( coarseA_, coarseA, AcceleratorRead);
accelerator_for(sf, fine->oSites(), CComplex::Nsimd(), {
@@ -207,13 +252,50 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
// z = A x + y
coalescedWrite(fineZ_[sf],coarseA_(sc)*fineX_(sf)+fineY_(sf));
#ifdef GRID_SIMT
typename vobj2::tensor_reduced::scalar_object cA;
typename vobj::scalar_object cAx;
#else
typename vobj2::tensor_reduced cA;
vobj cAx;
#endif
convertType(cA,TensorRemove(coarseA_(sc)));
auto prod = cA*fineX_(sf);
convertType(cAx,prod);
coalescedWrite(fineZ_[sf],cAx+fineY_(sf));
});
return;
}
template<class vobj,class CComplex>
inline void blockInnerProductD(Lattice<CComplex> &CoarseInner,
const Lattice<vobj> &fineX,
const Lattice<vobj> &fineY)
{
typedef iScalar<decltype(TensorRemove(innerProductD2(vobj(),vobj())))> dotp;
GridBase *coarse(CoarseInner.Grid());
GridBase *fine (fineX.Grid());
Lattice<dotp> fine_inner(fine); fine_inner.Checkerboard() = fineX.Checkerboard();
Lattice<dotp> coarse_inner(coarse);
// Precision promotion
fine_inner = localInnerProductD<vobj>(fineX,fineY);
blockSum(coarse_inner,fine_inner);
{
autoView( CoarseInner_ , CoarseInner,AcceleratorWrite);
autoView( coarse_inner_ , coarse_inner,AcceleratorRead);
accelerator_for(ss, coarse->oSites(), 1, {
convertType(CoarseInner_[ss], TensorRemove(coarse_inner_[ss]));
});
}
}
template<class vobj,class CComplex> // deprecate
inline void blockInnerProduct(Lattice<CComplex> &CoarseInner,
const Lattice<vobj> &fineX,
const Lattice<vobj> &fineY)
@@ -227,15 +309,17 @@ inline void blockInnerProduct(Lattice<CComplex> &CoarseInner,
Lattice<dotp> coarse_inner(coarse);
// Precision promotion?
auto CoarseInner_ = CoarseInner.View();
auto coarse_inner_ = coarse_inner.View();
fine_inner = localInnerProduct(fineX,fineY);
blockSum(coarse_inner,fine_inner);
{
autoView( CoarseInner_ , CoarseInner, AcceleratorWrite);
autoView( coarse_inner_ , coarse_inner, AcceleratorRead);
accelerator_for(ss, coarse->oSites(), 1, {
CoarseInner_[ss] = coarse_inner_[ss];
});
}
}
template<class vobj,class CComplex>
inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
{
@@ -266,8 +350,8 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
// Turn this around to loop threaded over sc and interior loop
// over sf would thread better
auto coarseData_ = coarseData.View();
auto fineData_ = fineData.View();
autoView( coarseData_ , coarseData, AcceleratorWrite);
autoView( fineData_ , fineData, AcceleratorRead);
accelerator_for(sc,coarse->oSites(),1,{
@@ -292,6 +376,7 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
return;
}
template<class vobj>
inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vobj> &picked,Coordinate coor)
{
@@ -313,8 +398,8 @@ inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vob
}
}
template<class vobj,class CComplex>
inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> > &Basis)
template<class CComplex,class VLattice>
inline void blockOrthonormalize(Lattice<CComplex> &ip,VLattice &Basis)
{
GridBase *coarse = ip.Grid();
GridBase *fine = Basis[0].Grid();
@@ -330,15 +415,22 @@ inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> >
for(int v=0;v<nbasis;v++) {
for(int u=0;u<v;u++) {
//Inner product & remove component
blockInnerProduct(ip,Basis[u],Basis[v]);
blockInnerProductD(ip,Basis[u],Basis[v]);
ip = -ip;
blockZAXPY<vobj,CComplex> (Basis[v],ip,Basis[u],Basis[v]);
blockZAXPY(Basis[v],ip,Basis[u],Basis[v]);
}
blockNormalise(ip,Basis[v]);
}
}
template<class vobj,class CComplex>
inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> > &Basis) // deprecated inaccurate naming
{
blockOrthonormalize(ip,Basis);
}
#if 0
// TODO: CPU optimized version here
template<class vobj,class CComplex,int nbasis>
inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
Lattice<vobj> &fineData,
@@ -360,8 +452,8 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
for(int d=0 ; d<_ndimension;d++){
block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
}
auto fineData_ = fineData.View();
auto coarseData_ = coarseData.View();
autoView( fineData_ , fineData, AcceleratorWrite);
autoView( coarseData_ , coarseData, AcceleratorRead);
// Loop with a cache friendly loop ordering
accelerator_for(sf,fine->oSites(),1,{
@@ -374,7 +466,7 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
for(int i=0;i<nbasis;i++) {
auto basis_ = Basis[i].View();
/* auto basis_ = Basis[i], );*/
if(i==0) fineData_[sf]=coarseData_[sc](i) *basis_[sf]);
else fineData_[sf]=fineData_[sf]+coarseData_[sc](i)*basis_[sf]);
}
@@ -383,24 +475,25 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
}
#else
template<class vobj,class CComplex,int nbasis>
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
Lattice<vobj> &fineData,
const std::vector<Lattice<vobj> > &Basis)
const VLattice &Basis)
{
GridBase * fine = fineData.Grid();
GridBase * coarse= coarseData.Grid();
fineData=Zero();
for(int i=0;i<nbasis;i++) {
Lattice<iScalar<CComplex> > ip = PeekIndex<0>(coarseData,i);
Lattice<CComplex> cip(coarse);
auto cip_ = cip.View();
auto ip_ = ip.View();
accelerator_forNB(sc,coarse->oSites(),CComplex::Nsimd(),{
coalescedWrite(cip_[sc], ip_(sc)());
});
blockZAXPY<vobj,CComplex >(fineData,cip,Basis[i],fineData);
//Lattice<CComplex> cip(coarse);
//autoView( cip_ , cip, AcceleratorWrite);
//autoView( ip_ , ip, AcceleratorRead);
//accelerator_forNB(sc,coarse->oSites(),CComplex::Nsimd(),{
// coalescedWrite(cip_[sc], ip_(sc)());
// });
//blockZAXPY<vobj,CComplex >(fineData,cip,Basis[i],fineData);
blockZAXPY(fineData,ip,Basis[i],fineData);
}
}
#endif
@@ -427,15 +520,17 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
assert(ig->lSites() == og->lSites());
}
autoView(in_v,in,CpuRead);
autoView(out_v,out,CpuWrite);
thread_for(idx, ig->lSites(),{
sobj s;
ssobj ss;
Coordinate lcoor(ni);
ig->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(s,in,lcoor);
peekLocalSite(s,in_v,lcoor);
ss=s;
pokeLocalSite(ss,out,lcoor);
pokeLocalSite(ss,out_v,lcoor);
});
}
@@ -470,8 +565,9 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
Coordinate rdt = Tg->_rdimensions;
Coordinate ist = Tg->_istride;
Coordinate ost = Tg->_ostride;
auto t_v = To.View();
auto f_v = From.View();
autoView( t_v , To, AcceleratorWrite);
autoView( f_v , From, AcceleratorRead);
accelerator_for(idx,Fg->lSites(),1,{
sobj s;
Coordinate Fcoor(nd);
@@ -494,8 +590,6 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
for(int w=0;w<words;w++){
tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd]; // FIXME IF RRII layout, type pun no worke
}
// peekLocalSite(s,From,Fcoor);
// pokeLocalSite(s,To ,Tcoor);
}
});
}
@@ -526,6 +620,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
}
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuRead);
autoView(higherDimv,higherDim,CpuWrite);
thread_for(idx,lg->lSites(),{
sobj s;
Coordinate lcoor(nl);
@@ -538,8 +634,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
hcoor[d]=lcoor[ddl++];
}
}
peekLocalSite(s,lowDim,lcoor);
pokeLocalSite(s,higherDim,hcoor);
peekLocalSite(s,lowDimv,lcoor);
pokeLocalSite(s,higherDimv,hcoor);
});
}
@@ -567,6 +663,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
}
}
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuWrite);
autoView(higherDimv,higherDim,CpuRead);
thread_for(idx,lg->lSites(),{
sobj s;
Coordinate lcoor(nl);
@@ -579,8 +677,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
hcoor[d]=lcoor[ddl++];
}
}
peekLocalSite(s,higherDim,hcoor);
pokeLocalSite(s,lowDim,lcoor);
peekLocalSite(s,higherDimv,hcoor);
pokeLocalSite(s,lowDimv,lcoor);
});
}
@@ -608,6 +706,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
}
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuRead);
autoView(higherDimv,higherDim,CpuWrite);
thread_for(idx,lg->lSites(),{
sobj s;
Coordinate lcoor(nl);
@@ -616,8 +716,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
if( lcoor[orthog] == slice_lo ) {
hcoor=lcoor;
hcoor[orthog] = slice_hi;
peekLocalSite(s,lowDim,lcoor);
pokeLocalSite(s,higherDim,hcoor);
peekLocalSite(s,lowDimv,lcoor);
pokeLocalSite(s,higherDimv,hcoor);
}
});
}
@@ -645,6 +745,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
}
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuWrite);
autoView(higherDimv,higherDim,CpuRead);
thread_for(idx,lg->lSites(),{
sobj s;
Coordinate lcoor(nl);
@@ -653,8 +755,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
if( lcoor[orthog] == slice_lo ) {
hcoor=lcoor;
hcoor[orthog] = slice_hi;
peekLocalSite(s,higherDim,hcoor);
pokeLocalSite(s,lowDim,lcoor);
peekLocalSite(s,higherDimv,hcoor);
pokeLocalSite(s,lowDimv,lcoor);
}
});
}
@@ -718,7 +820,7 @@ unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in)
}
//loop over outer index
auto in_v = in.View();
autoView( in_v , in, CpuRead);
thread_for(in_oidx,in_grid->oSites(),{
//Assemble vector of pointers to output elements
ExtractPointerArray<sobj> out_ptrs(in_nsimd);
@@ -811,7 +913,7 @@ vectorizeFromLexOrdArray( std::vector<sobj> &in, Lattice<vobj> &out)
icoor[lane].resize(ndim);
grid->iCoorFromIindex(icoor[lane],lane);
}
auto out_v = out.View();
autoView( out_v , out, CpuWrite);
thread_for(oidx, grid->oSites(),{
//Assemble vector of pointers to output elements
ExtractPointerArray<sobj> ptrs(nsimd);
@@ -914,7 +1016,7 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in)
std::vector<SobjOut> in_slex_conv(in_grid->lSites());
unvectorizeToLexOrdArray(in_slex_conv, in);
auto out_v = out.View();
autoView( out_v , out, CpuWrite);
thread_for(out_oidx,out_grid->oSites(),{
Coordinate out_ocoor(ndim);
out_grid->oCoorFromOindex(out_ocoor, out_oidx);

10
Grid/lattice/Lattice_transpose.h
View File

@@ -38,16 +38,18 @@ NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////////////////////////////////////////////////
// Transpose
////////////////////////////////////////////////////////////////////////////////////////////////////
/*
template<class vobj>
inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
auto ret_v = ret.View();
auto lhs_v = lhs.View();
autoView( ret_v, ret, AcceleratorWrite);
autoView( lhs_v, lhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss], transpose(lhs_v(ss)));
});
return ret;
};
*/
////////////////////////////////////////////////////////////////////////////////////////////////////
// Index level dependent transpose
@@ -56,8 +58,8 @@ template<int Index,class vobj>
inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(vobj()))>
{
Lattice<decltype(transposeIndex<Index>(vobj()))> ret(lhs.Grid());
auto ret_v = ret.View();
auto lhs_v = lhs.View();
autoView( ret_v, ret, AcceleratorWrite);
autoView( lhs_v, lhs, AcceleratorRead);
accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss] , transposeIndex<Index>(lhs_v(ss)));
});

16
Grid/lattice/Lattice_unary.h
View File

@@ -35,8 +35,8 @@ NAMESPACE_BEGIN(Grid);
template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
autoView( rhs, rhs_i, AcceleratorRead);
autoView( ret, ret_i, AcceleratorWrite);
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),1,{
ret[ss]=pow(rhs[ss],y);
@@ -45,8 +45,8 @@ template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
}
template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
autoView( rhs , rhs_i, AcceleratorRead);
autoView( ret , ret_i, AcceleratorWrite);
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],mod(rhs(ss),y));
@@ -56,8 +56,8 @@ template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
Lattice<obj> ret_i(rhs_i.Grid());
auto ret = ret_i.View();
auto rhs = rhs_i.View();
autoView( ret , ret_i, AcceleratorWrite);
autoView( rhs , rhs_i, AcceleratorRead);
ret.Checkerboard() = rhs_i.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],div(rhs(ss),y));
@@ -67,8 +67,8 @@ template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs_i, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
autoView( rhs , rhs_i, AcceleratorRead);
autoView( ret , ret_i, AcceleratorWrite);
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));

168
Grid/lattice/Lattice_view.h Normal file
View File

@@ -0,0 +1,168 @@
#pragma once
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////////////
// Base class which can be used by traits to pick up behaviour
///////////////////////////////////////////////////////////////////
class LatticeBase {};
/////////////////////////////////////////////////////////////////////////////////////////
// Conformable checks; same instance of Grid required
/////////////////////////////////////////////////////////////////////////////////////////
void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
{
assert(lhs == rhs);
}
////////////////////////////////////////////////////////////////////////////
// Minimal base class containing only data valid to access from accelerator
// _odata will be a managed pointer in CUDA
////////////////////////////////////////////////////////////////////////////
// Force access to lattice through a view object.
// prevents writing of code that will not offload to GPU, but perhaps annoyingly
// strict since host could could in principle direct access through the lattice object
// Need to decide programming model.
#define LATTICE_VIEW_STRICT
template<class vobj> class LatticeAccelerator : public LatticeBase
{
protected:
//public:
GridBase *_grid;
int checkerboard;
vobj *_odata; // A managed pointer
uint64_t _odata_size;
ViewAdvise advise;
public:
accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr), advise(AdviseDefault) { };
accelerator_inline uint64_t oSites(void) const { return _odata_size; };
accelerator_inline int Checkerboard(void) const { return checkerboard; };
accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
accelerator_inline ViewAdvise Advise(void) const { return advise; };
accelerator_inline ViewAdvise &Advise(void) { return this->advise; }; // can assign advise on a container, not a view
accelerator_inline void Conformable(GridBase * &grid) const
{
if (grid) conformable(grid, _grid);
else grid = _grid;
};
// Host only
GridBase * getGrid(void) const { return _grid; };
};
/////////////////////////////////////////////////////////////////////////////////////////
// A View class which provides accessor to the data.
// This will be safe to call from accelerator_for and is trivially copy constructible
// The copy constructor for this will need to be used by device lambda functions
/////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
class LatticeView : public LatticeAccelerator<vobj>
{
public:
// Rvalue
ViewMode mode;
void * cpu_ptr;
#ifdef GRID_SIMT
accelerator_inline const typename vobj::scalar_object operator()(size_t i) const {
return coalescedRead(this->_odata[i]);
}
#else
accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
#endif
accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
accelerator_inline vobj & operator[](size_t i) { return this->_odata[i]; };
accelerator_inline uint64_t begin(void) const { return 0;};
accelerator_inline uint64_t end(void) const { return this->_odata_size; };
accelerator_inline uint64_t size(void) const { return this->_odata_size; };
LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me){}
LatticeView(const LatticeView<vobj> &refer_to_me) = default; // Trivially copyable
LatticeView(const LatticeAccelerator<vobj> &refer_to_me,ViewMode mode) : LatticeAccelerator<vobj> (refer_to_me)
{
this->ViewOpen(mode);
}
// Host functions
void ViewOpen(ViewMode mode)
{ // Translate the pointer, could save a copy. Could use a "Handle" and not save _odata originally in base
// std::cout << "View Open"<<std::hex<<this->_odata<<std::dec <<std::endl;
this->cpu_ptr = (void *)this->_odata;
this->mode = mode;
this->_odata =(vobj *)
MemoryManager::ViewOpen(this->cpu_ptr,
this->_odata_size*sizeof(vobj),
mode,
this->advise);
}
void ViewClose(void)
{ // Inform the manager
// std::cout << "View Close"<<std::hex<<this->cpu_ptr<<std::dec <<std::endl;
MemoryManager::ViewClose(this->cpu_ptr,this->mode);
}
};
// Little autoscope assister
template<class View>
class ViewCloser
{
View v; // Take a copy of view and call view close when I go out of scope automatically
public:
ViewCloser(View &_v) : v(_v) {};
~ViewCloser() { v.ViewClose(); }
};
#define autoView(l_v,l,mode) \
auto l_v = l.View(mode); \
ViewCloser<decltype(l_v)> _autoView##l_v(l_v);
/////////////////////////////////////////////////////////////////////////////////////////
// Lattice expression types used by ET to assemble the AST
//
// Need to be able to detect code paths according to the whether a lattice object or not
// so introduce some trait type things
/////////////////////////////////////////////////////////////////////////////////////////
class LatticeExpressionBase {};
template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
template<class T> struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
template <typename Op, typename _T1>
class LatticeUnaryExpression : public LatticeExpressionBase
{
public:
typedef typename ViewMap<_T1>::Type T1;
Op op;
T1 arg1;
LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
};
template <typename Op, typename _T1, typename _T2>
class LatticeBinaryExpression : public LatticeExpressionBase
{
public:
typedef typename ViewMap<_T1>::Type T1;
typedef typename ViewMap<_T2>::Type T2;
Op op;
T1 arg1;
T2 arg2;
LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
};
template <typename Op, typename _T1, typename _T2, typename _T3>
class LatticeTrinaryExpression : public LatticeExpressionBase
{
public:
typedef typename ViewMap<_T1>::Type T1;
typedef typename ViewMap<_T2>::Type T2;
typedef typename ViewMap<_T3>::Type T3;
Op op;
T1 arg1;
T2 arg2;
T3 arg3;
LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
};
NAMESPACE_END(Grid);

