portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-19 00:07:05 +01:00
Code Releases Activity

Merge branch 'develop' into sycl

Browse Source
This commit is contained in:
Peter Boyle
2020-06-09 04:00:12 -04:00
parent 616d3dd737 ffbb3fc02c
commit cdf0a04fc5
85 changed files with 2632 additions and 1334 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
Grid/algorithms/CoarsenedMatrix.h
View File

@ -252,19 +252,14 @@ 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();
autoView( in_v , in, AcceleratorRead);
autoView( out_v , out, AcceleratorWrite);
typedef LatticeView<Cobj> Aview;
@ -278,12 +273,7 @@ public:
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;
@ -310,18 +300,11 @@ public:
}
coalescedWrite(out_v[ss](b),res);
});
usecs +=usecond();
double nrm_usec=-usecond();
RealD Nout= norm2(out);
nrm_usec+=usecond();
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer[p].ViewClose();
return Nout;
};
RealD Mdag (const CoarseVector &in, CoarseVector &out)
void Mdag (const CoarseVector &in, CoarseVector &out)
{
if(hermitian) {
// corresponds to Petrov-Galerkin coarsening
@ -332,7 +315,6 @@ public:
G5C(tmp, in);
M(tmp, out);
G5C(out, out);
return norm2(out);
}
};
void MdirComms(const CoarseVector &in)
@ -553,8 +535,6 @@ public:
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)); });
}
}

627
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);
}
@ -208,338 +207,305 @@ 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> {
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) {
Field tmp(in.Grid());
tmp.Checkerboard() = in.Checkerboard();
ni=Mpc(in,tmp);
no=MpcDag(tmp,out);
}
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
out.Checkerboard() = in.Checkerboard();
MpcDagMpc(in,out,n1,n2);
}
virtual void HermOp(const Field &in, Field &out){
RealD n1,n2;
HermOpAndNorm(in,out,n1,n2);
}
void Op (const Field &in, Field &out){
Mpc(in,out);
}
void AdjOp (const Field &in, Field &out){
MpcDag(in,out);
}
// Support for coarsening to a multigrid
void OpDiag (const Field &in, Field &out) {
assert(0); // must coarsen the unpreconditioned system
}
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 SchurDiagMooeeOperator : public SchurOperatorBase<Field> {
public:
Matrix &_Mat;
SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD 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);
}
virtual RealD 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);
}
};
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) {
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);
}
virtual RealD 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);
}
};
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) {
Field tmp(in.Grid());
_Mat.MooeeInv(in,out);
_Mat.Meooe(out,tmp);
_Mat.MooeeInv(tmp,out);
_Mat.Meooe(out,tmp);
return axpy_norm(out,-1.0,tmp,in);
}
virtual RealD MpcDag (const Field &in, Field &out){
Field tmp(in.Grid());
_Mat.MeooeDag(in,out);
_Mat.MooeeInvDag(out,tmp);
_Mat.MeooeDag(tmp,out);
_Mat.MooeeInvDag(out,tmp);
return axpy_norm(out,-1.0,tmp,in);
}
};
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) {
Field tmp(in.Grid());
tmp.Checkerboard() = in.Checkerboard();
ni = Mpc(in,tmp);
no = MpcDag(tmp,out);
}
virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
assert(0);
}
virtual void HermOp(const Field& in, Field& out) {
assert(0);
}
void Op(const Field& in, Field& out) {
Mpc(in, out);
}
void AdjOp(const Field& in, Field& out) {
MpcDag(in, out);
}
// Support for coarsening to a multigrid
void OpDiag(const Field& in, Field& out) {
assert(0); // must coarsen the unpreconditioned system
}
void OpDir(const Field& in, Field& out, int dir, int disp) {
assert(0);
}
};
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) {
Field tmp(in.Grid());
tmp.Checkerboard() = !in.Checkerboard();
_Mat.Meooe(in, tmp);
_Mat.MooeeInv(tmp, out);
_Mat.Meooe(out, tmp);
_Mat.Mooee(in, out);
return axpy_norm(out, -1.0, tmp, out);
}
virtual RealD 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);
}
};
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) {
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);
}
virtual RealD 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);
}
};
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) {
Field tmp(in.Grid());
_Mat.MooeeInv(in, out);
_Mat.Meooe(out, tmp);
_Mat.MooeeInv(tmp, out);
_Mat.Meooe(out, tmp);
return axpy_norm(out, -1.0, tmp, in);
}
virtual RealD MpcDag(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MeooeDag(in, out);
_Mat.MooeeInvDag(out, tmp);
_Mat.MeooeDag(tmp, out);
_Mat.MooeeInvDag(out, tmp);
return axpy_norm(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> {
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;
}
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;
}
template<class Field>
class SchurOperatorBase : public LinearOperatorBase<Field> {
public:
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();
Mpc(in,tmp);
MpcDag(tmp,out);
}
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
ncall++;
tMpc-=usecond();
n2 = Mpc(in,out);
tMpc+=usecond();
tIP-=usecond();
ComplexD dot= innerProduct(in,out);
tIP+=usecond();
n1 = real(dot);
out.Checkerboard() = in.Checkerboard();
MpcDagMpc(in,out);
ComplexD dot= innerProduct(in,out);
n1=real(dot);
n2=norm2(out);
}
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();
out.Checkerboard() = in.Checkerboard();
MpcDagMpc(in,out);
}
virtual RealD Mpc (const Field &in, Field &out)
{
void Op (const Field &in, Field &out){
Mpc(in,out);
}
void AdjOp (const Field &in, Field &out){
MpcDag(in,out);
}
// Support for coarsening to a multigrid
void OpDiag (const Field &in, Field &out) {
assert(0); // must coarsen the unpreconditioned system
}
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 SchurDiagMooeeOperator : public SchurOperatorBase<Field> {
public:
Matrix &_Mat;
SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
virtual void Mpc (const Field &in, Field &out) {
Field tmp(in.Grid());
tmp.Checkerboard() = !in.Checkerboard();
_Mat.Meooe(in,tmp);
_Mat.MooeeInv(tmp,out);
_Mat.Meooe(out,tmp);
_Mat.Mooee(in,out);
axpy(out,-1.0,tmp,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);
axpy(out,-1.0,tmp,out);
}
};
template<class Matrix,class Field>
class SchurDiagOneOperator : public SchurOperatorBase<Field> {
protected:
Matrix &_Mat;
public:
SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
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);
axpy(out,-1.0,tmp,in);
}
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);
axpy(out,-1.0,tmp,in);
}
};
template<class Matrix,class Field>
class SchurDiagTwoOperator : public SchurOperatorBase<Field> {
protected:
Matrix &_Mat;
public:
SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
virtual void Mpc (const Field &in, Field &out) {
Field tmp(in.Grid());
_Mat.MooeeInv(in,out);
_Mat.Meooe(out,tmp);
_Mat.MooeeInv(tmp,out);
_Mat.Meooe(out,tmp);
axpy(out,-1.0,tmp,in);
}
virtual void MpcDag (const Field &in, Field &out){
Field tmp(in.Grid());
_Mat.MeooeDag(in,out);
_Mat.MooeeInvDag(out,tmp);
_Mat.MeooeDag(tmp,out);
_Mat.MooeeInvDag(out,tmp);
axpy(out,-1.0,tmp,in);
}
};
template<class Field>
class NonHermitianSchurOperatorBase : public LinearOperatorBase<Field>
{
public:
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();
Mpc(in,tmp);
MpcDag(tmp,out);
}
virtual void HermOpAndNorm(const Field& in, Field& out, RealD& n1, RealD& n2) {
assert(0);
}
virtual void HermOp(const Field& in, Field& out) {
assert(0);
}
void Op(const Field& in, Field& out) {
Mpc(in, out);
}
void AdjOp(const Field& in, Field& out) {
MpcDag(in, out);
}
// Support for coarsening to a multigrid
void OpDiag(const Field& in, Field& out) {
assert(0); // must coarsen the unpreconditioned system
}
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>
{
public:
Matrix& _Mat;
NonHermitianSchurDiagMooeeOperator(Matrix& Mat): _Mat(Mat){};
virtual void Mpc(const Field& in, Field& out) {
Field tmp(in.Grid());
tmp.Checkerboard() = !in.Checkerboard();
_Mat.Meooe(in, tmp);
_Mat.MooeeInv(tmp, out);
_Mat.Meooe(out, tmp);
_Mat.Mooee(in, out);
axpy(out, -1.0, tmp, 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);
axpy(out, -1.0, tmp, out);
}
};
template<class Matrix,class Field>
class NonHermitianSchurDiagOneOperator : public NonHermitianSchurOperatorBase<Field>
{
protected:
Matrix &_Mat;
public:
NonHermitianSchurDiagOneOperator (Matrix& Mat): _Mat(Mat){};
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);
axpy(out, -1.0, tmp, in);
}
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);
axpy(out, -1.0, tmp, in);
}
};
template<class Matrix, class Field>
class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Field>
{
protected:
Matrix& _Mat;
public:
NonHermitianSchurDiagTwoOperator(Matrix& Mat): _Mat(Mat){};
virtual void Mpc(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MooeeInv(in, out);
_Mat.Meooe(out, tmp);
_Mat.MooeeInv(tmp, out);
_Mat.Meooe(out, tmp);
axpy(out, -1.0, tmp, in);
}
virtual void MpcDag(const Field& in, Field& out) {
Field tmp(in.Grid());
_Mat.MeooeDag(in, out);
_Mat.MooeeInvDag(out, tmp);
_Mat.MeooeDag(tmp, out);
_Mat.MooeeInvDag(out, tmp);
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> {
protected:
Matrix &_Mat;
Field tmp;
RealD mass;
public:
SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid())
{
assert( _Mat.isTrivialEE() );
mass = _Mat.Mass();
}
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
Mpc(in,out);
ComplexD dot= innerProduct(in,out);
n1 = real(dot);
n2 =0.0;
}
virtual void HermOp(const Field &in, Field &out){
Mpc(in,out);
// _Mat.Meooe(in,out);
// _Mat.Meooe(out,tmp);
// axpby(out,-1.0,mass*mass,tmp,in);
}
virtual void Mpc (const Field &in, Field &out)
{
Field tmp(in.Grid());
Field tmp2(in.Grid());
// _Mat.Mooee(in,out);
// _Mat.Mooee(out,tmp);
// std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
_Mat.Mooee(in,out);
_Mat.Mooee(out,tmp);
// std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
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;

