portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

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

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

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

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

26
Grid/DisableWarnings.h
View File

@ -30,8 +30,34 @@ directory
#ifndef DISABLE_WARNINGS_H
#define DISABLE_WARNINGS_H
#if defined __GNUC__ && __GNUC__>=6
#pragma GCC diagnostic ignored "-Wignored-attributes"
#endif
//disables and intel compiler specific warning (in json.hpp)
#pragma warning disable 488
#ifdef __NVCC__
//disables nvcc specific warning in json.hpp
#pragma clang diagnostic ignored "-Wdeprecated-register"
#pragma diag_suppress unsigned_compare_with_zero
#pragma diag_suppress cast_to_qualified_type
//disables nvcc specific warning in many files
#pragma diag_suppress esa_on_defaulted_function_ignored
#pragma diag_suppress extra_semicolon
//Eigen only
#endif
// Disable vectorisation in Eigen on the Power8/9 and PowerPC
#ifdef __ALTIVEC__
#define EIGEN_DONT_VECTORIZE
#endif
#ifdef __VSX__
#define EIGEN_DONT_VECTORIZE
#endif
#endif

10
Grid/GridCore.h
View File

@ -38,16 +38,19 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_BASE_H
#define GRID_BASE_H
#include <Grid/GridStd.h>
#include <Grid/DisableWarnings.h>
#include <Grid/Namespace.h>
#include <Grid/GridStd.h>
#include <Grid/threads/Pragmas.h>
#include <Grid/perfmon/Timer.h>
#include <Grid/perfmon/PerfCount.h>
#include <Grid/util/Util.h>
#include <Grid/log/Log.h>
#include <Grid/allocator/AlignedAllocator.h>
#include <Grid/simd/Simd.h>
#include <Grid/serialisation/Serialisation.h>
#include <Grid/threads/Threads.h>
#include <Grid/util/Util.h>
#include <Grid/serialisation/Serialisation.h>
#include <Grid/util/Sha.h>
#include <Grid/communicator/Communicator.h>
#include <Grid/cartesian/Cartesian.h>
@ -57,5 +60,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <Grid/stencil/Stencil.h>
#include <Grid/parallelIO/BinaryIO.h>
#include <Grid/algorithms/Algorithms.h>
NAMESPACE_CHECK(GridCore)
#endif

1
Grid/GridQCDcore.h
View File

@ -38,5 +38,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <Grid/qcd/spin/Spin.h>
#include <Grid/qcd/utils/Utils.h>
#include <Grid/qcd/representations/Representations.h>
NAMESPACE_CHECK(GridQCDCore);
#endif

1
Grid/GridStd.h
View File

@ -7,6 +7,7 @@
#include <cassert>
#include <complex>
#include <vector>
#include <array>
#include <string>
#include <iostream>
#include <iomanip>

27
Grid/Grid_Eigen_Dense.h
View File

@ -1,14 +1,41 @@
#include <Grid/GridCore.h>
#pragma once
// Force Eigen to use MKL if Grid has been configured with --enable-mkl
#ifdef USE_MKL
#define EIGEN_USE_MKL_ALL
#endif
#if defined __GNUC__
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wdeprecated-declarations"
#endif
/* NVCC save and restore compile environment*/
#ifdef __NVCC__
#pragma push
#pragma diag_suppress code_is_unreachable
#pragma push_macro("__CUDA_ARCH__")
#pragma push_macro("__NVCC__")
#pragma push_macro("__CUDACC__")
#undef __NVCC__
#undef __CUDACC__
#undef __CUDA_ARCH__
#define __NVCC__REDEFINE__
#endif
#include <Grid/Eigen/Dense>
#include <Grid/Eigen/unsupported/CXX11/Tensor>
/* NVCC restore */
#ifdef __NVCC__REDEFINE__
#pragma pop_macro("__CUDACC__")
#pragma pop_macro("__NVCC__")
#pragma pop_macro("__CUDA_ARCH__")
#pragma pop
#endif
#if defined __GNUC__
#pragma GCC diagnostic pop
#endif

1
Grid/Grid_Eigen_Tensor.h Normal file
View File

@ -0,0 +1 @@
#include <Grid/Grid_Eigen_Dense.h>

38
Grid/Namespace.h Normal file
View File

@ -0,0 +1,38 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/Namespace.h
Copyright (C) 2016
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#pragma once
#include <type_traits>
#include <cassert>
#define NAMESPACE_BEGIN(A) namespace A {
#define NAMESPACE_END(A) }
#define GRID_NAMESPACE_BEGIN NAMESPACE_BEGIN(Grid)
#define GRID_NAMESPACE_END NAMESPACE_END(Grid)
#define NAMESPACE_CHECK(x) struct namespaceTEST##x {}; static_assert(std::is_same<namespaceTEST##x, ::namespaceTEST##x>::value,"Not in :: at" );

780
Grid/algorithms/CoarsenedMatrix.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -26,501 +26,487 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_ALGORITHM_COARSENED_MATRIX_H
#define GRID_ALGORITHM_COARSENED_MATRIX_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
class Geometry {
// int dimension;
public:
int npoint;
std::vector<int> directions ;
std::vector<int> displacements;
class Geometry {
// int dimension;
public:
int npoint;
std::vector<int> directions ;
std::vector<int> displacements;
Geometry(int _d) {
int base = (_d==5) ? 1:0;
int base = (_d==5) ? 1:0;
// make coarse grid stencil for 4d , not 5d
if ( _d==5 ) _d=4;
// make coarse grid stencil for 4d , not 5d
if ( _d==5 ) _d=4;
npoint = 2*_d+1;
directions.resize(npoint);
displacements.resize(npoint);
for(int d=0;d<_d;d++){
directions[2*d ] = d+base;
directions[2*d+1] = d+base;
displacements[2*d ] = +1;
displacements[2*d+1] = -1;
}
directions [2*_d]=0;
displacements[2*_d]=0;
npoint = 2*_d+1;
directions.resize(npoint);
displacements.resize(npoint);
for(int d=0;d<_d;d++){
directions[2*d ] = d+base;
directions[2*d+1] = d+base;
displacements[2*d ] = +1;
displacements[2*d+1] = -1;
}
directions [2*_d]=0;
displacements[2*_d]=0;
//// report back
std::cout<<GridLogMessage<<"directions :";
for(int d=0;d<npoint;d++) std::cout<< directions[d]<< " ";
std::cout <<std::endl;
std::cout<<GridLogMessage<<"displacements :";
for(int d=0;d<npoint;d++) std::cout<< displacements[d]<< " ";
std::cout<<std::endl;
}
//// report back
std::cout<<GridLogMessage<<"directions :";
for(int d=0;d<npoint;d++) std::cout<< directions[d]<< " ";
std::cout <<std::endl;
std::cout<<GridLogMessage<<"displacements :";
for(int d=0;d<npoint;d++) std::cout<< displacements[d]<< " ";
std::cout<<std::endl;
}
/*
// Original cleaner code
Geometry(int _d) : dimension(_d), npoint(2*_d+1), directions(npoint), displacements(npoint) {
for(int d=0;d<dimension;d++){
directions[2*d ] = d;
directions[2*d+1] = d;
displacements[2*d ] = +1;
displacements[2*d+1] = -1;
}
directions [2*dimension]=0;
displacements[2*dimension]=0;
}
std::vector<int> GetDelta(int point) {
std::vector<int> delta(dimension,0);
delta[directions[point]] = displacements[point];
return delta;
};
*/
/*
// Original cleaner code
Geometry(int _d) : dimension(_d), npoint(2*_d+1), directions(npoint), displacements(npoint) {
for(int d=0;d<dimension;d++){
directions[2*d ] = d;
directions[2*d+1] = d;
displacements[2*d ] = +1;
displacements[2*d+1] = -1;
}
directions [2*dimension]=0;
displacements[2*dimension]=0;
}
std::vector<int> GetDelta(int point) {
std::vector<int> delta(dimension,0);
delta[directions[point]] = displacements[point];
return delta;
};
*/
};
template<class Fobj,class CComplex,int nbasis>
class Aggregation {
public:
typedef iVector<CComplex,nbasis > siteVector;
typedef Lattice<siteVector> CoarseVector;
typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
template<class Fobj,class CComplex,int nbasis>
class Aggregation {
public:
typedef iVector<CComplex,nbasis > siteVector;
typedef Lattice<siteVector> CoarseVector;
typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
typedef Lattice< CComplex > CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj > FineField;
GridBase *CoarseGrid;
GridBase *FineGrid;
std::vector<Lattice<Fobj> > subspace;
int checkerboard;
typedef Lattice< CComplex > CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj > FineField;
GridBase *CoarseGrid;
GridBase *FineGrid;
std::vector<Lattice<Fobj> > subspace;
int checkerboard;
int Checkerboard(void){return checkerboard;}
Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) :
CoarseGrid(_CoarseGrid),
FineGrid(_FineGrid),
subspace(nbasis,_FineGrid),
checkerboard(_checkerboard)
{
};
FineGrid(_FineGrid),
subspace(nbasis,_FineGrid),
checkerboard(_checkerboard)
{
};
void Orthogonalise(void){
CoarseScalar InnerProd(CoarseGrid);
std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
blockOrthogonalise(InnerProd,subspace);
std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
blockOrthogonalise(InnerProd,subspace);
// std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
// CheckOrthogonal();
}
void CheckOrthogonal(void){
CoarseVector iProj(CoarseGrid);
CoarseVector eProj(CoarseGrid);
for(int i=0;i<nbasis;i++){
blockProject(iProj,subspace[i],subspace);
eProj=zero;
parallel_for(int ss=0;ss<CoarseGrid->oSites();ss++){
eProj._odata[ss](i)=CComplex(1.0);
}
eProj=eProj - iProj;
std::cout<<GridLogMessage<<"Orthog check error "<<i<<" " << norm2(eProj)<<std::endl;
}
std::cout<<GridLogMessage <<"CheckOrthog done"<<std::endl;
void Orthogonalise(void){
CoarseScalar InnerProd(CoarseGrid);
std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
blockOrthogonalise(InnerProd,subspace);
std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
blockOrthogonalise(InnerProd,subspace);
// std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
// CheckOrthogonal();
}
void CheckOrthogonal(void){
CoarseVector iProj(CoarseGrid);
CoarseVector eProj(CoarseGrid);
for(int i=0;i<nbasis;i++){
blockProject(iProj,subspace[i],subspace);
eProj=Zero();
thread_for(ss, CoarseGrid->oSites(),{
eProj[ss](i)=CComplex(1.0);
});
eProj=eProj - iProj;
std::cout<<GridLogMessage<<"Orthog check error "<<i<<" " << norm2(eProj)<<std::endl;
}
void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
blockProject(CoarseVec,FineVec,subspace);
}
void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
FineVec.checkerboard = subspace[0].checkerboard;
blockPromote(CoarseVec,FineVec,subspace);
}
void CreateSubspaceRandom(GridParallelRNG &RNG){
for(int i=0;i<nbasis;i++){
random(RNG,subspace[i]);
std::cout<<GridLogMessage<<" norm subspace["<<i<<"] "<<norm2(subspace[i])<<std::endl;
}
Orthogonalise();
std::cout<<GridLogMessage <<"CheckOrthog done"<<std::endl;
}
void ProjectToSubspace(CoarseVector &CoarseVec,const FineField &FineVec){
blockProject(CoarseVec,FineVec,subspace);
}
void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
FineVec.Checkerboard() = subspace[0].Checkerboard();
blockPromote(CoarseVec,FineVec,subspace);
}
void CreateSubspaceRandom(GridParallelRNG &RNG){
for(int i=0;i<nbasis;i++){
random(RNG,subspace[i]);
std::cout<<GridLogMessage<<" norm subspace["<<i<<"] "<<norm2(subspace[i])<<std::endl;
}
Orthogonalise();
}
/*
/*
virtual void CreateSubspaceLanczos(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis)
{
// Run a Lanczos with sloppy convergence
const int Nstop = nn;
const int Nk = nn+20;
const int Np = nn+20;
const int Nm = Nk+Np;
const int MaxIt= 10000;
RealD resid = 1.0e-3;
// Run a Lanczos with sloppy convergence
const int Nstop = nn;
const int Nk = nn+20;
const int Np = nn+20;
const int Nm = Nk+Np;
const int MaxIt= 10000;
RealD resid = 1.0e-3;
Chebyshev<FineField> Cheb(0.5,64.0,21);
ImplicitlyRestartedLanczos<FineField> IRL(hermop,Cheb,Nstop,Nk,Nm,resid,MaxIt);
// IRL.lock = 1;
Chebyshev<FineField> Cheb(0.5,64.0,21);
ImplicitlyRestartedLanczos<FineField> IRL(hermop,Cheb,Nstop,Nk,Nm,resid,MaxIt);
// IRL.lock = 1;
FineField noise(FineGrid); gaussian(RNG,noise);
FineField tmp(FineGrid);
std::vector<RealD> eval(Nm);
std::vector<FineField> evec(Nm,FineGrid);
FineField noise(FineGrid); gaussian(RNG,noise);
FineField tmp(FineGrid);
std::vector<RealD> eval(Nm);
std::vector<FineField> evec(Nm,FineGrid);
int Nconv;
IRL.calc(eval,evec,
noise,
Nconv);
int Nconv;
IRL.calc(eval,evec,
noise,
Nconv);
// pull back nn vectors
for(int b=0;b<nn;b++){
// pull back nn vectors
for(int b=0;b<nn;b++){
subspace[b] = evec[b];
subspace[b] = evec[b];
std::cout << GridLogMessage <<"subspace["<<b<<"] = "<<norm2(subspace[b])<<std::endl;
std::cout << GridLogMessage <<"subspace["<<b<<"] = "<<norm2(subspace[b])<<std::endl;
hermop.Op(subspace[b],tmp);
std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(tmp)<<std::endl;
hermop.Op(subspace[b],tmp);
std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(tmp)<<std::endl;
noise = tmp - sqrt(eval[b])*subspace[b] ;
noise = tmp - sqrt(eval[b])*subspace[b] ;
std::cout<<GridLogMessage << " lambda_"<<b<<" = "<< eval[b] <<" ; [ M - Lambda ]_"<<b<<" vec_"<<b<<" = " <<norm2(noise)<<std::endl;
std::cout<<GridLogMessage << " lambda_"<<b<<" = "<< eval[b] <<" ; [ M - Lambda ]_"<<b<<" vec_"<<b<<" = " <<norm2(noise)<<std::endl;
noise = tmp + eval[b]*subspace[b] ;
noise = tmp + eval[b]*subspace[b] ;
std::cout<<GridLogMessage << " lambda_"<<b<<" = "<< eval[b] <<" ; [ M - Lambda ]_"<<b<<" vec_"<<b<<" = " <<norm2(noise)<<std::endl;
std::cout<<GridLogMessage << " lambda_"<<b<<" = "<< eval[b] <<" ; [ M - Lambda ]_"<<b<<" vec_"<<b<<" = " <<norm2(noise)<<std::endl;
}
Orthogonalise();
for(int b=0;b<nn;b++){
std::cout << GridLogMessage <<"subspace["<<b<<"] = "<<norm2(subspace[b])<<std::endl;
}
}
*/
virtual void CreateSubspace(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
Orthogonalise();
for(int b=0;b<nn;b++){
std::cout << GridLogMessage <<"subspace["<<b<<"] = "<<norm2(subspace[b])<<std::endl;
}
}
*/
virtual void CreateSubspace(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
RealD scale;
RealD scale;
ConjugateGradient<FineField> CG(1.0e-2,10000);
FineField noise(FineGrid);
FineField Mn(FineGrid);
ConjugateGradient<FineField> CG(1.0e-2,10000);
FineField noise(FineGrid);
FineField Mn(FineGrid);
for(int b=0;b<nn;b++){
for(int b=0;b<nn;b++){
subspace[b] = zero;
gaussian(RNG,noise);
subspace[b] = Zero();
gaussian(RNG,noise);
scale = std::pow(norm2(noise),-0.5);
noise=noise*scale;
hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
for(int i=0;i<1;i++){
CG(hermop,noise,subspace[b]);
noise = subspace[b];
scale = std::pow(norm2(noise),-0.5);
noise=noise*scale;
hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
for(int i=0;i<1;i++){
CG(hermop,noise,subspace[b]);
noise = subspace[b];
scale = std::pow(norm2(noise),-0.5);
noise=noise*scale;
}
hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
subspace[b] = noise;
}
Orthogonalise();
hermop.Op(noise,Mn); std::cout<<GridLogMessage << "filtered["<<b<<"] <f|MdagM|f> "<<norm2(Mn)<<std::endl;
subspace[b] = noise;
}
};
// Fine Object == (per site) type of fine field
// nbasis == number of deflation vectors
template<class Fobj,class CComplex,int nbasis>
class CoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > > {
public:
Orthogonalise();
}
};
// Fine Object == (per site) type of fine field
// nbasis == number of deflation vectors
template<class Fobj,class CComplex,int nbasis>
class CoarsenedMatrix : public SparseMatrixBase<Lattice<iVector<CComplex,nbasis > > > {
public:
typedef iVector<CComplex,nbasis > siteVector;
typedef Lattice<siteVector> CoarseVector;
typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
typedef iVector<CComplex,nbasis > siteVector;
typedef Lattice<siteVector> CoarseVector;
typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
typedef Lattice< CComplex > CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj > FineField;
typedef Lattice< CComplex > CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj > FineField;
////////////////////
// Data members
////////////////////
Geometry geom;
GridBase * _grid;
CartesianStencil<siteVector,siteVector> Stencil;
////////////////////
// Data members
////////////////////
Geometry geom;
GridBase * _grid;
std::vector<CoarseMatrix> A;
CartesianStencil<siteVector,siteVector,int> Stencil;
std::vector<CoarseMatrix> A;
///////////////////////
// Interface
///////////////////////
GridBase * Grid(void) { return _grid; }; // this is all the linalg routines need to know
///////////////////////
// Interface
///////////////////////
GridBase * Grid(void) { return _grid; }; // this is all the linalg routines need to know
RealD M (const CoarseVector &in, CoarseVector &out){
RealD M (const CoarseVector &in, CoarseVector &out){
conformable(_grid,in._grid);
conformable(in._grid,out._grid);
conformable(_grid,in.Grid());
conformable(in.Grid(),out.Grid());
SimpleCompressor<siteVector> compressor;
Stencil.HaloExchange(in,compressor);
SimpleCompressor<siteVector> compressor;
Stencil.HaloExchange(in,compressor);
auto in_v = in.View();
auto out_v = in.View();
thread_for(ss,Grid()->oSites(),{
siteVector res = Zero();
siteVector nbr;
int ptype;
StencilEntry *SE;
for(int point=0;point<geom.npoint;point++){
parallel_for(int ss=0;ss<Grid()->oSites();ss++){
siteVector res = zero;
siteVector nbr;
int ptype;
StencilEntry *SE;
for(int point=0;point<geom.npoint;point++){
SE=Stencil.GetEntry(ptype,point,ss);
SE=Stencil.GetEntry(ptype,point,ss);
if(SE->_is_local&&SE->_permute) {
permute(nbr,in._odata[SE->_offset],ptype);
} else if(SE->_is_local) {
nbr = in._odata[SE->_offset];
} else {
nbr = Stencil.CommBuf()[SE->_offset];
}
res = res + A[point]._odata[ss]*nbr;
if(SE->_is_local&&SE->_permute) {
permute(nbr,in_v[SE->_offset],ptype);
} else if(SE->_is_local) {
nbr = in_v[SE->_offset];
} else {
nbr = Stencil.CommBuf()[SE->_offset];
}
vstream(out._odata[ss],res);
auto A_point = A[point].View();
res = res + A_point[ss]*nbr;
}
return norm2(out);
};
vstream(out_v[ss],res);
});
return norm2(out);
};
RealD Mdag (const CoarseVector &in, CoarseVector &out){
// // corresponds to Petrov-Galerkin coarsening
// return M(in,out);
RealD Mdag (const CoarseVector &in, CoarseVector &out){
// // corresponds to Petrov-Galerkin coarsening
// return M(in,out);
// corresponds to Galerkin coarsening
CoarseVector tmp(Grid());
G5C(tmp, in);
M(tmp, out);
G5C(out, out);
return norm2(out);
};
// corresponds to Galerkin coarsening
CoarseVector tmp(Grid());
G5C(tmp, in);
M(tmp, out);
G5C(out, out);
return norm2(out);
};
void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
conformable(_grid,in.Grid());
conformable(in.Grid(),out.Grid());
SimpleCompressor<siteVector> compressor;
Stencil.HaloExchange(in,compressor);
auto point = [dir, disp](){
if(dir == 0 and disp == 0)
return 8;
else
return (4 * dir + 1 - disp) / 2;
}();
void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
conformable(_grid,in._grid);
conformable(in._grid,out._grid);
SimpleCompressor<siteVector> compressor;
Stencil.HaloExchange(in,compressor);
auto point = [dir, disp](){
if(dir == 0 and disp == 0)
return 8;
else
return (4 * dir + 1 - disp) / 2;
}();
parallel_for(int ss=0;ss<Grid()->oSites();ss++){
siteVector res = zero;
siteVector nbr;
int ptype;
StencilEntry *SE;
SE=Stencil.GetEntry(ptype,point,ss);
if(SE->_is_local&&SE->_permute) {
permute(nbr,in._odata[SE->_offset],ptype);
} else if(SE->_is_local) {
nbr = in._odata[SE->_offset];
} else {
nbr = Stencil.CommBuf()[SE->_offset];
}
res = res + A[point]._odata[ss]*nbr;
vstream(out._odata[ss],res);
auto out_v = out.View();
auto in_v = in.View();
thread_for(ss,Grid()->oSites(),{
siteVector res = Zero();
siteVector nbr;
int ptype;
StencilEntry *SE;
SE=Stencil.GetEntry(ptype,point,ss);
if(SE->_is_local&&SE->_permute) {
permute(nbr,in_v[SE->_offset],ptype);
} else if(SE->_is_local) {
nbr = in_v[SE->_offset];
} else {
nbr = Stencil.CommBuf()[SE->_offset];
}
};
void Mdiag(const CoarseVector &in, CoarseVector &out){
Mdir(in, out, 0, 0); // use the self coupling (= last) point of the stencil
};
auto A_point = A[point].View();
res = res + A_point[ss]*nbr;
vstream(out_v[ss],res);
});
};
CoarsenedMatrix(GridCartesian &CoarseGrid) :
void Mdiag(const CoarseVector &in, CoarseVector &out){
Mdir(in, out, 0, 0); // use the self coupling (= last) point of the stencil
};
_grid(&CoarseGrid),
geom(CoarseGrid._ndimension),
Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements),
A(geom.npoint,&CoarseGrid)
{
};
CoarsenedMatrix(GridCartesian &CoarseGrid) :
void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
Aggregation<Fobj,CComplex,nbasis> & Subspace){
_grid(&CoarseGrid),
geom(CoarseGrid._ndimension),
Stencil(&CoarseGrid,geom.npoint,Even,geom.directions,geom.displacements,0),
A(geom.npoint,&CoarseGrid)
{
};
FineField iblock(FineGrid); // contributions from within this block
FineField oblock(FineGrid); // contributions from outwith this block
void CoarsenOperator(GridBase *FineGrid,LinearOperatorBase<Lattice<Fobj> > &linop,
Aggregation<Fobj,CComplex,nbasis> & Subspace){
FineField phi(FineGrid);
FineField tmp(FineGrid);
FineField zz(FineGrid); zz=zero;
FineField Mphi(FineGrid);
FineField iblock(FineGrid); // contributions from within this block
FineField oblock(FineGrid); // contributions from outwith this block
Lattice<iScalar<vInteger> > coor(FineGrid);
FineField phi(FineGrid);
FineField tmp(FineGrid);
FineField zz(FineGrid); zz=Zero();
FineField Mphi(FineGrid);
CoarseVector iProj(Grid());
CoarseVector oProj(Grid());
CoarseScalar InnerProd(Grid());
Lattice<iScalar<vInteger> > coor(FineGrid);
// Orthogonalise the subblocks over the basis
blockOrthogonalise(InnerProd,Subspace.subspace);
CoarseVector iProj(Grid());
CoarseVector oProj(Grid());
CoarseScalar InnerProd(Grid());
// Compute the matrix elements of linop between this orthonormal
// set of vectors.
int self_stencil=-1;
for(int p=0;p<geom.npoint;p++){
A[p]=zero;
if( geom.displacements[p]==0){
self_stencil=p;
}
// Orthogonalise the subblocks over the basis
blockOrthogonalise(InnerProd,Subspace.subspace);
// Compute the matrix elements of linop between this orthonormal
// set of vectors.
int self_stencil=-1;
for(int p=0;p<geom.npoint;p++){
A[p]=Zero();
if( geom.displacements[p]==0){
self_stencil=p;
}
assert(self_stencil!=-1);
}
assert(self_stencil!=-1);
for(int i=0;i<nbasis;i++){
phi=Subspace.subspace[i];
for(int i=0;i<nbasis;i++){
phi=Subspace.subspace[i];
std::cout<<GridLogMessage<<"("<<i<<").."<<std::endl;
std::cout<<GridLogMessage<<"("<<i<<").."<<std::endl;
for(int p=0;p<geom.npoint;p++){
for(int p=0;p<geom.npoint;p++){
int dir = geom.directions[p];
int disp = geom.displacements[p];
int dir = geom.directions[p];
int disp = geom.displacements[p];
Integer block=(FineGrid->_rdimensions[dir])/(Grid()->_rdimensions[dir]);
Integer block=(FineGrid->_rdimensions[dir])/(Grid()->_rdimensions[dir]);
LatticeCoordinate(coor,dir);
LatticeCoordinate(coor,dir);
if ( disp==0 ){
linop.OpDiag(phi,Mphi);
}
else {
linop.OpDir(phi,Mphi,dir,disp);
}
////////////////////////////////////////////////////////////////////////
// Pick out contributions coming from this cell and neighbour cell
////////////////////////////////////////////////////////////////////////
if ( disp==0 ) {
iblock = Mphi;
oblock = zero;
} else if ( disp==1 ) {
oblock = where(mod(coor,block)==(block-1),Mphi,zz);
iblock = where(mod(coor,block)!=(block-1),Mphi,zz);
} else if ( disp==-1 ) {
oblock = where(mod(coor,block)==(Integer)0,Mphi,zz);
iblock = where(mod(coor,block)!=(Integer)0,Mphi,zz);
} else {
assert(0);
}
Subspace.ProjectToSubspace(iProj,iblock);
Subspace.ProjectToSubspace(oProj,oblock);
// blockProject(iProj,iblock,Subspace.subspace);
// blockProject(oProj,oblock,Subspace.subspace);
parallel_for(int ss=0;ss<Grid()->oSites();ss++){
for(int j=0;j<nbasis;j++){
if( disp!= 0 ) {
A[p]._odata[ss](j,i) = oProj._odata[ss](j);
}
A[self_stencil]._odata[ss](j,i) = A[self_stencil]._odata[ss](j,i) + iProj._odata[ss](j);
}
}
if ( disp==0 ){
linop.OpDiag(phi,Mphi);
}
else {
linop.OpDir(phi,Mphi,dir,disp);
}
////////////////////////////////////////////////////////////////////////
// Pick out contributions coming from this cell and neighbour cell
////////////////////////////////////////////////////////////////////////
if ( disp==0 ) {
iblock = Mphi;
oblock = Zero();
} else if ( disp==1 ) {
oblock = where(mod(coor,block)==(block-1),Mphi,zz);
iblock = where(mod(coor,block)!=(block-1),Mphi,zz);
} else if ( disp==-1 ) {
oblock = where(mod(coor,block)==(Integer)0,Mphi,zz);
iblock = where(mod(coor,block)!=(Integer)0,Mphi,zz);
} else {
assert(0);
}
Subspace.ProjectToSubspace(iProj,iblock);
Subspace.ProjectToSubspace(oProj,oblock);
// blockProject(iProj,iblock,Subspace.subspace);
// blockProject(oProj,oblock,Subspace.subspace);
auto iProj_v = iProj.View() ;
auto oProj_v = oProj.View() ;
auto A_p = A[p].View();
auto A_self = A[self_stencil].View();
thread_for(ss, Grid()->oSites(),{
for(int j=0;j<nbasis;j++){
if( disp!= 0 ) {
A_p[ss](j,i) = oProj_v[ss](j);
}
A_self[ss](j,i) = A_self[ss](j,i) + iProj_v[ss](j);
}
});
}
}
#if 0
///////////////////////////
// test code worth preserving in if block
///////////////////////////
std::cout<<GridLogMessage<< " Computed matrix elements "<< self_stencil <<std::endl;
for(int p=0;p<geom.npoint;p++){
std::cout<<GridLogMessage<< "A["<<p<<"]" << std::endl;
std::cout<<GridLogMessage<< A[p] << std::endl;
}
std::cout<<GridLogMessage<< " picking by block0 "<< self_stencil <<std::endl;
///////////////////////////
// test code worth preserving in if block
///////////////////////////
std::cout<<GridLogMessage<< " Computed matrix elements "<< self_stencil <<std::endl;
for(int p=0;p<geom.npoint;p++){
std::cout<<GridLogMessage<< "A["<<p<<"]" << std::endl;
std::cout<<GridLogMessage<< A[p] << std::endl;
}
std::cout<<GridLogMessage<< " picking by block0 "<< self_stencil <<std::endl;
phi=Subspace.subspace[0];
std::vector<int> bc(FineGrid->_ndimension,0);
phi=Subspace.subspace[0];
std::vector<int> bc(FineGrid->_ndimension,0);
blockPick(Grid(),phi,tmp,bc); // Pick out a block
linop.Op(tmp,Mphi); // Apply big dop
blockProject(iProj,Mphi,Subspace.subspace); // project it and print it
std::cout<<GridLogMessage<< " Computed matrix elements from block zero only "<<std::endl;
std::cout<<GridLogMessage<< iProj <<std::endl;
std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
blockPick(Grid(),phi,tmp,bc); // Pick out a block
linop.Op(tmp,Mphi); // Apply big dop
blockProject(iProj,Mphi,Subspace.subspace); // project it and print it
std::cout<<GridLogMessage<< " Computed matrix elements from block zero only "<<std::endl;
std::cout<<GridLogMessage<< iProj <<std::endl;
std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
#endif
// ForceHermitian();
// AssertHermitian();
// ForceDiagonal();
}
void ForceHermitian(void) {
for(int d=0;d<4;d++){
int dd=d+1;
A[2*d] = adj(Cshift(A[2*d+1],dd,1));
}
void ForceDiagonal(void) {
std::cout<<GridLogMessage<<"**************************************************"<<std::endl;
std::cout<<GridLogMessage<<"**** Forcing coarse operator to be diagonal ****"<<std::endl;
std::cout<<GridLogMessage<<"**************************************************"<<std::endl;
for(int p=0;p<8;p++){
A[p]=zero;
}
GridParallelRNG RNG(Grid()); RNG.SeedFixedIntegers(std::vector<int>({55,72,19,17,34}));
Lattice<iScalar<CComplex> > val(Grid()); random(RNG,val);
Complex one(1.0);
iMatrix<CComplex,nbasis> ident; ident=one;
val = val*adj(val);
val = val + 1.0;
A[8] = val*ident;
// for(int s=0;s<Grid()->oSites();s++) {
// A[8]._odata[s]=val._odata[s];
// }
}
void ForceHermitian(void) {
for(int d=0;d<4;d++){
int dd=d+1;
A[2*d] = adj(Cshift(A[2*d+1],dd,1));
}
// A[8] = 0.5*(A[8] + adj(A[8]));
}
void AssertHermitian(void) {
CoarseMatrix AA (Grid());
CoarseMatrix AAc (Grid());
CoarseMatrix Diff (Grid());
for(int d=0;d<4;d++){
// A[8] = 0.5*(A[8] + adj(A[8]));
}
void AssertHermitian(void) {
CoarseMatrix AA (Grid());
CoarseMatrix AAc (Grid());
CoarseMatrix Diff (Grid());
for(int d=0;d<4;d++){
int dd=d+1;
AAc = Cshift(A[2*d+1],dd,1);
AA = A[2*d];
int dd=d+1;
AAc = Cshift(A[2*d+1],dd,1);
AA = A[2*d];
Diff = AA - adj(AAc);
Diff = AA - adj(AAc);
std::cout<<GridLogMessage<<"Norm diff dim "<<d<<" "<< norm2(Diff)<<std::endl;
std::cout<<GridLogMessage<<"Norm dim "<<d<<" "<< norm2(AA)<<std::endl;
std::cout<<GridLogMessage<<"Norm diff dim "<<d<<" "<< norm2(Diff)<<std::endl;
std::cout<<GridLogMessage<<"Norm dim "<<d<<" "<< norm2(AA)<<std::endl;
}
Diff = A[8] - adj(A[8]);
std::cout<<GridLogMessage<<"Norm diff local "<< norm2(Diff)<<std::endl;
std::cout<<GridLogMessage<<"Norm local "<< norm2(A[8])<<std::endl;
}
Diff = A[8] - adj(A[8]);
std::cout<<GridLogMessage<<"Norm diff local "<< norm2(Diff)<<std::endl;
std::cout<<GridLogMessage<<"Norm local "<< norm2(A[8])<<std::endl;
}
};
};
}
NAMESPACE_END(Grid);
#endif