2
Grid/parallelIO/BinaryIO.h
View File

@@ -341,7 +341,7 @@ class BinaryIO {
int ieee32big = (format == std::string("IEEE32BIG"));
int ieee32 = (format == std::string("IEEE32"));
int ieee64big = (format == std::string("IEEE64BIG"));
int ieee64 = (format == std::string("IEEE64"));
int ieee64 = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
assert(ieee64||ieee32|ieee64big||ieee32big);
assert((ieee64+ieee32+ieee64big+ieee32big)==1);
//////////////////////////////////////////////////////////////////////////////

24
Grid/parallelIO/MetaData.h
View File

@@ -301,6 +301,30 @@ struct GaugeSimpleUnmunger {
};
};
template<class fobj,class sobj>
struct GaugeDoubleStoredMunger{
void operator()(fobj &in, sobj &out) {
for (int mu = 0; mu < Nds; mu++) {
for (int i = 0; i < Nc; i++) {
for (int j = 0; j < Nc; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}}
}
};
};
template <class fobj, class sobj>
struct GaugeDoubleStoredUnmunger {
void operator()(sobj &in, fobj &out) {
for (int mu = 0; mu < Nds; mu++) {
for (int i = 0; i < Nc; i++) {
for (int j = 0; j < Nc; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}}
}
};
};
template<class fobj,class sobj>
struct Gauge3x2munger{
void operator() (fobj &in,sobj &out){

2
Grid/parallelIO/NerscIO.h
View File

@@ -146,7 +146,7 @@ public:
int ieee32big = (format == std::string("IEEE32BIG"));
int ieee32 = (format == std::string("IEEE32"));
int ieee64big = (format == std::string("IEEE64BIG"));
int ieee64 = (format == std::string("IEEE64"));
int ieee64 = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
// depending on datatype, set up munger;

224
Grid/parallelIO/OpenQcdIO.h Normal file
View File

@@ -0,0 +1,224 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/parallelIO/OpenQcdIO.h
Copyright (C) 2015 - 2020
Author: Daniel Richtmann <daniel.richtmann@ur.de>
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 */
#pragma once
NAMESPACE_BEGIN(Grid);
struct OpenQcdHeader : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(OpenQcdHeader,
int, Nt,
int, Nx,
int, Ny,
int, Nz,
double, plaq);
};
class OpenQcdIO : public BinaryIO {
public:
static constexpr double normalisationFactor = Nc; // normalisation difference: grid 18, openqcd 6
static inline int readHeader(std::string file, GridBase* grid, FieldMetaData& field) {
OpenQcdHeader header;
{
std::ifstream fin(file, std::ios::in | std::ios::binary);
fin.read(reinterpret_cast<char*>(&header), sizeof(OpenQcdHeader));
assert(!fin.fail());
field.data_start = fin.tellg();
fin.close();
}
header.plaq /= normalisationFactor;
// sanity check (should trigger on endian issues)
assert(0 < header.Nt && header.Nt <= 1024);
assert(0 < header.Nx && header.Nx <= 1024);
assert(0 < header.Ny && header.Ny <= 1024);
assert(0 < header.Nz && header.Nz <= 1024);
field.dimension[0] = header.Nx;
field.dimension[1] = header.Ny;
field.dimension[2] = header.Nz;
field.dimension[3] = header.Nt;
std::cout << GridLogDebug << "header: " << header << std::endl;
std::cout << GridLogDebug << "grid dimensions: " << grid->_fdimensions << std::endl;
std::cout << GridLogDebug << "file dimensions: " << field.dimension << std::endl;
assert(grid->_ndimension == Nd);
for(int d = 0; d < Nd; d++)
assert(grid->_fdimensions[d] == field.dimension[d]);
field.plaquette = header.plaq;
return field.data_start;
}
template<class vsimd>
static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd>>& Umu,
FieldMetaData& header,
std::string file) {
typedef Lattice<iDoubleStoredColourMatrix<vsimd>> DoubleStoredGaugeField;
assert(Ns == 4 and Nd == 4 and Nc == 3);
auto grid = dynamic_cast<GridCartesian*>(Umu.Grid());
assert(grid != nullptr); assert(grid->_ndimension == Nd);
uint64_t offset = readHeader(file, Umu.Grid(), header);
FieldMetaData clone(header);
std::string format("IEEE64"); // they always store little endian double precsision
uint32_t nersc_csum, scidac_csuma, scidac_csumb;
GridCartesian* grid_openqcd = createOpenQcdGrid(grid);
GridRedBlackCartesian* grid_rb = SpaceTimeGrid::makeFourDimRedBlackGrid(grid);
typedef DoubleStoredColourMatrixD fobj;
typedef typename DoubleStoredGaugeField::vector_object::scalar_object sobj;
typedef typename DoubleStoredGaugeField::vector_object::Realified::scalar_type word;
word w = 0;
std::vector<fobj> iodata(grid_openqcd->lSites()); // Munge, checksum, byte order in here
std::vector<sobj> scalardata(grid->lSites());
IOobject(w, grid_openqcd, iodata, file, offset, format, BINARYIO_READ | BINARYIO_LEXICOGRAPHIC,
nersc_csum, scidac_csuma, scidac_csumb);
GridStopWatch timer;
timer.Start();
DoubleStoredGaugeField Umu_ds(grid);
auto munge = GaugeDoubleStoredMunger<DoubleStoredColourMatrixD, DoubleStoredColourMatrix>();
Coordinate ldim = grid->LocalDimensions();
thread_for(idx_g, grid->lSites(), {
Coordinate coor;
grid->LocalIndexToLocalCoor(idx_g, coor);
bool isOdd = grid_rb->CheckerBoard(coor) == Odd;
if(!isOdd) continue;
int idx_o = (coor[Tdir] * ldim[Xdir] * ldim[Ydir] * ldim[Zdir]
+ coor[Xdir] * ldim[Ydir] * ldim[Zdir]
+ coor[Ydir] * ldim[Zdir]
+ coor[Zdir])/2;
munge(iodata[idx_o], scalardata[idx_g]);
});
grid->Barrier(); timer.Stop();
std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: munge overhead " << timer.Elapsed() << std::endl;
timer.Reset(); timer.Start();
vectorizeFromLexOrdArray(scalardata, Umu_ds);
grid->Barrier(); timer.Stop();
std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: vectorize overhead " << timer.Elapsed() << std::endl;
timer.Reset(); timer.Start();
undoDoubleStore(Umu, Umu_ds);
grid->Barrier(); timer.Stop();
std::cout << Grid::GridLogMessage << "OpenQcdIO::readConfiguration: redistribute overhead " << timer.Elapsed() << std::endl;
GaugeStatistics(Umu, clone);
RealD plaq_diff = fabs(clone.plaquette - header.plaquette);
// clang-format off
std::cout << GridLogMessage << "OpenQcd Configuration " << file
<< " plaquette " << clone.plaquette
<< " header " << header.plaquette
<< " difference " << plaq_diff
<< std::endl;
// clang-format on
RealD precTol = (getPrecision<vsimd>::value == 1) ? 2e-7 : 2e-15;
RealD tol = precTol * std::sqrt(grid->_Nprocessors); // taken from RQCD chroma code
if(plaq_diff >= tol)
std::cout << " Plaquette mismatch (diff = " << plaq_diff << ", tol = " << tol << ")" << std::endl;
assert(plaq_diff < tol);
std::cout << GridLogMessage << "OpenQcd Configuration " << file << " and plaquette agree" << std::endl;
}
template<class vsimd>
static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd>>& Umu,
std::string file) {
std::cout << GridLogError << "Writing to openQCD file format is not implemented" << std::endl;
exit(EXIT_FAILURE);
}
private:
static inline GridCartesian* createOpenQcdGrid(GridCartesian* grid) {
// exploit GridCartesian to be able to still use IOobject
Coordinate gdim = grid->GlobalDimensions();
Coordinate ldim = grid->LocalDimensions();
Coordinate pcoor = grid->ThisProcessorCoor();
// openqcd does rb on the z direction
gdim[Zdir] /= 2;
ldim[Zdir] /= 2;
// and has the order T X Y Z (from slowest to fastest)
std::swap(gdim[Xdir], gdim[Zdir]);
std::swap(ldim[Xdir], ldim[Zdir]);
std::swap(pcoor[Xdir], pcoor[Zdir]);
GridCartesian* ret = SpaceTimeGrid::makeFourDimGrid(gdim, grid->_simd_layout, grid->ProcessorGrid());
ret->_ldimensions = ldim;
ret->_processor_coor = pcoor;
return ret;
}
template<class vsimd>
static inline void undoDoubleStore(Lattice<iLorentzColourMatrix<vsimd>>& Umu,
Lattice<iDoubleStoredColourMatrix<vsimd>> const& Umu_ds) {
conformable(Umu.Grid(), Umu_ds.Grid());
Lattice<iColourMatrix<vsimd>> U(Umu.Grid());
// they store T+, T-, X+, X-, Y+, Y-, Z+, Z-
for(int mu_g = 0; mu_g < Nd; ++mu_g) {
int mu_o = (mu_g + 1) % Nd;
U = PeekIndex<LorentzIndex>(Umu_ds, 2 * mu_o)
+ Cshift(PeekIndex<LorentzIndex>(Umu_ds, 2 * mu_o + 1), mu_g, +1);
PokeIndex<LorentzIndex>(Umu, U, mu_g);
}
}
};
NAMESPACE_END(Grid);

281
Grid/parallelIO/OpenQcdIOChromaReference.h Normal file
View File

@@ -0,0 +1,281 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/parallelIO/OpenQcdIOChromaReference.h
Copyright (C) 2015 - 2020
Author: Daniel Richtmann <daniel.richtmann@ur.de>
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 */
#pragma once
#include <ios>
#include <iostream>
#include <limits>
#include <iomanip>
#include <mpi.h>
#include <ostream>
#include <string>
#define CHECK {std::cerr << __FILE__ << " @l " << __LINE__ << ": CHECK" << grid->ThisRank() << std::endl;}
#define CHECK_VAR(a) { std::cerr << __FILE__ << "@l" << __LINE__ << " on "<< grid->ThisRank() << ": " << __func__ << " " << #a << "=" << (a) << std::endl; }
// #undef CHECK
// #define CHECK
NAMESPACE_BEGIN(Grid);
class ParRdr {
private:
bool const swap;
MPI_Status status;
MPI_File fp;
int err;
MPI_Datatype oddSiteType;
MPI_Datatype fileViewType;
GridBase* grid;
public:
ParRdr(MPI_Comm comm, std::string const& filename, GridBase* gridPtr)
: swap(false)
, grid(gridPtr) {
err = MPI_File_open(comm, const_cast<char*>(filename.c_str()), MPI_MODE_RDONLY, MPI_INFO_NULL, &fp);
assert(err == MPI_SUCCESS);
}
virtual ~ParRdr() { MPI_File_close(&fp); }
inline void errInfo(int const err, std::string const& func) {
static char estring[MPI_MAX_ERROR_STRING];
int eclass = -1, len = 0;
MPI_Error_class(err, &eclass);
MPI_Error_string(err, estring, &len);
std::cerr << func << " - Error " << eclass << ": " << estring << std::endl;
}
int readHeader(FieldMetaData& field) {
assert((grid->_ndimension == Nd) && (Nd == 4));
assert(Nc == 3);
OpenQcdHeader header;
readBlock(reinterpret_cast<char*>(&header), 0, sizeof(OpenQcdHeader), MPI_CHAR);
header.plaq /= 3.; // TODO change this into normalizationfactor
// sanity check (should trigger on endian issues) TODO remove?
assert(0 < header.Nt && header.Nt <= 1024);
assert(0 < header.Nx && header.Nx <= 1024);
assert(0 < header.Ny && header.Ny <= 1024);
assert(0 < header.Nz && header.Nz <= 1024);
field.dimension[0] = header.Nx;
field.dimension[1] = header.Ny;
field.dimension[2] = header.Nz;
field.dimension[3] = header.Nt;
for(int d = 0; d < Nd; d++)
assert(grid->FullDimensions()[d] == field.dimension[d]);
field.plaquette = header.plaq;
field.data_start = sizeof(OpenQcdHeader);
return field.data_start;
}
void readBlock(void* const dest, uint64_t const pos, uint64_t const nbytes, MPI_Datatype const datatype) {
err = MPI_File_read_at_all(fp, pos, dest, nbytes, datatype, &status);
errInfo(err, "MPI_File_read_at_all");
// CHECK_VAR(err)
int read = -1;
MPI_Get_count(&status, datatype, &read);
// CHECK_VAR(read)
assert(nbytes == (uint64_t)read);
assert(err == MPI_SUCCESS);
}
void createTypes() {
constexpr int elem_size = Nd * 2 * 2 * Nc * Nc * sizeof(double); // 2_complex 2_fwdbwd
err = MPI_Type_contiguous(elem_size, MPI_BYTE, &oddSiteType); assert(err == MPI_SUCCESS);
err = MPI_Type_commit(&oddSiteType); assert(err == MPI_SUCCESS);
Coordinate const L = grid->GlobalDimensions();
Coordinate const l = grid->LocalDimensions();
Coordinate const i = grid->ThisProcessorCoor();
Coordinate sizes({L[2] / 2, L[1], L[0], L[3]});
Coordinate subsizes({l[2] / 2, l[1], l[0], l[3]});
Coordinate starts({i[2] * l[2] / 2, i[1] * l[1], i[0] * l[0], i[3] * l[3]});
err = MPI_Type_create_subarray(grid->_ndimension, &sizes[0], &subsizes[0], &starts[0], MPI_ORDER_FORTRAN, oddSiteType, &fileViewType); assert(err == MPI_SUCCESS);
err = MPI_Type_commit(&fileViewType); assert(err == MPI_SUCCESS);
}
void freeTypes() {
err = MPI_Type_free(&fileViewType); assert(err == MPI_SUCCESS);
err = MPI_Type_free(&oddSiteType); assert(err == MPI_SUCCESS);
}
bool readGauge(std::vector<ColourMatrixD>& domain_buff, FieldMetaData& meta) {
auto hdr_offset = readHeader(meta);
CHECK
createTypes();
err = MPI_File_set_view(fp, hdr_offset, oddSiteType, fileViewType, "native", MPI_INFO_NULL); errInfo(err, "MPI_File_set_view0"); assert(err == MPI_SUCCESS);
CHECK
int const domainSites = grid->lSites();
domain_buff.resize(Nd * domainSites); // 2_fwdbwd * 4_Nd * domainSites / 2_onlyodd
// the actual READ
constexpr uint64_t cm_size = 2 * Nc * Nc * sizeof(double); // 2_complex
constexpr uint64_t os_size = Nd * 2 * cm_size; // 2_fwdbwd
constexpr uint64_t max_elems = std::numeric_limits<int>::max(); // int adressable elems: floor is fine
uint64_t const n_os = domainSites / 2;
for(uint64_t os_idx = 0; os_idx < n_os;) {
uint64_t const read_os = os_idx + max_elems <= n_os ? max_elems : n_os - os_idx;
uint64_t const cm = os_idx * Nd * 2;
readBlock(&(domain_buff[cm]), os_idx, read_os, oddSiteType);
os_idx += read_os;
}
CHECK
err = MPI_File_set_view(fp, 0, MPI_BYTE, MPI_BYTE, "native", MPI_INFO_NULL);
errInfo(err, "MPI_File_set_view1");
assert(err == MPI_SUCCESS);
freeTypes();
std::cout << GridLogMessage << "read sum: " << n_os * os_size << " bytes" << std::endl;
return true;
}
};
class OpenQcdIOChromaReference : public BinaryIO {
public:
template<class vsimd>
static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd>>& Umu,
Grid::FieldMetaData& header,
std::string file) {
typedef Lattice<iDoubleStoredColourMatrix<vsimd>> DoubledGaugeField;
assert(Ns == 4 and Nd == 4 and Nc == 3);
auto grid = Umu.Grid();
typedef ColourMatrixD fobj;
std::vector<fobj> iodata(
Nd * grid->lSites()); // actual size = 2*Nd*lsites but have only lsites/2 sites in file
{
ParRdr rdr(MPI_COMM_WORLD, file, grid);
rdr.readGauge(iodata, header);
} // equivalent to using binaryio
std::vector<iDoubleStoredColourMatrix<typename vsimd::scalar_type>> Umu_ds_scalar(grid->lSites());
copyToLatticeObject(Umu_ds_scalar, iodata, grid); // equivalent to munging
DoubledGaugeField Umu_ds(grid);
vectorizeFromLexOrdArray(Umu_ds_scalar, Umu_ds);
redistribute(Umu, Umu_ds); // equivalent to undoDoublestore
FieldMetaData clone(header);
GaugeStatistics(Umu, clone);
RealD plaq_diff = fabs(clone.plaquette - header.plaquette);
// clang-format off
std::cout << GridLogMessage << "OpenQcd Configuration " << file
<< " plaquette " << clone.plaquette
<< " header " << header.plaquette
<< " difference " << plaq_diff
<< std::endl;
// clang-format on
RealD precTol = (getPrecision<vsimd>::value == 1) ? 2e-7 : 2e-15;
RealD tol = precTol * std::sqrt(grid->_Nprocessors); // taken from RQCD chroma code
if(plaq_diff >= tol)
std::cout << " Plaquette mismatch (diff = " << plaq_diff << ", tol = " << tol << ")" << std::endl;
assert(plaq_diff < tol);
std::cout << GridLogMessage << "OpenQcd Configuration " << file << " and plaquette agree" << std::endl;
}
private:
template<class vsimd>
static inline void redistribute(Lattice<iLorentzColourMatrix<vsimd>>& Umu,
Lattice<iDoubleStoredColourMatrix<vsimd>> const& Umu_ds) {
Grid::conformable(Umu.Grid(), Umu_ds.Grid());
Lattice<iColourMatrix<vsimd>> U(Umu.Grid());
U = PeekIndex<LorentzIndex>(Umu_ds, 2) + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 3), 0, +1); PokeIndex<LorentzIndex>(Umu, U, 0);
U = PeekIndex<LorentzIndex>(Umu_ds, 4) + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 5), 1, +1); PokeIndex<LorentzIndex>(Umu, U, 1);
U = PeekIndex<LorentzIndex>(Umu_ds, 6) + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 7), 2, +1); PokeIndex<LorentzIndex>(Umu, U, 2);
U = PeekIndex<LorentzIndex>(Umu_ds, 0) + Cshift(PeekIndex<LorentzIndex>(Umu_ds, 1), 3, +1); PokeIndex<LorentzIndex>(Umu, U, 3);
}
static inline void copyToLatticeObject(std::vector<DoubleStoredColourMatrix>& u_fb,
std::vector<ColourMatrixD> const& node_buff,
GridBase* grid) {
assert(node_buff.size() == Nd * grid->lSites());
Coordinate const& l = grid->LocalDimensions();
Coordinate coord(Nd);
int& x = coord[0];
int& y = coord[1];
int& z = coord[2];
int& t = coord[3];
int buff_idx = 0;
for(t = 0; t < l[3]; ++t) // IMPORTANT: openQCD file ordering
for(x = 0; x < l[0]; ++x)
for(y = 0; y < l[1]; ++y)
for(z = 0; z < l[2]; ++z) {
if((t + z + y + x) % 2 == 0) continue;
int local_idx;
Lexicographic::IndexFromCoor(coord, local_idx, grid->LocalDimensions());
for(int mu = 0; mu < 2 * Nd; ++mu)
for(int c1 = 0; c1 < Nc; ++c1) {
for(int c2 = 0; c2 < Nc; ++c2) {
u_fb[local_idx](mu)()(c1,c2) = node_buff[mu+buff_idx]()()(c1,c2);
}
}
buff_idx += 2 * Nd;
}
assert(node_buff.size() == buff_idx);
}
};
NAMESPACE_END(Grid);