28
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);
@ -258,14 +256,28 @@ public:
// out = ()*T0 + Coeffs[1]*T1;
axpby(out,0.5*Coeffs[0],Coeffs[1],T0,T1);
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));
axpy(out,Coeffs[n],*Tnp,out);
if ( Coeffs[n] != 0.0) {
axpy(out,Coeffs[n],*Tnp,out);
}
#endif
// Cycle pointers to avoid copies
Field *swizzle = Tnm;
Tnm =Tn;

212
Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@ -37,218 +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)
{
GridBase* grid = basis[0].Grid();
typedef typename Field::vector_object vobj;
typedef decltype(basis[0].View(CpuWrite)) View;
Vector<View> basis_v; basis_v.reserve(basis.size());
for(int k=0;k<basis.size();k++) basis_v.push_back(basis[k].View(CpuWrite));
View *basis_vp = &basis_v[0];
#if 1
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
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(Bp[ss]);
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_vp[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_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
});
}
for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
#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)
{
GridBase* grid = basis[0].Grid();
typedef typename Field::vector_object vobj;
typedef decltype(basis[0].View(AcceleratorWrite)) View;
result.Checkerboard() = basis[0].Checkerboard();
autoView(result_v,result, AcceleratorWrite);
Vector<View> basis_v; basis_v.reserve(basis.size());
View * basis_vp = &basis_v[0];
for(int k=0;k<basis.size();k++) basis_v.push_back(basis[k].View(AcceleratorRead));
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(basis_vp[k0][ss]);
B=Zero();
for(int k=k0; k<k1; ++k){
B +=Qt_j[k] * coalescedRead(basis_vp[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);
}
}
/////////////////////////////////////////////////////////////
// Implicitly restarted lanczos
/////////////////////////////////////////////////////////////

66
Grid/allocator/AlignedAllocator.cc
View File

@ -6,72 +6,6 @@ NAMESPACE_BEGIN(Grid);
MemoryStats *MemoryProfiler::stats = nullptr;
bool MemoryProfiler::debug = false;
#ifdef GRID_CUDA
#define SMALL_LIMIT (0)
#else
#define SMALL_LIMIT (4096)
#endif
#ifdef ALLOCATION_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__

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

@ -255,6 +255,10 @@ 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;

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

4
Grid/lattice/Lattice.h
View File

@ -36,7 +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>
@ -44,4 +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>

3
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
@ -99,7 +100,7 @@ const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg)
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

1
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

2
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
@ -91,6 +92,7 @@ public:
// The view is trivially copy constructible and may be copied to an accelerator device
// in device lambdas
/////////////////////////////////////////////////////////////////////////////////
LatticeView<vobj> View (ViewMode mode) const
{
LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this),mode);

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);

6
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());
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());
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)));
});

132
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
@ -64,6 +65,37 @@ inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
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);
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_object sum(const vobj *arg, Integer osites)
@ -74,6 +106,15 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
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)
@ -101,44 +142,49 @@ 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;
ComplexD nrm;
GridBase *grid = left.Grid();
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites = grid->oSites();
// Might make all code paths go this way.
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;
typedef decltype(innerProduct(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);
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));
})
// GPU - SIMT lane compliance...
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
#if defined(GRID_CUDA)||defined(GRID_HIP)
nrm = TensorRemove(sumD_gpu(inner_tmp_v,sites));
#else
nrm = TensorRemove(sum_cpu(inner_tmp_v,sites));
#endif
grid->GlobalSum(nrm);
nrm = TensorRemove(sumD(inner_tmp_v,sites));
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
@ -181,17 +227,51 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
coalescedWrite(z_v[ss],tmp);
});
#if defined(GRID_CUDA)||defined(GRID_HIP)
nrm = real(TensorRemove(sumD_gpu(inner_tmp_v,sites)));
#else
// Already promoted to double
nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
#endif
nrm = real(TensorRemove(sumD(inner_tmp_v,sites)));
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(innerProduct(vobj(),vobj())) inner_t;
typedef decltype(innerProduct(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, nsimd,{
auto left_tmp = left_v(ss);
coalescedWrite(inner_tmp_v[ss],innerProduct(left_tmp,right_v(ss)));
coalescedWrite(norm_tmp_v[ss],innerProduct(left_tmp,left_tmp));
});
}
tmp[0] = TensorRemove(sumD(inner_tmp_v,sites));
tmp[1] = TensorRemove(sumD(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)
->typename decltype(expr.op.func(eval(0,expr.arg1)))::scalar_object