453
Grid/algorithms/FFT.h
View File

@ -1,5 +1,5 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,8 +24,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef _GRID_FFT_H_
#define _GRID_FFT_H_
@ -38,64 +38,64 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#endif
namespace Grid {
NAMESPACE_BEGIN(Grid);
template<class scalar> struct FFTW { };
template<class scalar> struct FFTW { };
#ifdef HAVE_FFTW
template<> struct FFTW<ComplexD> {
public:
template<> struct FFTW<ComplexD> {
public:
typedef fftw_complex FFTW_scalar;
typedef fftw_plan FFTW_plan;
typedef fftw_complex FFTW_scalar;
typedef fftw_plan FFTW_plan;
static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
FFTW_scalar *in, const int *inembed,
int istride, int idist,
FFTW_scalar *out, const int *onembed,
int ostride, int odist,
int sign, unsigned flags) {
return ::fftw_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
}
static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
FFTW_scalar *in, const int *inembed,
int istride, int idist,
FFTW_scalar *out, const int *onembed,
int ostride, int odist,
int sign, unsigned flags) {
return ::fftw_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
}
static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
::fftw_flops(p,add,mul,fmas);
}
static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
::fftw_flops(p,add,mul,fmas);
}
inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
::fftw_execute_dft(p,in,out);
}
inline static void fftw_destroy_plan(const FFTW_plan p) {
::fftw_destroy_plan(p);
}
};
inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
::fftw_execute_dft(p,in,out);
}
inline static void fftw_destroy_plan(const FFTW_plan p) {
::fftw_destroy_plan(p);
}
};
template<> struct FFTW<ComplexF> {
public:
template<> struct FFTW<ComplexF> {
public:
typedef fftwf_complex FFTW_scalar;
typedef fftwf_plan FFTW_plan;
typedef fftwf_complex FFTW_scalar;
typedef fftwf_plan FFTW_plan;
static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
FFTW_scalar *in, const int *inembed,
int istride, int idist,
FFTW_scalar *out, const int *onembed,
int ostride, int odist,
int sign, unsigned flags) {
return ::fftwf_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
}
static FFTW_plan fftw_plan_many_dft(int rank, const int *n,int howmany,
FFTW_scalar *in, const int *inembed,
int istride, int idist,
FFTW_scalar *out, const int *onembed,
int ostride, int odist,
int sign, unsigned flags) {
return ::fftwf_plan_many_dft(rank,n,howmany,in,inembed,istride,idist,out,onembed,ostride,odist,sign,flags);
}
static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
::fftwf_flops(p,add,mul,fmas);
}
static void fftw_flops(const FFTW_plan p,double *add, double *mul, double *fmas){
::fftwf_flops(p,add,mul,fmas);
}
inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
::fftwf_execute_dft(p,in,out);
}
inline static void fftw_destroy_plan(const FFTW_plan p) {
::fftwf_destroy_plan(p);
}
};
inline static void fftw_execute_dft(const FFTW_plan p,FFTW_scalar *in,FFTW_scalar *out) {
::fftwf_execute_dft(p,in,out);
}
inline static void fftw_destroy_plan(const FFTW_plan p) {
::fftwf_destroy_plan(p);
}
};
#endif
@ -104,203 +104,188 @@ namespace Grid {
#define FFTW_BACKWARD (+1)
#endif
class FFT {
private:
class FFT {
private:
GridCartesian *vgrid;
GridCartesian *sgrid;
GridCartesian *vgrid;
GridCartesian *sgrid;
int Nd;
double flops;
double flops_call;
uint64_t usec;
int Nd;
double flops;
double flops_call;
uint64_t usec;
std::vector<int> dimensions;
std::vector<int> processors;
std::vector<int> processor_coor;
Coordinate dimensions;
Coordinate processors;
Coordinate processor_coor;
public:
public:
static const int forward=FFTW_FORWARD;
static const int backward=FFTW_BACKWARD;
static const int forward=FFTW_FORWARD;
static const int backward=FFTW_BACKWARD;
double Flops(void) {return flops;}
double MFlops(void) {return flops/usec;}
double USec(void) {return (double)usec;}
double Flops(void) {return flops;}
double MFlops(void) {return flops/usec;}
double USec(void) {return (double)usec;}
FFT ( GridCartesian * grid ) :
FFT ( GridCartesian * grid ) :
vgrid(grid),
Nd(grid->_ndimension),
dimensions(grid->_fdimensions),
processors(grid->_processors),
processor_coor(grid->_processor_coor)
{
flops=0;
usec =0;
std::vector<int> layout(Nd,1);
sgrid = new GridCartesian(dimensions,layout,processors);
};
~FFT ( void) {
delete sgrid;
}
template<class vobj>
void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,std::vector<int> mask,int sign){
conformable(result._grid,vgrid);
conformable(source._grid,vgrid);
Lattice<vobj> tmp(vgrid);
tmp = source;
for(int d=0;d<Nd;d++){
if( mask[d] ) {
FFT_dim(result,tmp,d,sign);
tmp=result;
}
}
}
template<class vobj>
void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
std::vector<int> mask(Nd,1);
FFT_dim_mask(result,source,mask,sign);
}
template<class vobj>
void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int sign){
#ifndef HAVE_FFTW
assert(0);
#else
conformable(result._grid,vgrid);
conformable(source._grid,vgrid);
int L = vgrid->_ldimensions[dim];
int G = vgrid->_fdimensions[dim];
std::vector<int> layout(Nd,1);
std::vector<int> pencil_gd(vgrid->_fdimensions);
pencil_gd[dim] = G*processors[dim];
// Pencil global vol LxLxGxLxL per node
GridCartesian pencil_g(pencil_gd,layout,processors);
// Construct pencils
typedef typename vobj::scalar_object sobj;
typedef typename sobj::scalar_type scalar;
Lattice<sobj> pgbuf(&pencil_g);
typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
typedef typename FFTW<scalar>::FFTW_plan FFTW_plan;
int Ncomp = sizeof(sobj)/sizeof(scalar);
int Nlow = 1;
for(int d=0;d<dim;d++){
Nlow*=vgrid->_ldimensions[d];
}
int rank = 1; /* 1d transforms */
int n[] = {G}; /* 1d transforms of length G */
int howmany = Ncomp;
int odist,idist,istride,ostride;
idist = odist = 1; /* Distance between consecutive FT's */
istride = ostride = Ncomp*Nlow; /* distance between two elements in the same FT */
int *inembed = n, *onembed = n;
scalar div;
if ( sign == backward ) div = 1.0/G;
else if ( sign == forward ) div = 1.0;
else assert(0);
FFTW_plan p;
{
FFTW_scalar *in = (FFTW_scalar *)&pgbuf._odata[0];
FFTW_scalar *out= (FFTW_scalar *)&pgbuf._odata[0];
p = FFTW<scalar>::fftw_plan_many_dft(rank,n,howmany,
in,inembed,
istride,idist,
out,onembed,
ostride, odist,
sign,FFTW_ESTIMATE);
}
// Barrel shift and collect global pencil
std::vector<int> lcoor(Nd), gcoor(Nd);
result = source;
int pc = processor_coor[dim];
for(int p=0;p<processors[dim];p++) {
PARALLEL_REGION
{
std::vector<int> cbuf(Nd);
sobj s;
PARALLEL_FOR_LOOP_INTERN
for(int idx=0;idx<sgrid->lSites();idx++) {
sgrid->LocalIndexToLocalCoor(idx,cbuf);
peekLocalSite(s,result,cbuf);
cbuf[dim]+=((pc+p) % processors[dim])*L;
// cbuf[dim]+=p*L;
pokeLocalSite(s,pgbuf,cbuf);
}
}
if (p != processors[dim] - 1)
{
result = Cshift(result,dim,L);
}
}
// Loop over orthog coords
int NN=pencil_g.lSites();
GridStopWatch timer;
timer.Start();
PARALLEL_REGION
{
std::vector<int> cbuf(Nd);
PARALLEL_FOR_LOOP_INTERN
for(int idx=0;idx<NN;idx++) {
pencil_g.LocalIndexToLocalCoor(idx, cbuf);
if ( cbuf[dim] == 0 ) { // restricts loop to plane at lcoor[dim]==0
FFTW_scalar *in = (FFTW_scalar *)&pgbuf._odata[idx];
FFTW_scalar *out= (FFTW_scalar *)&pgbuf._odata[idx];
FFTW<scalar>::fftw_execute_dft(p,in,out);
}
}
}
timer.Stop();
// performance counting
double add,mul,fma;
FFTW<scalar>::fftw_flops(p,&add,&mul,&fma);
flops_call = add+mul+2.0*fma;
usec += timer.useconds();
flops+= flops_call*NN;
// writing out result
PARALLEL_REGION
{
std::vector<int> clbuf(Nd), cgbuf(Nd);
sobj s;
PARALLEL_FOR_LOOP_INTERN
for(int idx=0;idx<sgrid->lSites();idx++) {
sgrid->LocalIndexToLocalCoor(idx,clbuf);
cgbuf = clbuf;
cgbuf[dim] = clbuf[dim]+L*pc;
peekLocalSite(s,pgbuf,cgbuf);
pokeLocalSite(s,result,clbuf);
}
}
result = result*div;
// destroying plan
FFTW<scalar>::fftw_destroy_plan(p);
#endif
}
{
flops=0;
usec =0;
Coordinate layout(Nd,1);
sgrid = new GridCartesian(dimensions,layout,processors);
};
}
~FFT ( void) {
delete sgrid;
}
template<class vobj>
void FFT_dim_mask(Lattice<vobj> &result,const Lattice<vobj> &source,Coordinate mask,int sign){
conformable(result.Grid(),vgrid);
conformable(source.Grid(),vgrid);
Lattice<vobj> tmp(vgrid);
tmp = source;
for(int d=0;d<Nd;d++){
if( mask[d] ) {
FFT_dim(result,tmp,d,sign);
tmp=result;
}
}
}
template<class vobj>
void FFT_all_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int sign){
Coordinate mask(Nd,1);
FFT_dim_mask(result,source,mask,sign);
}
template<class vobj>
void FFT_dim(Lattice<vobj> &result,const Lattice<vobj> &source,int dim, int sign){
#ifndef HAVE_FFTW
assert(0);
#else
conformable(result.Grid(),vgrid);
conformable(source.Grid(),vgrid);
int L = vgrid->_ldimensions[dim];
int G = vgrid->_fdimensions[dim];
Coordinate layout(Nd,1);
Coordinate pencil_gd(vgrid->_fdimensions);
pencil_gd[dim] = G*processors[dim];
// Pencil global vol LxLxGxLxL per node
GridCartesian pencil_g(pencil_gd,layout,processors);
// Construct pencils
typedef typename vobj::scalar_object sobj;
typedef typename sobj::scalar_type scalar;
Lattice<sobj> pgbuf(&pencil_g);
auto pgbuf_v = pgbuf.View();
typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
typedef typename FFTW<scalar>::FFTW_plan FFTW_plan;
int Ncomp = sizeof(sobj)/sizeof(scalar);
int Nlow = 1;
for(int d=0;d<dim;d++){
Nlow*=vgrid->_ldimensions[d];
}
int rank = 1; /* 1d transforms */
int n[] = {G}; /* 1d transforms of length G */
int howmany = Ncomp;
int odist,idist,istride,ostride;
idist = odist = 1; /* Distance between consecutive FT's */
istride = ostride = Ncomp*Nlow; /* distance between two elements in the same FT */
int *inembed = n, *onembed = n;
scalar div;
if ( sign == backward ) div = 1.0/G;
else if ( sign == forward ) div = 1.0;
else assert(0);
FFTW_plan p;
{
FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[0];
FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[0];
p = FFTW<scalar>::fftw_plan_many_dft(rank,n,howmany,
in,inembed,
istride,idist,
out,onembed,
ostride, odist,
sign,FFTW_ESTIMATE);
}
// Barrel shift and collect global pencil
Coordinate lcoor(Nd), gcoor(Nd);
result = source;
int pc = processor_coor[dim];
for(int p=0;p<processors[dim];p++) {
thread_for(idx, sgrid->lSites(),{
Coordinate cbuf(Nd);
sobj s;
sgrid->LocalIndexToLocalCoor(idx,cbuf);
peekLocalSite(s,result,cbuf);
cbuf[dim]+=((pc+p) % processors[dim])*L;
// cbuf[dim]+=p*L;
pokeLocalSite(s,pgbuf,cbuf);
});
if (p != processors[dim] - 1) {
result = Cshift(result,dim,L);
}
}
// Loop over orthog coords
int NN=pencil_g.lSites();
GridStopWatch timer;
timer.Start();
thread_for( idx,NN,{
Coordinate cbuf(Nd);
pencil_g.LocalIndexToLocalCoor(idx, cbuf);
if ( cbuf[dim] == 0 ) { // restricts loop to plane at lcoor[dim]==0
FFTW_scalar *in = (FFTW_scalar *)&pgbuf_v[idx];
FFTW_scalar *out= (FFTW_scalar *)&pgbuf_v[idx];
FFTW<scalar>::fftw_execute_dft(p,in,out);
}
});
timer.Stop();
// performance counting
double add,mul,fma;
FFTW<scalar>::fftw_flops(p,&add,&mul,&fma);
flops_call = add+mul+2.0*fma;
usec += timer.useconds();
flops+= flops_call*NN;
// writing out result
thread_for(idx,sgrid->lSites(),{
Coordinate clbuf(Nd), cgbuf(Nd);
sobj s;
sgrid->LocalIndexToLocalCoor(idx,clbuf);
cgbuf = clbuf;
cgbuf[dim] = clbuf[dim]+L*pc;
peekLocalSite(s,pgbuf,cgbuf);
pokeLocalSite(s,result,clbuf);
});
result = result*div;
// destroying plan
FFTW<scalar>::fftw_destroy_plan(p);
#endif
}
};
NAMESPACE_END(Grid);
#endif