6
Grid/perfmon/PerfCount.h
View File

@@ -44,7 +44,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <sys/syscall.h>
#endif
#ifdef __x86_64__
#ifdef GRID_NVCC
#ifdef GRID_CUDA
accelerator_inline uint64_t __rdtsc(void) { return 0; }
accelerator_inline uint64_t __rdpmc(int ) { return 0; }
#else
@@ -95,7 +95,8 @@ inline uint64_t cyclecount(void){
}
#elif defined __x86_64__
inline uint64_t cyclecount(void){
return __rdtsc();
uint64_t ret = __rdtsc();
return (uint64_t)ret;
}
#else
@@ -111,7 +112,6 @@ class PerformanceCounter {
private:
typedef struct {
public:
uint32_t type;
uint64_t config;
const char *name;

6
Grid/perfmon/Timer.h
View File

@@ -110,15 +110,15 @@ public:
#endif
accumulator = std::chrono::duration_cast<GridUsecs>(start-start);
}
GridTime Elapsed(void) {
GridTime Elapsed(void) const {
assert(running == false);
return std::chrono::duration_cast<GridTime>( accumulator );
}
uint64_t useconds(void){
uint64_t useconds(void) const {
assert(running == false);
return (uint64_t) accumulator.count();
}
bool isRunning(void){
bool isRunning(void) const {
return running;
}
};

2
Grid/pugixml/pugixml.cc
View File

@@ -12773,7 +12773,7 @@ namespace pugi
#undef PUGI__THROW_ERROR
#undef PUGI__CHECK_ERROR
#ifdef GRID_NVCC
#ifdef GRID_CUDA
#pragma pop
#endif

40
Grid/qcd/QCD.h
View File

@@ -133,23 +133,23 @@ typedef iSpinColourMatrix<vComplex > vSpinColourMatrix;
typedef iSpinColourMatrix<vComplexF> vSpinColourMatrixF;
typedef iSpinColourMatrix<vComplexD> vSpinColourMatrixD;
// SpinColourSpinColour matrix
typedef iSpinColourSpinColourMatrix<Complex > SpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<ComplexF > SpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<ComplexD > SpinColourSpinColourMatrixD;
// SpinColourSpinColour matrix
typedef iSpinColourSpinColourMatrix<Complex > SpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<ComplexF > SpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<ComplexD > SpinColourSpinColourMatrixD;
typedef iSpinColourSpinColourMatrix<vComplex > vSpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<vComplexF> vSpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<vComplexD> vSpinColourSpinColourMatrixD;
typedef iSpinColourSpinColourMatrix<vComplex > vSpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<vComplexF> vSpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<vComplexD> vSpinColourSpinColourMatrixD;
// SpinColourSpinColour matrix
typedef iSpinColourSpinColourMatrix<Complex > SpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<ComplexF > SpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<ComplexD > SpinColourSpinColourMatrixD;
// SpinColourSpinColour matrix
typedef iSpinColourSpinColourMatrix<Complex > SpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<ComplexF > SpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<ComplexD > SpinColourSpinColourMatrixD;
typedef iSpinColourSpinColourMatrix<vComplex > vSpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<vComplexF> vSpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<vComplexD> vSpinColourSpinColourMatrixD;
typedef iSpinColourSpinColourMatrix<vComplex > vSpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<vComplexF> vSpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<vComplexD> vSpinColourSpinColourMatrixD;
// LorentzColour
typedef iLorentzColourMatrix<Complex > LorentzColourMatrix;
@@ -443,9 +443,9 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
//////////////////////////////////////////////
// Fermion <-> propagator assignements
//////////////////////////////////////////////
//template <class Prop, class Ferm>
template <class Fimpl>
void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
//template <class Prop, class Ferm>
template <class Fimpl>
void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
{
for(int j = 0; j < Ns; ++j)
{
@@ -460,9 +460,9 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
}
}
//template <class Prop, class Ferm>
template <class Fimpl>
void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
//template <class Prop, class Ferm>
template <class Fimpl>
void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
{
for(int j = 0; j < Ns; ++j)
{

30
Grid/qcd/action/fermion/CayleyFermion5D.h
View File

@@ -40,8 +40,8 @@ public:
public:
// override multiply
virtual RealD M (const FermionField &in, FermionField &out);
virtual RealD Mdag (const FermionField &in, FermionField &out);
virtual void M (const FermionField &in, FermionField &out);
virtual void Mdag (const FermionField &in, FermionField &out);
// half checkerboard operations
virtual void Meooe (const FermionField &in, FermionField &out);
@@ -141,7 +141,33 @@ public:
Vector<iSinglet<Simd> > MatpInvDag;
Vector<iSinglet<Simd> > MatmInvDag;
///////////////////////////////////////////////////////////////
// Conserved current utilities
///////////////////////////////////////////////////////////////
// Virtual can't template
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx);
void ContractJ5q(PropagatorField &q_in,ComplexField &J5q);
void ContractJ5q(FermionField &q_in,ComplexField &J5q);
///////////////////////////////////////////////////////////////
// Constructors
///////////////////////////////////////////////////////////////
CayleyFermion5D(GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,

4
Grid/qcd/action/fermion/ContinuedFractionFermion5D.h
View File

@@ -41,8 +41,8 @@ public:
public:
// override multiply
virtual RealD M (const FermionField &in, FermionField &out);
virtual RealD Mdag (const FermionField &in, FermionField &out);
virtual void M (const FermionField &in, FermionField &out);
virtual void Mdag (const FermionField &in, FermionField &out);
// half checkerboard operaions
virtual void Meooe (const FermionField &in, FermionField &out);

4
Grid/qcd/action/fermion/DomainWallEOFAFermion.h
View File

@@ -53,8 +53,8 @@ public:
virtual void DtildeInv (const FermionField& in, FermionField& out);
// override multiply
virtual RealD M (const FermionField& in, FermionField& out);
virtual RealD Mdag (const FermionField& in, FermionField& out);
virtual void M (const FermionField& in, FermionField& out);
virtual void Mdag (const FermionField& in, FermionField& out);
// half checkerboard operations
virtual void Mooee (const FermionField& in, FermionField& out);

13
Grid/qcd/action/fermion/DomainWallVec5dImpl.h
View File

@@ -115,18 +115,21 @@ public:
PokeIndex<LorentzIndex>(Uadj, U, mu);
}
for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
autoView(Umu_v,Umu,CpuRead);
autoView(Uadj_v,Uadj,CpuRead);
autoView(Uds_v,Uds,CpuWrite);
thread_for( lidx, GaugeGrid->lSites(), {
Coordinate lcoor;
GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
peekLocalSite(ScalarUmu, Umu, lcoor);
peekLocalSite(ScalarUmu, Umu_v, lcoor);
for (int mu = 0; mu < 4; mu++) ScalarUds(mu) = ScalarUmu(mu);
peekLocalSite(ScalarUmu, Uadj, lcoor);
peekLocalSite(ScalarUmu, Uadj_v, lcoor);
for (int mu = 0; mu < 4; mu++) ScalarUds(mu + 4) = ScalarUmu(mu);
pokeLocalSite(ScalarUds, Uds, lcoor);
}
pokeLocalSite(ScalarUds, Uds_v, lcoor);
});
}
inline void InsertForce4D(GaugeField &mat, FermionField &Btilde,FermionField &A, int mu)

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

@@ -57,6 +57,7 @@ NAMESPACE_CHECK(WilsonClover);
#include <Grid/qcd/action/fermion/WilsonFermion5D.h> // 5d base used by all 5d overlap types
NAMESPACE_CHECK(Wilson5D);
#include <Grid/qcd/action/fermion/NaiveStaggeredFermion.h>
#include <Grid/qcd/action/fermion/ImprovedStaggeredFermion.h>
#include <Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h>
NAMESPACE_CHECK(Staggered);
@@ -282,11 +283,15 @@ typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD;
typedef NaiveStaggeredFermion<StaggeredImplR> NaiveStaggeredFermionR;
typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
typedef ImprovedStaggeredFermion5D<StaggeredImplR> ImprovedStaggeredFermion5DR;
typedef ImprovedStaggeredFermion5D<StaggeredImplF> ImprovedStaggeredFermion5DF;
typedef ImprovedStaggeredFermion5D<StaggeredImplD> ImprovedStaggeredFermion5DD;
#ifndef GRID_NVCC
#ifndef GRID_CUDA
typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplR> ImprovedStaggeredFermionVec5dR;
typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplF> ImprovedStaggeredFermionVec5dF;
typedef ImprovedStaggeredFermion5D<StaggeredVec5dImplD> ImprovedStaggeredFermionVec5dD;

13
Grid/qcd/action/fermion/FermionOperator.h
View File

@@ -58,8 +58,8 @@ public:
virtual GridBase *GaugeRedBlackGrid(void) =0;
// override multiply
virtual RealD M (const FermionField &in, FermionField &out)=0;
virtual RealD Mdag (const FermionField &in, FermionField &out)=0;
virtual void M (const FermionField &in, FermionField &out)=0;
virtual void Mdag (const FermionField &in, FermionField &out)=0;
// half checkerboard operaions
virtual void Meooe (const FermionField &in, FermionField &out)=0;
@@ -86,7 +86,6 @@ public:
virtual void DhopDerivEO(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)=0;
virtual void DhopDerivOE(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)=0;
virtual void Mdiag (const FermionField &in, FermionField &out) { Mooee(in,out);}; // Same as Mooee applied to both CB's
virtual void Mdir (const FermionField &in, FermionField &out,int dir,int disp)=0; // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
virtual void MdirAll(const FermionField &in, std::vector<FermionField> &out)=0; // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
@@ -148,15 +147,19 @@ public:
virtual void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu)=0;
unsigned int mu)
{assert(0);};
virtual void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx)=0;
ComplexField &lattice_cmplx)
{assert(0);};
// Only reimplemented in Wilson5D
// Default to just a zero correlation function

53
Grid/qcd/action/fermion/GparityWilsonImpl.h
View File

@@ -38,6 +38,7 @@ public:
static const bool isFundamental = Representation::isFundamental;
static const int Nhcs = Options::Nhcs;
static const bool LsVectorised=false;
static const bool isGparity=true;
typedef ConjugateGaugeImpl< GaugeImplTypes<S,Dimension> > Gimpl;
INHERIT_GIMPL_TYPES(Gimpl);
@@ -46,7 +47,7 @@ public:
typedef typename Options::template PrecisionMapper<Simd>::LowerPrecVector SimdL;
template <typename vtype> using iImplSpinor = iVector<iVector<iVector<vtype, Dimension>, Ns>, Ngp>;
template <typename vtype> using iImplPropagator = iVector<iMatrix<iMatrix<vtype, Dimension>, Ns>, Ngp>;
template <typename vtype> using iImplPropagator = iMatrix<iMatrix<iMatrix<vtype, Dimension>, Ns>, Ngp>;
template <typename vtype> using iImplHalfSpinor = iVector<iVector<iVector<vtype, Dimension>, Nhs>, Ngp>;
template <typename vtype> using iImplHalfCommSpinor = iVector<iVector<iVector<vtype, Dimension>, Nhcs>, Ngp>;
template <typename vtype> using iImplDoubledGaugeField = iVector<iVector<iScalar<iMatrix<vtype, Dimension> >, Nds>, Ngp>;
@@ -80,6 +81,7 @@ public:
{
assert(0);
}
template<class _Spinor>
static accelerator_inline void multLink(_Spinor &phi,
const SiteDoubledGaugeField &U,
@@ -94,11 +96,11 @@ public:
int sl = St._simd_layout[direction];
Coordinate icoor;
#ifdef __CUDA_ARCH__
#ifdef GRID_SIMT
_Spinor tmp;
const int Nsimd =SiteDoubledGaugeField::Nsimd();
int s = SIMTlane(Nsimd);
int s = acceleratorSIMTlane(Nsimd);
St.iCoorFromIindex(icoor,s);
int mmu = mu % Nd;
@@ -191,6 +193,16 @@ public:
#endif
}
template<class _SpinorField>
inline void multLinkField(_SpinorField & out,
const DoubledGaugeField &Umu,
const _SpinorField & phi,
int mu)
{
assert(0);
}
template <class ref>
static accelerator_inline void loadLinkElement(Simd &reg, ref &memory)
{
@@ -221,14 +233,16 @@ public:
Uconj = where(coor==neglink,-Uconj,Uconj);
}
auto U_v = U.View();
auto Uds_v = Uds.View();
auto Uconj_v = Uconj.View();
auto Utmp_v= Utmp.View();
{
autoView( U_v , U, CpuRead);
autoView( Uconj_v , Uconj, CpuRead);
autoView( Uds_v , Uds, CpuWrite);
autoView( Utmp_v, Utmp, CpuWrite);
thread_foreach(ss,U_v,{
Uds_v[ss](0)(mu) = U_v[ss]();
Uds_v[ss](1)(mu) = Uconj_v[ss]();
});
}
U = adj(Cshift(U ,mu,-1)); // correct except for spanning the boundary
Uconj = adj(Cshift(Uconj,mu,-1));
@@ -238,19 +252,25 @@ public:
Utmp = where(coor==0,Uconj,Utmp);
}
{
autoView( Uds_v , Uds, CpuWrite);
autoView( Utmp_v, Utmp, CpuWrite);
thread_foreach(ss,Utmp_v,{
Uds_v[ss](0)(mu+4) = Utmp_v[ss]();
});
}
Utmp = Uconj;
if ( Params.twists[mu] ) {
Utmp = where(coor==0,U,Utmp);
}
{
autoView( Uds_v , Uds, CpuWrite);
autoView( Utmp_v, Utmp, CpuWrite);
thread_foreach(ss,Utmp_v,{
Uds_v[ss](1)(mu+4) = Utmp_v[ss]();
});
}
}
}
@@ -260,11 +280,14 @@ public:
GaugeLinkField link(mat.Grid());
// use lorentz for flavour as hack.
auto tmp = TraceIndex<SpinIndex>(outerProduct(Btilde, A));
auto link_v = link.View();
auto tmp_v = tmp.View();
{
autoView( link_v , link, CpuWrite);
autoView( tmp_v , tmp, CpuRead);
thread_foreach(ss,tmp_v,{
link_v[ss]() = tmp_v[ss](0, 0) + conjugate(tmp_v[ss](1, 1));
});
}
PokeIndex<LorentzIndex>(mat, link, mu);
return;
}
@@ -294,9 +317,10 @@ public:
GaugeLinkField tmp(mat.Grid());
tmp = Zero();
auto tmp_v = tmp.View();
auto Atilde_v = Atilde.View();
auto Btilde_v = Btilde.View();
{
autoView( tmp_v , tmp, CpuWrite);
autoView( Atilde_v , Atilde, CpuRead);
autoView( Btilde_v , Btilde, CpuRead);
thread_for(ss,tmp.Grid()->oSites(),{
for (int s = 0; s < Ls; s++) {
int sF = s + Ls * ss;
@@ -304,6 +328,7 @@ public:
tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
}
});
}
PokeIndex<LorentzIndex>(mat, tmp, mu);
return;
}

6
Grid/qcd/action/fermion/ImprovedStaggeredFermion.h
View File

@@ -71,8 +71,8 @@ public:
// override multiply; cut number routines if pass dagger argument
// and also make interface more uniformly consistent
//////////////////////////////////////////////////////////////////
RealD M(const FermionField &in, FermionField &out);
RealD Mdag(const FermionField &in, FermionField &out);
void M(const FermionField &in, FermionField &out);
void Mdag(const FermionField &in, FermionField &out);
/////////////////////////////////////////////////////////
// half checkerboard operations
@@ -185,10 +185,12 @@ public:
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &srct,
Current curr_type,
unsigned int mu,
unsigned int tmin,