2
Grid/lattice/Lattice_trace.h
View File

@ -37,6 +37,7 @@ NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////////////////////////////////////////////////
// Trace
////////////////////////////////////////////////////////////////////////////////////////////////////
/*
template<class vobj>
inline auto trace(const Lattice<vobj> &lhs) -> Lattice<decltype(trace(vobj()))>
{
@ -48,6 +49,7 @@ inline auto trace(const Lattice<vobj> &lhs) -> Lattice<decltype(trace(vobj()))>
});
return ret;
};
*/
////////////////////////////////////////////////////////////////////////////////////////////////////
// Trace Index level dependent operation

266
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
@ -83,32 +84,138 @@ 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)[SIMTlane(vComplexF::Nsimd())] = in;
}
accelerator_inline void convertType(vComplexD & out, const ComplexD & in) {
((ComplexD*)&out)[SIMTlane(vComplexD::Nsimd())] = in;
}
accelerator_inline void convertType(vComplexD2 & out, const ComplexD & in) {
((ComplexD*)&out)[SIMTlane(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()[0],rhs.View()[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 Lattice<vobj> &fineData,
const VLattice &Basis)
{
GridBase * fine = fineData.Grid();
GridBase * coarse= coarseData.Grid();
Lattice<CComplex> ip(coarse);
Lattice<iScalar<CComplex>> ip(coarse);
Lattice<vobj> fineDataRed = fineData;
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>
inline void blockZAXPY(Lattice<vobj> &fineZ,
const Lattice<CComplex> &coarseA,
const Lattice<vobj> &fineX,
const Lattice<vobj> &fineY)
template<class vobj,class vobj2,class CComplex>
inline void blockZAXPY(Lattice<vobj> &fineZ,
const Lattice<CComplex> &coarseA,
const Lattice<vobj2> &fineX,
const Lattice<vobj> &fineY)
{
GridBase * fine = fineZ.Grid();
GridBase * coarse= coarseA.Grid();
@ -120,7 +227,7 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
conformable(fineX,fineZ);
int _ndimension = coarse->_ndimension;
Coordinate block_r (_ndimension);
// FIXME merge with subdivide checking routine as this is redundant
@ -135,23 +242,60 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
autoView( coarseA_, coarseA, AcceleratorRead);
accelerator_for(sf, fine->oSites(), CComplex::Nsimd(), {
int sc;
Coordinate coor_c(_ndimension);
Coordinate coor_f(_ndimension);
Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
int sc;
Coordinate coor_c(_ndimension);
Coordinate coor_f(_ndimension);
// z = A x + y
coalescedWrite(fineZ_[sf],coarseA_(sc)*fineX_(sf)+fineY_(sf));
Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
});
// z = A x + y
#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(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)
@ -167,12 +311,15 @@ inline void blockInnerProduct(Lattice<CComplex> &CoarseInner,
// Precision promotion?
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];
});
{
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)
{
@ -185,7 +332,7 @@ inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
// useful in multigrid project;
// Generic name : Coarsen?
template<class vobj>
inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
{
GridBase * fine = fineData.Grid();
GridBase * coarse= coarseData.Grid();
@ -193,9 +340,9 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
subdivides(coarse,fine); // require they map
int _ndimension = coarse->_ndimension;
Coordinate block_r (_ndimension);
for(int d=0 ; d<_ndimension;d++){
block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
}
@ -208,27 +355,28 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
accelerator_for(sc,coarse->oSites(),1,{
// One thread per sub block
Coordinate coor_c(_ndimension);
Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions); // Block coordinate
coarseData_[sc]=Zero();
// One thread per sub block
Coordinate coor_c(_ndimension);
Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions); // Block coordinate
coarseData_[sc]=Zero();
for(int sb=0;sb<blockVol;sb++){
int sf;
Coordinate coor_b(_ndimension);
Coordinate coor_f(_ndimension);
Lexicographic::CoorFromIndex(coor_b,sb,block_r); // Block sub coordinate
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);
for(int sb=0;sb<blockVol;sb++){
coarseData_[sc]=coarseData_[sc]+fineData_[sf];
}
int sf;
Coordinate coor_b(_ndimension);
Coordinate coor_f(_ndimension);
Lexicographic::CoorFromIndex(coor_b,sb,block_r); // Block sub coordinate
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);
});
coarseData_[sc]=coarseData_[sc]+fineData_[sf];
}
});
return;
}
template<class vobj>
inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vobj> &picked,Coordinate coor)
{
@ -250,8 +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();
@ -259,23 +407,30 @@ inline void blockOrthogonalise(Lattice<CComplex> &ip,std::vector<Lattice<vobj> >
int nbasis = Basis.size() ;
// checks
subdivides(coarse,fine);
subdivides(coarse,fine);
for(int i=0;i<nbasis;i++){
conformable(Basis[i].Grid(),fine);
}
for(int v=0;v<nbasis;v++) {
for(int u=0;u<v;u++) {
//Inner product & remove component
blockInnerProduct(ip,Basis[u],Basis[v]);
//Inner product & remove component
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,
@ -320,17 +475,17 @@ 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);
autoView( cip_ , cip, AcceleratorWrite);
autoView( ip_ , ip, AcceleratorRead);
@ -407,6 +562,7 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
Coordinate rdt = Tg->_rdimensions;
Coordinate ist = Tg->_istride;
Coordinate ost = Tg->_ostride;
autoView( t_v , To, AcceleratorWrite);
autoView( f_v , From, AcceleratorRead);
accelerator_for(idx,Fg->lSites(),1,{

4
Grid/lattice/Lattice_transpose.h
View File

@ -38,6 +38,7 @@ NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////////////////////////////////////////////////
// Transpose
////////////////////////////////////////////////////////////////////////////////////////////////////
/*
template<class vobj>
inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
@ -48,7 +49,8 @@ inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
});
return ret;
};
*/
////////////////////////////////////////////////////////////////////////////////////////////////////
// Index level dependent transpose
////////////////////////////////////////////////////////////////////////////////////////////////////

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/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;
}
};

4
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);

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);

9
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;
@ -86,15 +86,14 @@ 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
virtual void MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) { assert(0);};
virtual void MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) { assert(0);};
virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary,std::vector<double> twist)
virtual void FreePropagator(const FermionField &in,FermionField &out,RealD mass,std::vector<Complex> boundary,std::vector<double> twist)
{
FFT theFFT((GridCartesian *) in.Grid());

4
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

5
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);

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);

4
Grid/qcd/action/fermion/NaiveStaggeredFermion.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

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);

5
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);

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

@ -78,8 +78,8 @@ 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

5
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);};

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;

6
Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h
View File

@ -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

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){

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));
}
/********************************************************************

29
Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h
View File

@ -470,21 +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 {
@ -492,7 +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 {
@ -501,27 +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);
}

44
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;
@ -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());

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));
}
/********************************************************************

8
Grid/qcd/action/fermion/implementation/NaiveStaggeredFermionImplementation.h
View File

@ -128,17 +128,17 @@ void NaiveStaggeredFermion<Impl>::ImportGauge(const GaugeField &_U)
/////////////////////////////
template <class Impl>
RealD NaiveStaggeredFermion<Impl>::M(const FermionField &in, FermionField &out) {
void NaiveStaggeredFermion<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 NaiveStaggeredFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
void NaiveStaggeredFermion<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>

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>

14
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>
@ -132,14 +130,14 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
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>