571
Grid/algorithms/LinearOperator.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,153 +24,152 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_ALGORITHM_LINEAR_OP_H
#define GRID_ALGORITHM_LINEAR_OP_H
*************************************************************************************/
/* END LEGAL */
#pragma once
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////////////////
// LinearOperators Take a something and return a something.
/////////////////////////////////////////////////////////////////////////////////////////////
//
// Hopefully linearity is satisfied and the AdjOp is indeed the Hermitian conjugateugate (transpose if real):
//SBase
// i) F(a x + b y) = aF(x) + b F(y).
// ii) <x|Op|y> = <y|AdjOp|x>^\ast
//
// Would be fun to have a test linearity & Herm Conj function!
/////////////////////////////////////////////////////////////////////////////////////////////
template<class Field> class LinearOperatorBase {
public:
/////////////////////////////////////////////////////////////////////////////////////////////
// LinearOperators Take a something and return a something.
/////////////////////////////////////////////////////////////////////////////////////////////
//
// Hopefully linearity is satisfied and the AdjOp is indeed the Hermitian Conjugateugate (transpose if real):
//SBase
// i) F(a x + b y) = aF(x) + b F(y).
// ii) <x|Op|y> = <y|AdjOp|x>^\ast
//
// Would be fun to have a test linearity & Herm Conj function!
/////////////////////////////////////////////////////////////////////////////////////////////
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
// 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
virtual void Op (const Field &in, Field &out) = 0; // Abstract base
virtual void AdjOp (const Field &in, Field &out) = 0; // Abstract base
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2) = 0;
virtual void HermOp(const Field &in, Field &out)=0;
};
virtual void Op (const Field &in, Field &out) = 0; // Abstract base
virtual void AdjOp (const Field &in, Field &out) = 0; // Abstract base
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
virtual void HermOp(const Field &in, Field &out)=0;
};
/////////////////////////////////////////////////////////////////////////////////////////////
// By sharing the class for Sparse Matrix across multiple operator wrappers, we can share code
// between RB and non-RB variants. Sparse matrix is like the fermion action def, and then
// the wrappers implement the specialisation of "Op" and "AdjOp" to the cases minimising
// replication of code.
//
// I'm not entirely happy with implementation; to share the Schur code between herm and non-herm
// while still having a "OpAndNorm" in the abstract base I had to implement it in both cases
// with an assert trap in the non-herm. This isn't right; there must be a better C++ way to
// do it, but I fear it required multiple inheritance and mixed in abstract base classes
/////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////////////////
// By sharing the class for Sparse Matrix across multiple operator wrappers, we can share code
// between RB and non-RB variants. Sparse matrix is like the fermion action def, and then
// the wrappers implement the specialisation of "Op" and "AdjOp" to the cases minimising
// replication of code.
//
// I'm not entirely happy with implementation; to share the Schur code between herm and non-herm
// while still having a "OpAndNorm" in the abstract base I had to implement it in both cases
// with an assert trap in the non-herm. This isn't right; there must be a better C++ way to
// do it, but I fear it required multiple inheritance and mixed in abstract base classes
/////////////////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////
// Construct herm op from non-herm matrix
////////////////////////////////////////////////////////////////////
template<class Matrix,class Field>
class MdagMLinearOperator : public LinearOperatorBase<Field> {
Matrix &_Mat;
public:
MdagMLinearOperator(Matrix &Mat): _Mat(Mat){};
////////////////////////////////////////////////////////////////////
// Construct herm op from non-herm matrix
////////////////////////////////////////////////////////////////////
template<class Matrix,class Field>
class MdagMLinearOperator : public LinearOperatorBase<Field> {
Matrix &_Mat;
public:
MdagMLinearOperator(Matrix &Mat): _Mat(Mat){};
// Support for coarsening to a multigrid
void OpDiag (const Field &in, Field &out) {
_Mat.Mdiag(in,out);
}
void OpDir (const Field &in, Field &out,int dir,int disp) {
_Mat.Mdir(in,out,dir,disp);
}
void Op (const Field &in, Field &out){
_Mat.M(in,out);
}
void AdjOp (const Field &in, Field &out){
_Mat.Mdag(in,out);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
_Mat.MdagM(in,out,n1,n2);
}
void HermOp(const Field &in, Field &out){
RealD n1,n2;
HermOpAndNorm(in,out,n1,n2);
}
};
// Support for coarsening to a multigrid
void OpDiag (const Field &in, Field &out) {
_Mat.Mdiag(in,out);
}
void OpDir (const Field &in, Field &out,int dir,int disp) {
_Mat.Mdir(in,out,dir,disp);
}
void Op (const Field &in, Field &out){
_Mat.M(in,out);
}
void AdjOp (const Field &in, Field &out){
_Mat.Mdag(in,out);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
_Mat.MdagM(in,out,n1,n2);
}
void HermOp(const Field &in, Field &out){
RealD n1,n2;
HermOpAndNorm(in,out,n1,n2);
}
};
////////////////////////////////////////////////////////////////////
// Construct herm op and shift it for mgrid smoother
////////////////////////////////////////////////////////////////////
template<class Matrix,class Field>
class ShiftedMdagMLinearOperator : public LinearOperatorBase<Field> {
Matrix &_Mat;
RealD _shift;
public:
ShiftedMdagMLinearOperator(Matrix &Mat,RealD shift): _Mat(Mat), _shift(shift){};
// Support for coarsening to a multigrid
void OpDiag (const Field &in, Field &out) {
_Mat.Mdiag(in,out);
assert(0);
}
void OpDir (const Field &in, Field &out,int dir,int disp) {
_Mat.Mdir(in,out,dir,disp);
assert(0);
}
void Op (const Field &in, Field &out){
_Mat.M(in,out);
assert(0);
}
void AdjOp (const Field &in, Field &out){
_Mat.Mdag(in,out);
assert(0);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
_Mat.MdagM(in,out,n1,n2);
out = out + _shift*in;
////////////////////////////////////////////////////////////////////
// Construct herm op and shift it for mgrid smoother
////////////////////////////////////////////////////////////////////
template<class Matrix,class Field>
class ShiftedMdagMLinearOperator : public LinearOperatorBase<Field> {
Matrix &_Mat;
RealD _shift;
public:
ShiftedMdagMLinearOperator(Matrix &Mat,RealD shift): _Mat(Mat), _shift(shift){};
// Support for coarsening to a multigrid
void OpDiag (const Field &in, Field &out) {
_Mat.Mdiag(in,out);
assert(0);
}
void OpDir (const Field &in, Field &out,int dir,int disp) {
_Mat.Mdir(in,out,dir,disp);
assert(0);
}
void Op (const Field &in, Field &out){
_Mat.M(in,out);
assert(0);
}
void AdjOp (const Field &in, Field &out){
_Mat.Mdag(in,out);
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);
n1=real(dot);
n2=norm2(out);
}
void HermOp(const Field &in, Field &out){
RealD n1,n2;
HermOpAndNorm(in,out,n1,n2);
}
};
ComplexD dot;
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);
}
};
////////////////////////////////////////////////////////////////////
// Wrap an already herm matrix
////////////////////////////////////////////////////////////////////
template<class Matrix,class Field>
class HermitianLinearOperator : public LinearOperatorBase<Field> {
Matrix &_Mat;
public:
HermitianLinearOperator(Matrix &Mat): _Mat(Mat){};
// Support for coarsening to a multigrid
void OpDiag (const Field &in, Field &out) {
_Mat.Mdiag(in,out);
}
void OpDir (const Field &in, Field &out,int dir,int disp) {
_Mat.Mdir(in,out,dir,disp);
}
void Op (const Field &in, Field &out){
_Mat.M(in,out);
}
void AdjOp (const Field &in, Field &out){
_Mat.M(in,out);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
_Mat.M(in,out);
////////////////////////////////////////////////////////////////////
// Wrap an already herm matrix
////////////////////////////////////////////////////////////////////
template<class Matrix,class Field>
class HermitianLinearOperator : public LinearOperatorBase<Field> {
Matrix &_Mat;
public:
HermitianLinearOperator(Matrix &Mat): _Mat(Mat){};
// Support for coarsening to a multigrid
void OpDiag (const Field &in, Field &out) {
_Mat.Mdiag(in,out);
}
void OpDir (const Field &in, Field &out,int dir,int disp) {
_Mat.Mdir(in,out,dir,disp);
}
void Op (const Field &in, Field &out){
_Mat.M(in,out);
}
void AdjOp (const Field &in, Field &out){
_Mat.M(in,out);
}
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
_Mat.M(in,out);
ComplexD dot= innerProduct(in,out); n1=real(dot);
n2=norm2(out);
}
void HermOp(const Field &in, Field &out){
_Mat.M(in,out);
}
};
ComplexD dot= innerProduct(in,out); n1=real(dot);
n2=norm2(out);
}
void HermOp(const Field &in, Field &out){
_Mat.M(in,out);
}
};
//////////////////////////////////////////////////////////
// Even Odd Schur decomp operators; there are several
@ -183,13 +182,13 @@ namespace Grid {
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;
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;
out.Checkerboard() = in.Checkerboard();
MpcDagMpc(in,out,n1,n2);
}
virtual void HermOp(const Field &in, Field &out){
@ -216,20 +215,20 @@ namespace Grid {
Matrix &_Mat;
SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD Mpc (const Field &in, Field &out) {
Field tmp(in._grid);
tmp.checkerboard = !in.checkerboard;
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;
//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);
Field tmp(in.Grid());
_Mat.MeooeDag(in,tmp);
_Mat.MooeeInvDag(tmp,out);
@ -247,7 +246,7 @@ namespace Grid {
SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD Mpc (const Field &in, Field &out) {
Field tmp(in._grid);
Field tmp(in.Grid());
_Mat.Meooe(in,out);
_Mat.MooeeInv(out,tmp);
@ -257,7 +256,7 @@ namespace Grid {
return axpy_norm(out,-1.0,tmp,in);
}
virtual RealD MpcDag (const Field &in, Field &out){
Field tmp(in._grid);
Field tmp(in.Grid());
_Mat.MooeeInvDag(in,out);
_Mat.MeooeDag(out,tmp);
@ -275,7 +274,7 @@ namespace Grid {
SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
virtual RealD Mpc (const Field &in, Field &out) {
Field tmp(in._grid);
Field tmp(in.Grid());
_Mat.MooeeInv(in,out);
_Mat.Meooe(out,tmp);
@ -285,7 +284,7 @@ namespace Grid {
return axpy_norm(out,-1.0,tmp,in);
}
virtual RealD MpcDag (const Field &in, Field &out){
Field tmp(in._grid);
Field tmp(in.Grid());
_Mat.MeooeDag(in,out);
_Mat.MooeeInvDag(out,tmp);
@ -315,7 +314,7 @@ namespace Grid {
double tMeo;
double taxpby_norm;
uint64_t ncall;
public:
public:
void Report(void)
{
std::cout << GridLogMessage << " HermOpAndNorm.Mpc "<< tMpc/ncall<<" usec "<<std::endl;
@ -333,17 +332,17 @@ namespace Grid {
taxpby_norm=0;
ncall=0;
}
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
ncall++;
tMpc-=usecond();
n2 = Mpc(in,out);
n2 = Mpc(in,out);
tMpc+=usecond();
tIP-=usecond();
ComplexD dot= innerProduct(in,out);
ComplexD dot= innerProduct(in,out);
tIP+=usecond();
n1 = real(dot);
}
virtual void HermOp(const Field &in, Field &out){
n1 = real(dot);
}
virtual void HermOp(const Field &in, Field &out){
ncall++;
tMpc-=usecond();
_Mat.Meooe(in,out);
@ -352,135 +351,145 @@ namespace Grid {
taxpby_norm-=usecond();
axpby(out,-1.0,mass*mass,tmp,in);
taxpby_norm+=usecond();
}
virtual RealD Mpc (const Field &in, Field &out) {
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;
}
virtual RealD MpcDag (const Field &in, Field &out){
return Mpc(in,out);
}
virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
assert(0);// Never need with staggered
}
};
template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
}
virtual RealD Mpc (const Field &in, Field &out)
{
Field tmp(in.Grid());
Field tmp2(in.Grid());
// std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
_Mat.Mooee(in,out);
_Mat.Mooee(out,tmp);
// std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
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;
}
virtual RealD MpcDag (const Field &in, Field &out){
return Mpc(in,out);
}
virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
assert(0);// Never need with staggered
}
};
template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
/////////////////////////////////////////////////////////////
// Base classes for functions of operators
/////////////////////////////////////////////////////////////
template<class Field> class OperatorFunction {
public:
virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) = 0;
virtual void operator() (LinearOperatorBase<Field> &Linop, const std::vector<Field> &in,std::vector<Field> &out) {
assert(in.size()==out.size());
for(int k=0;k<in.size();k++){
(*this)(Linop,in[k],out[k]);
}
};
};
template<class Field> class LinearFunction {
public:
virtual void operator() (const Field &in, Field &out) = 0;
};
template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
public:
void operator() (const Field &in, Field &out){
out = in;
};
};
/////////////////////////////////////////////////////////////
// Base classes for Multishift solvers for operators
/////////////////////////////////////////////////////////////
template<class Field> class OperatorMultiFunction {
public:
virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, std::vector<Field> &out) = 0;
};
// FIXME : To think about
// Chroma functionality list defining LinearOperator
/*
virtual void operator() (T& chi, const T& psi, enum PlusMinus isign) const = 0;
virtual void operator() (T& chi, const T& psi, enum PlusMinus isign, Real epsilon) const
virtual const Subset& subset() const = 0;
virtual unsigned long nFlops() const { return 0; }
virtual void deriv(P& ds_u, const T& chi, const T& psi, enum PlusMinus isign) const
class UnprecLinearOperator : public DiffLinearOperator<T,P,Q>
const Subset& subset() const {return all;}
};
*/
////////////////////////////////////////////////////////////////////////////////////////////
// Hermitian operator Linear function and operator function
////////////////////////////////////////////////////////////////////////////////////////////
template<class Field>
class HermOpOperatorFunction : public OperatorFunction<Field> {
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
Linop.HermOp(in,out);
};
};
template<typename Field>
class PlainHermOp : public LinearFunction<Field> {
public:
LinearOperatorBase<Field> &_Linop;
PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop)
{}
void operator()(const Field& in, Field& out) {
_Linop.HermOp(in,out);
}
};
template<typename Field>
class FunctionHermOp : public LinearFunction<Field> {
public:
OperatorFunction<Field> & _poly;
LinearOperatorBase<Field> &_Linop;
FunctionHermOp(OperatorFunction<Field> & poly,LinearOperatorBase<Field>& linop)
: _poly(poly), _Linop(linop) {};
void operator()(const Field& in, Field& out) {
_poly(_Linop,in,out);
}
};
template<class Field>
class Polynomial : public OperatorFunction<Field> {
private:
std::vector<RealD> Coeffs;
public:
Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
// Implement the required interface
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
Field AtoN(in._grid);
Field Mtmp(in._grid);
AtoN = in;
out = AtoN*Coeffs[0];
for(int n=1;n<Coeffs.size();n++){
Mtmp = AtoN;
Linop.HermOp(Mtmp,AtoN);
out=out+AtoN*Coeffs[n];
}
};
/////////////////////////////////////////////////////////////
// Base classes for functions of operators
/////////////////////////////////////////////////////////////
template<class Field> class OperatorFunction {
public:
virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) = 0;
virtual void operator() (LinearOperatorBase<Field> &Linop, const std::vector<Field> &in,std::vector<Field> &out) {
assert(in.size()==out.size());
for(int k=0;k<in.size();k++){
(*this)(Linop,in[k],out[k]);
}
};
};
}
template<class Field> class LinearFunction {
public:
virtual void operator() (const Field &in, Field &out) = 0;
};
#endif
template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
public:
void operator() (const Field &in, Field &out){
out = in;
};
};
/////////////////////////////////////////////////////////////
// Base classes for Multishift solvers for operators
/////////////////////////////////////////////////////////////
template<class Field> class OperatorMultiFunction {
public:
virtual void operator() (LinearOperatorBase<Field> &Linop, const Field &in, std::vector<Field> &out) = 0;
};
// FIXME : To think about
// Chroma functionality list defining LinearOperator
/*
virtual void operator() (T& chi, const T& psi, enum PlusMinus isign) const = 0;
virtual void operator() (T& chi, const T& psi, enum PlusMinus isign, Real epsilon) const
virtual const Subset& subset() const = 0;
virtual unsigned long nFlops() const { return 0; }
virtual void deriv(P& ds_u, const T& chi, const T& psi, enum PlusMinus isign) const
class UnprecLinearOperator : public DiffLinearOperator<T,P,Q>
const Subset& subset() const {return all;}
};
*/
////////////////////////////////////////////////////////////////////////////////////////////
// Hermitian operator Linear function and operator function
////////////////////////////////////////////////////////////////////////////////////////////
template<class Field>
class HermOpOperatorFunction : public OperatorFunction<Field> {
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
Linop.HermOp(in,out);
};
};
template<typename Field>
class PlainHermOp : public LinearFunction<Field> {
public:
LinearOperatorBase<Field> &_Linop;
PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop)
{}
void operator()(const Field& in, Field& out) {
_Linop.HermOp(in,out);
}
};
template<typename Field>
class FunctionHermOp : public LinearFunction<Field> {
public:
OperatorFunction<Field> & _poly;
LinearOperatorBase<Field> &_Linop;
FunctionHermOp(OperatorFunction<Field> & poly,LinearOperatorBase<Field>& linop)
: _poly(poly), _Linop(linop) {};
void operator()(const Field& in, Field& out) {
_poly(_Linop,in,out);
}
};
template<class Field>
class Polynomial : public OperatorFunction<Field> {
private:
std::vector<RealD> Coeffs;
public:
using OperatorFunction<Field>::operator();
Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
// Implement the required interface
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
Field AtoN(in.Grid());
Field Mtmp(in.Grid());
AtoN = in;
out = AtoN*Coeffs[0];
for(int n=1;n<Coeffs.size();n++){
Mtmp = AtoN;
Linop.HermOp(Mtmp,AtoN);
out=out+AtoN*Coeffs[n];
}
};
};
NAMESPACE_END(Grid);