7
Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h
View File

@@ -1,4 +1,3 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@@ -74,8 +73,8 @@ public:
GridBase *FermionRedBlackGrid(void) { return _FiveDimRedBlackGrid;}
// full checkerboard operations; leave unimplemented as abstract for now
RealD M (const FermionField &in, FermionField &out);
RealD Mdag (const FermionField &in, FermionField &out);
void M (const FermionField &in, FermionField &out);
void Mdag (const FermionField &in, FermionField &out);
// half checkerboard operations
void Meooe (const FermionField &in, FermionField &out);
@@ -217,10 +216,12 @@ public:
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu,
unsigned int tmin,

4
Grid/qcd/action/fermion/MobiusEOFAFermion.h
View File

@@ -56,8 +56,8 @@ public:
virtual void DtildeInv (const FermionField& in, FermionField& out);
// override multiply
virtual RealD M (const FermionField& in, FermionField& out);
virtual RealD Mdag (const FermionField& in, FermionField& out);
virtual void M (const FermionField& in, FermionField& out);
virtual void Mdag (const FermionField& in, FermionField& out);
// half checkerboard operations
virtual void Mooee (const FermionField& in, FermionField& out);

2
Grid/qcd/action/fermion/MobiusFermion.h
View File

@@ -59,7 +59,7 @@ public:
{
RealD eps = 1.0;
std::cout<<GridLogMessage << "MobiusFermion (b="<<b<<",c="<<c<<") with Ls= "<<this->Ls<<" Tanh approx"<<std::endl;
// std::cout<<GridLogMessage << "MobiusFermion (b="<<b<<",c="<<c<<") with Ls= "<<this->Ls<<" Tanh approx"<<std::endl;
Approx::zolotarev_data *zdata = Approx::higham(eps,this->Ls);// eps is ignored for higham
assert(zdata->n==this->Ls);

194
Grid/qcd/action/fermion/NaiveStaggeredFermion.h Normal file
View File

@@ -0,0 +1,194 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/ImprovedStaggered.h
Copyright (C) 2015
Author: Azusa Yamaguchi, Peter Boyle
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 */
#ifndef GRID_QCD_NAIVE_STAG_FERMION_H
#define GRID_QCD_NAIVE_STAG_FERMION_H
NAMESPACE_BEGIN(Grid);
class NaiveStaggeredFermionStatic {
public:
static const std::vector<int> directions;
static const std::vector<int> displacements;
static const int npoint = 8;
};
template <class Impl>
class NaiveStaggeredFermion : public StaggeredKernels<Impl>, public NaiveStaggeredFermionStatic {
public:
INHERIT_IMPL_TYPES(Impl);
typedef StaggeredKernels<Impl> Kernels;
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
////////////////////////////////////////
// Performance monitoring
////////////////////////////////////////
void Report(void);
void ZeroCounters(void);
double DhopTotalTime;
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
double DhopComputeTime2;
double DhopFaceTime;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////
GridBase *GaugeGrid(void) { return _grid; }
GridBase *GaugeRedBlackGrid(void) { return _cbgrid; }
GridBase *FermionGrid(void) { return _grid; }
GridBase *FermionRedBlackGrid(void) { return _cbgrid; }
//////////////////////////////////////////////////////////////////
// override multiply; cut number routines if pass dagger argument
// and also make interface more uniformly consistent
//////////////////////////////////////////////////////////////////
void M(const FermionField &in, FermionField &out);
void Mdag(const FermionField &in, FermionField &out);
/////////////////////////////////////////////////////////
// half checkerboard operations
/////////////////////////////////////////////////////////
void Meooe(const FermionField &in, FermionField &out);
void MeooeDag(const FermionField &in, FermionField &out);
void Mooee(const FermionField &in, FermionField &out);
void MooeeDag(const FermionField &in, FermionField &out);
void MooeeInv(const FermionField &in, FermionField &out);
void MooeeInvDag(const FermionField &in, FermionField &out);
////////////////////////
// Derivative interface
////////////////////////
// Interface calls an internal routine
void DhopDeriv (GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
void DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
void DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag);
///////////////////////////////////////////////////////////////
// non-hermitian hopping term; half cb or both
///////////////////////////////////////////////////////////////
void Dhop (const FermionField &in, FermionField &out, int dag);
void DhopOE(const FermionField &in, FermionField &out, int dag);
void DhopEO(const FermionField &in, FermionField &out, int dag);
///////////////////////////////////////////////////////////////
// Multigrid assistance; force term uses too
///////////////////////////////////////////////////////////////
void Mdir(const FermionField &in, FermionField &out, int dir, int disp);
void MdirAll(const FermionField &in, std::vector<FermionField> &out);
void DhopDir(const FermionField &in, FermionField &out, int dir, int disp);
///////////////////////////////////////////////////////////////
// Extra methods added by derived
///////////////////////////////////////////////////////////////
void DerivInternal(StencilImpl &st,
DoubledGaugeField &U,
GaugeField &mat,
const FermionField &A, const FermionField &B, int dag);
void DhopInternal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
const FermionField &in, FermionField &out, int dag);
void DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
const FermionField &in, FermionField &out, int dag);
void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
const FermionField &in, FermionField &out, int dag);
//////////////////////////////////////////////////////////////////////////
// Grid own interface Constructor
//////////////////////////////////////////////////////////////////////////
NaiveStaggeredFermion(GaugeField &_U, GridCartesian &Fgrid,
GridRedBlackCartesian &Hgrid, RealD _mass,
RealD _c1, RealD _u0,
const ImplParams &p = ImplParams());
NaiveStaggeredFermion(GridCartesian &Fgrid,
GridRedBlackCartesian &Hgrid, RealD _mass,
RealD _c1, RealD _u0,
const ImplParams &p = ImplParams());
// DoubleStore impl dependent
void ImportGauge (const GaugeField &_U );
DoubledGaugeField &GetU(void) { return Umu ; } ;
void CopyGaugeCheckerboards(void);
///////////////////////////////////////////////////////////////
// Data members require to support the functionality
///////////////////////////////////////////////////////////////
// protected:
public:
// any other parameters of action ???
virtual int isTrivialEE(void) { return 1; };
virtual RealD Mass(void) { return mass; }
RealD mass;
RealD u0;
RealD c1;
GridBase *_grid;
GridBase *_cbgrid;
// Defines the stencils for even and odd
StencilImpl Stencil;
StencilImpl StencilEven;
StencilImpl StencilOdd;
// Copy of the gauge field , with even and odd subsets
DoubledGaugeField Umu;
DoubledGaugeField UmuEven;
DoubledGaugeField UmuOdd;
LebesgueOrder Lebesgue;
LebesgueOrder LebesgueEvenOdd;
///////////////////////////////////////////////////////////////
// Conserved current utilities
///////////////////////////////////////////////////////////////
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &srct,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx);
};
typedef NaiveStaggeredFermion<StaggeredImplF> NaiveStaggeredFermionF;
typedef NaiveStaggeredFermion<StaggeredImplD> NaiveStaggeredFermionD;
NAMESPACE_END(Grid);
#endif

4
Grid/qcd/action/fermion/PartialFractionFermion5D.h
View File

@@ -47,8 +47,8 @@ public:
void M_internal(const FermionField &in, FermionField &out,int dag);
// override multiply
virtual RealD M (const FermionField &in, FermionField &out);
virtual RealD Mdag (const FermionField &in, FermionField &out);
virtual void M (const FermionField &in, FermionField &out);
virtual void Mdag (const FermionField &in, FermionField &out);
// half checkerboard operaions
virtual void Meooe (const FermionField &in, FermionField &out);

47
Grid/qcd/action/fermion/StaggeredKernels.h
View File

@@ -47,23 +47,34 @@ template<class Impl> class StaggeredKernels : public FermionOperator<Impl> , pub
INHERIT_IMPL_TYPES(Impl);
typedef FermionOperator<Impl> Base;
public:
public:
void DhopImproved(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U, DoubledGaugeField &UUU,
const FermionField &in, FermionField &out, int dag, int interior,int exterior);
void DhopNaive(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in, FermionField &out, int dag, int interior,int exterior);
void DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp);
protected:
///////////////////////////////////////////////////////////////////////////////////////
// Generic Nc kernels
///////////////////////////////////////////////////////////////////////////////////////
void DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
template<int Naik> accelerator_inline
void DhopSiteGeneric(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag);
void DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo,
template<int Naik> accelerator_inline
void DhopSiteGenericInt(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag);
void DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo,
template<int Naik> accelerator_inline
void DhopSiteGenericExt(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -71,15 +82,18 @@ public:
///////////////////////////////////////////////////////////////////////////////////////
// Nc=3 specific kernels
///////////////////////////////////////////////////////////////////////////////////////
void DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
template<int Naik> accelerator_inline
void DhopSiteHand(StencilView &st,
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag);
void DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
template<int Naik> accelerator_inline
void DhopSiteHandInt(StencilView &st,
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag);
void DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
template<int Naik> accelerator_inline
void DhopSiteHandExt(StencilView &st,
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag);
@@ -87,27 +101,10 @@ public:
///////////////////////////////////////////////////////////////////////////////////////
// Asm Nc=3 specific kernels
///////////////////////////////////////////////////////////////////////////////////////
void DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
void DhopSiteAsm(StencilView &st,
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag);
///////////////////////////////////////////////////////////////////////////////////////////////////
// Generic interface; fan out to right routine
///////////////////////////////////////////////////////////////////////////////////////////////////
void DhopSite(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out, int interior=1,int exterior=1);
void DhopSiteDag(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out, int interior=1,int exterior=1);
void DhopSite(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor * buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out, int dag, int interior,int exterior);
public:

15
Grid/qcd/action/fermion/StaggeredVec5dImpl.h
View File

@@ -113,20 +113,7 @@ public:
inline void InsertGaugeField(DoubledGaugeField &U_ds,const GaugeLinkField &U,int mu)
{
GridBase *GaugeGrid = U_ds.Grid();
thread_for(lidx, GaugeGrid->lSites(),{
SiteScalarGaugeLink ScalarU;
SiteDoubledGaugeField ScalarUds;
Coordinate lcoor;
GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
peekLocalSite(ScalarUds, U_ds, lcoor);
peekLocalSite(ScalarU, U, lcoor);
ScalarUds(mu) = ScalarU();
});
assert(0);
}
inline void DoubleStore(GridBase *GaugeGrid,
DoubledGaugeField &UUUds, // for Naik term

42
Grid/qcd/action/fermion/WilsonCloverFermion.h
View File

@@ -109,9 +109,8 @@ public:
ImportGauge(_Umu);
}
virtual RealD M(const FermionField &in, FermionField &out);
virtual RealD Mdag(const FermionField &in, FermionField &out);
virtual void M(const FermionField &in, FermionField &out);
virtual void Mdag(const FermionField &in, FermionField &out);
virtual void Mooee(const FermionField &in, FermionField &out);
virtual void MooeeDag(const FermionField &in, FermionField &out);
virtual void MooeeInv(const FermionField &in, FermionField &out);
@@ -258,15 +257,16 @@ private:
CloverFieldType CloverTermDagEven, CloverTermDagOdd; // Clover term Dag EO
CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
public:
// eventually these can be compressed into 6x6 blocks instead of the 12x12
// using the DeGrand-Rossi basis for the gamma matrices
CloverFieldType fillCloverYZ(const GaugeLinkField &F)
{
CloverFieldType T(F.Grid());
T = Zero();
auto T_v = T.View();
auto F_v = F.View();
thread_for(i, CloverTerm.Grid()->oSites(),
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
@@ -282,9 +282,9 @@ private:
CloverFieldType T(F.Grid());
T = Zero();
auto T_v = T.View();
auto F_v = F.View();
thread_for(i, CloverTerm.Grid()->oSites(),
autoView(T_v, T,AcceleratorWrite);
autoView(F_v, F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = -F_v[i]()();
T_v[i]()(1, 0) = F_v[i]()();
@@ -300,9 +300,9 @@ private:
CloverFieldType T(F.Grid());
T = Zero();
auto T_v = T.View();
auto F_v = F.View();
thread_for(i, CloverTerm.Grid()->oSites(),
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
T_v[i]()(1, 1) = timesI(F_v[i]()());
@@ -318,9 +318,9 @@ private:
CloverFieldType T(F.Grid());
T = Zero();
auto T_v = T.View();
auto F_v = F.View();
thread_for(i, CloverTerm.Grid()->oSites(),
autoView( T_v , T, AcceleratorWrite);
autoView( F_v , F, AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = timesI(F_v[i]()());
T_v[i]()(1, 0) = timesI(F_v[i]()());
@@ -336,9 +336,9 @@ private:
CloverFieldType T(F.Grid());
T = Zero();
auto T_v = T.View();
auto F_v = F.View();
thread_for(i, CloverTerm.Grid()->oSites(),
autoView( T_v ,T,AcceleratorWrite);
autoView( F_v ,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = -(F_v[i]()());
T_v[i]()(1, 0) = (F_v[i]()());
@@ -355,9 +355,9 @@ private:
T = Zero();
auto T_v = T.View();
auto F_v = F.View();
thread_for(i, CloverTerm.Grid()->oSites(),
autoView( T_v , T,AcceleratorWrite);
autoView( F_v , F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 0) = timesI(F_v[i]()());
T_v[i]()(1, 1) = timesMinusI(F_v[i]()());

6
Grid/qcd/action/fermion/WilsonFermion.h
View File

@@ -92,8 +92,8 @@ public:
// override multiply; cut number routines if pass dagger argument
// and also make interface more uniformly consistent
//////////////////////////////////////////////////////////////////
virtual RealD M(const FermionField &in, FermionField &out);
virtual RealD Mdag(const FermionField &in, FermionField &out);
virtual void M(const FermionField &in, FermionField &out);
virtual void Mdag(const FermionField &in, FermionField &out);
/////////////////////////////////////////////////////////
// half checkerboard operations
@@ -193,10 +193,12 @@ public:
void ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu);
void SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu,
unsigned int tmin,

23
Grid/qcd/action/fermion/WilsonFermion5D.h
View File

@@ -1,4 +1,3 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@@ -99,8 +98,8 @@ public:
GridBase *FermionRedBlackGrid(void) { return _FiveDimRedBlackGrid;}
// full checkerboard operations; leave unimplemented as abstract for now
virtual RealD M (const FermionField &in, FermionField &out){assert(0); return 0.0;};
virtual RealD Mdag (const FermionField &in, FermionField &out){assert(0); return 0.0;};
virtual void M (const FermionField &in, FermionField &out){assert(0);};
virtual void Mdag (const FermionField &in, FermionField &out){assert(0);};
// half checkerboard operations; leave unimplemented as abstract for now
virtual void Meooe (const FermionField &in, FermionField &out){assert(0);};
@@ -218,24 +217,6 @@ public:
// 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,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx);
void ContractJ5q(PropagatorField &q_in,ComplexField &J5q);
void ContractJ5q(FermionField &q_in,ComplexField &J5q);
};

25
Grid/qcd/action/fermion/WilsonImpl.h
View File

@@ -41,6 +41,7 @@ public:
static const int Dimension = Representation::Dimension;
static const bool isFundamental = Representation::isFundamental;
static const bool LsVectorised=false;
static const bool isGparity=false;
static const int Nhcs = Options::Nhcs;
typedef PeriodicGaugeImpl<GaugeImplTypes<S, Dimension > > Gimpl;
@@ -99,6 +100,19 @@ public:
multLink(phi,U,chi,mu);
}
template<class _SpinorField>
inline void multLinkField(_SpinorField & out,
const DoubledGaugeField &Umu,
const _SpinorField & phi,
int mu)
{
autoView( out_v, out, AcceleratorWrite);
autoView( phi_v, phi, AcceleratorRead);
autoView( Umu_v, Umu, AcceleratorRead);
accelerator_for(sss,out.Grid()->oSites(),1,{
multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
});
}
template <class ref>
static accelerator_inline void loadLinkElement(Simd &reg, ref &memory)
@@ -177,18 +191,19 @@ public:
int Ls=Btilde.Grid()->_fdimensions[0];
GaugeLinkField tmp(mat.Grid());
tmp = Zero();
auto tmp_v = tmp.View();
auto Btilde_v = Btilde.View();
auto Atilde_v = Atilde.View();
thread_for(sss,tmp.Grid()->oSites(),{
{
autoView( tmp_v , tmp, AcceleratorWrite);
autoView( Btilde_v , Btilde, AcceleratorRead);
autoView( Atilde_v , Atilde, AcceleratorRead);
accelerator_for(sss,tmp.Grid()->oSites(),1,{
int sU=sss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
tmp_v[sU] = tmp_v[sU]+ traceIndex<SpinIndex>(outerProduct(Btilde_v[sF],Atilde_v[sF])); // ordering here
}
});
}
PokeIndex<LorentzIndex>(mat,tmp,mu);
}
};

35
Grid/qcd/action/fermion/WilsonKernels.h
View File

@@ -66,41 +66,6 @@ public:
static void DhopDirKernel(StencilImpl &st, DoubledGaugeField &U,SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out, int dirdisp, int gamma);
//////////////////////////////////////////////////////////////////////////////
// Utilities for inserting Wilson conserved current.
//////////////////////////////////////////////////////////////////////////////
static void ContractConservedCurrentSiteFwd(const SitePropagator &q_in_1,
const SitePropagator &q_in_2,
SitePropagator &q_out,
DoubledGaugeFieldView &U,
unsigned int sU,
unsigned int mu,
bool switch_sign = false);
static void ContractConservedCurrentSiteBwd(const SitePropagator &q_in_1,
const SitePropagator &q_in_2,
SitePropagator &q_out,
DoubledGaugeFieldView &U,
unsigned int sU,
unsigned int mu,
bool switch_sign = false);
static void SeqConservedCurrentSiteFwd(const SitePropagator &q_in,
SitePropagator &q_out,
DoubledGaugeFieldView &U,
unsigned int sU,
unsigned int mu,
vPredicate t_mask,
bool switch_sign = false);
static void SeqConservedCurrentSiteBwd(const SitePropagator &q_in,
SitePropagator &q_out,
DoubledGaugeFieldView &U,
unsigned int sU,
unsigned int mu,
vPredicate t_mask,
bool switch_sign = false);
private:
static accelerator_inline void DhopDirK(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor * buf,