52
Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
View File

@ -102,21 +102,24 @@ void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
/////////////////////////////
template <class Impl>
RealD WilsonFermion<Impl>::M(const FermionField &in, FermionField &out) {
void WilsonFermion<Impl>::M(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerNo);
return axpy_norm(out, diag_mass, in, out);
axpy(out, diag_mass, in, out);
}
template <class Impl>
RealD WilsonFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
void WilsonFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerYes);
return axpy_norm(out, diag_mass, in, out);
axpy(out, diag_mass, in, out);
}
template <class Impl>
void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out)
{
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerNo);
} else {
@ -125,7 +128,8 @@ void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
}
template <class Impl>
void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out)
{
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerYes);
} else {
@ -134,26 +138,30 @@ void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
}
template <class Impl>
void WilsonFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
void WilsonFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
typename FermionField::scalar_type scal(diag_mass);
out = scal * in;
}
template <class Impl>
void WilsonFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
void WilsonFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
Mooee(in, out);
}
template<class Impl>
void WilsonFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
void WilsonFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
out = (1.0/(diag_mass))*in;
}
template<class Impl>
void WilsonFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out) {
void WilsonFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
{
out.Checkerboard() = in.Checkerboard();
MooeeInv(in,out);
}
@ -249,7 +257,8 @@ void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
}
template <class Impl>
void WilsonFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
void WilsonFermion<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());
@ -260,7 +269,8 @@ void WilsonFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, cons
}
template <class Impl>
void WilsonFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
void WilsonFermion<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()); not general, leaving as a comment (Guido)
@ -274,7 +284,8 @@ void WilsonFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, co
}
template <class Impl>
void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
void WilsonFermion<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());
@ -287,7 +298,8 @@ void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, co
}
template <class Impl>
void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) {
void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
{
conformable(in.Grid(), _grid); // verifies full grid
conformable(in.Grid(), out.Grid());
@ -297,7 +309,8 @@ void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int da
}
template <class Impl>
void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) {
void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
{
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
@ -308,7 +321,8 @@ void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int
}
template <class Impl>
void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag) {
void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
{
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
@ -386,7 +400,8 @@ template <class Impl>
void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag) {
FermionField &out, int dag)
{
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor(dag);
@ -436,7 +451,8 @@ template <class Impl>
void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag) {
FermionField &out, int dag)
{
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor(dag);
st.HaloExchange(in, compressor);

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

@ -447,19 +447,19 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSite); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite); return;}
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSite); printf("."); return;}
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSite); return;}
#endif
} else if( interior ) {
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALLNB(GenericDhopSiteInt); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteInt); return;}
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteInt); printf("-"); return;}
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteInt); return;}
#endif
} else if( exterior ) {
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteExt); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt); return;}
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteExt); printf("+"); return;}
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteExt); return;}
#endif
}
assert(0 && " Kernel optimisation case not covered ");

2
Grid/qcd/action/gauge/GaugeImplementations.h
View File

@ -59,7 +59,7 @@ public:
}
static inline GaugeLinkField
CovShiftIdentityBackward(const GaugeLinkField &Link, int mu) {
return Cshift(adj(Link), mu, -1);
return Cshift(closure(adj(Link)), mu, -1);
}
static inline GaugeLinkField
CovShiftIdentityForward(const GaugeLinkField &Link, int mu) {

4
Grid/qcd/hmc/HMC_aggregate.h
View File

@ -39,6 +39,10 @@ directory
#include <Grid/parallelIO/IldgIOtypes.h>
#include <Grid/parallelIO/IldgIO.h>
#include <Grid/parallelIO/NerscIO.h>
#include <Grid/parallelIO/OpenQcdIO.h>
#if !defined(GRID_COMMS_NONE)
#include <Grid/parallelIO/OpenQcdIOChromaReference.h>
#endif
NAMESPACE_CHECK(Ildg);
#include <Grid/qcd/hmc/checkpointers/CheckPointers.h>

1
Grid/qcd/modules/Registration.h
View File

@ -80,6 +80,7 @@ static Registrar<OneFlavourRatioEOFModule<FermionImplementationPolicy>,
static Registrar< ConjugateGradientModule<WilsonFermionR::FermionField>,
HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __CGWFmodXMLInit("ConjugateGradient");
static Registrar< BiCGSTABModule<WilsonFermionR::FermionField>,
HMC_SolverModuleFactory<solver_string, WilsonFermionR::FermionField, Serialiser> > __BiCGWFmodXMLInit("BiCGSTAB");
static Registrar< ConjugateResidualModule<WilsonFermionR::FermionField>,