30
Grid/algorithms/Preconditioner.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,24 +23,24 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_PRECONDITIONER_H
#define GRID_PRECONDITIONER_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
template<class Field> class Preconditioner : public LinearFunction<Field> {
virtual void operator()(const Field &src, Field & psi)=0;
};
template<class Field> class Preconditioner : public LinearFunction<Field> {
virtual void operator()(const Field &src, Field & psi)=0;
};
template<class Field> class TrivialPrecon : public Preconditioner<Field> {
public:
void operator()(const Field &src, Field & psi){
psi = src;
}
TrivialPrecon(void){};
};
template<class Field> class TrivialPrecon : public Preconditioner<Field> {
public:
void operator()(const Field &src, Field & psi){
psi = src;
}
TrivialPrecon(void){};
};
}
NAMESPACE_END(Grid);
#endif

72
Grid/algorithms/SparseMatrix.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,38 +23,38 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_ALGORITHM_SPARSE_MATRIX_H
#define GRID_ALGORITHM_SPARSE_MATRIX_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////////////////
// Interface defining what I expect of a general sparse matrix, such as a Fermion action
/////////////////////////////////////////////////////////////////////////////////////////////
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 Mdiag (const Field &in, Field &out)=0;
virtual void Mdir (const Field &in, Field &out,int dir, int disp)=0;
};
/////////////////////////////////////////////////////////////////////////////////////////////
// Interface defining what I expect of a general sparse matrix, such as a Fermion action
/////////////////////////////////////////////////////////////////////////////////////////////
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 Mdiag (const Field &in, Field &out)=0;
virtual void Mdir (const Field &in, Field &out,int dir, int disp)=0;
};
/////////////////////////////////////////////////////////////////////////////////////////////
// Interface augmented by a red black sparse matrix, such as a Fermion action
/////////////////////////////////////////////////////////////////////////////////////////////
template<class Field> class CheckerBoardedSparseMatrixBase : public SparseMatrixBase<Field> {
public:
virtual GridBase *RedBlackGrid(void)=0;
/////////////////////////////////////////////////////////////////////////////////////////////
// Interface augmented by a red black sparse matrix, such as a Fermion action
/////////////////////////////////////////////////////////////////////////////////////////////
template<class Field> class CheckerBoardedSparseMatrixBase : public SparseMatrixBase<Field> {
public:
virtual GridBase *RedBlackGrid(void)=0;
//////////////////////////////////////////////////////////////////////
// Query the even even properties to make algorithmic decisions
@ -63,17 +63,17 @@ namespace Grid {
virtual int ConstEE(void) { return 1; }; // Disable assumptions unless overridden
virtual int isTrivialEE(void) { return 0; }; // by a derived class that knows better
// half checkerboard operaions
virtual void Meooe (const Field &in, Field &out)=0;
virtual void Mooee (const Field &in, Field &out)=0;
virtual void MooeeInv (const Field &in, Field &out)=0;
// half checkerboard operaions
virtual void Meooe (const Field &in, Field &out)=0;
virtual void Mooee (const Field &in, Field &out)=0;
virtual void MooeeInv (const Field &in, Field &out)=0;
virtual void MeooeDag (const Field &in, Field &out)=0;
virtual void MooeeDag (const Field &in, Field &out)=0;
virtual void MooeeInvDag (const Field &in, Field &out)=0;
virtual void MeooeDag (const Field &in, Field &out)=0;
virtual void MooeeDag (const Field &in, Field &out)=0;
virtual void MooeeInvDag (const Field &in, Field &out)=0;
};
};
}
NAMESPACE_END(Grid);
#endif

654
Grid/algorithms/approx/Chebyshev.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -25,14 +25,14 @@ Author: Christoph Lehner <clehner@bnl.gov>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CHEBYSHEV_H
#define GRID_CHEBYSHEV_H
#include <Grid/algorithms/LinearOperator.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
struct ChebyParams : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(ChebyParams,
@ -41,337 +41,339 @@ struct ChebyParams : Serializable {
int, Npoly);
};
////////////////////////////////////////////////////////////////////////////////////////////
// Generic Chebyshev approximations
////////////////////////////////////////////////////////////////////////////////////////////
template<class Field>
class Chebyshev : public OperatorFunction<Field> {
private:
std::vector<RealD> Coeffs;
int order;
RealD hi;
RealD lo;
////////////////////////////////////////////////////////////////////////////////////////////
// Generic Chebyshev approximations
////////////////////////////////////////////////////////////////////////////////////////////
template<class Field>
class Chebyshev : public OperatorFunction<Field> {
private:
using OperatorFunction<Field>::operator();
public:
void csv(std::ostream &out){
RealD diff = hi-lo;
RealD delta = (hi-lo)*1.0e-9;
for (RealD x=lo; x<hi; x+=delta) {
delta*=1.1;
RealD f = approx(x);
out<< x<<" "<<f<<std::endl;
}
return;
std::vector<RealD> Coeffs;
int order;
RealD hi;
RealD lo;
public:
void csv(std::ostream &out){
RealD diff = hi-lo;
RealD delta = diff*1.0e-9;
for (RealD x=lo; x<hi; x+=delta) {
delta*=1.1;
RealD f = approx(x);
out<< x<<" "<<f<<std::endl;
}
return;
}
// Convenience for plotting the approximation
void PlotApprox(std::ostream &out) {
out<<"Polynomial approx ["<<lo<<","<<hi<<"]"<<std::endl;
for(RealD x=lo;x<hi;x+=(hi-lo)/50.0){
out <<x<<"\t"<<approx(x)<<std::endl;
}
};
Chebyshev(){};
Chebyshev(ChebyParams p){ Init(p.alpha,p.beta,p.Npoly);};
Chebyshev(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD) ) {Init(_lo,_hi,_order,func);};
Chebyshev(RealD _lo,RealD _hi,int _order) {Init(_lo,_hi,_order);};
////////////////////////////////////////////////////////////////////////////////////////////////////
// c.f. numerical recipes "chebft"/"chebev". This is sec 5.8 "Chebyshev approximation".
////////////////////////////////////////////////////////////////////////////////////////////////////
// CJ: the one we need for Lanczos
void Init(RealD _lo,RealD _hi,int _order)
{
lo=_lo;
hi=_hi;
order=_order;
if(order < 2) exit(-1);
Coeffs.resize(order);
Coeffs.assign(0.,order);
Coeffs[order-1] = 1.;
};
void Init(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD))
{
lo=_lo;
hi=_hi;
order=_order;
if(order < 2) exit(-1);
Coeffs.resize(order);
for(int j=0;j<order;j++){
RealD s=0;
for(int k=0;k<order;k++){
RealD y=std::cos(M_PI*(k+0.5)/order);
RealD x=0.5*(y*(hi-lo)+(hi+lo));
RealD f=func(x);
s=s+f*std::cos( j*M_PI*(k+0.5)/order );
}
Coeffs[j] = s * 2.0/order;
}
};
void JacksonSmooth(void){
RealD M=order;
RealD alpha = M_PI/(M+2);
RealD lmax = std::cos(alpha);
RealD sumUsq =0;
std::vector<RealD> U(M);
std::vector<RealD> a(M);
std::vector<RealD> g(M);
for(int n=0;n<=M;n++){
U[n] = std::sin((n+1)*std::acos(lmax))/std::sin(std::acos(lmax));
sumUsq += U[n]*U[n];
}
sumUsq = std::sqrt(sumUsq);
for(int i=1;i<=M;i++){
a[i] = U[i]/sumUsq;
}
g[0] = 1.0;
for(int m=1;m<=M;m++){
g[m] = 0;
for(int i=0;i<=M-m;i++){
g[m]+= a[i]*a[m+i];
}
}
for(int m=1;m<=M;m++){
Coeffs[m]*=g[m];
}
}
RealD approx(RealD x) // Convenience for plotting the approximation
{
RealD Tn;
RealD Tnm;
RealD Tnp;
RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
RealD T0=1;
RealD T1=y;
RealD sum;
sum = 0.5*Coeffs[0]*T0;
sum+= Coeffs[1]*T1;
Tn =T1;
Tnm=T0;
for(int i=2;i<order;i++){
Tnp=2*y*Tn-Tnm;
Tnm=Tn;
Tn =Tnp;
sum+= Tn*Coeffs[i];
}
return sum;
};
RealD approxD(RealD x)
{
RealD Un;
RealD Unm;
RealD Unp;
RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
RealD U0=1;
RealD U1=2*y;
RealD sum;
sum = Coeffs[1]*U0;
sum+= Coeffs[2]*U1*2.0;
Un =U1;
Unm=U0;
for(int i=2;i<order-1;i++){
Unp=2*y*Un-Unm;
Unm=Un;
Un =Unp;
sum+= Un*Coeffs[i+1]*(i+1.0);
}
return sum/(0.5*(hi-lo));
};
RealD approxInv(RealD z, RealD x0, int maxiter, RealD resid) {
RealD x = x0;
RealD eps;
int i;
for (i=0;i<maxiter;i++) {
eps = approx(x) - z;
if (fabs(eps / z) < resid)
return x;
x = x - eps / approxD(x);
}
return std::numeric_limits<double>::quiet_NaN();
}
// Implement the required interface
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
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();
Field T0(grid); T0 = in;
Field T1(grid);
Field T2(grid);
Field y(grid);
Field *Tnm = &T0;
Field *Tn = &T1;
Field *Tnp = &T2;
// Tn=T1 = (xscale M + mscale)in
RealD xscale = 2.0/(hi-lo);
RealD mscale = -(hi+lo)/(hi-lo);
Linop.HermOp(T0,y);
T1=y*xscale+in*mscale;
// sum = .5 c[0] T0 + c[1] T1
out = (0.5*Coeffs[0])*T0 + Coeffs[1]*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);
// Cycle pointers to avoid copies
Field *swizzle = Tnm;
Tnm =Tn;
Tn =Tnp;
Tnp =swizzle;
}
// Convenience for plotting the approximation
void PlotApprox(std::ostream &out) {
out<<"Polynomial approx ["<<lo<<","<<hi<<"]"<<std::endl;
for(RealD x=lo;x<hi;x+=(hi-lo)/50.0){
out <<x<<"\t"<<approx(x)<<std::endl;
}
};
Chebyshev(){};
Chebyshev(ChebyParams p){ Init(p.alpha,p.beta,p.Npoly);};
Chebyshev(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD) ) {Init(_lo,_hi,_order,func);};
Chebyshev(RealD _lo,RealD _hi,int _order) {Init(_lo,_hi,_order);};
template<class Field>
class ChebyshevLanczos : public Chebyshev<Field> {
private:
std::vector<RealD> Coeffs;
int order;
RealD alpha;
RealD beta;
RealD mu;
////////////////////////////////////////////////////////////////////////////////////////////////////
// c.f. numerical recipes "chebft"/"chebev". This is sec 5.8 "Chebyshev approximation".
////////////////////////////////////////////////////////////////////////////////////////////////////
// CJ: the one we need for Lanczos
void Init(RealD _lo,RealD _hi,int _order)
{
lo=_lo;
hi=_hi;
order=_order;
if(order < 2) exit(-1);
Coeffs.resize(order);
Coeffs.assign(0.,order);
Coeffs[order-1] = 1.;
};
public:
ChebyshevLanczos(RealD _alpha,RealD _beta,RealD _mu,int _order) :
void Init(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD))
{
lo=_lo;
hi=_hi;
order=_order;
if(order < 2) exit(-1);
Coeffs.resize(order);
for(int j=0;j<order;j++){
RealD s=0;
for(int k=0;k<order;k++){
RealD y=std::cos(M_PI*(k+0.5)/order);
RealD x=0.5*(y*(hi-lo)+(hi+lo));
RealD f=func(x);
s=s+f*std::cos( j*M_PI*(k+0.5)/order );
}
Coeffs[j] = s * 2.0/order;
}
};
void JacksonSmooth(void){
RealD M=order;
RealD alpha = M_PI/(M+2);
RealD lmax = std::cos(alpha);
RealD sumUsq =0;
std::vector<RealD> U(M);
std::vector<RealD> a(M);
std::vector<RealD> g(M);
for(int n=0;n<=M;n++){
U[n] = std::sin((n+1)*std::acos(lmax))/std::sin(std::acos(lmax));
sumUsq += U[n]*U[n];
}
sumUsq = std::sqrt(sumUsq);
for(int i=1;i<=M;i++){
a[i] = U[i]/sumUsq;
}
g[0] = 1.0;
for(int m=1;m<=M;m++){
g[m] = 0;
for(int i=0;i<=M-m;i++){
g[m]+= a[i]*a[m+i];
}
}
for(int m=1;m<=M;m++){
Coeffs[m]*=g[m];
}
}
RealD approx(RealD x) // Convenience for plotting the approximation
{
RealD Tn;
RealD Tnm;
RealD Tnp;
RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
RealD T0=1;
RealD T1=y;
RealD sum;
sum = 0.5*Coeffs[0]*T0;
sum+= Coeffs[1]*T1;
Tn =T1;
Tnm=T0;
for(int i=2;i<order;i++){
Tnp=2*y*Tn-Tnm;
Tnm=Tn;
Tn =Tnp;
sum+= Tn*Coeffs[i];
}
return sum;
};
RealD approxD(RealD x)
{
RealD Un;
RealD Unm;
RealD Unp;
RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
RealD U0=1;
RealD U1=2*y;
RealD sum;
sum = Coeffs[1]*U0;
sum+= Coeffs[2]*U1*2.0;
Un =U1;
Unm=U0;
for(int i=2;i<order-1;i++){
Unp=2*y*Un-Unm;
Unm=Un;
Un =Unp;
sum+= Un*Coeffs[i+1]*(i+1.0);
}
return sum/(0.5*(hi-lo));
};
RealD approxInv(RealD z, RealD x0, int maxiter, RealD resid) {
RealD x = x0;
RealD eps;
int i;
for (i=0;i<maxiter;i++) {
eps = approx(x) - z;
if (fabs(eps / z) < resid)
return x;
x = x - eps / approxD(x);
}
return std::numeric_limits<double>::quiet_NaN();
}
// Implement the required interface
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
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();
Field T0(grid); T0 = in;
Field T1(grid);
Field T2(grid);
Field y(grid);
Field *Tnm = &T0;
Field *Tn = &T1;
Field *Tnp = &T2;
// Tn=T1 = (xscale M + mscale)in
RealD xscale = 2.0/(hi-lo);
RealD mscale = -(hi+lo)/(hi-lo);
Linop.HermOp(T0,y);
T1=y*xscale+in*mscale;
// sum = .5 c[0] T0 + c[1] T1
out = (0.5*Coeffs[0])*T0 + Coeffs[1]*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);
// Cycle pointers to avoid copies
Field *swizzle = Tnm;
Tnm =Tn;
Tn =Tnp;
Tnp =swizzle;
}
}
};
template<class Field>
class ChebyshevLanczos : public Chebyshev<Field> {
private:
std::vector<RealD> Coeffs;
int order;
RealD alpha;
RealD beta;
RealD mu;
public:
ChebyshevLanczos(RealD _alpha,RealD _beta,RealD _mu,int _order) :
alpha(_alpha),
beta(_beta),
mu(_mu)
{
order=_order;
Coeffs.resize(order);
for(int i=0;i<_order;i++){
Coeffs[i] = 0.0;
}
Coeffs[order-1]=1.0;
};
void csv(std::ostream &out){
for (RealD x=-1.2*alpha; x<1.2*alpha; x+=(2.0*alpha)/10000) {
RealD f = approx(x);
out<< x<<" "<<f<<std::endl;
}
return;
}
RealD approx(RealD xx) // Convenience for plotting the approximation
{
RealD Tn;
RealD Tnm;
RealD Tnp;
Real aa = alpha * alpha;
Real bb = beta * beta;
RealD x = ( 2.0 * (xx-mu)*(xx-mu) - (aa+bb) ) / (aa-bb);
RealD y= x;
RealD T0=1;
RealD T1=y;
RealD sum;
sum = 0.5*Coeffs[0]*T0;
sum+= Coeffs[1]*T1;
Tn =T1;
Tnm=T0;
for(int i=2;i<order;i++){
Tnp=2*y*Tn-Tnm;
Tnm=Tn;
Tn =Tnp;
sum+= Tn*Coeffs[i];
}
return sum;
};
// shift_Multiply in Rudy's code
void AminusMuSq(LinearOperatorBase<Field> &Linop, const Field &in, Field &out)
{
GridBase *grid=in._grid;
Field tmp(grid);
RealD aa= alpha*alpha;
RealD bb= beta * beta;
Linop.HermOp(in,out);
out = out - mu*in;
Linop.HermOp(out,tmp);
tmp = tmp - mu * out;
out = (2.0/ (aa-bb) ) * tmp - ((aa+bb)/(aa-bb))*in;
};
// Implement the required interface
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
GridBase *grid=in._grid;
int vol=grid->gSites();
Field T0(grid); T0 = in;
Field T1(grid);
Field T2(grid);
Field y(grid);
Field *Tnm = &T0;
Field *Tn = &T1;
Field *Tnp = &T2;
// Tn=T1 = (xscale M )*in
AminusMuSq(Linop,T0,T1);
// sum = .5 c[0] T0 + c[1] T1
out = (0.5*Coeffs[0])*T0 + Coeffs[1]*T1;
for(int n=2;n<order;n++){
AminusMuSq(Linop,*Tn,y);
*Tnp=2.0*y-(*Tnm);
out=out+Coeffs[n]* (*Tnp);
// Cycle pointers to avoid copies
Field *swizzle = Tnm;
Tnm =Tn;
Tn =Tnp;
Tnp =swizzle;
}
beta(_beta),
mu(_mu)
{
order=_order;
Coeffs.resize(order);
for(int i=0;i<_order;i++){
Coeffs[i] = 0.0;
}
Coeffs[order-1]=1.0;
};
}
void csv(std::ostream &out){
for (RealD x=-1.2*alpha; x<1.2*alpha; x+=(2.0*alpha)/10000) {
RealD f = approx(x);
out<< x<<" "<<f<<std::endl;
}
return;
}
RealD approx(RealD xx) // Convenience for plotting the approximation
{
RealD Tn;
RealD Tnm;
RealD Tnp;
Real aa = alpha * alpha;
Real bb = beta * beta;
RealD x = ( 2.0 * (xx-mu)*(xx-mu) - (aa+bb) ) / (aa-bb);
RealD y= x;
RealD T0=1;
RealD T1=y;
RealD sum;
sum = 0.5*Coeffs[0]*T0;
sum+= Coeffs[1]*T1;
Tn =T1;
Tnm=T0;
for(int i=2;i<order;i++){
Tnp=2*y*Tn-Tnm;
Tnm=Tn;
Tn =Tnp;
sum+= Tn*Coeffs[i];
}
return sum;
};
// shift_Multiply in Rudy's code
void AminusMuSq(LinearOperatorBase<Field> &Linop, const Field &in, Field &out)
{
GridBase *grid=in.Grid();
Field tmp(grid);
RealD aa= alpha*alpha;
RealD bb= beta * beta;
Linop.HermOp(in,out);
out = out - mu*in;
Linop.HermOp(out,tmp);
tmp = tmp - mu * out;
out = (2.0/ (aa-bb) ) * tmp - ((aa+bb)/(aa-bb))*in;
};
// Implement the required interface
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
GridBase *grid=in.Grid();
int vol=grid->gSites();
Field T0(grid); T0 = in;
Field T1(grid);
Field T2(grid);
Field y(grid);
Field *Tnm = &T0;
Field *Tn = &T1;
Field *Tnp = &T2;
// Tn=T1 = (xscale M )*in
AminusMuSq(Linop,T0,T1);
// sum = .5 c[0] T0 + c[1] T1
out = (0.5*Coeffs[0])*T0 + Coeffs[1]*T1;
for(int n=2;n<order;n++){
AminusMuSq(Linop,*Tn,y);
*Tnp=2.0*y-(*Tnm);
out=out+Coeffs[n]* (*Tnp);
// Cycle pointers to avoid copies
Field *swizzle = Tnm;
Tnm =Tn;
Tn =Tnp;
Tnp =swizzle;
}
}
};
NAMESPACE_END(Grid);
#endif