8
Grid/qcd/action/fermion/WilsonTMFermion5D.h
View File

@@ -120,7 +120,8 @@ class WilsonTMFermion5D : public WilsonFermion5D<Impl>
}
}
virtual RealD M(const FermionField &in, FermionField &out) {
virtual void M(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
this->Dhop(in, out, DaggerNo);
FermionField tmp(out.Grid());
@@ -129,11 +130,12 @@ class WilsonTMFermion5D : public WilsonFermion5D<Impl>
ComplexD b(0.0,this->mu[s]);
axpbg5y_ssp(tmp,a,in,b,in,s,s);
}
return axpy_norm(out, 1.0, tmp, out);
axpy(out, 1.0, tmp, out);
}
// needed for fast PV
void update(const std::vector<RealD>& _mass, const std::vector<RealD>& _mu) {
void update(const std::vector<RealD>& _mass, const std::vector<RealD>& _mu)
{
assert(_mass.size() == _mu.size());
assert(_mass.size() == this->FermionGrid()->_fdimensions[0]);
this->mass = _mass;

358
Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h
View File

@@ -180,7 +180,7 @@ template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls : " << MooeeInvCalls << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
#ifdef GRID_NVCC
#ifdef GRID_CUDA
RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
@@ -323,7 +323,7 @@ void CayleyFermion5D<Impl>::MeooeDag5D (const FermionField &psi, FermionField
}
template<class Impl>
RealD CayleyFermion5D<Impl>::M (const FermionField &psi, FermionField &chi)
void CayleyFermion5D<Impl>::M (const FermionField &psi, FermionField &chi)
{
FermionField Din(psi.Grid());
@@ -335,11 +335,10 @@ RealD CayleyFermion5D<Impl>::M (const FermionField &psi, FermionField &chi)
axpby(chi,1.0,1.0,chi,psi);
M5D(psi,chi);
return(norm2(chi));
}
template<class Impl>
RealD CayleyFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
void CayleyFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
{
// Under adjoint
//D1+ D1- P- -> D1+^dag P+ D2-^dag
@@ -354,7 +353,6 @@ RealD CayleyFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
M5Ddag(psi,chi);
// ((b D_W + D_w hop terms +1) on s-diag
axpby (chi,1.0,1.0,chi,psi);
return norm2(chi);
}
// half checkerboard operations
@@ -588,6 +586,356 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
// this->MooeeInternalCompute(1,inv,MatpInvDag,MatmInvDag);
}
template <class Impl>
void CayleyFermion5D<Impl>::ContractJ5q(FermionField &q_in,ComplexField &J5q)
{
conformable(this->GaugeGrid(), J5q.Grid());
conformable(q_in.Grid(), this->FermionGrid());
Gamma G5(Gamma::Algebra::Gamma5);
// 4d field
int Ls = this->Ls;
FermionField psi(this->GaugeGrid());
FermionField p_plus (this->GaugeGrid());
FermionField p_minus(this->GaugeGrid());
FermionField p(this->GaugeGrid());
ExtractSlice(p_plus , q_in, Ls/2-1 , 0);
ExtractSlice(p_minus, q_in, Ls/2 , 0);
p_plus = p_plus + G5*p_plus;
p_minus= p_minus - G5*p_minus;
p=0.5*(p_plus+p_minus);
J5q = localInnerProduct(p,p);
}
template <class Impl>
void CayleyFermion5D<Impl>::ContractJ5q(PropagatorField &q_in,ComplexField &J5q)
{
conformable(this->GaugeGrid(), J5q.Grid());
conformable(q_in.Grid(), this->FermionGrid());
Gamma G5(Gamma::Algebra::Gamma5);
// 4d field
int Ls = this->Ls;
PropagatorField psi(this->GaugeGrid());
PropagatorField p_plus (this->GaugeGrid());
PropagatorField p_minus(this->GaugeGrid());
PropagatorField p(this->GaugeGrid());
ExtractSlice(p_plus , q_in, Ls/2-1 , 0);
ExtractSlice(p_minus, q_in, Ls/2 , 0);
p_plus = p_plus + G5*p_plus;
p_minus= p_minus - G5*p_minus;
p=0.5*(p_plus+p_minus);
J5q = localInnerProduct(p,p);
}
#define Pp(Q) (0.5*(Q+g5*Q))
#define Pm(Q) (0.5*(Q-g5*Q))
#define Q_4d(Q) (Pm((Q)[0]) + Pp((Q)[Ls-1]))
#define TopRowWithSource(Q) (phys_src + (1.0-mass)*Q_4d(Q))
template <class Impl>
void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu)
{
#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT,
Gamma::Algebra::Gamma5
};
auto UGrid= this->GaugeGrid();
auto FGrid= this->FermionGrid();
RealD sgn=1.0;
if ( curr_type == Current::Axial ) sgn = -1.0;
int Ls = this->Ls;
std::vector<PropagatorField> L_Q(Ls,UGrid);
std::vector<PropagatorField> R_Q(Ls,UGrid);
for(int s=0;s<Ls;s++){
ExtractSlice(L_Q[s], q_in_1, s , 0);
ExtractSlice(R_Q[s], q_in_2, s , 0);
}
Gamma g5(Gamma::Algebra::Gamma5);
PropagatorField C(UGrid);
PropagatorField p5d(UGrid);
PropagatorField us_p5d(UGrid);
PropagatorField gp5d(UGrid);
PropagatorField gus_p5d(UGrid);
PropagatorField L_TmLsGq0(UGrid);
PropagatorField L_TmLsTmp(UGrid);
PropagatorField R_TmLsGq0(UGrid);
PropagatorField R_TmLsTmp(UGrid);
{
PropagatorField TermA(UGrid);
PropagatorField TermB(UGrid);
PropagatorField TermC(UGrid);
PropagatorField TermD(UGrid);
TermA = (Pp(Q_4d(L_Q)));
TermB = (Pm(Q_4d(L_Q)));
TermC = (Pm(TopRowWithSource(L_Q)));
TermD = (Pp(TopRowWithSource(L_Q)));
L_TmLsGq0 = (TermD - TermA + TermB);
L_TmLsTmp = (TermC - TermB + TermA);
TermA = (Pp(Q_4d(R_Q)));
TermB = (Pm(Q_4d(R_Q)));
TermC = (Pm(TopRowWithSource(R_Q)));
TermD = (Pp(TopRowWithSource(R_Q)));
R_TmLsGq0 = (TermD - TermA + TermB);
R_TmLsTmp = (TermC - TermB + TermA);
}
std::vector<PropagatorField> R_TmLsGq(Ls,UGrid);
std::vector<PropagatorField> L_TmLsGq(Ls,UGrid);
for(int s=0;s<Ls;s++){
R_TmLsGq[s] = (Pm((R_Q)[(s)]) + Pp((R_Q)[((s)-1+Ls)%Ls]));
L_TmLsGq[s] = (Pm((L_Q)[(s)]) + Pp((L_Q)[((s)-1+Ls)%Ls]));
}
Gamma gmu=Gamma(Gmu[mu]);
q_out = Zero();
PropagatorField tmp(UGrid);
for(int s=0;s<Ls;s++){
int sp = (s+1)%Ls;
int sr = Ls-1-s;
int srp= (sr+1)%Ls;
// Mobius parameters
auto b=this->bs[s];
auto c=this->cs[s];
auto bpc = 1.0/(b+c); // -0.5 factor in gauge links
if (s == 0) {
p5d =(b*Pm(L_TmLsGq[Ls-1])+ c*Pp(L_TmLsGq[Ls-1]) + b*Pp(L_TmLsTmp) + c*Pm(L_TmLsTmp ));
tmp =(b*Pm(R_TmLsGq0) + c*Pp(R_TmLsGq0 ) + b*Pp(R_TmLsGq[1]) + c*Pm(R_TmLsGq[1]));
} else if (s == Ls-1) {
p5d =(b*Pm(L_TmLsGq0) + c*Pp(L_TmLsGq0 ) + b*Pp(L_TmLsGq[1]) + c*Pm(L_TmLsGq[1]));
tmp =(b*Pm(R_TmLsGq[Ls-1])+ c*Pp(R_TmLsGq[Ls-1]) + b*Pp(R_TmLsTmp) + c*Pm(R_TmLsTmp ));
} else {
p5d =(b*Pm(L_TmLsGq[sr]) + c*Pp(L_TmLsGq[sr])+ b*Pp(L_TmLsGq[srp])+ c*Pm(L_TmLsGq[srp]));
tmp =(b*Pm(R_TmLsGq[s]) + c*Pp(R_TmLsGq[s]) + b*Pp(R_TmLsGq[sp ])+ c*Pm(R_TmLsGq[sp]));
}
tmp = Cshift(tmp,mu,1);
Impl::multLinkField(us_p5d,this->Umu,tmp,mu);
gp5d=g5*p5d*g5;
gus_p5d=gmu*us_p5d;
C = bpc*(adj(gp5d)*us_p5d);
C-= bpc*(adj(gp5d)*gus_p5d);
if (s == 0) {
p5d =(b*Pm(R_TmLsGq0) + c*Pp(R_TmLsGq0 ) + b*Pp(R_TmLsGq[1]) + c*Pm(R_TmLsGq[1]));
tmp =(b*Pm(L_TmLsGq[Ls-1])+ c*Pp(L_TmLsGq[Ls-1]) + b*Pp(L_TmLsTmp) + c*Pm(L_TmLsTmp ));
} else if (s == Ls-1) {
p5d =(b*Pm(R_TmLsGq[Ls-1])+ c*Pp(R_TmLsGq[Ls-1]) + b*Pp(R_TmLsTmp) + c*Pm(R_TmLsTmp ));
tmp =(b*Pm(L_TmLsGq0) + c*Pp(L_TmLsGq0 ) + b*Pp(L_TmLsGq[1]) + c*Pm(L_TmLsGq[1]));
} else {
p5d =(b*Pm(R_TmLsGq[s]) + c*Pp(R_TmLsGq[s]) + b*Pp(R_TmLsGq[sp ])+ c*Pm(R_TmLsGq[sp]));
tmp =(b*Pm(L_TmLsGq[sr]) + c*Pp(L_TmLsGq[sr]) + b*Pp(L_TmLsGq[srp])+ c*Pm(L_TmLsGq[srp]));
}
tmp = Cshift(tmp,mu,1);
Impl::multLinkField(us_p5d,this->Umu,tmp,mu);
gp5d=gmu*p5d;
gus_p5d=g5*us_p5d*g5;
C-= bpc*(adj(gus_p5d)*gp5d);
C-= bpc*(adj(gus_p5d)*p5d);
if (s < Ls/2) q_out += sgn*C;
else q_out += C;
}
#endif
}
template <class Impl>
void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &phys_src,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &ph)// Complex phase factor
{
assert(mu>=0);
assert(mu<Nd);
#if 0
int tshift = (mu == Nd-1) ? 1 : 0;
////////////////////////////////////////////////
// SHAMIR CASE
////////////////////////////////////////////////
int Ls = this->Ls;
auto UGrid= this->GaugeGrid();
auto FGrid= this->FermionGrid();
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT
};
Gamma gmu=Gamma(Gmu[mu]);
PropagatorField L_Q(UGrid);
PropagatorField R_Q(UGrid);
PropagatorField tmp(UGrid);
PropagatorField Utmp(UGrid);
LatticeInteger zz (UGrid); zz=0.0;
LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
for (int s=0;s<Ls;s++) {
RealD G_s = (curr_type == Current::Axial ) ? ((s < Ls/2) ? -1 : 1) : 1;
ExtractSlice(R_Q, q_in, s , 0);
tmp = Cshift(R_Q,mu,1);
Impl::multLinkField(Utmp,this->Umu,tmp,mu);
tmp = G_s*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
tmp = where((lcoor>=tmin),tmp,zz); // Mask the time
tmp = where((lcoor<=tmax),tmp,zz);
L_Q = tmp;
tmp = R_Q*ph;
tmp = Cshift(tmp,mu,-1);
Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd);// Adjoint link
tmp = -G_s*( Utmp + gmu*Utmp );
tmp = where((lcoor>=tmin+tshift),tmp,zz); // Mask the time
tmp = where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
L_Q= L_Q+tmp;
InsertSlice(L_Q, q_out, s , 0);
}
#endif
#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
int tshift = (mu == Nd-1) ? 1 : 0;
////////////////////////////////////////////////
// GENERAL CAYLEY CASE
////////////////////////////////////////////////
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT,
Gamma::Algebra::Gamma5
};
Gamma gmu=Gamma(Gmu[mu]);
Gamma g5(Gamma::Algebra::Gamma5);
int Ls = this->Ls;
auto UGrid= this->GaugeGrid();
auto FGrid= this->FermionGrid();
std::vector<PropagatorField> R_Q(Ls,UGrid);
PropagatorField L_Q(UGrid);
PropagatorField tmp(UGrid);
PropagatorField Utmp(UGrid);
LatticeInteger zz (UGrid); zz=0.0;
LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
for(int s=0;s<Ls;s++){
ExtractSlice(R_Q[s], q_in, s , 0);
}
PropagatorField R_TmLsGq0(UGrid);
PropagatorField R_TmLsTmp(UGrid);
{
PropagatorField TermA(UGrid);
PropagatorField TermB(UGrid);
PropagatorField TermC(UGrid);
PropagatorField TermD(UGrid);
TermA = (Pp(Q_4d(R_Q)));
TermB = (Pm(Q_4d(R_Q)));
TermC = (Pm(TopRowWithSource(R_Q)));
TermD = (Pp(TopRowWithSource(R_Q)));
R_TmLsGq0 = (TermD - TermA + TermB);
R_TmLsTmp = (TermC - TermB + TermA);
}
std::vector<PropagatorField> R_TmLsGq(Ls,UGrid);
for(int s=0;s<Ls;s++){
R_TmLsGq[s] = (Pm((R_Q)[(s)]) + Pp((R_Q)[((s)-1+Ls)%Ls]));
}
std::vector<RealD> G_s(Ls,1.0);
if ( curr_type == Current::Axial ) {
for(int s=0;s<Ls/2;s++){
G_s[s] = -1.0;
}
}
for(int s=0;s<Ls;s++){
int sp = (s+1)%Ls;
int sr = Ls-1-s;
int srp= (sr+1)%Ls;
// Mobius parameters
auto b=this->bs[s];
auto c=this->cs[s];
// auto bpc = G_s[s]*1.0/(b+c); // -0.5 factor in gauge links
if (s == 0) {
tmp =(b*Pm(R_TmLsGq0) + c*Pp(R_TmLsGq0 ) + b*Pp(R_TmLsGq[1]) + c*Pm(R_TmLsGq[1]));
} else if (s == Ls-1) {
tmp =(b*Pm(R_TmLsGq[Ls-1])+ c*Pp(R_TmLsGq[Ls-1]) + b*Pp(R_TmLsTmp) + c*Pm(R_TmLsTmp ));
} else {
tmp =(b*Pm(R_TmLsGq[s]) + c*Pp(R_TmLsGq[s]) + b*Pp(R_TmLsGq[sp ])+ c*Pm(R_TmLsGq[sp]));
}
tmp = Cshift(tmp,mu,1);
Impl::multLinkField(Utmp,this->Umu,tmp,mu);
tmp = G_s[s]*( Utmp*ph - gmu*Utmp*ph ); // Forward hop
tmp = where((lcoor>=tmin),tmp,zz); // Mask the time
L_Q = where((lcoor<=tmax),tmp,zz); // Position of current complicated
if (s == 0) {
tmp =(b*Pm(R_TmLsGq0) + c*Pp(R_TmLsGq0 ) + b*Pp(R_TmLsGq[1]) + c*Pm(R_TmLsGq[1]));
} else if (s == Ls-1) {
tmp =(b*Pm(R_TmLsGq[Ls-1])+ c*Pp(R_TmLsGq[Ls-1]) + b*Pp(R_TmLsTmp) + c*Pm(R_TmLsTmp ));
} else {
tmp =(b*Pm(R_TmLsGq[s]) + c*Pp(R_TmLsGq[s]) + b*Pp(R_TmLsGq[sp])+ c*Pm(R_TmLsGq[sp]));
}
tmp = tmp *ph;
tmp = Cshift(tmp,mu,-1);
Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
tmp = -G_s[s]*( Utmp + gmu*Utmp );
tmp = where((lcoor>=tmin+tshift),tmp,zz); // Mask the time
L_Q += where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
InsertSlice(L_Q, q_out, s , 0);
}
#endif
}
#undef Pp
#undef Pm
#undef Q_4d
#undef TopRowWithSource
#if 0
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInternalCompute(int dag, int inv,

20
Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h
View File

@@ -50,9 +50,9 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i,AcceleratorRead);
autoView(phi , phi_i,AcceleratorRead);
autoView(chi , chi_i,AcceleratorWrite);
assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
@@ -93,9 +93,9 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i,AcceleratorRead);
autoView(phi , phi_i,AcceleratorRead);
autoView(chi , chi_i,AcceleratorWrite);
assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
@@ -131,8 +131,8 @@ CayleyFermion5D<Impl>::MooeeInv (const FermionField &psi_i, FermionField &chi
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i,AcceleratorRead);
autoView(chi , chi_i,AcceleratorWrite);
int Ls=this->Ls;
@@ -193,8 +193,8 @@ CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi
GridBase *grid=psi_i.Grid();
int Ls=this->Ls;
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i,AcceleratorRead);
autoView(chi , chi_i,AcceleratorWrite);
auto plee = & lee [0];
auto pdee = & dee [0];