156
Grid/qcd/utils/BaryonUtils.h
View File

@ -46,7 +46,7 @@ public:
typedef typename SpinMatrixField::vector_object sobj;
static const int epsilon[6][3] ;
static const Complex epsilon_sgn[6];
static const Real epsilon_sgn[6];
private:
template <class mobj, class robj>
@ -151,14 +151,18 @@ public:
template <class FImpl>
const int BaryonUtils<FImpl>::epsilon[6][3] = {{0,1,2},{1,2,0},{2,0,1},{0,2,1},{2,1,0},{1,0,2}};
template <class FImpl>
/*template <class FImpl>
const Complex BaryonUtils<FImpl>::epsilon_sgn[6] = {Complex(1),
Complex(1),
Complex(1),
Complex(-1),
Complex(-1),
Complex(-1)};
*/
template <class FImpl>
const Real BaryonUtils<FImpl>::epsilon_sgn[6] = {1.,1.,1.,-1.,-1.,-1.};
//This is the old version
template <class FImpl>
template <class mobj, class robj>
void BaryonUtils<FImpl>::baryon_site(const mobj &D1,
@ -173,12 +177,14 @@ void BaryonUtils<FImpl>::baryon_site(const mobj &D1,
robj &result)
{
Gamma g4(Gamma::Algebra::GammaT); //needed for parity P_\pm = 0.5*(1 \pm \gamma_4)
Gamma g4(Gamma::Algebra::GammaT); //needed for parity P_\pm = 0.5*(1 \pm \gamma_4)
auto gD1a = GammaA_left * GammaA_right * D1;
auto gD1b = GammaA_left * g4 * GammaA_right * D1;
auto pD1 = 0.5* (gD1a + (double)parity * gD1b);
auto pD1 = 0.5* (gD1a + (Real)parity * gD1b);
auto gD3 = GammaB_right * D3;
auto D2g = D2 * GammaB_left;
auto pD1g = pD1 * GammaB_left;
auto gD3g = gD3 * GammaB_left;
for (int ie_left=0; ie_left < 6 ; ie_left++){
int a_left = epsilon[ie_left][0]; //a
@ -188,59 +194,78 @@ void BaryonUtils<FImpl>::baryon_site(const mobj &D1,
int a_right = epsilon[ie_right][0]; //a'
int b_right = epsilon[ie_right][1]; //b'
int c_right = epsilon[ie_right][2]; //c'
Real ee = epsilon_sgn[ie_left] * epsilon_sgn[ie_right];
//This is the \delta_{456}^{123} part
if (wick_contraction[0]){
auto D2g = D2 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() += epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1()(gamma_left,gamma_left)(c_right,c_left)*D2g()(alpha_right,beta_left)(a_right,a_left)*gD3()(alpha_right,beta_left)(b_right,b_left);
}}}
auto eepD1 = ee * pD1()(gamma_left,gamma_left)(c_right,c_left);
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
auto D2g_ab = D2g()(alpha_right,beta_left)(a_right,a_left);
auto gD3_ab = gD3()(alpha_right,beta_left)(b_right,b_left);
result()()() += eepD1*D2g_ab*gD3_ab;
}}
}
}
//This is the \delta_{456}^{231} part
if (wick_contraction[1]){
auto pD1g = pD1 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() += epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1g()(gamma_left,beta_left)(c_right,a_left)*D2()(alpha_right,beta_left)(a_right,b_left)*gD3()(alpha_right,gamma_left)(b_right,c_left);
}}}
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
auto gD3_ag = gD3()(alpha_right,gamma_left)(b_right,c_left);
for (int beta_left=0; beta_left<Ns; beta_left++){
auto eepD1g_gb = ee * pD1g()(gamma_left,beta_left)(c_right,a_left);
auto D2_ab = D2()(alpha_right,beta_left)(a_right,b_left);
result()()() += eepD1g_gb*D2_ab*gD3_ag;
}
}}
}
//This is the \delta_{456}^{312} part
if (wick_contraction[2]){
auto gD3g = gD3 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() += epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1()(gamma_left,beta_left)(c_right,b_left)*D2()(alpha_right,gamma_left)(a_right,c_left)*gD3g()(alpha_right,beta_left)(b_right,a_left);
}}}
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
auto D2_ag = D2()(alpha_right,gamma_left)(a_right,c_left);
for (int beta_left=0; beta_left<Ns; beta_left++){
auto eepD1_gb = ee * pD1()(gamma_left,beta_left)(c_right,b_left);
auto gD3g_ab = gD3g()(alpha_right,beta_left)(b_right,a_left);
result()()() += eepD1_gb*D2_ag*gD3g_ab;
}
}}
}
//This is the \delta_{456}^{132} part
if (wick_contraction[3]){
auto gD3g = gD3 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() -= epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1()(gamma_left,gamma_left)(c_right,c_left)*D2()(alpha_right,beta_left)(a_right,b_left)*gD3g()(alpha_right,beta_left)(b_right,a_left);
}}}
auto eepD1 = ee * pD1()(gamma_left,gamma_left)(c_right,c_left);
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
auto D2_ab = D2()(alpha_right,beta_left)(a_right,b_left);
auto gD3g_ab = gD3g()(alpha_right,beta_left)(b_right,a_left);
result()()() -= eepD1*D2_ab*gD3g_ab;
}}
}
}
//This is the \delta_{456}^{321} part
if (wick_contraction[4]){
auto D2g = D2 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() -= epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1()(gamma_left,beta_left)(c_right,b_left)*D2g()(alpha_right,beta_left)(a_right,a_left)*gD3()(alpha_right,gamma_left)(b_right,c_left);
}}}
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
auto gD3_ag = gD3()(alpha_right,gamma_left)(b_right,c_left);
for (int beta_left=0; beta_left<Ns; beta_left++){
auto eepD1_gb = ee * pD1()(gamma_left,beta_left)(c_right,b_left);
auto D2g_ab = D2g()(alpha_right,beta_left)(a_right,a_left);
result()()() -= eepD1_gb*D2g_ab*gD3_ag;
}
}}
}
//This is the \delta_{456}^{213} part
if (wick_contraction[5]){
auto pD1g = pD1 * GammaB_left;
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
for (int beta_left=0; beta_left<Ns; beta_left++){
for (int gamma_left=0; gamma_left<Ns; gamma_left++){
result()()() -= epsilon_sgn[ie_left] * epsilon_sgn[ie_right] * pD1g()(gamma_left,beta_left)(c_right,a_left)*D2()(alpha_right,gamma_left)(a_right,c_left)*gD3()(alpha_right,beta_left)(b_right,b_left);
}}}
for (int alpha_right=0; alpha_right<Ns; alpha_right++){
auto D2_ag = D2()(alpha_right,gamma_left)(a_right,c_left);
for (int beta_left=0; beta_left<Ns; beta_left++){
auto eepD1g_gb = ee * pD1g()(gamma_left,beta_left)(c_right,a_left);
auto gD3_ab = gD3()(alpha_right,beta_left)(b_right,b_left);
result()()() -= eepD1g_gb*D2_ag*gD3_ab;
}
}}
}
}
}
@ -259,11 +284,15 @@ void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
const int parity,
ComplexField &baryon_corr)
{
assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
std::cout << "Contraction <" << quarks_right[0] << quarks_right[1] << quarks_right[2] << "|" << quarks_left[0] << quarks_left[1] << quarks_left[2] << ">" << std::endl;
std::cout << "GammaA (left) " << (GammaA_left.g) << std::endl;
std::cout << "GammaB (left) " << (GammaB_left.g) << std::endl;
std::cout << "GammaA (right) " << (GammaA_right.g) << std::endl;
std::cout << "GammaB (right) " << (GammaB_right.g) << std::endl;
std::cout << "GammaA (left) " << (GammaA_left.g) << std::endl;
std::cout << "GammaB (left) " << (GammaB_left.g) << std::endl;
std::cout << "GammaA (right) " << (GammaA_right.g) << std::endl;
std::cout << "GammaB (right) " << (GammaB_right.g) << std::endl;
assert(parity==1 || parity == -1 && "Parity must be +1 or -1");
@ -278,18 +307,32 @@ void BaryonUtils<FImpl>::ContractBaryons(const PropagatorField &q1_left,
autoView( v2 , q2_left, CpuRead);
autoView( v3 , q3_left, CpuRead);
// accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
thread_for(ss,grid->oSites(),{
//for(int ss=0; ss < grid->oSites(); ss++){
Real bytes =0.;
bytes += grid->oSites() * (432.*sizeof(vComplex) + 126.*sizeof(int) + 36.*sizeof(Real));
for (int ie=0; ie < 6 ; ie++){
if(ie==0 or ie==3){
bytes += grid->oSites() * (4.*sizeof(int) + 4752.*sizeof(vComplex)) * wick_contraction[ie];
}
else{
bytes += grid->oSites() * (64.*sizeof(int) + 5184.*sizeof(vComplex)) * wick_contraction[ie];
}
}
Real t=0.;
t =-usecond();
accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
auto D1 = v1[ss];
auto D2 = v2[ss];
auto D3 = v3[ss];
vobj result=Zero();
baryon_site(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contraction,result);
vbaryon_corr[ss] = result;
} );//end loop over lattice sites
t += usecond();
std::cout << std::setw(10) << bytes/t*1.0e6/1024/1024/1024 << " GB/s " << std::endl;
}
template <class FImpl>
template <class mobj, class robj>
@ -305,11 +348,15 @@ void BaryonUtils<FImpl>::ContractBaryons_Sliced(const mobj &D1,
const int parity,
robj &result)
{
assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
std::cout << "Contraction <" << quarks_right[0] << quarks_right[1] << quarks_right[2] << "|" << quarks_left[0] << quarks_left[1] << quarks_left[2] << ">" << std::endl;
std::cout << "GammaA (left) " << (GammaA_left.g) << std::endl;
std::cout << "GammaB (left) " << (GammaB_left.g) << std::endl;
std::cout << "GammaA (right) " << (GammaA_right.g) << std::endl;
std::cout << "GammaB (right) " << (GammaB_right.g) << std::endl;
std::cout << "GammaA (left) " << (GammaA_left.g) << std::endl;
std::cout << "GammaB (left) " << (GammaB_left.g) << std::endl;
std::cout << "GammaA (right) " << (GammaA_right.g) << std::endl;
std::cout << "GammaB (right) " << (GammaB_right.g) << std::endl;
assert(parity==1 || parity == -1 && "Parity must be +1 or -1");
@ -317,8 +364,8 @@ void BaryonUtils<FImpl>::ContractBaryons_Sliced(const mobj &D1,
for (int ie=0; ie < 6 ; ie++)
wick_contraction[ie] = (quarks_left[0] == quarks_right[epsilon[ie][0]] && quarks_left[1] == quarks_right[epsilon[ie][1]] && quarks_left[2] == quarks_right[epsilon[ie][2]]) ? 1 : 0;
result=Zero();
baryon_site(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contraction,result);
result=Zero();
baryon_site<decltype(D1),decltype(result)>(D1,D2,D3,GammaA_left,GammaB_left,GammaA_right,GammaB_right,parity,wick_contraction,result);
}
/***********************************************************************
@ -558,6 +605,10 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_Eye(const PropagatorField &qq_loop,
const std::string op,
SpinMatrixField &stn_corr)
{
assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
GridBase *grid = qs_ti.Grid();
autoView( vcorr, stn_corr, CpuWrite);
@ -565,8 +616,7 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_Eye(const PropagatorField &qq_loop,
autoView( vd_tf , qd_tf, CpuRead);
autoView( vs_ti , qs_ti, CpuRead);
// accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
thread_for(ss,grid->oSites(),{
accelerator_for(ss, grid->oSites(), grid->Nsimd(), {
auto Dq_loop = vq_loop[ss];
auto Dd_tf = vd_tf[ss];
auto Ds_ti = vs_ti[ss];
@ -595,6 +645,10 @@ void BaryonUtils<FImpl>::Sigma_to_Nucleon_NonEye(const PropagatorField &qq_ti,
const std::string op,
SpinMatrixField &stn_corr)
{
assert(Ns==4 && "Baryon code only implemented for N_spin = 4");
assert(Nc==3 && "Baryon code only implemented for N_colour = 3");
GridBase *grid = qs_ti.Grid();
autoView( vcorr , stn_corr, CpuWrite);