228
Grid/algorithms/approx/Forecast.h
View File

@ -26,127 +26,127 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
/* END LEGAL */
#ifndef INCLUDED_FORECAST_H
#define INCLUDED_FORECAST_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
// Abstract base class.
// Takes a matrix (Mat), a source (phi), and a vector of Fields (chi)
// and returns a forecasted solution to the system D*psi = phi (psi).
template<class Matrix, class Field>
class Forecast
// Abstract base class.
// Takes a matrix (Mat), a source (phi), and a vector of Fields (chi)
// and returns a forecasted solution to the system D*psi = phi (psi).
template<class Matrix, class Field>
class Forecast
{
public:
virtual Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& chi) = 0;
};
// Implementation of Brower et al.'s chronological inverter (arXiv:hep-lat/9509012),
// used to forecast solutions across poles of the EOFA heatbath.
//
// Modified from CPS (cps_pp/src/util/dirac_op/d_op_base/comsrc/minresext.C)
template<class Matrix, class Field>
class ChronoForecast : public Forecast<Matrix,Field>
{
public:
Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& prev_solns)
{
public:
virtual Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& chi) = 0;
int degree = prev_solns.size();
Field chi(phi); // forecasted solution
// Trivial cases
if(degree == 0){ chi = Zero(); return chi; }
else if(degree == 1){ return prev_solns[0]; }
// RealD dot;
ComplexD xp;
Field r(phi); // residual
Field Mv(phi);
std::vector<Field> v(prev_solns); // orthonormalized previous solutions
std::vector<Field> MdagMv(degree,phi);
// Array to hold the matrix elements
std::vector<std::vector<ComplexD>> G(degree, std::vector<ComplexD>(degree));
// Solution and source vectors
std::vector<ComplexD> a(degree);
std::vector<ComplexD> b(degree);
// Orthonormalize the vector basis
for(int i=0; i<degree; i++){
v[i] *= 1.0/std::sqrt(norm2(v[i]));
for(int j=i+1; j<degree; j++){ v[j] -= innerProduct(v[i],v[j]) * v[i]; }
}
// Perform sparse matrix multiplication and construct rhs
for(int i=0; i<degree; i++){
b[i] = innerProduct(v[i],phi);
Mat.M(v[i],Mv);
Mat.Mdag(Mv,MdagMv[i]);
G[i][i] = innerProduct(v[i],MdagMv[i]);
}
// Construct the matrix
for(int j=0; j<degree; j++){
for(int k=j+1; k<degree; k++){
G[j][k] = innerProduct(v[j],MdagMv[k]);
G[k][j] = conjugate(G[j][k]);
}}
// Gauss-Jordan elimination with partial pivoting
for(int i=0; i<degree; i++){
// Perform partial pivoting
int k = i;
for(int j=i+1; j<degree; j++){ if(abs(G[j][j]) > abs(G[k][k])){ k = j; } }
if(k != i){
xp = b[k];
b[k] = b[i];
b[i] = xp;
for(int j=0; j<degree; j++){
xp = G[k][j];
G[k][j] = G[i][j];
G[i][j] = xp;
}
}
// Convert matrix to upper triangular form
for(int j=i+1; j<degree; j++){
xp = G[j][i]/G[i][i];
b[j] -= xp * b[i];
for(int k=0; k<degree; k++){ G[j][k] -= xp*G[i][k]; }
}
}
// Use Gaussian elimination to solve equations and calculate initial guess
chi = Zero();
r = phi;
for(int i=degree-1; i>=0; i--){
a[i] = 0.0;
for(int j=i+1; j<degree; j++){ a[i] += G[i][j] * a[j]; }
a[i] = (b[i]-a[i])/G[i][i];
chi += a[i]*v[i];
r -= a[i]*MdagMv[i];
}
RealD true_r(0.0);
ComplexD tmp;
for(int i=0; i<degree; i++){
tmp = -b[i];
for(int j=0; j<degree; j++){ tmp += G[i][j]*a[j]; }
tmp = conjugate(tmp)*tmp;
true_r += std::sqrt(tmp.real());
}
RealD error = std::sqrt(norm2(r)/norm2(phi));
std::cout << GridLogMessage << "ChronoForecast: |res|/|src| = " << error << std::endl;
return chi;
};
};
// Implementation of Brower et al.'s chronological inverter (arXiv:hep-lat/9509012),
// used to forecast solutions across poles of the EOFA heatbath.
//
// Modified from CPS (cps_pp/src/util/dirac_op/d_op_base/comsrc/minresext.C)
template<class Matrix, class Field>
class ChronoForecast : public Forecast<Matrix,Field>
{
public:
Field operator()(Matrix &Mat, const Field& phi, const std::vector<Field>& prev_solns)
{
int degree = prev_solns.size();
Field chi(phi); // forecasted solution
// Trivial cases
if(degree == 0){ chi = zero; return chi; }
else if(degree == 1){ return prev_solns[0]; }
RealD dot;
ComplexD xp;
Field r(phi); // residual
Field Mv(phi);
std::vector<Field> v(prev_solns); // orthonormalized previous solutions
std::vector<Field> MdagMv(degree,phi);
// Array to hold the matrix elements
std::vector<std::vector<ComplexD>> G(degree, std::vector<ComplexD>(degree));
// Solution and source vectors
std::vector<ComplexD> a(degree);
std::vector<ComplexD> b(degree);
// Orthonormalize the vector basis
for(int i=0; i<degree; i++){
v[i] *= 1.0/std::sqrt(norm2(v[i]));
for(int j=i+1; j<degree; j++){ v[j] -= innerProduct(v[i],v[j]) * v[i]; }
}
// Perform sparse matrix multiplication and construct rhs
for(int i=0; i<degree; i++){
b[i] = innerProduct(v[i],phi);
Mat.M(v[i],Mv);
Mat.Mdag(Mv,MdagMv[i]);
G[i][i] = innerProduct(v[i],MdagMv[i]);
}
// Construct the matrix
for(int j=0; j<degree; j++){
for(int k=j+1; k<degree; k++){
G[j][k] = innerProduct(v[j],MdagMv[k]);
G[k][j] = std::conj(G[j][k]);
}}
// Gauss-Jordan elimination with partial pivoting
for(int i=0; i<degree; i++){
// Perform partial pivoting
int k = i;
for(int j=i+1; j<degree; j++){ if(std::abs(G[j][j]) > std::abs(G[k][k])){ k = j; } }
if(k != i){
xp = b[k];
b[k] = b[i];
b[i] = xp;
for(int j=0; j<degree; j++){
xp = G[k][j];
G[k][j] = G[i][j];
G[i][j] = xp;
}
}
// Convert matrix to upper triangular form
for(int j=i+1; j<degree; j++){
xp = G[j][i]/G[i][i];
b[j] -= xp * b[i];
for(int k=0; k<degree; k++){ G[j][k] -= xp*G[i][k]; }
}
}
// Use Gaussian elimination to solve equations and calculate initial guess
chi = zero;
r = phi;
for(int i=degree-1; i>=0; i--){
a[i] = 0.0;
for(int j=i+1; j<degree; j++){ a[i] += G[i][j] * a[j]; }
a[i] = (b[i]-a[i])/G[i][i];
chi += a[i]*v[i];
r -= a[i]*MdagMv[i];
}
RealD true_r(0.0);
ComplexD tmp;
for(int i=0; i<degree; i++){
tmp = -b[i];
for(int j=0; j<degree; j++){ tmp += G[i][j]*a[j]; }
tmp = std::conj(tmp)*tmp;
true_r += std::sqrt(tmp.real());
}
RealD error = std::sqrt(norm2(r)/norm2(phi));
std::cout << GridLogMessage << "ChronoForecast: |res|/|src| = " << error << std::endl;
return chi;
};
};
}
NAMESPACE_END(Grid);
#endif

5
Grid/algorithms/approx/MultiShiftFunction.cc
View File

@ -27,7 +27,8 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
/* END LEGAL */
#include <Grid/GridCore.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
double MultiShiftFunction::approx(double x)
{
double a = norm;
@ -53,4 +54,4 @@ void MultiShiftFunction::csv(std::ostream &out)
}
return;
}
}
NAMESPACE_END(Grid);

4
Grid/algorithms/approx/MultiShiftFunction.h
View File

@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#ifndef MULTI_SHIFT_FUNCTION
#define MULTI_SHIFT_FUNCTION
namespace Grid {
NAMESPACE_BEGIN(Grid);
class MultiShiftFunction {
public:
@ -63,5 +63,5 @@ public:
}
};
}
NAMESPACE_END(Grid);
#endif

3
Grid/algorithms/approx/Remez.cc
View File

@ -298,7 +298,7 @@ void AlgRemez::stpini(bigfloat *step) {
// Search for error maxima and minima
void AlgRemez::search(bigfloat *step) {
bigfloat a, q, xm, ym, xn, yn, xx0, xx1;
int i, j, meq, emsign, ensign, steps;
int i, meq, emsign, ensign, steps;
meq = neq + 1;
bigfloat *yy = new bigfloat[meq];
@ -306,7 +306,6 @@ void AlgRemez::search(bigfloat *step) {
bigfloat eclose = 1.0e30;
bigfloat farther = 0l;
j = 1;
xx0 = apstrt;
for (i = 0; i < meq; i++) {

11
Grid/algorithms/approx/Zolotarev.cc
View File

@ -58,8 +58,8 @@
/* Compute the partial fraction expansion coefficients (alpha) from the
* factored form */
namespace Grid {
namespace Approx {
NAMESPACE_BEGIN(Grid);
NAMESPACE_BEGIN(Approx);
static void construct_partfrac(izd *z) {
int dn = z -> dn, dd = z -> dd, type = z -> type;
@ -516,7 +516,9 @@ zolotarev_data* higham(PRECISION epsilon, int n) {
free(d);
return zd;
}
}}
NAMESPACE_END(Approx);
NAMESPACE_END(Grid);
#ifdef TEST
@ -585,6 +587,7 @@ static PRECISION zolotarev_cayley_eval(PRECISION x, zolotarev_data* rdata) {
return (ONE - T) / (ONE + T);
}
/* Test program. Apart from printing out the parameters for R(x) it produces
* the following data files for plotting (unless NPLOT is defined):
*
@ -723,5 +726,5 @@ int main(int argc, char** argv) {
return EXIT_SUCCESS;
}
#endif /* TEST */

9
Grid/algorithms/approx/Zolotarev.h
View File

@ -1,13 +1,13 @@
/* -*- Mode: C; comment-column: 22; fill-column: 79; -*- */
#ifdef __cplusplus
namespace Grid {
namespace Approx {
#include <Grid/Namespace.h>
NAMESPACE_BEGIN(Grid);
NAMESPACE_BEGIN(Approx);
#endif
#define HVERSION Header Time-stamp: <14-OCT-2004 09:26:51.00 adk@MISSCONTRARY>
#ifndef ZOLOTAREV_INTERNAL
#ifndef PRECISION
#define PRECISION double
@ -83,5 +83,6 @@ void zolotarev_free(zolotarev_data *zdata);
#endif
#ifdef __cplusplus
}}
NAMESPACE_END(Approx);
NAMESPACE_END(Grid);
#endif

4
Grid/algorithms/approx/bigfloat.h
View File

@ -10,10 +10,12 @@
#ifndef INCLUDED_BIGFLOAT_H
#define INCLUDED_BIGFLOAT_H
#define __GMP_WITHIN_CONFIGURE
#include <gmp.h>
#include <mpf2mpfr.h>
#include <mpfr.h>
#undef __GMP_WITHIN_CONFIGURE
class bigfloat {
private:

4
Grid/algorithms/iterative/AdefGeneric.h
View File

@ -90,8 +90,8 @@ class TwoLevelFlexiblePcg : public LinearFunction<Field>
void operator() (const Field &src, Field &psi){
void operator() (const Field &src, Field &psi){
psi.checkerboard = src.checkerboard;
grid = src._grid;
psi.Checkerboard() = src.Checkerboard();
grid = src.Grid();
RealD f;
RealD rtzp,rtz,a,d,b;

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

@ -27,11 +27,9 @@ See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_BLOCK_CONJUGATE_GRADIENT_H
#define GRID_BLOCK_CONJUGATE_GRADIENT_H
#pragma once
namespace Grid {
NAMESPACE_BEGIN(Grid);
enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS, BlockCGVec, BlockCGrQVec };
@ -154,12 +152,12 @@ virtual void operator()(LinearOperatorBase<Field> &Linop, const std::vector<Fiel
void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
{
int Orthog = blockDim; // First dimension is block dim; this is an assumption
Nblock = B._grid->_fdimensions[Orthog];
Nblock = B.Grid()->_fdimensions[Orthog];
/* FAKE */
Nblock=8;
std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
X.checkerboard = B.checkerboard;
X.Checkerboard() = B.Checkerboard();
conformable(X, B);
Field tmp(B);
@ -334,11 +332,11 @@ void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
{
int Orthog = blockDim; // First dimension is block dim
Nblock = Src._grid->_fdimensions[Orthog];
Nblock = Src.Grid()->_fdimensions[Orthog];
std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
Psi.checkerboard = Src.checkerboard;
Psi.Checkerboard() = Src.Checkerboard();
conformable(Psi, Src);
Field P(Src);
@ -478,7 +476,7 @@ void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<
for(int b=0;b<Nblock;b++){
tmp[b] = Y[b];
for(int bp=0;bp<Nblock;bp++) {
tmp[b] = tmp[b] + (scale*m(bp,b))*X[bp];
tmp[b] = tmp[b] + scomplex(scale*m(bp,b))*X[bp];
}
}
for(int b=0;b<Nblock;b++){
@ -488,9 +486,9 @@ void MaddMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<
void MulMatrix(std::vector<Field> &AP, Eigen::MatrixXcd &m , const std::vector<Field> &X){
// Should make this cache friendly with site outermost, parallel_for
for(int b=0;b<Nblock;b++){
AP[b] = zero;
AP[b] = Zero();
for(int bp=0;bp<Nblock;bp++) {
AP[b] += (m(bp,b))*X[bp];
AP[b] += scomplex(m(bp,b))*X[bp];
}
}
}
@ -517,7 +515,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
std::cout<<GridLogMessage<<" Block Conjugate Gradient Vec rQ : Nblock "<<Nblock<<std::endl;
for(int b=0;b<Nblock;b++){
X[b].checkerboard = B[b].checkerboard;
X[b].Checkerboard() = B[b].Checkerboard();
conformable(X[b], B[b]);
conformable(X[b], X[0]);
}
@ -690,9 +688,7 @@ void BlockCGrQsolveVec(LinearOperatorBase<Field> &Linop, const std::vector<Field
IterationsToComplete = k;
}
};
}
#endif
NAMESPACE_END(Grid);