24
Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h
View File

@@ -65,9 +65,9 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
autoView(psi, psi_i,CpuRead);
autoView(phi, phi_i,CpuRead);
autoView(chi, chi_i,CpuWrite);
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
const int nsimd= Simd::Nsimd();
@@ -213,9 +213,9 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi=psi_i.View();
auto phi=phi_i.View();
auto chi=chi_i.View();
autoView(psi,psi_i,CpuRead);
autoView(phi,phi_i,CpuRead);
autoView(chi,chi_i,CpuWrite);
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
int nsimd= Simd::Nsimd();
@@ -357,8 +357,8 @@ CayleyFermion5D<Impl>::MooeeInternalAsm(const FermionField &psi_i, FermionField
Vector<iSinglet<Simd> > &Matm)
{
EnableIf<Impl::LsVectorised&&EnableBool,int> sfinae=0;
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i,CpuRead);
autoView(chi , chi_i,CpuWrite);
#ifndef AVX512
{
SiteHalfSpinor BcastP;
@@ -535,8 +535,8 @@ CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField
EnableIf<Impl::LsVectorised,int> sfinae=0;
#ifndef AVX512
{
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i,CpuRead);
autoView(chi , chi_i,CpuWrite);
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
@@ -586,8 +586,8 @@ CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField
}
#else
{
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i,CpuRead);
autoView(chi , chi_i,CpuWrite);
// pointers
// MASK_REGS;
#define Chi_00 %zmm0

7
Grid/qcd/action/fermion/implementation/ContinuedFractionFermion5DImplementation.h
View File

@@ -94,7 +94,7 @@ void ContinuedFractionFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,Ap
template<class Impl>
RealD ContinuedFractionFermion5D<Impl>::M (const FermionField &psi, FermionField &chi)
void ContinuedFractionFermion5D<Impl>::M (const FermionField &psi, FermionField &chi)
{
int Ls = this->Ls;
@@ -116,15 +116,14 @@ RealD ContinuedFractionFermion5D<Impl>::M (const FermionField &psi, F
}
sign=-sign;
}
return norm2(chi);
}
template<class Impl>
RealD ContinuedFractionFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
void ContinuedFractionFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
{
// This matrix is already hermitian. (g5 Dw) = Dw dag g5 = (g5 Dw)dag
// The rest of matrix is symmetric.
// Can ignore "dag"
return M(psi,chi);
M(psi,chi);
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::Mdir (const FermionField &psi, FermionField &chi,int dir,int disp){

20
Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionCache.h
View File

@@ -46,9 +46,9 @@ void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionFi
chi_i.Checkerboard() = psi_i.Checkerboard();
int Ls = this->Ls;
GridBase* grid = psi_i.Grid();
auto phi = phi_i.View();
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView( phi , phi_i, AcceleratorRead);
autoView( psi , psi_i, AcceleratorRead);
autoView( chi , chi_i, AcceleratorWrite);
assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
@@ -82,9 +82,9 @@ void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const Fermio
GridBase* grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
autoView( psi , psi_i, AcceleratorRead);
autoView( phi , phi_i, AcceleratorRead);
autoView( chi , chi_i, AcceleratorWrite);
assert(phi.Checkerboard() == psi.Checkerboard());
auto pdiag = &diag[0];
auto pupper = &upper[0];
@@ -116,8 +116,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionFie
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi=psi_i.View();
auto chi=chi_i.View();
autoView( psi, psi_i, AcceleratorRead);
autoView( chi, chi_i, AcceleratorWrite);
int Ls = this->Ls;
auto plee = & this->lee[0];
@@ -172,8 +172,8 @@ void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, Fermion
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView( psi, psi_i, AcceleratorRead);
autoView( chi, chi_i, AcceleratorWrite);
int Ls = this->Ls;
auto plee = & this->lee[0];

6
Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionImplementation.h
View File

@@ -89,7 +89,7 @@ void DomainWallEOFAFermion<Impl>::DtildeInv(const FermionField& psi, FermionFiel
/*****************************************************************************************************/
template<class Impl>
RealD DomainWallEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
void DomainWallEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
@@ -97,11 +97,10 @@ RealD DomainWallEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
this->DW(Din, chi, DaggerNo);
axpby(chi, 1.0, 1.0, chi, psi);
this->M5D(psi, chi);
return(norm2(chi));
}
template<class Impl>
RealD DomainWallEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
void DomainWallEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
@@ -109,7 +108,6 @@ RealD DomainWallEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& c
this->MeooeDag5D(Din, chi);
this->M5Ddag(psi, chi);
axpby(chi, 1.0, 1.0, chi, psi);
return(norm2(chi));
}
/********************************************************************

165
Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h
View File

@@ -221,10 +221,10 @@ void ImprovedStaggeredFermion5D<Impl>::DhopDir(const FermionField &in, FermionFi
Compressor compressor;
Stencil.HaloExchange(in,compressor);
auto Umu_v = Umu.View();
auto UUUmu_v = UUUmu.View();
auto in_v = in.View();
auto out_v = out.View();
autoView( Umu_v , Umu, CpuRead);
autoView( UUUmu_v , UUUmu, CpuRead);
autoView( in_v , in, CpuRead);
autoView( out_v , out, CpuWrite);
thread_for( ss,Umu.Grid()->oSites(),{
for(int s=0;s<Ls;s++){
int sU=ss;
@@ -281,11 +281,9 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
DoubledGaugeField & U,DoubledGaugeField & UUU,
const FermionField &in, FermionField &out,int dag)
{
#ifdef GRID_OMP
if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
else
#endif
DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
}
@@ -294,9 +292,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
DoubledGaugeField & U,DoubledGaugeField & UUU,
const FermionField &in, FermionField &out,int dag)
{
#ifdef GRID_OMP
// assert((dag==DaggerNo) ||(dag==DaggerYes));
Compressor compressor;
int LLs = in.Grid()->_rdimensions[0];
@@ -305,99 +301,42 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl &
DhopFaceTime-=usecond();
st.Prepare();
st.HaloGather(in,compressor);
DhopFaceTime+=usecond();
DhopCommTime -=usecond();
std::vector<std::vector<CommsRequest_t> > requests;
st.CommunicateBegin(requests);
// st.HaloExchangeOptGather(in,compressor); // Wilson compressor
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
double ctime=0;
double ptime=0;
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
// Remove explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
DhopComputeTime-=usecond();
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
double start = usecond();
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = U.Grid()->oSites(); // 4d vol
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
int interior=1;
int exterior=0;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
// do the compute
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
// Interior = 1; Exterior = 0; must implement for staggered
Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,1,0); //<---------
}
} else {
for (int ss = myblock; ss < myblock+myn; ++ss) {
// Interior = 1; Exterior = 0;
int sU = ss;
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,1,0); //<------------
}
}
ptime = usecond() - start;
} else {
double start = usecond();
st.CommunicateThreaded();
ctime = usecond() - start;
}
}
DhopCommTime += ctime;
DhopComputeTime+=ptime;
// First to enter, last to leave timing
st.CollateThreads();
DhopComputeTime+=usecond();
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
DhopComputeTime2-=usecond();
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
thread_for( ss,sz,{
int sU = st.surface_list[ss];
Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,0,1); //<----------
});
} else {
int sz=st.surface_list.size();
thread_for( ss,sz,{
int sU = st.surface_list[ss];
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,0,1);//<----------
});
DhopComputeTime2-=usecond();
{
int interior=0;
int exterior=1;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime2+=usecond();
#else
assert(0);
#endif
}
template<class Impl>
@@ -408,8 +347,6 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
Compressor compressor;
int LLs = in.Grid()->_rdimensions[0];
//double t1=usecond();
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
@@ -418,28 +355,13 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
DhopComputeTime -= usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
thread_for( ss,U.Grid()->oSites(),{
int sU=ss;
Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), LLs, sU,in_v, out_v);
});
} else {
thread_for( ss,U.Grid()->oSites(),{
int sU=ss;
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v);
});
{
int interior=1;
int exterior=1;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
//double t2=usecond();
//std::cout << __FILE__ << " " << __func__ << " Total Time " << DhopTotalTime << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Total Time Org " << t2-t1 << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Comml Time " << DhopCommTime << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Compute Time " << DhopComputeTime << std::endl;
}
/*CHANGE END*/
@@ -548,21 +470,24 @@ void ImprovedStaggeredFermion5D<Impl>::MdirAll(const FermionField &in, std::vect
assert(0);
}
template <class Impl>
RealD ImprovedStaggeredFermion5D<Impl>::M(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion5D<Impl>::M(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerNo);
return axpy_norm(out, mass, in, out);
axpy(out, mass, in, out);
}
template <class Impl>
RealD ImprovedStaggeredFermion5D<Impl>::Mdag(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion5D<Impl>::Mdag(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerYes);
return axpy_norm(out, mass, in, out);
axpy(out, mass, in, out);
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::Meooe(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion5D<Impl>::Meooe(const FermionField &in, FermionField &out)
{
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerNo);
} else {
@@ -570,7 +495,8 @@ void ImprovedStaggeredFermion5D<Impl>::Meooe(const FermionField &in, FermionFiel
}
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion5D<Impl>::MeooeDag(const FermionField &in, FermionField &out)
{
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerYes);
} else {
@@ -579,27 +505,30 @@ void ImprovedStaggeredFermion5D<Impl>::MeooeDag(const FermionField &in, FermionF
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::Mooee(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion5D<Impl>::Mooee(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
typename FermionField::scalar_type scal(mass);
out = scal * in;
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion5D<Impl>::MooeeDag(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Mooee(in, out);
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion5D<Impl>::MooeeInv(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
out = (1.0 / (mass)) * in;
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MooeeInvDag(const FermionField &in,
FermionField &out) {
void ImprovedStaggeredFermion5D<Impl>::MooeeInvDag(const FermionField &in,FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
MooeeInv(in, out);
}
@@ -611,6 +540,7 @@ template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu)
{
@@ -620,6 +550,7 @@ void ImprovedStaggeredFermion5D<Impl>::ContractConservedCurrent(PropagatorField
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu,
unsigned int tmin,

164
Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermionImplementation.h
View File

@@ -171,21 +171,24 @@ void ImprovedStaggeredFermion<Impl>::ImportGauge(const GaugeField &_Uthin,const
/////////////////////////////
template <class Impl>
RealD ImprovedStaggeredFermion<Impl>::M(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion<Impl>::M(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerNo);
return axpy_norm(out, mass, in, out);
axpy(out, mass, in, out);
}
template <class Impl>
RealD ImprovedStaggeredFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerYes);
return axpy_norm(out, mass, in, out);
axpy(out, mass, in, out);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion<Impl>::Meooe(const FermionField &in, FermionField &out)
{
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerNo);
} else {
@@ -193,7 +196,8 @@ void ImprovedStaggeredFermion<Impl>::Meooe(const FermionField &in, FermionField
}
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out)
{
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerYes);
} else {
@@ -202,27 +206,30 @@ void ImprovedStaggeredFermion<Impl>::MeooeDag(const FermionField &in, FermionFie
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
typename FermionField::scalar_type scal(mass);
out = scal * in;
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Mooee(in, out);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
void ImprovedStaggeredFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
out = (1.0 / (mass)) * in;
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MooeeInvDag(const FermionField &in,
FermionField &out) {
void ImprovedStaggeredFermion<Impl>::MooeeInvDag(const FermionField &in,FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
MooeeInv(in, out);
}
@@ -234,7 +241,8 @@ void ImprovedStaggeredFermion<Impl>::MooeeInvDag(const FermionField &in,
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U, DoubledGaugeField &UUU,
GaugeField & mat,
const FermionField &A, const FermionField &B, int dag) {
const FermionField &A, const FermionField &B, int dag)
{
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor;
@@ -250,10 +258,10 @@ void ImprovedStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGauge
////////////////////////
// Call the single hop
////////////////////////
auto U_v = U.View();
auto UUU_v = UUU.View();
auto B_v = B.View();
auto Btilde_v = Btilde.View();
autoView( U_v , U, CpuRead);
autoView( UUU_v , UUU, CpuRead);
autoView( B_v , B, CpuWrite);
autoView( Btilde_v , Btilde, CpuWrite);
thread_for(sss,B.Grid()->oSites(),{
Kernels::DhopDirKernel(st, U_v, UUU_v, st.CommBuf(), sss, sss, B_v, Btilde_v, mu,1);
});
@@ -284,8 +292,8 @@ void ImprovedStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGauge
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
void ImprovedStaggeredFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
{
conformable(U.Grid(), _grid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
@@ -296,8 +304,8 @@ void ImprovedStaggeredFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionFie
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
void ImprovedStaggeredFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
{
conformable(U.Grid(), _cbgrid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
@@ -310,8 +318,8 @@ void ImprovedStaggeredFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionF
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
void ImprovedStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
{
conformable(U.Grid(), _cbgrid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
@@ -378,10 +386,10 @@ void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionFiel
Compressor compressor;
Stencil.HaloExchange(in, compressor);
auto Umu_v = Umu.View();
auto UUUmu_v = UUUmu.View();
auto in_v = in.View();
auto out_v = out.View();
autoView( Umu_v , Umu, CpuRead);
autoView( UUUmu_v , UUUmu, CpuRead);
autoView( in_v , in, CpuRead);
autoView( out_v , out, CpuWrite);
thread_for( sss, in.Grid()->oSites(),{
Kernels::DhopDirKernel(Stencil, Umu_v, UUUmu_v, Stencil.CommBuf(), sss, sss, in_v, out_v, dir, disp);
});
@@ -395,11 +403,9 @@ void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder
const FermionField &in,
FermionField &out, int dag)
{
#ifdef GRID_OMP
if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
else
#endif
DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
}
template <class Impl>
@@ -409,7 +415,6 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
const FermionField &in,
FermionField &out, int dag)
{
#ifdef GRID_OMP
Compressor compressor;
int len = U.Grid()->oSites();
@@ -418,60 +423,30 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
DhopFaceTime -= usecond();
st.Prepare();
st.HaloGather(in,compressor);
st.CommsMergeSHM(compressor);
DhopFaceTime += usecond();
DhopCommTime -=usecond();
std::vector<std::vector<CommsRequest_t> > requests;
st.CommunicateBegin(requests);
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);
DhopFaceTime+= usecond();
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
// Removed explicit thread comms
//////////////////////////////////////////////////////////////////////////////////////////////////////
DhopComputeTime -= usecond();
#pragma omp parallel
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = len;
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
// do the compute
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
// Interior = 1; Exterior = 0; must implement for staggered
Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,1,0);
}
} else {
for (int ss = myblock; ss < myblock+myn; ++ss) {
// Interior = 1; Exterior = 0;
int sU = ss;
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,1,0);
}
}
} else {
st.CommunicateThreaded();
}
int interior=1;
int exterior=0;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
// First to enter, last to leave timing
DhopFaceTime -= usecond();
st.CommsMerge(compressor);
@@ -479,28 +454,11 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
DhopComputeTime2 -= usecond();
{
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
thread_for(ss,sz,{
int sU = st.surface_list[ss];
Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,0,1);
});
} else {
int sz=st.surface_list.size();
thread_for(ss,sz,{
int sU = st.surface_list[ss];
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,0,1);
});
}
int interior=0;
int exterior=1;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime2 += usecond();
#else
assert(0);
#endif
}
@@ -520,19 +478,11 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Le
st.HaloExchange(in, compressor);
DhopCommTime += usecond();
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
DhopComputeTime -= usecond();
if (dag == DaggerYes) {
thread_for(sss, in.Grid()->oSites(),{
Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
});
} else {
thread_for(sss, in.Grid()->oSites(),{
Kernels::DhopSite(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
});
{
int interior=1;
int exterior=1;
Kernels::DhopImproved(st,lo,U,UUU,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
@@ -600,6 +550,7 @@ template <class Impl>
void ImprovedStaggeredFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu)
{
@@ -609,6 +560,7 @@ void ImprovedStaggeredFermion<Impl>::ContractConservedCurrent(PropagatorField &q
template <class Impl>
void ImprovedStaggeredFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu,
unsigned int tmin,

40
Grid/qcd/action/fermion/implementation/MobiusEOFAFermionCache.h
View File

@@ -44,9 +44,9 @@ void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i, AcceleratorRead);
autoView(phi , phi_i, AcceleratorRead);
autoView(chi , chi_i, AcceleratorWrite);
assert(phi.Checkerboard() == psi.Checkerboard());
@@ -84,9 +84,9 @@ void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const Fermion
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i, AcceleratorRead);
autoView(phi , phi_i, AcceleratorRead);
autoView(chi , chi_i, AcceleratorWrite);
auto pm = this->pm;
int shift_s = (pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
@@ -132,9 +132,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionFie
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i, AcceleratorRead);
autoView(phi , phi_i, AcceleratorRead);
autoView(chi , chi_i, AcceleratorWrite);
assert(phi.Checkerboard() == psi.Checkerboard());
@@ -174,9 +174,9 @@ void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const Ferm
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
int shift_s = (this->pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i, AcceleratorRead);
autoView(phi , phi_i, AcceleratorRead);
autoView(chi , chi_i, AcceleratorWrite);
assert(phi.Checkerboard() == psi.Checkerboard());
@@ -226,8 +226,8 @@ void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i, AcceleratorRead);
autoView(chi , chi_i, AcceleratorWrite);
auto plee = & this->lee [0];
auto pdee = & this->dee [0];
@@ -286,8 +286,8 @@ void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionF
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i, AcceleratorRead);
autoView(chi , chi_i, AcceleratorWrite);
auto pm = this->pm;
auto plee = & this->lee [0];
@@ -354,8 +354,8 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionFiel
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i, AcceleratorRead);
autoView(chi , chi_i, AcceleratorWrite);
auto plee = & this->lee [0];
auto pdee = & this->dee [0];
@@ -410,8 +410,8 @@ void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, Fermi
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
autoView(psi , psi_i, AcceleratorRead);
autoView(chi , chi_i, AcceleratorWrite);
int Ls = this->Ls;
auto pm = this->pm;