40
Grid/qcd/utils/CovariantCshift.h
View File

@ -52,6 +52,26 @@ namespace PeriodicBC {
tmp = adj(Link)*field;
return Cshift(tmp,mu,-1);// moves towards positive mu
}
template<class gauge,typename Op, typename T1> auto
CovShiftForward(const Lattice<gauge> &Link,
int mu,
const LatticeUnaryExpression<Op,T1> &expr)
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))>
{
Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> arg(expr);
return CovShiftForward(Link,mu,arg);
}
template<class gauge,typename Op, typename T1> auto
CovShiftBackward(const Lattice<gauge> &Link,
int mu,
const LatticeUnaryExpression<Op,T1> &expr)
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))>
{
Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> arg(expr);
return CovShiftForward(Link,mu,arg);
}
}
@ -122,6 +142,26 @@ namespace ConjugateBC {
return Cshift(tmp,mu,-1);// moves towards positive mu
}
template<class gauge,typename Op, typename T1> auto
CovShiftForward(const Lattice<gauge> &Link,
int mu,
const LatticeUnaryExpression<Op,T1> &expr)
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))>
{
Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> arg(expr);
return CovShiftForward(Link,mu,arg);
}
template<class gauge,typename Op, typename T1> auto
CovShiftBackward(const Lattice<gauge> &Link,
int mu,
const LatticeUnaryExpression<Op,T1> &expr)
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))>
{
Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> arg(expr);
return CovShiftForward(Link,mu,arg);
}
}

4
Grid/qcd/utils/WilsonLoops.h
View File

@ -485,7 +485,7 @@ public:
// Up staple ___ ___
// | |
tmp = Cshift(adj(U[nu]), nu, -1);
tmp = Cshift(closure(adj(U[nu])), nu, -1);
tmp = adj(U2[mu]) * tmp;
tmp = Cshift(tmp, mu, -2);
@ -519,7 +519,7 @@ public:
//
// | |
tmp = Cshift(adj(U2[nu]), nu, -2);
tmp = Cshift(closure(adj(U2[nu])), nu, -2);
tmp = Gimpl::CovShiftBackward(U[mu], mu, tmp);
tmp = U2[nu] * Cshift(tmp, nu, 2);
Stap += Cshift(tmp, mu, 1);

6
Grid/serialisation/BaseIO.h
View File

@ -87,11 +87,7 @@ namespace Grid {
template<typename Scalar_, typename Dimensions_, int Options_, typename IndexType>
struct is_tensor_fixed<Eigen::TensorFixedSize<Scalar_, Dimensions_, Options_, IndexType>>
: public std::true_type {};
template<typename Scalar_, typename Dimensions_, int Options_, typename IndexType,
int MapOptions_, template <class> class MapPointer_>
struct is_tensor_fixed<Eigen::TensorMap<Eigen::TensorFixedSize<Scalar_, Dimensions_,
Options_, IndexType>, MapOptions_, MapPointer_>>
: public std::true_type {};
template<typename T> struct is_tensor_fixed<Eigen::TensorMap<T>> : public std::true_type {};
// Is this a variable-size Eigen tensor
template<typename T, typename V = void> struct is_tensor_variable : public std::false_type {};

5
Grid/serialisation/MacroMagic.h
View File

@ -114,7 +114,8 @@ THE SOFTWARE.
#define GRID_MACRO_WRITE_MEMBER(A,B) ::Grid::write(WR,#B,obj. B);
#define GRID_SERIALIZABLE_CLASS_MEMBERS(cname,...)\
std::string SerialisableClassName(void) const {return std::string(#cname);} \
static inline std::string SerialisableClassName(void) {return std::string(#cname);} \
static constexpr bool isEnum = false; \
GRID_MACRO_EVAL(GRID_MACRO_MAP(GRID_MACRO_MEMBER,__VA_ARGS__))\
template <typename T>\
static inline void write(Writer<T> &WR,const std::string &s, const cname &obj){ \
@ -162,6 +163,8 @@ public:\
public:\
accelerator name(void) : value_(undefname) {}; \
accelerator name(int value): value_(value) {}; \
static inline std::string SerialisableClassName(void) {return std::string(#name);}\
static constexpr bool isEnum = true; \
template <typename T>\
static inline void write(::Grid::Writer<T> &WR,const std::string &s, const name &obj) \
{\

4
Grid/serialisation/VectorUtils.h
View File

@ -432,12 +432,10 @@ namespace Grid {
std::vector<T> strToVec(const std::string s)
{
std::istringstream sstr(s);
T buf;
std::vector<T> v;
while(!sstr.eof())
for(T buf; sstr >> buf;)
{
sstr >> buf;
v.push_back(buf);
}

2
Grid/tensors/Tensor_class.h
View File

@ -6,6 +6,7 @@ Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Michael Marshall <michael.marshall@ed.ac.au>
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
@ -55,6 +56,7 @@ class GridTensorBase {};
using Complexified = typename Traits::Complexified; \
using Realified = typename Traits::Realified; \
using DoublePrecision = typename Traits::DoublePrecision; \
using DoublePrecision2= typename Traits::DoublePrecision2; \
static constexpr int TensorLevel = Traits::TensorLevel
template <class vtype>