42
Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h
View File

@ -34,6 +34,8 @@ namespace Grid {
template<class Field>
class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<Field> {
public:
using OperatorFunction<Field>::operator();
bool ErrorOnNoConverge; // Throw an assert when CAGMRES fails to converge,
// defaults to true
@ -52,10 +54,10 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
std::vector<ComplexD> y;
std::vector<ComplexD> gamma;
std::vector<ComplexD> c;
std::vector<ComplexD> s;
CommunicationAvoidingGeneralisedMinimalResidual(RealD tol,
Integer maxit,
@ -76,7 +78,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
std::cout << GridLogWarning << "This algorithm currently doesn't differ from regular GMRES" << std::endl;
psi.checkerboard = src.checkerboard;
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
RealD guess = norm2(psi);
@ -86,7 +88,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Field r(src._grid);
Field r(src.Grid());
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: guess " << guess << std::endl;
@ -142,11 +144,11 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
RealD cp = 0;
Field w(src._grid);
Field r(src._grid);
Field w(src.Grid());
Field r(src.Grid());
// this should probably be made a class member so that it is only allocated once, not in every restart
std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
std::vector<Field> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
MatrixTimer.Start();
LinOp.Op(psi, w);
@ -157,7 +159,9 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
gamma[0] = sqrt(norm2(r));
v[0] = (1. / gamma[0]) * r;
ComplexD scale = 1.0/gamma[0];
v[0] = scale * r;
LinalgTimer.Stop();
for (int i=0; i<RestartLength; i++) {
@ -168,7 +172,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
qrUpdate(i);
cp = std::norm(gamma[i+1]);
cp = norm(gamma[i+1]);
std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
@ -194,11 +198,11 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
w = w - ComplexD(H(iter, i)) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
@ -206,13 +210,13 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
auto tmp = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
H(iter, i) = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
@ -221,7 +225,7 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
gamma[iter] = conjugate(c[iter]) * gamma[iter];
QrTimer.Stop();
}
@ -231,8 +235,8 @@ class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
y[i] = y[i] - ComplexD(H(k, i)) * y[k];
y[i] = y[i] / ComplexD(H(i, i));
}
for (int i = 0; i <= iter; i++)

41
Grid/algorithms/iterative/ConjugateGradient.h
View File

@ -27,11 +27,11 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
/* END LEGAL */
#ifndef GRID_CONJUGATE_GRADIENT_H
#define GRID_CONJUGATE_GRADIENT_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
@ -40,7 +40,10 @@ namespace Grid {
template <class Field>
class ConjugateGradient : public OperatorFunction<Field> {
public:
public:
using OperatorFunction<Field>::operator();
bool ErrorOnNoConverge; // throw an assert when the CG fails to converge.
// Defaults true.
RealD Tolerance;
@ -48,17 +51,18 @@ class ConjugateGradient : public OperatorFunction<Field> {
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
: Tolerance(tol),
MaxIterations(maxit),
ErrorOnNoConverge(err_on_no_conv){};
: Tolerance(tol),
MaxIterations(maxit),
ErrorOnNoConverge(err_on_no_conv){};
void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
psi.Checkerboard() = src.Checkerboard();
psi.checkerboard = src.checkerboard;
conformable(psi, src);
RealD cp, c, a, d, b, ssq, qq, b_pred;
RealD cp, c, a, d, b, ssq, qq;
//RealD b_pred;
Field p(src);
Field mmp(src);
@ -70,7 +74,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
Linop.HermOpAndNorm(psi, mmp, d, b);
r = src - mmp;
p = r;
@ -127,10 +131,13 @@ class ConjugateGradient : public OperatorFunction<Field> {
b = cp / c;
LinearCombTimer.Start();
parallel_for(int ss=0;ss<src._grid->oSites();ss++){
vstream(psi[ss], a * p[ss] + psi[ss]);
vstream(p [ss], b * p[ss] + r[ss]);
}
auto psi_v = psi.View();
auto p_v = p.View();
auto r_v = r.View();
accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(psi_v[ss], a * p_v(ss) + psi_v(ss));
coalescedWrite(p_v[ss] , b * p_v(ss) + r_v (ss));
});
LinearCombTimer.Stop();
LinalgTimer.Stop();
@ -143,12 +150,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
Linop.HermOpAndNorm(psi, mmp, d, qq);
p = mmp - src;
RealD srcnorm = sqrt(norm2(src));
RealD resnorm = sqrt(norm2(p));
RealD srcnorm = std::sqrt(norm2(src));
RealD resnorm = std::sqrt(norm2(p));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k << std::endl;
std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
@ -174,5 +181,5 @@ class ConjugateGradient : public OperatorFunction<Field> {
}
};
}
NAMESPACE_END(Grid);
#endif

151
Grid/algorithms/iterative/ConjugateGradientMixedPrec.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,13 +23,12 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
//Mixed precision restarted defect correction CG
template<class FieldD,class FieldF,
@ -67,98 +66,96 @@ namespace Grid {
guesser = &g;
}
void operator() (const FieldD &src_d_in, FieldD &sol_d){
TotalInnerIterations = 0;
void operator() (const FieldD &src_d_in, FieldD &sol_d){
TotalInnerIterations = 0;
GridStopWatch TotalTimer;
TotalTimer.Start();
GridStopWatch TotalTimer;
TotalTimer.Start();
int cb = src_d_in.checkerboard;
sol_d.checkerboard = cb;
int cb = src_d_in.Checkerboard();
sol_d.Checkerboard() = cb;
RealD src_norm = norm2(src_d_in);
RealD stop = src_norm * Tolerance*Tolerance;
RealD src_norm = norm2(src_d_in);
RealD stop = src_norm * Tolerance*Tolerance;
GridBase* DoublePrecGrid = src_d_in._grid;
FieldD tmp_d(DoublePrecGrid);
tmp_d.checkerboard = cb;
GridBase* DoublePrecGrid = src_d_in.Grid();
FieldD tmp_d(DoublePrecGrid);
tmp_d.Checkerboard() = cb;
FieldD tmp2_d(DoublePrecGrid);
tmp2_d.checkerboard = cb;
FieldD tmp2_d(DoublePrecGrid);
tmp2_d.Checkerboard() = cb;
FieldD src_d(DoublePrecGrid);
src_d = src_d_in; //source for next inner iteration, computed from residual during operation
FieldD src_d(DoublePrecGrid);
src_d = src_d_in; //source for next inner iteration, computed from residual during operation
RealD inner_tol = InnerTolerance;
RealD inner_tol = InnerTolerance;
FieldF src_f(SinglePrecGrid);
src_f.checkerboard = cb;
FieldF src_f(SinglePrecGrid);
src_f.Checkerboard() = cb;
FieldF sol_f(SinglePrecGrid);
sol_f.checkerboard = cb;
FieldF sol_f(SinglePrecGrid);
sol_f.Checkerboard() = cb;
ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
CG_f.ErrorOnNoConverge = false;
ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
CG_f.ErrorOnNoConverge = false;
GridStopWatch InnerCGtimer;
GridStopWatch InnerCGtimer;
GridStopWatch PrecChangeTimer;
GridStopWatch PrecChangeTimer;
Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
//Compute double precision rsd and also new RHS vector.
Linop_d.HermOp(sol_d, tmp_d);
RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
//Compute double precision rsd and also new RHS vector.
Linop_d.HermOp(sol_d, tmp_d);
RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
if(norm < OuterLoopNormMult * stop){
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
break;
}
while(norm * inner_tol * inner_tol < stop) inner_tol *= 2; // inner_tol = sqrt(stop/norm) ??
PrecChangeTimer.Start();
precisionChange(src_f, src_d);
PrecChangeTimer.Stop();
zeroit(sol_f);
//Optionally improve inner solver guess (eg using known eigenvectors)
if(guesser != NULL)
(*guesser)(src_f, sol_f);
//Inner CG
CG_f.Tolerance = inner_tol;
InnerCGtimer.Start();
CG_f(Linop_f, src_f, sol_f);
InnerCGtimer.Stop();
TotalInnerIterations += CG_f.IterationsToComplete;
//Convert sol back to double and add to double prec solution
PrecChangeTimer.Start();
precisionChange(tmp_d, sol_f);
PrecChangeTimer.Stop();
axpy(sol_d, 1.0, tmp_d, sol_d);
if(norm < OuterLoopNormMult * stop){
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
break;
}
//Final trial CG
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting final patch-up double-precision solve"<<std::endl;
ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
CG_d(Linop_d, src_d_in, sol_d);
TotalFinalStepIterations = CG_d.IterationsToComplete;
while(norm * inner_tol * inner_tol < stop) inner_tol *= 2; // inner_tol = sqrt(stop/norm) ??
TotalTimer.Stop();
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
PrecChangeTimer.Start();
precisionChange(src_f, src_d);
PrecChangeTimer.Stop();
sol_f = Zero();
//Optionally improve inner solver guess (eg using known eigenvectors)
if(guesser != NULL)
(*guesser)(src_f, sol_f);
//Inner CG
CG_f.Tolerance = inner_tol;
InnerCGtimer.Start();
CG_f(Linop_f, src_f, sol_f);
InnerCGtimer.Stop();
TotalInnerIterations += CG_f.IterationsToComplete;
//Convert sol back to double and add to double prec solution
PrecChangeTimer.Start();
precisionChange(tmp_d, sol_f);
PrecChangeTimer.Stop();
axpy(sol_d, 1.0, tmp_d, sol_d);
}
};
//Final trial CG
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting final patch-up double-precision solve"<<std::endl;
ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
CG_d(Linop_d, src_d_in, sol_d);
TotalFinalStepIterations = CG_d.IterationsToComplete;
TotalTimer.Stop();
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
}
};
}
NAMESPACE_END(Grid);
#endif

423
Grid/algorithms/iterative/ConjugateGradientMultiShift.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,147 +24,150 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CONJUGATE_MULTI_SHIFT_GRADIENT_H
#define GRID_CONJUGATE_MULTI_SHIFT_GRADIENT_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
// single input vec, single output vec.
/////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
// single input vec, single output vec.
/////////////////////////////////////////////////////////////
template<class Field>
class ConjugateGradientMultiShift : public OperatorMultiFunction<Field>,
public OperatorFunction<Field>
{
template<class Field>
class ConjugateGradientMultiShift : public OperatorMultiFunction<Field>,
public OperatorFunction<Field>
{
public:
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
int verbose;
MultiShiftFunction shifts;
ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) :
MaxIterations(maxit),
shifts(_shifts)
{
verbose=1;
using OperatorFunction<Field>::operator();
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
int verbose;
MultiShiftFunction shifts;
ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) :
MaxIterations(maxit),
shifts(_shifts)
{
verbose=1;
}
void operator() (LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
{
GridBase *grid = src.Grid();
int nshift = shifts.order;
std::vector<Field> results(nshift,grid);
(*this)(Linop,src,results,psi);
}
void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &results, Field &psi)
{
int nshift = shifts.order;
(*this)(Linop,src,results);
psi = shifts.norm*src;
for(int i=0;i<nshift;i++){
psi = psi + shifts.residues[i]*results[i];
}
void operator() (LinearOperatorBase<Field> &Linop, const Field &src, Field &psi)
{
GridBase *grid = src._grid;
int nshift = shifts.order;
std::vector<Field> results(nshift,grid);
(*this)(Linop,src,results,psi);
}
void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &results, Field &psi)
{
int nshift = shifts.order;
(*this)(Linop,src,results);
psi = shifts.norm*src;
for(int i=0;i<nshift;i++){
psi = psi + shifts.residues[i]*results[i];
return;
}
return;
}
void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
{
GridBase *grid = src.Grid();
////////////////////////////////////////////////////////////////////////
// Convenience references to the info stored in "MultiShiftFunction"
////////////////////////////////////////////////////////////////////////
int nshift = shifts.order;
void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector<Field> &psi)
{
GridBase *grid = src._grid;
////////////////////////////////////////////////////////////////////////
// Convenience references to the info stored in "MultiShiftFunction"
////////////////////////////////////////////////////////////////////////
int nshift = shifts.order;
std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
std::vector<RealD> &mresidual(shifts.tolerances);
std::vector<RealD> alpha(nshift,1.0);
std::vector<Field> ps(nshift,grid);// Search directions
std::vector<RealD> &mass(shifts.poles); // Make references to array in "shifts"
std::vector<RealD> &mresidual(shifts.tolerances);
std::vector<RealD> alpha(nshift,1.0);
std::vector<Field> ps(nshift,grid);// Search directions
assert(psi.size()==nshift);
assert(mass.size()==nshift);
assert(mresidual.size()==nshift);
assert(psi.size()==nshift);
assert(mass.size()==nshift);
assert(mresidual.size()==nshift);
// dynamic sized arrays on stack; 2d is a pain with vector
RealD bs[nshift];
RealD rsq[nshift];
RealD z[nshift][2];
int converged[nshift];
// dynamic sized arrays on stack; 2d is a pain with vector
RealD bs[nshift];
RealD rsq[nshift];
RealD z[nshift][2];
int converged[nshift];
const int primary =0;
const int primary =0;
//Primary shift fields CG iteration
RealD a,b,c,d;
RealD cp,bp,qq; //prev
//Primary shift fields CG iteration
RealD a,b,c,d;
RealD cp,bp,qq; //prev
// Matrix mult fields
Field r(grid);
Field p(grid);
Field tmp(grid);
Field mmp(grid);
// Matrix mult fields
Field r(grid);
Field p(grid);
Field tmp(grid);
Field mmp(grid);
// Check lightest mass
for(int s=0;s<nshift;s++){
assert( mass[s]>= mass[primary] );
converged[s]=0;
}
// Check lightest mass
for(int s=0;s<nshift;s++){
assert( mass[s]>= mass[primary] );
converged[s]=0;
}
// Wire guess to zero
// Residuals "r" are src
// First search direction "p" is also src
cp = norm2(src);
for(int s=0;s<nshift;s++){
rsq[s] = cp * mresidual[s] * mresidual[s];
std::cout<<GridLogMessage<<"ConjugateGradientMultiShift: shift "<<s
<<" target resid "<<rsq[s]<<std::endl;
ps[s] = src;
}
// r and p for primary
r=src;
p=src;
// Wire guess to zero
// Residuals "r" are src
// First search direction "p" is also src
cp = norm2(src);
for(int s=0;s<nshift;s++){
rsq[s] = cp * mresidual[s] * mresidual[s];
std::cout<<GridLogMessage<<"ConjugateGradientMultiShift: shift "<<s
<<" target resid "<<rsq[s]<<std::endl;
ps[s] = src;
}
// r and p for primary
r=src;
p=src;
//MdagM+m[0]
Linop.HermOpAndNorm(p,mmp,d,qq);
axpy(mmp,mass[0],p,mmp);
RealD rn = norm2(p);
d += rn*mass[0];
//MdagM+m[0]
Linop.HermOpAndNorm(p,mmp,d,qq);
axpy(mmp,mass[0],p,mmp);
RealD rn = norm2(p);
d += rn*mass[0];
// have verified that inner product of
// p and mmp is equal to d after this since
// the d computation is tricky
// qq = real(innerProduct(p,mmp));
// std::cout<<GridLogMessage << "debug equal ? qq "<<qq<<" d "<< d<<std::endl;
// have verified that inner product of
// p and mmp is equal to d after this since
// the d computation is tricky
// qq = real(innerProduct(p,mmp));
// std::cout<<GridLogMessage << "debug equal ? qq "<<qq<<" d "<< d<<std::endl;
b = -cp /d;
b = -cp /d;
// Set up the various shift variables
int iz=0;
z[0][1-iz] = 1.0;
z[0][iz] = 1.0;
bs[0] = b;
for(int s=1;s<nshift;s++){
z[s][1-iz] = 1.0;
z[s][iz] = 1.0/( 1.0 - b*(mass[s]-mass[0]));
bs[s] = b*z[s][iz];
}
// Set up the various shift variables
int iz=0;
z[0][1-iz] = 1.0;
z[0][iz] = 1.0;
bs[0] = b;
for(int s=1;s<nshift;s++){
z[s][1-iz] = 1.0;
z[s][iz] = 1.0/( 1.0 - b*(mass[s]-mass[0]));
bs[s] = b*z[s][iz];
}
// r += b[0] A.p[0]
// c= norm(r)
c=axpy_norm(r,b,mmp,r);
// r += b[0] A.p[0]
// c= norm(r)
c=axpy_norm(r,b,mmp,r);
for(int s=0;s<nshift;s++) {
axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
}
for(int s=0;s<nshift;s++) {
axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
}
///////////////////////////////////////
// Timers
///////////////////////////////////////
@ -175,37 +178,37 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
GridStopWatch SolverTimer;
SolverTimer.Start();
// Iteration loop
int k;
// Iteration loop
int k;
for (k=1;k<=MaxIterations;k++){
for (k=1;k<=MaxIterations;k++){
a = c /cp;
a = c /cp;
AXPYTimer.Start();
axpy(p,a,p,r);
axpy(p,a,p,r);
AXPYTimer.Stop();
// Note to self - direction ps is iterated seperately
// for each shift. Does not appear to have any scope
// for avoiding linear algebra in "single" case.
//
// However SAME r is used. Could load "r" and update
// ALL ps[s]. 2/3 Bandwidth saving
// New Kernel: Load r, vector of coeffs, vector of pointers ps
// Note to self - direction ps is iterated seperately
// for each shift. Does not appear to have any scope
// for avoiding linear algebra in "single" case.
//
// However SAME r is used. Could load "r" and update
// ALL ps[s]. 2/3 Bandwidth saving
// New Kernel: Load r, vector of coeffs, vector of pointers ps
AXPYTimer.Start();
for(int s=0;s<nshift;s++){
if ( ! converged[s] ) {
if (s==0){
axpy(ps[s],a,ps[s],r);
} else{
RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
axpby(ps[s],z[s][iz],as,r,ps[s]);
for(int s=0;s<nshift;s++){
if ( ! converged[s] ) {
if (s==0){
axpy(ps[s],a,ps[s],r);
} else{
RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
axpby(ps[s],z[s][iz],as,r,ps[s]);
}
}
}
}
AXPYTimer.Stop();
cp=c;
cp=c;
MatrixTimer.Start();
//Linop.HermOpAndNorm(p,mmp,d,qq); // d is used
// The below is faster on KNL
@ -215,89 +218,89 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
MatrixTimer.Stop();
AXPYTimer.Start();
axpy(mmp,mass[0],p,mmp);
axpy(mmp,mass[0],p,mmp);
AXPYTimer.Stop();
RealD rn = norm2(p);
d += rn*mass[0];
RealD rn = norm2(p);
d += rn*mass[0];
bp=b;
b=-cp/d;
bp=b;
b=-cp/d;
AXPYTimer.Start();
c=axpy_norm(r,b,mmp,r);
c=axpy_norm(r,b,mmp,r);
AXPYTimer.Stop();
// Toggle the recurrence history
bs[0] = b;
iz = 1-iz;
// Toggle the recurrence history
bs[0] = b;
iz = 1-iz;
ShiftTimer.Start();
for(int s=1;s<nshift;s++){
if((!converged[s])){
RealD z0 = z[s][1-iz];
RealD z1 = z[s][iz];
z[s][iz] = z0*z1*bp
/ (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b));
bs[s] = b*z[s][iz]/z0; // NB sign rel to Mike
for(int s=1;s<nshift;s++){
if((!converged[s])){
RealD z0 = z[s][1-iz];
RealD z1 = z[s][iz];
z[s][iz] = z0*z1*bp
/ (b*a*(z1-z0) + z1*bp*(1- (mass[s]-mass[0])*b));
bs[s] = b*z[s][iz]/z0; // NB sign rel to Mike
}
}
}
ShiftTimer.Stop();
for(int s=0;s<nshift;s++){
int ss = s;
// Scope for optimisation here in case of "single".
// Could load psi[0] and pull all ps[s] in.
// if ( single ) ss=primary;
// Bandwith saving in single case is Ls * 3 -> 2+Ls, so ~ 3x saving
// Pipelined CG gain:
//
// New Kernel: Load r, vector of coeffs, vector of pointers ps
// New Kernel: Load psi[0], vector of coeffs, vector of pointers ps
// If can predict the coefficient bs then we can fuse these and avoid write reread cyce
// on ps[s].
// Before: 3 x npole + 3 x npole
// After : 2 x npole (ps[s]) => 3x speed up of multishift CG.
for(int s=0;s<nshift;s++){
int ss = s;
// Scope for optimisation here in case of "single".
// Could load psi[0] and pull all ps[s] in.
// if ( single ) ss=primary;
// Bandwith saving in single case is Ls * 3 -> 2+Ls, so ~ 3x saving
// Pipelined CG gain:
//
// New Kernel: Load r, vector of coeffs, vector of pointers ps
// New Kernel: Load psi[0], vector of coeffs, vector of pointers ps
// If can predict the coefficient bs then we can fuse these and avoid write reread cyce
// on ps[s].
// Before: 3 x npole + 3 x npole
// After : 2 x npole (ps[s]) => 3x speed up of multishift CG.
if( (!converged[s]) ) {
axpy(psi[ss],-bs[s]*alpha[s],ps[s],psi[ss]);
}
}
// Convergence checks
int all_converged = 1;
for(int s=0;s<nshift;s++){
if ( (!converged[s]) ){
RealD css = c * z[s][iz]* z[s][iz];
if(css<rsq[s]){
if ( ! converged[s] )
std::cout<<GridLogMessage<<"ConjugateGradientMultiShift k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
converged[s]=1;
} else {
all_converged=0;
if( (!converged[s]) ) {
axpy(psi[ss],-bs[s]*alpha[s],ps[s],psi[ss]);
}
}
}
if ( all_converged ){
// Convergence checks
int all_converged = 1;
for(int s=0;s<nshift;s++){
if ( (!converged[s]) ){
RealD css = c * z[s][iz]* z[s][iz];
if(css<rsq[s]){
if ( ! converged[s] )
std::cout<<GridLogMessage<<"ConjugateGradientMultiShift k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
converged[s]=1;
} else {
all_converged=0;
}
}
}
if ( all_converged ){
SolverTimer.Stop();
std::cout<<GridLogMessage<< "CGMultiShift: All shifts have converged iteration "<<k<<std::endl;
std::cout<<GridLogMessage<< "CGMultiShift: Checking solutions"<<std::endl;
std::cout<<GridLogMessage<< "CGMultiShift: All shifts have converged iteration "<<k<<std::endl;
std::cout<<GridLogMessage<< "CGMultiShift: Checking solutions"<<std::endl;
// Check answers
for(int s=0; s < nshift; s++) {
Linop.HermOpAndNorm(psi[s],mmp,d,qq);
axpy(tmp,mass[s],psi[s],mmp);
axpy(r,-alpha[s],src,tmp);
RealD rn = norm2(r);
RealD cn = norm2(src);
std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(rn/cn)<<std::endl;
}
// Check answers
for(int s=0; s < nshift; s++) {
Linop.HermOpAndNorm(psi[s],mmp,d,qq);
axpy(tmp,mass[s],psi[s],mmp);
axpy(r,-alpha[s],src,tmp);
RealD rn = norm2(r);
RealD cn = norm2(src);
std::cout<<GridLogMessage<<"CGMultiShift: shift["<<s<<"] true residual "<<std::sqrt(rn/cn)<<std::endl;
}
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
@ -307,16 +310,16 @@ void operator() (LinearOperatorBase<Field> &Linop, const Field &src, std::vector
IterationsToComplete = k;
return;
}
return;
}
}
// ugly hack
std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
// assert(0);
}
// ugly hack
std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
// assert(0);
}
};
}
};
NAMESPACE_END(Grid);
#endif