6
Grid/qcd/action/fermion/implementation/MobiusEOFAFermionImplementation.h
View File

@@ -166,7 +166,7 @@ void MobiusEOFAFermion<Impl>::DtildeInv(const FermionField& psi, FermionField& c
/*****************************************************************************************************/
template<class Impl>
RealD MobiusEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
void MobiusEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
@@ -174,11 +174,10 @@ RealD MobiusEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
this->DW(Din, chi, DaggerNo);
axpby(chi, 1.0, 1.0, chi, psi);
this->M5D(psi, chi);
return(norm2(chi));
}
template<class Impl>
RealD MobiusEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
void MobiusEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
@@ -186,7 +185,6 @@ RealD MobiusEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
this->MeooeDag5D(Din, chi);
this->M5Ddag(psi, chi);
axpby(chi, 1.0, 1.0, chi, psi);
return(norm2(chi));
}
/********************************************************************

499
Grid/qcd/action/fermion/implementation/NaiveStaggeredFermionImplementation.h Normal file
View File

@@ -0,0 +1,499 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
Copyright (C) 2015
Author: Azusa Yamaguchi, Peter Boyle
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>
#pragma once
NAMESPACE_BEGIN(Grid);
/////////////////////////////////
// Constructor and gauge import
/////////////////////////////////
template <class Impl>
NaiveStaggeredFermion<Impl>::NaiveStaggeredFermion(GridCartesian &Fgrid, GridRedBlackCartesian &Hgrid,
RealD _mass,
RealD _c1, RealD _u0,
const ImplParams &p)
: Kernels(p),
_grid(&Fgrid),
_cbgrid(&Hgrid),
Stencil(&Fgrid, npoint, Even, directions, displacements,p),
StencilEven(&Hgrid, npoint, Even, directions, displacements,p), // source is Even
StencilOdd(&Hgrid, npoint, Odd, directions, displacements,p), // source is Odd
mass(_mass),
Lebesgue(_grid),
LebesgueEvenOdd(_cbgrid),
Umu(&Fgrid),
UmuEven(&Hgrid),
UmuOdd(&Hgrid),
_tmp(&Hgrid)
{
int vol4;
int LLs=1;
c1=_c1;
u0=_u0;
vol4= _grid->oSites();
Stencil.BuildSurfaceList(LLs,vol4);
vol4= _cbgrid->oSites();
StencilEven.BuildSurfaceList(LLs,vol4);
StencilOdd.BuildSurfaceList(LLs,vol4);
}
template <class Impl>
NaiveStaggeredFermion<Impl>::NaiveStaggeredFermion(GaugeField &_U, GridCartesian &Fgrid,
GridRedBlackCartesian &Hgrid, RealD _mass,
RealD _c1, RealD _u0,
const ImplParams &p)
: NaiveStaggeredFermion(Fgrid,Hgrid,_mass,_c1,_u0,p)
{
ImportGauge(_U);
}
////////////////////////////////////////////////////////////
// Momentum space propagator should be
// https://arxiv.org/pdf/hep-lat/9712010.pdf
//
// mom space action.
// gamma_mu i ( c1 sin pmu + c2 sin 3 pmu ) + m
//
// must track through staggered flavour/spin reduction in literature to
// turn to free propagator for the one component chi field, a la page 4/5
// of above link to implmement fourier based solver.
////////////////////////////////////////////////////////////
template <class Impl>
void NaiveStaggeredFermion<Impl>::CopyGaugeCheckerboards(void)
{
pickCheckerboard(Even, UmuEven, Umu);
pickCheckerboard(Odd, UmuOdd , Umu);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::ImportGauge(const GaugeField &_U)
{
GaugeLinkField U(GaugeGrid());
DoubledGaugeField _UUU(GaugeGrid());
////////////////////////////////////////////////////////
// Double Store should take two fields for Naik and one hop separately.
// Discard teh Naik as Naive
////////////////////////////////////////////////////////
Impl::DoubleStore(GaugeGrid(), _UUU, Umu, _U, _U );
////////////////////////////////////////////////////////
// Apply scale factors to get the right fermion Kinetic term
// Could pass coeffs into the double store to save work.
// 0.5 ( U p(x+mu) - Udag(x-mu) p(x-mu) )
////////////////////////////////////////////////////////
for (int mu = 0; mu < Nd; mu++) {
U = PeekIndex<LorentzIndex>(Umu, mu);
PokeIndex<LorentzIndex>(Umu, U*( 0.5*c1/u0), mu );
U = PeekIndex<LorentzIndex>(Umu, mu+4);
PokeIndex<LorentzIndex>(Umu, U*(-0.5*c1/u0), mu+4);
}
CopyGaugeCheckerboards();
}
/////////////////////////////
// Implement the interface
/////////////////////////////
template <class Impl>
void NaiveStaggeredFermion<Impl>::M(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerNo);
axpy(out, mass, in, out);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerYes);
axpy(out, mass, in, out);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerNo);
} else {
DhopOE(in, out, DaggerNo);
}
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerYes);
} else {
DhopOE(in, out, DaggerYes);
}
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
typename FermionField::scalar_type scal(mass);
out = scal * in;
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Mooee(in, out);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
out = (1.0 / (mass)) * in;
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
MooeeInv(in, out);
}
///////////////////////////////////
// Internal
///////////////////////////////////
template <class Impl>
void NaiveStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
GaugeField & mat,
const FermionField &A, const FermionField &B, int dag)
{
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor;
FermionField Btilde(B.Grid());
FermionField Atilde(B.Grid());
Atilde = A;
st.HaloExchange(B, compressor);
for (int mu = 0; mu < Nd; mu++) {
////////////////////////
// Call the single hop
////////////////////////
autoView( U_v , U, CpuRead);
autoView( B_v , B, CpuWrite);
autoView( Btilde_v , Btilde, CpuWrite);
thread_for(sss,B.Grid()->oSites(),{
Kernels::DhopDirKernel(st, U_v, U_v, st.CommBuf(), sss, sss, B_v, Btilde_v, mu,1);
});
assert(0);// need to figure out the force interface with a blasted three link term.
}
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _grid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
mat.Checkerboard() = U.Checkerboard();
DerivInternal(Stencil, Umu, mat, U, V, dag);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _cbgrid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
assert(V.Checkerboard() == Even);
assert(U.Checkerboard() == Odd);
mat.Checkerboard() = Odd;
DerivInternal(StencilEven, UmuOdd, mat, U, V, dag);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _cbgrid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
assert(V.Checkerboard() == Odd);
assert(U.Checkerboard() == Even);
mat.Checkerboard() = Even;
DerivInternal(StencilOdd, UmuEven, mat, U, V, dag);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=2;
conformable(in.Grid(), _grid); // verifies full grid
conformable(in.Grid(), out.Grid());
out.Checkerboard() = in.Checkerboard();
DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=1;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
assert(in.Checkerboard() == Even);
out.Checkerboard() = Odd;
DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, in, out, dag);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=1;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
assert(in.Checkerboard() == Odd);
out.Checkerboard() = Even;
DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, in, out, dag);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp)
{
DhopDir(in, out, dir, disp);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::MdirAll(const FermionField &in, std::vector<FermionField> &out)
{
assert(0); // Not implemented yet
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp)
{
Compressor compressor;
Stencil.HaloExchange(in, compressor);
autoView( Umu_v , Umu, CpuRead);
autoView( in_v , in, CpuRead);
autoView( out_v , out, CpuWrite);
// thread_for( sss, in.Grid()->oSites(),{
// Kernels::DhopDirKernel(Stencil, Umu_v, Stencil.CommBuf(), sss, sss, in_v, out_v, dir, disp);
// });
assert(0);
};
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag)
{
if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,in,out,dag);
else
DhopInternalSerialComms(st,lo,U,in,out,dag);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag)
{
Compressor compressor;
int len = U.Grid()->oSites();
DhopTotalTime -= usecond();
DhopFaceTime -= usecond();
st.Prepare();
st.HaloGather(in,compressor);
DhopFaceTime += usecond();
DhopCommTime -=usecond();
std::vector<std::vector<CommsRequest_t> > requests;
st.CommunicateBegin(requests);
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);
DhopFaceTime+= usecond();
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Removed explicit thread comms
//////////////////////////////////////////////////////////////////////////////////////////////////////
DhopComputeTime -= usecond();
{
int interior=1;
int exterior=0;
Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
// First to enter, last to leave timing
DhopFaceTime -= usecond();
st.CommsMerge(compressor);
DhopFaceTime -= usecond();
DhopComputeTime2 -= usecond();
{
int interior=0;
int exterior=1;
Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
}
DhopComputeTime2 += usecond();
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag)
{
assert((dag == DaggerNo) || (dag == DaggerYes));
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
Compressor compressor;
st.HaloExchange(in, compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
{
int interior=1;
int exterior=1;
Kernels::DhopNaive(st,lo,U,in,out,dag,interior,exterior);
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
};
////////////////////////////////////////////////////////////////
// Reporting
////////////////////////////////////////////////////////////////
template<class Impl>
void NaiveStaggeredFermion<Impl>::Report(void)
{
Coordinate latt = _grid->GlobalDimensions();
RealD volume = 1; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_grid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "NaiveStaggeredFermion Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "NaiveStaggeredFermion StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "NaiveStaggeredFermion StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void NaiveStaggeredFermion<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DhopFaceTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
////////////////////////////////////////////////////////
// Conserved current - not yet implemented.
////////////////////////////////////////////////////////
template <class Impl>
void NaiveStaggeredFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu)
{
assert(0);
}
template <class Impl>
void NaiveStaggeredFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
PropagatorField &src,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx)
{
assert(0);
}
NAMESPACE_END(Grid);

6
Grid/qcd/action/fermion/implementation/PartialFractionFermion5DImplementation.h
View File

@@ -269,16 +269,14 @@ void PartialFractionFermion5D<Impl>::M_internal(const FermionField &psi, Fermi
}
template<class Impl>
RealD PartialFractionFermion5D<Impl>::M (const FermionField &in, FermionField &out)
void PartialFractionFermion5D<Impl>::M (const FermionField &in, FermionField &out)
{
M_internal(in,out,DaggerNo);
return norm2(out);
}
template<class Impl>
RealD PartialFractionFermion5D<Impl>::Mdag (const FermionField &in, FermionField &out)
void PartialFractionFermion5D<Impl>::Mdag (const FermionField &in, FermionField &out)
{
M_internal(in,out,DaggerYes);
return norm2(out);
}
template<class Impl>

43
Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h
View File

@@ -618,10 +618,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
NAMESPACE_BEGIN(Grid);
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
void StaggeredKernels<Impl>::DhopSiteAsm(StencilView &st,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
SiteSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out,int dag)
{
assert(0);
@@ -680,12 +680,13 @@ void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
gauge2 =(uint64_t)&UU[sU]( Z ); \
gauge3 =(uint64_t)&UU[sU]( T );
// This is the single precision 5th direction vectorised kernel
#include <Grid/simd/Intel512single.h>
template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilView &st,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
SiteSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out,int dag)
{
#ifdef AVX512
@@ -702,9 +703,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
// for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// int sF=s+LLs*sU;
{
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
@@ -736,10 +738,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl
}
#include <Grid/simd/Intel512double.h>
template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilView &st,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
SiteSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out, int dag)
{
#ifdef AVX512
@@ -756,8 +758,9 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// for(int s=0;s<LLs;s++){
// int sF=s+LLs*sU;
{
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
@@ -821,10 +824,10 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
// This is the single precision 5th direction vectorised kernel
#include <Grid/simd/Intel512single.h>
template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilView &st,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
SiteSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out,int dag)
{
#ifdef AVX512
@@ -841,9 +844,9 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// for(int s=0;s<LLs;s++){
// int sF=s+LLs*sU;
{
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHIa(addr0,addr1);
@@ -890,10 +893,10 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
}
#include <Grid/simd/Intel512double.h>
template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilView &st,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
SiteSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out,int dag)
{
#ifdef AVX512
@@ -910,9 +913,9 @@ template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st,
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// for(int s=0;s<LLs;s++){
// int sF=s+LLs*sU;
{
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHIa(addr0,addr1);

42
Grid/qcd/action/fermion/implementation/StaggeredKernelsHand.h
View File

@@ -146,9 +146,10 @@ NAMESPACE_BEGIN(Grid);
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
template <int Naik>
void StaggeredKernels<Impl>::DhopSiteHand(StencilView &st,
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag)
{
typedef typename Simd::scalar_type S;
@@ -181,8 +182,9 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
StencilEntry *SE;
int skew;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// for(int s=0;s<LLs;s++){
// int sF=s+LLs*sU;
{
skew = 0;
HAND_STENCIL_LEG_BEGIN(Xp,3,skew,even);
@@ -193,6 +195,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
HAND_STENCIL_LEG (U,Ym,2,skew,odd);
HAND_STENCIL_LEG (U,Zm,1,skew,even);
HAND_STENCIL_LEG (U,Tm,0,skew,odd);
if (Naik) {
skew = 8;
HAND_STENCIL_LEG(UUU,Xp,3,skew,even);
HAND_STENCIL_LEG(UUU,Yp,2,skew,odd);
@@ -202,7 +205,7 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
HAND_STENCIL_LEG(UUU,Ym,2,skew,odd);
HAND_STENCIL_LEG(UUU,Zm,1,skew,even);
HAND_STENCIL_LEG(UUU,Tm,0,skew,odd);
}
if ( dag ) {
result()()(0) = - even_0 - odd_0;
result()()(1) = - even_1 - odd_1;
@@ -218,9 +221,10 @@ void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
template <int Naik>
void StaggeredKernels<Impl>::DhopSiteHandInt(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag)
{
typedef typename Simd::scalar_type S;
@@ -253,8 +257,9 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
StencilEntry *SE;
int skew;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// for(int s=0;s<LLs;s++){
// int sF=s+LLs*sU;
{
even_0 = Zero(); even_1 = Zero(); even_2 = Zero();
odd_0 = Zero(); odd_1 = Zero(); odd_2 = Zero();
@@ -268,6 +273,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
HAND_STENCIL_LEG_INT(U,Ym,2,skew,odd);
HAND_STENCIL_LEG_INT(U,Zm,1,skew,even);
HAND_STENCIL_LEG_INT(U,Tm,0,skew,odd);
if (Naik) {
skew = 8;
HAND_STENCIL_LEG_INT(UUU,Xp,3,skew,even);
HAND_STENCIL_LEG_INT(UUU,Yp,2,skew,odd);
@@ -277,7 +283,7 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
HAND_STENCIL_LEG_INT(UUU,Ym,2,skew,odd);
HAND_STENCIL_LEG_INT(UUU,Zm,1,skew,even);
HAND_STENCIL_LEG_INT(UUU,Tm,0,skew,odd);
}
// Assume every site must be connected to at least one interior point. No 1^4 subvols.
if ( dag ) {
result()()(0) = - even_0 - odd_0;
@@ -294,9 +300,10 @@ void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
template <int Naik>
void StaggeredKernels<Impl>::DhopSiteHandExt(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag)
{
typedef typename Simd::scalar_type S;
@@ -329,8 +336,9 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
StencilEntry *SE;
int skew;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// for(int s=0;s<LLs;s++){
// int sF=s+LLs*sU;
{
even_0 = Zero(); even_1 = Zero(); even_2 = Zero();
odd_0 = Zero(); odd_1 = Zero(); odd_2 = Zero();
@@ -344,6 +352,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
HAND_STENCIL_LEG_EXT(U,Ym,2,skew,odd);
HAND_STENCIL_LEG_EXT(U,Zm,1,skew,even);
HAND_STENCIL_LEG_EXT(U,Tm,0,skew,odd);
if (Naik) {
skew = 8;
HAND_STENCIL_LEG_EXT(UUU,Xp,3,skew,even);
HAND_STENCIL_LEG_EXT(UUU,Yp,2,skew,odd);
@@ -353,7 +362,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
HAND_STENCIL_LEG_EXT(UUU,Ym,2,skew,odd);
HAND_STENCIL_LEG_EXT(UUU,Zm,1,skew,even);
HAND_STENCIL_LEG_EXT(UUU,Tm,0,skew,odd);
}
// Add sum of all exterior connected stencil legs
if ( nmu ) {
if ( dag ) {
@@ -370,6 +379,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
}
}
/*
#define DHOP_SITE_HAND_INSTANTIATE(IMPL) \
template void StaggeredKernels<IMPL>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, \
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
@@ -385,7 +395,7 @@ void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
SiteSpinor *buf, int LLs, int sU, \
const FermionFieldView &in, FermionFieldView &out, int dag); \
*/
#undef LOAD_CHI
NAMESPACE_END(Grid);