74
Grid/tensors/Tensor_inner.h
View File

@ -8,6 +8,7 @@
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
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
@ -194,6 +195,79 @@ auto innerProductD (const iScalar<l>& lhs,const iScalar<r>& rhs) -> iScalar<decl
ret._internal = innerProductD(lhs._internal,rhs._internal);
return ret;
}
//////////////////////////////////////
// innerProductD2: precision promotion without inner sum
//////////////////////////////////////
accelerator_inline vComplexD2 TensorRemove(const vComplexD2 & x) { return x; };
accelerator_inline vRealD2 TensorRemove(const vRealD2 & x) { return x; };
accelerator_inline ComplexD innerProductD2(const ComplexF &l,const ComplexF &r){ return innerProduct(l,r); }
accelerator_inline ComplexD innerProductD2(const ComplexD &l,const ComplexD &r){ return innerProduct(l,r); }
accelerator_inline RealD innerProductD2(const RealD &l,const RealD &r){ return innerProduct(l,r); }
accelerator_inline RealD innerProductD2(const RealF &l,const RealF &r){ return innerProduct(l,r); }
accelerator_inline vComplexD innerProductD2(const vComplexD &l,const vComplexD &r){ return innerProduct(l,r); }
accelerator_inline vRealD innerProductD2(const vRealD &l,const vRealD &r){ return innerProduct(l,r); }
accelerator_inline vComplexD2 innerProductD2(const vComplexF &l,const vComplexF &r)
{
vComplexD la,lb;
vComplexD ra,rb;
Optimization::PrecisionChange::StoD(l.v,la.v,lb.v);
Optimization::PrecisionChange::StoD(r.v,ra.v,rb.v);
vComplexD2 ret;
ret._internal[0] = innerProduct(la,ra);
ret._internal[1] = innerProduct(lb,rb);
return ret;
}
accelerator_inline vRealD2 innerProductD2(const vRealF &l,const vRealF &r)
{
vRealD la,lb;
vRealD ra,rb;
Optimization::PrecisionChange::StoD(l.v,la.v,lb.v);
Optimization::PrecisionChange::StoD(r.v,ra.v,rb.v);
vRealD2 ret;
ret._internal[0]=innerProduct(la,ra);
ret._internal[1]=innerProduct(lb,rb);
return ret;
}
// Now do it for vector, matrix, scalar
template<class l,class r,int N> accelerator_inline
auto innerProductD2 (const iVector<l,N>& lhs,const iVector<r,N>& rhs) -> iScalar<decltype(innerProductD2(lhs._internal[0],rhs._internal[0]))>
{
typedef decltype(innerProductD2(lhs._internal[0],rhs._internal[0])) ret_t;
iScalar<ret_t> ret;
zeroit(ret);
for(int c1=0;c1<N;c1++){
ret._internal += innerProductD2(lhs._internal[c1],rhs._internal[c1]);
}
return ret;
}
template<class l,class r,int N> accelerator_inline
auto innerProductD2 (const iMatrix<l,N>& lhs,const iMatrix<r,N>& rhs) -> iScalar<decltype(innerProductD2(lhs._internal[0][0],rhs._internal[0][0]))>
{
typedef decltype(innerProductD2(lhs._internal[0][0],rhs._internal[0][0])) ret_t;
iScalar<ret_t> ret;
ret=Zero();
for(int c1=0;c1<N;c1++){
for(int c2=0;c2<N;c2++){
ret._internal+=innerProductD2(lhs._internal[c1][c2],rhs._internal[c1][c2]);
}}
return ret;
}
template<class l,class r> accelerator_inline
auto innerProductD2 (const iScalar<l>& lhs,const iScalar<r>& rhs) -> iScalar<decltype(innerProductD2(lhs._internal,rhs._internal))>
{
typedef decltype(innerProductD2(lhs._internal,rhs._internal)) ret_t;
iScalar<ret_t> ret;
ret._internal = innerProductD2(lhs._internal,rhs._internal);
return ret;
}
//////////////////////
// Keep same precison
//////////////////////