394
Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,234 +23,236 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
#define GRID_CONJUGATE_GRADIENT_RELIABLE_UPDATE_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
template<class FieldD,class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
public:
bool ErrorOnNoConverge; // throw an assert when the CG fails to converge.
// Defaults true.
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
Integer ReliableUpdatesPerformed;
template<class FieldD,class FieldF,
typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class ConjugateGradientReliableUpdate : public LinearFunction<FieldD> {
public:
bool ErrorOnNoConverge; // throw an assert when the CG fails to converge.
// Defaults true.
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
Integer ReliableUpdatesPerformed;
bool DoFinalCleanup; //Final DP cleanup, defaults to true
Integer IterationsToCleanup; //Final DP cleanup step iterations
bool DoFinalCleanup; //Final DP cleanup, defaults to true
Integer IterationsToCleanup; //Final DP cleanup step iterations
LinearOperatorBase<FieldF> &Linop_f;
LinearOperatorBase<FieldD> &Linop_d;
GridBase* SinglePrecGrid;
RealD Delta; //reliable update parameter
LinearOperatorBase<FieldF> &Linop_f;
LinearOperatorBase<FieldD> &Linop_d;
GridBase* SinglePrecGrid;
RealD Delta; //reliable update parameter
//Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
LinearOperatorBase<FieldF> *Linop_fallback;
RealD fallback_transition_tol;
//Optional ability to switch to a different linear operator once the tolerance reaches a certain point. Useful for single/half -> single/single
LinearOperatorBase<FieldF> *Linop_fallback;
RealD fallback_transition_tol;
ConjugateGradientReliableUpdate(RealD tol, Integer maxit, RealD _delta, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d, bool err_on_no_conv = true)
: Tolerance(tol),
MaxIterations(maxit),
Delta(_delta),
Linop_f(_Linop_f),
Linop_d(_Linop_d),
SinglePrecGrid(_sp_grid),
ErrorOnNoConverge(err_on_no_conv),
DoFinalCleanup(true),
Linop_fallback(NULL)
{};
ConjugateGradientReliableUpdate(RealD tol, Integer maxit, RealD _delta, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d, bool err_on_no_conv = true)
: Tolerance(tol),
MaxIterations(maxit),
Delta(_delta),
Linop_f(_Linop_f),
Linop_d(_Linop_d),
SinglePrecGrid(_sp_grid),
ErrorOnNoConverge(err_on_no_conv),
DoFinalCleanup(true),
Linop_fallback(NULL)
{};
void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
Linop_fallback = &_Linop_fallback;
fallback_transition_tol = _fallback_transition_tol;
}
void setFallbackLinop(LinearOperatorBase<FieldF> &_Linop_fallback, const RealD _fallback_transition_tol){
Linop_fallback = &_Linop_fallback;
fallback_transition_tol = _fallback_transition_tol;
}
void operator()(const FieldD &src, FieldD &psi) {
LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
bool using_fallback = false;
void operator()(const FieldD &src, FieldD &psi) {
LinearOperatorBase<FieldF> *Linop_f_use = &Linop_f;
bool using_fallback = false;
psi.checkerboard = src.checkerboard;
conformable(psi, src);
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
RealD cp, c, a, d, b, ssq, qq, b_pred;
RealD cp, c, a, d, b, ssq, qq, b_pred;
FieldD p(src);
FieldD mmp(src);
FieldD r(src);
FieldD p(src);
FieldD mmp(src);
FieldD r(src);
// Initial residual computation & set up
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
// Initial residual computation & set up
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
Linop_d.HermOpAndNorm(psi, mmp, d, b);
Linop_d.HermOpAndNorm(psi, mmp, d, b);
r = src - mmp;
p = r;
r = src - mmp;
p = r;
a = norm2(p);
cp = a;
ssq = norm2(src);
a = norm2(p);
cp = a;
ssq = norm2(src);
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: guess " << guess << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: src " << ssq << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: mp " << d << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: mmp " << b << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: cp,r " << cp << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: p " << a << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: guess " << guess << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: src " << ssq << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: mp " << d << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: mmp " << b << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: cp,r " << cp << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradientReliableUpdate: p " << a << std::endl;
RealD rsq = Tolerance * Tolerance * ssq;
RealD rsq = Tolerance * Tolerance * ssq;
// Check if guess is really REALLY good :)
// Check if guess is really REALLY good :)
if (cp <= rsq) {
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate guess was REALLY good\n";
std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
return;
}
//Single prec initialization
FieldF r_f(SinglePrecGrid);
r_f.Checkerboard() = r.Checkerboard();
precisionChange(r_f, r);
FieldF psi_f(r_f);
psi_f = Zero();
FieldF p_f(r_f);
FieldF mmp_f(r_f);
RealD MaxResidSinceLastRelUp = cp; //initial residual
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
GridStopWatch LinalgTimer;
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
SolverTimer.Start();
int k = 0;
int l = 0;
for (k = 1; k <= MaxIterations; k++) {
c = cp;
MatrixTimer.Start();
Linop_f_use->HermOpAndNorm(p_f, mmp_f, d, qq);
MatrixTimer.Stop();
LinalgTimer.Start();
a = c / d;
b_pred = a * (a * qq - d) / c;
cp = axpy_norm(r_f, -a, mmp_f, r_f);
b = cp / c;
// Fuse these loops ; should be really easy
psi_f = a * p_f + psi_f;
//p_f = p_f * b + r_f;
LinalgTimer.Stop();
std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: Iteration " << k
<< " residual " << cp << " target " << rsq << std::endl;
std::cout << GridLogDebug << "a = "<< a << " b_pred = "<< b_pred << " b = "<< b << std::endl;
std::cout << GridLogDebug << "qq = "<< qq << " d = "<< d << " c = "<< c << std::endl;
if(cp > MaxResidSinceLastRelUp){
std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: updating MaxResidSinceLastRelUp : " << MaxResidSinceLastRelUp << " -> " << cp << std::endl;
MaxResidSinceLastRelUp = cp;
}
// Stopping condition
if (cp <= rsq) {
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate guess was REALLY good\n";
std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
//Although not written in the paper, I assume that I have to add on the final solution
precisionChange(mmp, psi_f);
psi = psi + mmp;
SolverTimer.Stop();
Linop_d.HermOpAndNorm(psi, mmp, d, qq);
p = mmp - src;
RealD srcnorm = std::sqrt(norm2(src));
RealD resnorm = std::sqrt(norm2(p));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate Converged on iteration " << k << " after " << l << " reliable updates" << std::endl;
std::cout << GridLogMessage << "\tComputed residual " << std::sqrt(cp / ssq)<<std::endl;
std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
std::cout << GridLogMessage << "Time breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tLinalg " << LinalgTimer.Elapsed() <<std::endl;
IterationsToComplete = k;
ReliableUpdatesPerformed = l;
if(DoFinalCleanup){
//Do a final CG to cleanup
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate performing final cleanup.\n";
ConjugateGradient<FieldD> CG(Tolerance,MaxIterations);
CG.ErrorOnNoConverge = ErrorOnNoConverge;
CG(Linop_d,src,psi);
IterationsToCleanup = CG.IterationsToComplete;
}
else if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate complete.\n";
return;
}
else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
<< cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
precisionChange(mmp, psi_f);
psi = psi + mmp;
//Single prec initialization
FieldF r_f(SinglePrecGrid);
r_f.checkerboard = r.checkerboard;
precisionChange(r_f, r);
Linop_d.HermOpAndNorm(psi, mmp, d, qq);
r = src - mmp;
FieldF psi_f(r_f);
psi_f = zero;
psi_f = Zero();
precisionChange(r_f, r);
cp = norm2(r);
MaxResidSinceLastRelUp = cp;
FieldF p_f(r_f);
FieldF mmp_f(r_f);
RealD MaxResidSinceLastRelUp = cp; //initial residual
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
GridStopWatch LinalgTimer;
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
SolverTimer.Start();
int k = 0;
int l = 0;
for (k = 1; k <= MaxIterations; k++) {
c = cp;
MatrixTimer.Start();
Linop_f_use->HermOpAndNorm(p_f, mmp_f, d, qq);
MatrixTimer.Stop();
LinalgTimer.Start();
a = c / d;
b_pred = a * (a * qq - d) / c;
cp = axpy_norm(r_f, -a, mmp_f, r_f);
b = cp / c;
// Fuse these loops ; should be really easy
psi_f = a * p_f + psi_f;
//p_f = p_f * b + r_f;
LinalgTimer.Stop();
std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: Iteration " << k
<< " residual " << cp << " target " << rsq << std::endl;
std::cout << GridLogDebug << "a = "<< a << " b_pred = "<< b_pred << " b = "<< b << std::endl;
std::cout << GridLogDebug << "qq = "<< qq << " d = "<< d << " c = "<< c << std::endl;
if(cp > MaxResidSinceLastRelUp){
std::cout << GridLogIterative << "ConjugateGradientReliableUpdate: updating MaxResidSinceLastRelUp : " << MaxResidSinceLastRelUp << " -> " << cp << std::endl;
MaxResidSinceLastRelUp = cp;
}
b = cp/c;
// Stopping condition
if (cp <= rsq) {
//Although not written in the paper, I assume that I have to add on the final solution
precisionChange(mmp, psi_f);
psi = psi + mmp;
SolverTimer.Stop();
Linop_d.HermOpAndNorm(psi, mmp, d, qq);
p = mmp - src;
RealD srcnorm = sqrt(norm2(src));
RealD resnorm = sqrt(norm2(p));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate Converged on iteration " << k << " after " << l << " reliable updates" << std::endl;
std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
std::cout << GridLogMessage << "Time breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tLinalg " << LinalgTimer.Elapsed() <<std::endl;
IterationsToComplete = k;
ReliableUpdatesPerformed = l;
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate new residual " << cp << std::endl;
if(DoFinalCleanup){
//Do a final CG to cleanup
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate performing final cleanup.\n";
ConjugateGradient<FieldD> CG(Tolerance,MaxIterations);
CG.ErrorOnNoConverge = ErrorOnNoConverge;
CG(Linop_d,src,psi);
IterationsToCleanup = CG.IterationsToComplete;
}
else if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate complete.\n";
return;
}
else if(cp < Delta * MaxResidSinceLastRelUp) { //reliable update
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate "
<< cp << "(residual) < " << Delta << "(Delta) * " << MaxResidSinceLastRelUp << "(MaxResidSinceLastRelUp) on iteration " << k << " : performing reliable update\n";
precisionChange(mmp, psi_f);
psi = psi + mmp;
Linop_d.HermOpAndNorm(psi, mmp, d, qq);
r = src - mmp;
psi_f = zero;
precisionChange(r_f, r);
cp = norm2(r);
MaxResidSinceLastRelUp = cp;
b = cp/c;
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate new residual " << cp << std::endl;
l = l+1;
}
p_f = p_f * b + r_f; //update search vector after reliable update appears to help convergence
if(!using_fallback && Linop_fallback != NULL && cp < fallback_transition_tol){
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate switching to fallback linear operator on iteration " << k << " at residual " << cp << std::endl;
Linop_f_use = Linop_fallback;
using_fallback = true;
}
l = l+1;
}
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate did NOT converge"
<< std::endl;
p_f = p_f * b + r_f; //update search vector after reliable update appears to help convergence
if(!using_fallback && Linop_fallback != NULL && cp < fallback_transition_tol){
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate switching to fallback linear operator on iteration " << k << " at residual " << cp << std::endl;
Linop_f_use = Linop_fallback;
using_fallback = true;
}
}
std::cout << GridLogMessage << "ConjugateGradientReliableUpdate did NOT converge"
<< std::endl;
if (ErrorOnNoConverge) assert(0);
IterationsToComplete = k;
ReliableUpdatesPerformed = l;
}
};
if (ErrorOnNoConverge) assert(0);
IterationsToComplete = k;
ReliableUpdatesPerformed = l;
}
};
NAMESPACE_END(Grid);
#endif

132
Grid/algorithms/iterative/ConjugateResidual.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,88 +24,90 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CONJUGATE_RESIDUAL_H
#define GRID_CONJUGATE_RESIDUAL_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
// single input vec, single output vec.
/////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
// single input vec, single output vec.
/////////////////////////////////////////////////////////////
template<class Field>
class ConjugateResidual : public OperatorFunction<Field> {
public:
RealD Tolerance;
Integer MaxIterations;
int verbose;
template<class Field>
class ConjugateResidual : public OperatorFunction<Field> {
public:
using OperatorFunction<Field>::operator();
ConjugateResidual(RealD tol,Integer maxit) : Tolerance(tol), MaxIterations(maxit) {
verbose=0;
};
RealD Tolerance;
Integer MaxIterations;
int verbose;
void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
ConjugateResidual(RealD tol,Integer maxit) : Tolerance(tol), MaxIterations(maxit) {
verbose=0;
};
RealD a, b, c, d;
RealD cp, ssq,rsq;
void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
RealD a, b; // c, d;
RealD cp, ssq,rsq;
RealD rAr, rAAr, rArp;
RealD pAp, pAAp;
RealD rAr, rAAr, rArp;
RealD pAp, pAAp;
GridBase *grid = src._grid;
psi=zero;
Field r(grid), p(grid), Ap(grid), Ar(grid);
GridBase *grid = src.Grid();
psi=Zero();
Field r(grid), p(grid), Ap(grid), Ar(grid);
r=src;
p=src;
r=src;
p=src;
Linop.HermOpAndNorm(p,Ap,pAp,pAAp);
Linop.HermOpAndNorm(r,Ar,rAr,rAAr);
cp =norm2(r);
ssq=norm2(src);
rsq=Tolerance*Tolerance*ssq;
if (verbose) std::cout<<GridLogMessage<<"ConjugateResidual: iteration " <<0<<" residual "<<cp<< " target"<< rsq<<std::endl;
for(int k=1;k<MaxIterations;k++){
a = rAr/pAAp;
axpy(psi,a,p,psi);
cp = axpy_norm(r,-a,Ap,r);
rArp=rAr;
Linop.HermOpAndNorm(p,Ap,pAp,pAAp);
Linop.HermOpAndNorm(r,Ar,rAr,rAAr);
cp =norm2(r);
ssq=norm2(src);
rsq=Tolerance*Tolerance*ssq;
if (verbose) std::cout<<GridLogMessage<<"ConjugateResidual: iteration " <<0<<" residual "<<cp<< " target"<< rsq<<std::endl;
for(int k=1;k<MaxIterations;k++){
a = rAr/pAAp;
axpy(psi,a,p,psi);
cp = axpy_norm(r,-a,Ap,r);
rArp=rAr;
Linop.HermOpAndNorm(r,Ar,rAr,rAAr);
b =rAr/rArp;
b =rAr/rArp;
axpy(p,b,p,r);
pAAp=axpy_norm(Ap,b,Ap,Ar);
axpy(p,b,p,r);
pAAp=axpy_norm(Ap,b,Ap,Ar);
if(verbose) std::cout<<GridLogMessage<<"ConjugateResidual: iteration " <<k<<" residual "<<cp<< " target"<< rsq<<std::endl;
if(cp<rsq) {
Linop.HermOp(psi,Ap);
axpy(r,-1.0,src,Ap);
RealD true_resid = norm2(r)/ssq;
std::cout<<GridLogMessage<<"ConjugateResidual: Converged on iteration " <<k
<< " computed residual "<<sqrt(cp/ssq)
<< " true residual "<<sqrt(true_resid)
<< " target " <<Tolerance <<std::endl;
return;
}
if(verbose) std::cout<<GridLogMessage<<"ConjugateResidual: iteration " <<k<<" residual "<<cp<< " target"<< rsq<<std::endl;
if(cp<rsq) {
Linop.HermOp(psi,Ap);
axpy(r,-1.0,src,Ap);
RealD true_resid = norm2(r)/ssq;
std::cout<<GridLogMessage<<"ConjugateResidual: Converged on iteration " <<k
<< " computed residual "<<std::sqrt(cp/ssq)
<< " true residual "<<std::sqrt(true_resid)
<< " target " <<Tolerance <<std::endl;
return;
}
std::cout<<GridLogMessage<<"ConjugateResidual did NOT converge"<<std::endl;
assert(0);
}
};
}
std::cout<<GridLogMessage<<"ConjugateResidual did NOT converge"<<std::endl;
assert(0);
}
};
NAMESPACE_END(Grid);
#endif

12
Grid/algorithms/iterative/Deflation.h
View File

@ -33,7 +33,7 @@ namespace Grid {
template<class Field>
class ZeroGuesser: public LinearFunction<Field> {
public:
virtual void operator()(const Field &src, Field &guess) { guess = zero; };
virtual void operator()(const Field &src, Field &guess) { guess = Zero(); };
};
template<class Field>
class DoNothingGuesser: public LinearFunction<Field> {
@ -60,14 +60,14 @@ public:
DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
virtual void operator()(const Field &src,Field &guess) {
guess = zero;
guess = Zero();
assert(evec.size()==eval.size());
auto N = evec.size();
for (int i=0;i<N;i++) {
const Field& tmp = evec[i];
axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
}
guess.checkerboard = src.checkerboard;
guess.Checkerboard() = src.Checkerboard();
}
};
@ -90,15 +90,15 @@ public:
void operator()(const FineField &src,FineField &guess) {
int N = (int)evec_coarse.size();
CoarseField src_coarse(evec_coarse[0]._grid);
CoarseField guess_coarse(evec_coarse[0]._grid); guess_coarse = zero;
CoarseField src_coarse(evec_coarse[0].Grid());
CoarseField guess_coarse(evec_coarse[0].Grid()); guess_coarse = Zero();
blockProject(src_coarse,src,subspace);
for (int i=0;i<N;i++) {
const CoarseField & tmp = evec_coarse[i];
axpy(guess_coarse,TensorRemove(innerProduct(tmp,src_coarse)) / eval_coarse[i],tmp,guess_coarse);
}
blockPromote(guess_coarse,guess,subspace);
guess.checkerboard = src.checkerboard;
guess.Checkerboard() = src.Checkerboard();
};
};