210
Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h
View File

@@ -78,10 +78,12 @@ StaggeredKernels<Impl>::StaggeredKernels(const ImplParams &p) : Base(p){};
// Int, Ext, Int+Ext cases for comms overlap
////////////////////////////////////////////////////////////////////////////////////
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
template <int Naik>
void StaggeredKernels<Impl>::DhopSiteGeneric(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out, int dag) {
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out, int dag)
{
const SiteSpinor *chi_p;
SiteSpinor chi;
SiteSpinor Uchi;
@@ -89,8 +91,10 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
int ptype;
int skew;
for(int s=0;s<LLs;s++){
int sF=LLs*sU+s;
// for(int s=0;s<LLs;s++){
//
// int sF=LLs*sU+s;
{
skew = 0;
GENERIC_STENCIL_LEG(U,Xp,skew,Impl::multLink);
GENERIC_STENCIL_LEG(U,Yp,skew,Impl::multLinkAdd);
@@ -100,6 +104,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
GENERIC_STENCIL_LEG(U,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(U,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(U,Tm,skew,Impl::multLinkAdd);
if ( Naik ) {
skew=8;
GENERIC_STENCIL_LEG(UUU,Xp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Yp,skew,Impl::multLinkAdd);
@@ -109,6 +114,7 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
GENERIC_STENCIL_LEG(UUU,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Tm,skew,Impl::multLinkAdd);
}
if ( dag ) {
Uchi = - Uchi;
}
@@ -120,9 +126,10 @@ void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
// Only contributions from interior of our node
///////////////////////////////////////////////////
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo,
template <int Naik>
void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag) {
const SiteSpinor *chi_p;
SiteSpinor chi;
@@ -131,8 +138,9 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
int ptype;
int skew ;
for(int s=0;s<LLs;s++){
int sF=LLs*sU+s;
// for(int s=0;s<LLs;s++){
// int sF=LLs*sU+s;
{
skew = 0;
Uchi=Zero();
GENERIC_STENCIL_LEG_INT(U,Xp,skew,Impl::multLinkAdd);
@@ -143,6 +151,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
GENERIC_STENCIL_LEG_INT(U,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Tm,skew,Impl::multLinkAdd);
if ( Naik ) {
skew=8;
GENERIC_STENCIL_LEG_INT(UUU,Xp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Yp,skew,Impl::multLinkAdd);
@@ -152,6 +161,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
GENERIC_STENCIL_LEG_INT(UUU,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Tm,skew,Impl::multLinkAdd);
}
if ( dag ) {
Uchi = - Uchi;
}
@@ -164,9 +174,10 @@ void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &
// Only contributions from exterior of our node
///////////////////////////////////////////////////
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo,
template <int Naik>
void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilView &st,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
SiteSpinor *buf, int sF, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag) {
const SiteSpinor *chi_p;
// SiteSpinor chi;
@@ -176,8 +187,9 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
int nmu=0;
int skew ;
for(int s=0;s<LLs;s++){
int sF=LLs*sU+s;
// for(int s=0;s<LLs;s++){
// int sF=LLs*sU+s;
{
skew = 0;
Uchi=Zero();
GENERIC_STENCIL_LEG_EXT(U,Xp,skew,Impl::multLinkAdd);
@@ -188,6 +200,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
GENERIC_STENCIL_LEG_EXT(U,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Tm,skew,Impl::multLinkAdd);
if ( Naik ) {
skew=8;
GENERIC_STENCIL_LEG_EXT(UUU,Xp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Yp,skew,Impl::multLinkAdd);
@@ -197,7 +210,7 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
GENERIC_STENCIL_LEG_EXT(UUU,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
}
if ( nmu ) {
if ( dag ) {
out[sF] = out[sF] - Uchi;
@@ -211,72 +224,9 @@ void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &
////////////////////////////////////////////////////////////////////////////////////
// Driving / wrapping routine to select right kernel
////////////////////////////////////////////////////////////////////////////////////
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,
int interior,int exterior)
{
int dag=1;
DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
};
template <class Impl>
void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,
int interior,int exterior)
{
int dag=0;
DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
};
template <class Impl>
void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionFieldView &in, FermionFieldView &out,
int dag,int interior,int exterior)
{
switch(Opt) {
#ifdef AVX512
case OptInlineAsm:
if ( interior && exterior ) {
DhopSiteAsm(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else {
std::cout << GridLogError << "Cannot overlap comms and compute with Staggered assembly"<<std::endl;
assert(0);
}
break;
#endif
case OptHandUnroll:
if ( interior && exterior ) {
DhopSiteHand (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else if ( interior ) {
DhopSiteHandInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else if ( exterior ) {
DhopSiteHandExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
}
break;
case OptGeneric:
if ( interior && exterior ) {
DhopSiteGeneric (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else if ( interior ) {
DhopSiteGenericInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else if ( exterior ) {
DhopSiteGenericExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
}
break;
default:
std::cout<<"Oops Opt = "<<Opt<<std::endl;
assert(0);
break;
}
};
template <class Impl>
void StaggeredKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out, int dir, int disp)
void StaggeredKernels<Impl>::DhopDirKernel(StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor * buf,
int sF, int sU, const FermionFieldView &in, FermionFieldView &out, int dir,int disp)
{
// Disp should be either +1,-1,+3,-3
// What about "dag" ?
@@ -285,6 +235,108 @@ void StaggeredKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeFieldVi
assert(0);
}
#define KERNEL_CALLNB(A,improved) \
const uint64_t NN = Nsite*Ls; \
accelerator_forNB( ss, NN, Simd::Nsimd(), { \
int sF = ss; \
int sU = ss/Ls; \
ThisKernel:: template A<improved>(st_v,U_v,UUU_v,buf,sF,sU,in_v,out_v,dag); \
});
#define KERNEL_CALL(A,improved) KERNEL_CALLNB(A,improved); accelerator_barrier();
#define ASM_CALL(A) \
const uint64_t NN = Nsite*Ls; \
thread_for( ss, NN, { \
int sF = ss; \
int sU = ss/Ls; \
ThisKernel::A(st_v,U_v,UUU_v,buf,sF,sU,in_v,out_v,dag); \
});
template <class Impl>
void StaggeredKernels<Impl>::DhopImproved(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U, DoubledGaugeField &UUU,
const FermionField &in, FermionField &out, int dag, int interior,int exterior)
{
GridBase *FGrid=in.Grid();
GridBase *UGrid=U.Grid();
typedef StaggeredKernels<Impl> ThisKernel;
autoView( UUU_v , UUU, AcceleratorRead);
autoView( U_v , U, AcceleratorRead);
autoView( in_v , in, AcceleratorRead);
autoView( out_v , out, AcceleratorWrite);
autoView( st_v , st, AcceleratorRead);
SiteSpinor * buf = st.CommBuf();
int Ls=1;
if(FGrid->Nd()==UGrid->Nd()+1){
Ls = FGrid->_rdimensions[0];
}
int Nsite = UGrid->oSites();
if( interior && exterior ) {
if (Opt == OptGeneric ) { KERNEL_CALL(DhopSiteGeneric,1); return;}
#ifndef GRID_CUDA
if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHand,1); return;}
if (Opt == OptInlineAsm ) { ASM_CALL(DhopSiteAsm); return;}
#endif
} else if( interior ) {
if (Opt == OptGeneric ) { KERNEL_CALL(DhopSiteGenericInt,1); return;}
#ifndef GRID_CUDA
if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandInt,1); return;}
#endif
} else if( exterior ) {
if (Opt == OptGeneric ) { KERNEL_CALL(DhopSiteGenericExt,1); return;}
#ifndef GRID_CUDA
if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandExt,1); return;}
#endif
}
assert(0 && " Kernel optimisation case not covered ");
}
template <class Impl>
void StaggeredKernels<Impl>::DhopNaive(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in, FermionField &out, int dag, int interior,int exterior)
{
GridBase *FGrid=in.Grid();
GridBase *UGrid=U.Grid();
typedef StaggeredKernels<Impl> ThisKernel;
autoView( UUU_v , U, AcceleratorRead);
autoView( U_v , U, AcceleratorRead);
autoView( in_v , in, AcceleratorRead);
autoView( out_v , out, AcceleratorWrite);
autoView( st_v , st, AcceleratorRead);
SiteSpinor * buf = st.CommBuf();
int Ls=1;
if(FGrid->Nd()==UGrid->Nd()+1){
Ls = FGrid->_rdimensions[0];
}
int Nsite = UGrid->oSites();
if( interior && exterior ) {
if (Opt == OptGeneric ) { KERNEL_CALL(DhopSiteGeneric,0); return;}
#ifndef GRID_CUDA
if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHand,0); return;}
#endif
} else if( interior ) {
if (Opt == OptGeneric ) { KERNEL_CALL(DhopSiteGenericInt,0); return;}
#ifndef GRID_CUDA
if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandInt,0); return;}
#endif
} else if( exterior ) {
if (Opt == OptGeneric ) { KERNEL_CALL(DhopSiteGenericExt,0); return;}
#ifndef GRID_CUDA
if (Opt == OptHandUnroll ) { KERNEL_CALL(DhopSiteHandExt,0); return;}
#endif
}
}
#undef KERNEL_CALLNB
#undef KERNEL_CALL
#undef ASM_CALL
NAMESPACE_END(Grid);

23
Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h
View File

@@ -35,7 +35,7 @@ NAMESPACE_BEGIN(Grid);
// *NOT* EO
template <class Impl>
RealD WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
{
FermionField temp(out.Grid());
@@ -47,11 +47,10 @@ RealD WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
Mooee(in, temp);
out += temp;
return norm2(out);
}
template <class Impl>
RealD WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
{
FermionField temp(out.Grid());
@@ -63,7 +62,6 @@ RealD WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
MooeeDag(in, temp);
out += temp;
return norm2(out);
}
template <class Impl>
@@ -100,11 +98,13 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
Coordinate lcoor;
typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
for (int site = 0; site < lvol; site++)
{
autoView(CTv,CloverTerm,CpuRead);
autoView(CTIv,CloverTermInv,CpuWrite);
for (int site = 0; site < lvol; site++) {
grid->LocalIndexToLocalCoor(site, lcoor);
EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
peekLocalSite(Qx, CloverTerm, lcoor);
peekLocalSite(Qx, CTv, lcoor);
Qxinv = Zero();
//if (csw!=0){
for (int j = 0; j < Ns; j++)
@@ -125,21 +125,22 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
// if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
// }
pokeLocalSite(Qxinv, CloverTermInv, lcoor);
pokeLocalSite(Qxinv, CTIv, lcoor);
}
}
// Separate the even and odd parts
pickCheckerboard(Even, CloverTermEven, CloverTerm);
pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
pickCheckerboard(Even, CloverTermDagEven, adj(CloverTerm));
pickCheckerboard(Odd, CloverTermDagOdd, adj(CloverTerm));
pickCheckerboard(Even, CloverTermDagEven, closure(adj(CloverTerm)));
pickCheckerboard(Odd, CloverTermDagOdd, closure(adj(CloverTerm)));
pickCheckerboard(Even, CloverTermInvEven, CloverTermInv);
pickCheckerboard(Odd, CloverTermInvOdd, CloverTermInv);
pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
pickCheckerboard(Even, CloverTermInvDagEven, closure(adj(CloverTermInv)));
pickCheckerboard(Odd, CloverTermInvDagOdd, closure(adj(CloverTermInv)));
}
template <class Impl>

231
Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
View File

@@ -580,16 +580,21 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
cosha = (one + W*W + sk) / (abs(W)*2.0);
// FIXME Need a Lattice acosh
{
autoView(cosha_v,cosha,CpuRead);
autoView(a_v,a,CpuWrite);
for(int idx=0;idx<_grid->lSites();idx++){
Coordinate lcoor(Nd);
Tcomplex cc;
// RealD sgn;
_grid->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(cc,cosha,lcoor);
peekLocalSite(cc,cosha_v,lcoor);
assert((double)real(cc)>=1.0);
assert(fabs((double)imag(cc))<=1.0e-15);
cc = ScalComplex(::acosh(real(cc)),0.0);
pokeLocalSite(cc,a,lcoor);
pokeLocalSite(cc,a_v,lcoor);
}
}
Wea = ( exp( a) * abs(W) );
@@ -775,17 +780,20 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt(FermionField &out,const Fe
cosha = (one + W*W + sk) / (abs(W)*2.0);
// FIXME Need a Lattice acosh
{
autoView(cosha_v,cosha,CpuRead);
autoView(a_v,a,CpuWrite);
for(int idx=0;idx<_grid->lSites();idx++){
Coordinate lcoor(Nd);
Tcomplex cc;
// RealD sgn;
_grid->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(cc,cosha,lcoor);
peekLocalSite(cc,cosha_v,lcoor);
assert((double)real(cc)>=1.0);
assert(fabs((double)imag(cc))<=1.0e-15);
cc = ScalComplex(::acosh(real(cc)),0.0);
pokeLocalSite(cc,a,lcoor);
}
pokeLocalSite(cc,a_v,lcoor);
}}
Wea = ( exp( a) * abs(W) );
Wema= ( exp(-a) * abs(W) );
@@ -861,7 +869,6 @@ 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) \
@@ -877,220 +884,10 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHw(FermionField &out,const Fe
merge(qSiteRev, qSiteVec); \
}
// psi = chiralProjectPlus(Result_s[Ls/2-1]);
// psi+= chiralProjectMinus(Result_s[Ls/2]);
// PJ5q+=localInnerProduct(psi,psi);
template<class vobj>
Lattice<vobj> spProj5p(const Lattice<vobj> & in)
{
GridBase *grid=in.Grid();
Gamma G5(Gamma::Algebra::Gamma5);
Lattice<vobj> ret(grid);
auto ret_v = ret.View();
auto in_v = in.View();
thread_for(ss,grid->oSites(),{
ret_v[ss] = in_v[ss] + G5*in_v[ss];
});
return ret;
}
template<class vobj>
Lattice<vobj> spProj5m(const Lattice<vobj> & in)
{
Gamma G5(Gamma::Algebra::Gamma5);
GridBase *grid=in.Grid();
Lattice<vobj> ret(grid);
auto ret_v = ret.View();
auto in_v = in.View();
thread_for(ss,grid->oSites(),{
ret_v[ss] = in_v[ss] - G5*in_v[ss];
});
return ret;
}
template <class Impl>
void WilsonFermion5D<Impl>::ContractJ5q(FermionField &q_in,ComplexField &J5q)
{
conformable(GaugeGrid(), J5q.Grid());
conformable(q_in.Grid(), FermionGrid());
// 4d field
int Ls = this->Ls;
FermionField psi(GaugeGrid());
FermionField p_plus (GaugeGrid());
FermionField p_minus(GaugeGrid());
FermionField p(GaugeGrid());
ExtractSlice(p_plus , q_in, Ls/2 , 0);
ExtractSlice(p_minus, q_in, Ls/2-1 , 0);
p_plus = spProj5p(p_plus );
p_minus= spProj5m(p_minus);
p=p_plus+p_minus;
J5q = localInnerProduct(p,p);
}
template <class Impl>
void WilsonFermion5D<Impl>::ContractJ5q(PropagatorField &q_in,ComplexField &J5q)
{
conformable(GaugeGrid(), J5q.Grid());
conformable(q_in.Grid(), FermionGrid());
// 4d field
int Ls = this->Ls;
PropagatorField psi(GaugeGrid());
PropagatorField p_plus (GaugeGrid());
PropagatorField p_minus(GaugeGrid());
PropagatorField p(GaugeGrid());
ExtractSlice(p_plus , q_in, Ls/2 , 0);
ExtractSlice(p_minus, q_in, Ls/2-1 , 0);
p_plus = spProj5p(p_plus );
p_minus= spProj5m(p_minus);
p=p_plus+p_minus;
J5q = localInnerProduct(p,p);
}
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);
auto q_in_1_v = q_in_1.View();
auto q_in_2_v = q_in_2.View();
auto tmp1_v = tmp1.View();
auto tmp2_v = tmp2.View();
auto q_out_v = q_out.View();
auto Umu_v = Umu.View();
thread_for(sU, Umu.Grid()->oSites(),{
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_v[sF2], qSite2, Ls / LLs);
REVERSE_LS(tmp2_v[sF2], qmuSite2, Ls / LLs);
}
else
{
qSite2 = q_in_2_v[sF2];
qmuSite2 = tmp2_v[sF2];
}
Kernels::ContractConservedCurrentSiteFwd(tmp1_v[sF1],
qSite2,
q_out_v[sU],
Umu_v, sU, mu, axial_sign);
Kernels::ContractConservedCurrentSiteBwd(q_in_1_v[sF1],
qmuSite2,
q_out_v[sU],
Umu_v, 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,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx)
{
conformable(q_in.Grid(), FermionGrid());
conformable(q_in.Grid(), q_out.Grid());
PropagatorField tmp(GaugeGrid()),tmp2(GaugeGrid());
unsigned int tshift = (mu == Tp) ? 1 : 0;
unsigned int LLs = q_in.Grid()->_rdimensions[0];
unsigned int LLt = GridDefaultLatt()[Tp];
q_out = Zero();
LatticeInteger coords(_FourDimGrid);
LatticeCoordinate(coords, Tp);
auto q_out_v = q_out.View();
auto tmp2_v = tmp2.View();
auto coords_v= coords.View();
auto Umu_v = Umu.View();
for (unsigned int s = 0; s < LLs; ++s)
{
bool axial_sign = ((curr_type == Current::Axial) && (s < (LLs / 2)));
bool tadpole_sign = (curr_type == Current::Tadpole);
bool switch_sgn = tadpole_sign || axial_sign;
//forward direction: Need q(x + mu, s)*A(x)
ExtractSlice(tmp2, q_in, s, 0); //q(x,s)
tmp = Cshift(tmp2, mu, 1); //q(x+mu,s)
tmp2 = tmp*lattice_cmplx; //q(x+mu,s)*A(x)
thread_for(sU, Umu.Grid()->oSites(),{
// Compute the sequential conserved current insertion only if our simd
// object contains a timeslice we need.
vPredicate t_mask;
t_mask() = ((coords_v[sU] >= tmin) && (coords_v[sU] <= tmax));
Integer timeSlices = Reduce(t_mask());
if (timeSlices > 0)
{
unsigned int sF = sU * LLs + s;
Kernels::SeqConservedCurrentSiteFwd(tmp2_v[sU],
q_out_v[sF], Umu_v, sU,
mu, t_mask, switch_sgn);
}
});
//backward direction: Need q(x - mu, s)*A(x-mu)
ExtractSlice(tmp2, q_in, s, 0); //q(x,s)
tmp = lattice_cmplx*tmp2; //q(x,s)*A(x)
tmp2 = Cshift(tmp, mu, -1); //q(x-mu,s)*A(x-mu,s)
thread_for(sU, Umu.Grid()->oSites(),
{
vPredicate t_mask;
t_mask()= ((coords_v[sU] >= (tmin + tshift)) && (coords_v[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_v[sU] == t0 ));
Integer timeSlices = Reduce(t_mask());
if (timeSlices > 0) {
unsigned int sF = sU * LLs + s;
Kernels::SeqConservedCurrentSiteBwd(tmp2_v[sU],
q_out_v[sF], Umu_v, sU,
mu, t_mask, axial_sign);
}
});
}
}
NAMESPACE_END(Grid);

Some files were not shown because too many files have changed in this diff Show More