122
Grid/tensors/Tensor_traits.h
View File

@ -6,6 +6,7 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Christopher Kelly <ckelly@phys.columbia.edu>
Author: Michael Marshall <michael.marshall@ed.ac.au>
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
@ -37,6 +38,60 @@ NAMESPACE_BEGIN(Grid);
template<class T, int N> struct isGridTensor<iVector<T, N>> : public std::true_type { static constexpr bool notvalue = false; };
template<class T, int N> struct isGridTensor<iMatrix<T, N>> : public std::true_type { static constexpr bool notvalue = false; };
// Traits to identify scalars
template<typename T> struct isGridScalar : public std::false_type { static constexpr bool notvalue = true; };
template<class T> struct isGridScalar<iScalar<T>> : public std::true_type { static constexpr bool notvalue = false; };
// Store double-precision data in single-precision grids for precision promoted localInnerProductD
template<typename T>
class TypePair {
public:
T _internal[2];
TypePair<T>& operator=(const Grid::Zero& o) {
_internal[0] = Zero();
_internal[1] = Zero();
return *this;
}
TypePair<T> operator+(const TypePair<T>& o) const {
TypePair<T> r;
r._internal[0] = _internal[0] + o._internal[0];
r._internal[1] = _internal[1] + o._internal[1];
return r;
}
TypePair<T>& operator+=(const TypePair<T>& o) {
_internal[0] += o._internal[0];
_internal[1] += o._internal[1];
return *this;
}
friend accelerator_inline void add(TypePair<T>* ret, const TypePair<T>* a, const TypePair<T>* b) {
add(&ret->_internal[0],&a->_internal[0],&b->_internal[0]);
add(&ret->_internal[1],&a->_internal[1],&b->_internal[1]);
}
};
typedef TypePair<ComplexD> ComplexD2;
typedef TypePair<RealD> RealD2;
typedef TypePair<vComplexD> vComplexD2;
typedef TypePair<vRealD> vRealD2;
// Traits to identify fundamental data types
template<typename T> struct isGridFundamental : public std::false_type { static constexpr bool notvalue = true; };
template<> struct isGridFundamental<vComplexF> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<vComplexD> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<vRealF> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<vRealD> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<ComplexF> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<ComplexD> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<RealF> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<RealD> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<vComplexD2> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<vRealD2> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<ComplexD2> : public std::true_type { static constexpr bool notvalue = false; };
template<> struct isGridFundamental<RealD2> : public std::true_type { static constexpr bool notvalue = false; };
//////////////////////////////////////////////////////////////////////////////////
// Want to recurse: GridTypeMapper<Matrix<vComplexD> >::scalar_type == ComplexD.
// Use of a helper class like this allows us to template specialise and "dress"
@ -81,6 +136,7 @@ NAMESPACE_BEGIN(Grid);
typedef ComplexF Complexified;
typedef RealF Realified;
typedef RealD DoublePrecision;
typedef RealD2 DoublePrecision2;
};
template<> struct GridTypeMapper<RealD> : public GridTypeMapper_Base {
typedef RealD scalar_type;
@ -93,6 +149,20 @@ NAMESPACE_BEGIN(Grid);
typedef ComplexD Complexified;
typedef RealD Realified;
typedef RealD DoublePrecision;
typedef RealD DoublePrecision2;
};
template<> struct GridTypeMapper<RealD2> : public GridTypeMapper_Base {
typedef RealD2 scalar_type;
typedef RealD2 scalar_typeD;
typedef RealD2 vector_type;
typedef RealD2 vector_typeD;
typedef RealD2 tensor_reduced;
typedef RealD2 scalar_object;
typedef RealD2 scalar_objectD;
typedef ComplexD2 Complexified;
typedef RealD2 Realified;
typedef RealD2 DoublePrecision;
typedef RealD2 DoublePrecision2;
};
template<> struct GridTypeMapper<ComplexF> : public GridTypeMapper_Base {
typedef ComplexF scalar_type;
@ -105,6 +175,7 @@ NAMESPACE_BEGIN(Grid);
typedef ComplexF Complexified;
typedef RealF Realified;
typedef ComplexD DoublePrecision;
typedef ComplexD2 DoublePrecision2;
};
template<> struct GridTypeMapper<ComplexD> : public GridTypeMapper_Base {
typedef ComplexD scalar_type;
@ -117,6 +188,20 @@ NAMESPACE_BEGIN(Grid);
typedef ComplexD Complexified;
typedef RealD Realified;
typedef ComplexD DoublePrecision;
typedef ComplexD DoublePrecision2;
};
template<> struct GridTypeMapper<ComplexD2> : public GridTypeMapper_Base {
typedef ComplexD2 scalar_type;
typedef ComplexD2 scalar_typeD;
typedef ComplexD2 vector_type;
typedef ComplexD2 vector_typeD;
typedef ComplexD2 tensor_reduced;
typedef ComplexD2 scalar_object;
typedef ComplexD2 scalar_objectD;
typedef ComplexD2 Complexified;
typedef RealD2 Realified;
typedef ComplexD2 DoublePrecision;
typedef ComplexD2 DoublePrecision2;
};
template<> struct GridTypeMapper<Integer> : public GridTypeMapper_Base {
typedef Integer scalar_type;
@ -129,6 +214,7 @@ NAMESPACE_BEGIN(Grid);
typedef void Complexified;
typedef void Realified;
typedef void DoublePrecision;
typedef void DoublePrecision2;
};
template<> struct GridTypeMapper<vRealF> : public GridTypeMapper_Base {
@ -142,6 +228,7 @@ NAMESPACE_BEGIN(Grid);
typedef vComplexF Complexified;
typedef vRealF Realified;
typedef vRealD DoublePrecision;
typedef vRealD2 DoublePrecision2;
};
template<> struct GridTypeMapper<vRealD> : public GridTypeMapper_Base {
typedef RealD scalar_type;
@ -154,6 +241,20 @@ NAMESPACE_BEGIN(Grid);
typedef vComplexD Complexified;
typedef vRealD Realified;
typedef vRealD DoublePrecision;
typedef vRealD DoublePrecision2;
};
template<> struct GridTypeMapper<vRealD2> : public GridTypeMapper_Base {
typedef RealD2 scalar_type;
typedef RealD2 scalar_typeD;
typedef vRealD2 vector_type;
typedef vRealD2 vector_typeD;
typedef vRealD2 tensor_reduced;
typedef RealD2 scalar_object;
typedef RealD2 scalar_objectD;
typedef vComplexD2 Complexified;
typedef vRealD2 Realified;
typedef vRealD2 DoublePrecision;
typedef vRealD2 DoublePrecision2;
};
template<> struct GridTypeMapper<vRealH> : public GridTypeMapper_Base {
// Fixme this is incomplete until Grid supports fp16 or bfp16 arithmetic types
@ -167,6 +268,7 @@ NAMESPACE_BEGIN(Grid);
typedef vComplexH Complexified;
typedef vRealH Realified;
typedef vRealD DoublePrecision;
typedef vRealD DoublePrecision2;
};
template<> struct GridTypeMapper<vComplexH> : public GridTypeMapper_Base {
// Fixme this is incomplete until Grid supports fp16 or bfp16 arithmetic types
@ -180,6 +282,7 @@ NAMESPACE_BEGIN(Grid);
typedef vComplexH Complexified;
typedef vRealH Realified;
typedef vComplexD DoublePrecision;
typedef vComplexD DoublePrecision2;
};
template<> struct GridTypeMapper<vComplexF> : public GridTypeMapper_Base {
typedef ComplexF scalar_type;
@ -192,6 +295,7 @@ NAMESPACE_BEGIN(Grid);
typedef vComplexF Complexified;
typedef vRealF Realified;
typedef vComplexD DoublePrecision;
typedef vComplexD2 DoublePrecision2;
};
template<> struct GridTypeMapper<vComplexD> : public GridTypeMapper_Base {
typedef ComplexD scalar_type;
@ -204,6 +308,20 @@ NAMESPACE_BEGIN(Grid);
typedef vComplexD Complexified;
typedef vRealD Realified;
typedef vComplexD DoublePrecision;
typedef vComplexD DoublePrecision2;
};
template<> struct GridTypeMapper<vComplexD2> : public GridTypeMapper_Base {
typedef ComplexD2 scalar_type;
typedef ComplexD2 scalar_typeD;
typedef vComplexD2 vector_type;
typedef vComplexD2 vector_typeD;
typedef vComplexD2 tensor_reduced;
typedef ComplexD2 scalar_object;
typedef ComplexD2 scalar_objectD;
typedef vComplexD2 Complexified;
typedef vRealD2 Realified;
typedef vComplexD2 DoublePrecision;
typedef vComplexD2 DoublePrecision2;
};
template<> struct GridTypeMapper<vInteger> : public GridTypeMapper_Base {
typedef Integer scalar_type;
@ -216,6 +334,7 @@ NAMESPACE_BEGIN(Grid);
typedef void Complexified;
typedef void Realified;
typedef void DoublePrecision;
typedef void DoublePrecision2;
};
#define GridTypeMapper_RepeatedTypes \
@ -234,6 +353,7 @@ NAMESPACE_BEGIN(Grid);
using Complexified = iScalar<typename BaseTraits::Complexified>;
using Realified = iScalar<typename BaseTraits::Realified>;
using DoublePrecision = iScalar<typename BaseTraits::DoublePrecision>;
using DoublePrecision2= iScalar<typename BaseTraits::DoublePrecision2>;
static constexpr int Rank = BaseTraits::Rank + 1;
static constexpr std::size_t count = BaseTraits::count;
static constexpr int Dimension(int dim) {
@ -248,6 +368,7 @@ NAMESPACE_BEGIN(Grid);
using Complexified = iVector<typename BaseTraits::Complexified, N>;
using Realified = iVector<typename BaseTraits::Realified, N>;
using DoublePrecision = iVector<typename BaseTraits::DoublePrecision, N>;
using DoublePrecision2= iVector<typename BaseTraits::DoublePrecision2, N>;
static constexpr int Rank = BaseTraits::Rank + 1;
static constexpr std::size_t count = BaseTraits::count * N;
static constexpr int Dimension(int dim) {
@ -262,6 +383,7 @@ NAMESPACE_BEGIN(Grid);
using Complexified = iMatrix<typename BaseTraits::Complexified, N>;
using Realified = iMatrix<typename BaseTraits::Realified, N>;
using DoublePrecision = iMatrix<typename BaseTraits::DoublePrecision, N>;
using DoublePrecision2= iMatrix<typename BaseTraits::DoublePrecision2, N>;
static constexpr int Rank = BaseTraits::Rank + 2;
static constexpr std::size_t count = BaseTraits::count * N * N;
static constexpr int Dimension(int dim) {

1
Grid/util/Init.h
View File

@ -56,6 +56,7 @@ std::string GridCmdVectorIntToString(const VectorInt & vec);
void GridCmdOptionCSL(std::string str,std::vector<std::string> & vec);
template<class VectorInt>
void GridCmdOptionIntVector(std::string &str,VectorInt & vec);
void GridCmdOptionInt(std::string &str,int & val);
void GridParseLayout(char **argv,int argc,