40
Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h
View File

@ -34,6 +34,8 @@ namespace Grid {
template<class Field>
class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<Field> {
public:
using OperatorFunction<Field>::operator();
bool ErrorOnNoConverge; // Throw an assert when FCAGMRES fails to converge,
// defaults to true
@ -53,10 +55,10 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
std::vector<ComplexD> y;
std::vector<ComplexD> gamma;
std::vector<ComplexD> c;
std::vector<ComplexD> s;
LinearFunction<Field> &Preconditioner;
@ -81,7 +83,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
std::cout << GridLogWarning << "This algorithm currently doesn't differ from regular FGMRES" << std::endl;
psi.checkerboard = src.checkerboard;
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
RealD guess = norm2(psi);
@ -91,7 +93,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Field r(src._grid);
Field r(src.Grid());
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: guess " << guess << std::endl;
@ -149,12 +151,12 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
RealD cp = 0;
Field w(src._grid);
Field r(src._grid);
Field w(src.Grid());
Field r(src.Grid());
// these should probably be made class members so that they are only allocated once, not in every restart
std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
std::vector<Field> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
std::vector<Field> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
std::vector<Field> z(RestartLength + 1, src.Grid()); for (auto &elem : z) elem = Zero();
MatrixTimer.Start();
LinOp.Op(psi, w);
@ -176,7 +178,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
qrUpdate(i);
cp = std::norm(gamma[i+1]);
cp = norm(gamma[i+1]);
std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
@ -206,11 +208,11 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
w = w - ComplexD(H(iter, i)) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
@ -218,13 +220,13 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
auto tmp = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
H(iter, i) = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
@ -233,7 +235,7 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
gamma[iter] = conjugate(c[iter]) * gamma[iter];
QrTimer.Stop();
}
@ -243,8 +245,8 @@ class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorF
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
y[i] = y[i] - ComplexD(H(k, i)) * y[k];
y[i] = y[i] / ComplexD(H(i, i));
}
for (int i = 0; i <= iter; i++)

40
Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h
View File

@ -34,6 +34,8 @@ namespace Grid {
template<class Field>
class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
public:
using OperatorFunction<Field>::operator();
bool ErrorOnNoConverge; // Throw an assert when FGMRES fails to converge,
// defaults to true
@ -53,10 +55,10 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
std::vector<ComplexD> y;
std::vector<ComplexD> gamma;
std::vector<ComplexD> c;
std::vector<ComplexD> s;
LinearFunction<Field> &Preconditioner;
@ -79,7 +81,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
psi.checkerboard = src.checkerboard;
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
RealD guess = norm2(psi);
@ -89,7 +91,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Field r(src._grid);
Field r(src.Grid());
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: guess " << guess << std::endl;
@ -147,12 +149,12 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
RealD cp = 0;
Field w(src._grid);
Field r(src._grid);
Field w(src.Grid());
Field r(src.Grid());
// these should probably be made class members so that they are only allocated once, not in every restart
std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
std::vector<Field> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
std::vector<Field> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
std::vector<Field> z(RestartLength + 1, src.Grid()); for (auto &elem : z) elem = Zero();
MatrixTimer.Start();
LinOp.Op(psi, w);
@ -174,7 +176,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
qrUpdate(i);
cp = std::norm(gamma[i+1]);
cp = norm(gamma[i+1]);
std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
@ -204,11 +206,11 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
w = w - ComplexD(H(iter, i)) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
@ -216,13 +218,13 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
auto tmp = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
H(iter, i) = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
@ -231,7 +233,7 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
gamma[iter] = conjugate(c[iter]) * gamma[iter];
QrTimer.Stop();
}
@ -241,8 +243,8 @@ class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
y[i] = y[i] - ComplexD(H(k, i)) * y[k];
y[i] = y[i] / ComplexD(H(i, i));
}
for (int i = 0; i <= iter; i++)

38
Grid/algorithms/iterative/GeneralisedMinimalResidual.h
View File

@ -34,6 +34,8 @@ namespace Grid {
template<class Field>
class GeneralisedMinimalResidual : public OperatorFunction<Field> {
public:
using OperatorFunction<Field>::operator();
bool ErrorOnNoConverge; // Throw an assert when GMRES fails to converge,
// defaults to true
@ -52,10 +54,10 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
std::vector<ComplexD> y;
std::vector<ComplexD> gamma;
std::vector<ComplexD> c;
std::vector<ComplexD> s;
GeneralisedMinimalResidual(RealD tol,
Integer maxit,
@ -74,7 +76,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
psi.checkerboard = src.checkerboard;
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
RealD guess = norm2(psi);
@ -84,7 +86,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Field r(src._grid);
Field r(src.Grid());
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "GeneralisedMinimalResidual: guess " << guess << std::endl;
@ -140,11 +142,11 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
RealD cp = 0;
Field w(src._grid);
Field r(src._grid);
Field w(src.Grid());
Field r(src.Grid());
// this should probably be made a class member so that it is only allocated once, not in every restart
std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
std::vector<Field> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
MatrixTimer.Start();
LinOp.Op(psi, w);
@ -166,7 +168,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
qrUpdate(i);
cp = std::norm(gamma[i+1]);
cp = norm(gamma[i+1]);
std::cout << GridLogIterative << "GeneralisedMinimalResidual: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
@ -192,11 +194,11 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
w = w - ComplexD(H(iter, i)) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
@ -204,13 +206,13 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
auto tmp = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
H(iter, i) = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
@ -219,7 +221,7 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
gamma[iter] = conjugate(c[iter]) * gamma[iter];
QrTimer.Stop();
}
@ -229,8 +231,8 @@ class GeneralisedMinimalResidual : public OperatorFunction<Field> {
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
y[i] = y[i] - ComplexD(H(k, i)) * y[k];
y[i] = y[i] / ComplexD(H(i, i));
}
for (int i = 0; i <= iter; i++)

73
Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@ -35,7 +35,7 @@ Author: Christoph Lehner <clehner@bnl.gov>
//#include <zlib.h>
#include <sys/stat.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////
// Move following 100 LOC to lattice/Lattice_basis.h
@ -52,26 +52,31 @@ void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k)
template<class Field>
void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm)
{
typedef decltype(basis[0].View()) View;
auto tmp_v = basis[0].View();
std::vector<View> basis_v(basis.size(),tmp_v);
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0]._grid;
GridBase* grid = basis[0].Grid();
parallel_region
{
for(int k=0;k<basis.size();k++){
basis_v[k] = basis[k].View();
}
thread_region
{
std::vector < vobj , commAllocator<vobj> > B(Nm); // Thread private
parallel_for_internal(int ss=0;ss < grid->oSites();ss++){
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[k]._odata[ss];
B[j] +=Qt(j,k) * basis_v[k][ss];
}
}
for(int j=j0; j<j1; ++j){
basis[j]._odata[ss] = B[j];
basis_v[j][ss] = B[j];
}
}
});
}
}
@ -80,16 +85,18 @@ template<class Field>
void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm)
{
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0]._grid;
GridBase* grid = basis[0].Grid();
result.checkerboard = basis[0].checkerboard;
parallel_for(int ss=0;ss < grid->oSites();ss++){
vobj B = zero;
result.Checkerboard() = basis[0].Checkerboard();
auto result_v=result.View();
thread_for(ss, grid->oSites(),{
vobj B = Zero();
for(int k=k0; k<k1; ++k){
B +=Qt(j,k) * basis[k]._odata[ss];
auto basis_k = basis[k].View();
B +=Qt(j,k) * basis_k[ss];
}
result._odata[ss] = B;
}
result_v[ss] = B;
});
}
template<class Field>
@ -119,7 +126,7 @@ void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, s
assert(idx[i] > i); assert(j!=idx.size()); assert(idx[j]==i);
std::swap(_v[i]._odata,_v[idx[i]]._odata); // should use vector move constructor, no data copy
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];
@ -150,6 +157,19 @@ void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, boo
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
/////////////////////////////////////////////////////////////
@ -289,7 +309,7 @@ public:
template<typename T> static RealD normalise(T& v)
{
RealD nn = norm2(v);
nn = sqrt(nn);
nn = std::sqrt(nn);
v = v * (1.0/nn);
return nn;
}
@ -321,8 +341,8 @@ until convergence
*/
void calc(std::vector<RealD>& eval, std::vector<Field>& evec, const Field& src, int& Nconv, bool reverse=false)
{
GridBase *grid = src._grid;
assert(grid == evec[0]._grid);
GridBase *grid = src.Grid();
assert(grid == evec[0].Grid());
GridLogIRL.TimingMode(1);
std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
@ -446,7 +466,7 @@ until convergence
assert(k2<Nm); assert(k2<Nm); assert(k1>0);
basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
std::cout<<GridLogIRL <<"basisRotated by Qt"<<std::endl;
std::cout<<GridLogIRL <<"basisRotated by Qt *"<<k1-1<<","<<k2+1<<")"<<std::endl;
////////////////////////////////////////////////////
// Compressed vector f and beta(k2)
@ -454,7 +474,7 @@ until convergence
f *= Qt(k2-1,Nm-1);
f += lme[k2-1] * evec[k2];
beta_k = norm2(f);
beta_k = sqrt(beta_k);
beta_k = std::sqrt(beta_k);
std::cout<<GridLogIRL<<" beta(k) = "<<beta_k<<std::endl;
RealD betar = 1.0/beta_k;
@ -477,7 +497,7 @@ until convergence
std::cout << GridLogIRL << "Test convergence: rotate subset of vectors to test convergence " << std::endl;
Field B(grid); B.checkerboard = evec[0].checkerboard;
Field B(grid); B.Checkerboard() = evec[0].Checkerboard();
// power of two search pattern; not every evalue in eval2 is assessed.
int allconv =1;
@ -515,7 +535,7 @@ until convergence
converged:
{
Field B(grid); B.checkerboard = evec[0].checkerboard;
Field B(grid); B.Checkerboard() = evec[0].Checkerboard();
basisRotate(evec,Qt,0,Nk,0,Nk,Nm);
std::cout << GridLogIRL << " Rotated basis"<<std::endl;
Nconv=0;
@ -807,7 +827,7 @@ void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme,
// determination of 2x2 leading submatrix
RealD dsub = lmd[kmax-1]-lmd[kmax-2];
RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
RealD dd = std::sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
// (Dsh: shift)
@ -838,5 +858,6 @@ void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme,
abort();
}
};
}
NAMESPACE_END(Grid);
#endif

61
Grid/algorithms/iterative/LocalCoherenceLanczos.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,16 +24,15 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LOCAL_COHERENCE_IRL_H
#define GRID_LOCAL_COHERENCE_IRL_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
struct LanczosParams : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParams,
ChebyParams, Cheby,/*Chebyshev*/
int, Nstop, /*Vecs in Lanczos must converge Nstop < Nk < Nm*/
@ -46,7 +45,7 @@ struct LanczosParams : Serializable {
};
struct LocalCoherenceLanczosParams : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
bool, saveEvecs,
bool, doFine,
@ -59,7 +58,7 @@ struct LocalCoherenceLanczosParams : Serializable {
RealD , coarse_relax_tol,
std::vector<int>, blockSize,
std::string, config,
std::vector < std::complex<double> >, omega,
std::vector < ComplexD >, omega,
RealD, mass,
RealD, M5);
};
@ -83,14 +82,14 @@ public:
};
void operator()(const CoarseField& in, CoarseField& out) {
GridBase *FineGrid = subspace[0]._grid;
int checkerboard = subspace[0].checkerboard;
FineField fin (FineGrid); fin.checkerboard= checkerboard;
FineField fout(FineGrid); fout.checkerboard = checkerboard;
GridBase *FineGrid = subspace[0].Grid();
int checkerboard = subspace[0].Checkerboard();
FineField fin (FineGrid); fin.Checkerboard()= checkerboard;
FineField fout(FineGrid); fout.Checkerboard() = checkerboard;
blockPromote(in,fin,subspace); std::cout<<GridLogIRL<<"ProjectedHermop : Promote to fine"<<std::endl;
_Linop.HermOp(fin,fout); std::cout<<GridLogIRL<<"ProjectedHermop : HermOp (fine) "<<std::endl;
_Linop.HermOp(fin,fout); std::cout<<GridLogIRL<<"ProjectedHermop : HermOp (fine) "<<std::endl;
blockProject(out,fout,subspace); std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
}
};
@ -117,12 +116,12 @@ public:
{ };
void operator()(const CoarseField& in, CoarseField& out) {
GridBase *FineGrid = subspace[0]._grid;
int checkerboard = subspace[0].checkerboard;
FineField fin (FineGrid); fin.checkerboard =checkerboard;
FineField fout(FineGrid);fout.checkerboard =checkerboard;
GridBase *FineGrid = subspace[0].Grid();
int checkerboard = subspace[0].Checkerboard();
FineField fin (FineGrid); fin.Checkerboard() =checkerboard;
FineField fout(FineGrid);fout.Checkerboard() =checkerboard;
blockPromote(in,fin,subspace); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
_poly(_Linop,fin,fout); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Poly "<<std::endl;
@ -133,7 +132,7 @@ public:
template<class Fobj,class CComplex,int nbasis>
class ImplicitlyRestartedLanczosSmoothedTester : public ImplicitlyRestartedLanczosTester<Lattice<iVector<CComplex,nbasis > > >
{
public:
public:
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
@ -142,7 +141,7 @@ class ImplicitlyRestartedLanczosSmoothedTester : public ImplicitlyRestartedLanc
LinearFunction<CoarseField> & _Poly;
OperatorFunction<FineField> & _smoother;
LinearOperatorBase<FineField> &_Linop;
RealD _coarse_relax_tol;
RealD _coarse_relax_tol;
std::vector<FineField> &_subspace;
ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField> &Poly,
@ -182,10 +181,10 @@ class ImplicitlyRestartedLanczosSmoothedTester : public ImplicitlyRestartedLanc
}
int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
{
GridBase *FineGrid = _subspace[0]._grid;
int checkerboard = _subspace[0].checkerboard;
FineField fB(FineGrid);fB.checkerboard =checkerboard;
FineField fv(FineGrid);fv.checkerboard =checkerboard;
GridBase *FineGrid = _subspace[0].Grid();
int checkerboard = _subspace[0].Checkerboard();
FineField fB(FineGrid);fB.Checkerboard() =checkerboard;
FineField fv(FineGrid);fv.Checkerboard() =checkerboard;
blockPromote(B,fv,_subspace);
@ -305,11 +304,11 @@ public:
int Nk = nbasis;
subspace.resize(Nk,_FineGrid);
subspace[0]=1.0;
subspace[0].checkerboard=_checkerboard;
subspace[0].Checkerboard()=_checkerboard;
normalise(subspace[0]);
PlainHermOp<FineField> Op(_FineOp);
for(int k=1;k<Nk;k++){
subspace[k].checkerboard=_checkerboard;
subspace[k].Checkerboard()=_checkerboard;
Op(subspace[k-1],subspace[k]);
normalise(subspace[k]);
}
@ -360,7 +359,11 @@ public:
ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
FineField src(_FineGrid); src=1.0; src.checkerboard = _checkerboard;
FineField src(_FineGrid);
typedef typename FineField::scalar_type Scalar;
// src=1.0;
src=Scalar(1.0);
src.Checkerboard() = _checkerboard;
int Nconv;
IRL.calc(evals_fine,subspace,src,Nconv,false);
@ -402,5 +405,5 @@ public:
}
};
}
NAMESPACE_END(Grid);
#endif

9
Grid/algorithms/iterative/MinimalResidual.h
View File

@ -33,6 +33,8 @@ namespace Grid {
template<class Field> class MinimalResidual : public OperatorFunction<Field> {
public:
using OperatorFunction<Field>::operator();
bool ErrorOnNoConverge; // throw an assert when the MR fails to converge.
// Defaults true.
RealD Tolerance;
@ -46,11 +48,11 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
psi.checkerboard = src.checkerboard;
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
Complex a, c;
Real d;
ComplexD a, c;
RealD d;
Field Mr(src);
Field r(src);
@ -71,7 +73,6 @@ template<class Field> class MinimalResidual : public OperatorFunction<Field> {
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "MinimalResidual: guess " << guess << std::endl;
std::cout << GridLogIterative << "MinimalResidual: src " << ssq << std::endl;
std::cout << GridLogIterative << "MinimalResidual: mp " << d << std::endl;
std::cout << GridLogIterative << "MinimalResidual: cp,r " << cp << std::endl;
if (cp <= rsq) {

41
Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h
View File

@ -34,6 +34,9 @@ namespace Grid {
template<class FieldD, class FieldF, typename std::enable_if<getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction<FieldD> {
public:
using OperatorFunction<FieldD>::operator();
bool ErrorOnNoConverge; // Throw an assert when MPFGMRES fails to converge,
// defaults to true
@ -54,10 +57,10 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
std::vector<ComplexD> y;
std::vector<ComplexD> gamma;
std::vector<ComplexD> c;
std::vector<ComplexD> s;
GridBase* SinglePrecGrid;
@ -84,7 +87,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
void operator()(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi) {
psi.checkerboard = src.checkerboard;
psi.Checkerboard() = src.Checkerboard();
conformable(psi, src);
RealD guess = norm2(psi);
@ -94,7 +97,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
FieldD r(src._grid);
FieldD r(src.Grid());
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "MPFGMRES: guess " << guess << std::endl;
@ -154,12 +157,12 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
RealD cp = 0;
FieldD w(src._grid);
FieldD r(src._grid);
FieldD w(src.Grid());
FieldD r(src.Grid());
// these should probably be made class members so that they are only allocated once, not in every restart
std::vector<FieldD> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
std::vector<FieldD> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
std::vector<FieldD> v(RestartLength + 1, src.Grid()); for (auto &elem : v) elem = Zero();
std::vector<FieldD> z(RestartLength + 1, src.Grid()); for (auto &elem : z) elem = Zero();
MatrixTimer.Start();
LinOp.Op(psi, w);
@ -181,7 +184,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
qrUpdate(i);
cp = std::norm(gamma[i+1]);
cp = norm(gamma[i+1]);
std::cout << GridLogIterative << "MPFGMRES: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
@ -223,11 +226,11 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
w = w - ComplexD(H(iter, i)) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
v[iter + 1] = ComplexD(1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
@ -235,13 +238,13 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
auto tmp = -s[i] * ComplexD(H(iter, i)) + c[i] * ComplexD(H(iter, i + 1));
H(iter, i) = conjugate(c[i]) * ComplexD(H(iter, i)) + conjugate(s[i]) * ComplexD(H(iter, i + 1));
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
auto nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
@ -250,7 +253,7 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
gamma[iter] = conjugate(c[iter]) * gamma[iter];
QrTimer.Stop();
}
@ -260,8 +263,8 @@ class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
y[i] = y[i] - ComplexD(H(k, i)) * y[k];
y[i] = y[i] / ComplexD(H(i, i));
}
for (int i = 0; i <= iter; i++)

44
Grid/algorithms/iterative/NormalEquations.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,38 +23,38 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_NORMAL_EQUATIONS_H
#define GRID_NORMAL_EQUATIONS_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////////////////////////////////////////////////
// Take a matrix and form an NE solver calling a Herm solver
///////////////////////////////////////////////////////////////////////////////////////////////////////
template<class Field> class NormalEquations : public OperatorFunction<Field>{
private:
SparseMatrixBase<Field> & _Matrix;
OperatorFunction<Field> & _HermitianSolver;
///////////////////////////////////////////////////////////////////////////////////////////////////////
// Take a matrix and form an NE solver calling a Herm solver
///////////////////////////////////////////////////////////////////////////////////////////////////////
template<class Field> class NormalEquations : public OperatorFunction<Field>{
private:
SparseMatrixBase<Field> & _Matrix;
OperatorFunction<Field> & _HermitianSolver;
public:
public:
/////////////////////////////////////////////////////
// Wrap the usual normal equations trick
/////////////////////////////////////////////////////
/////////////////////////////////////////////////////
// Wrap the usual normal equations trick
/////////////////////////////////////////////////////
NormalEquations(SparseMatrixBase<Field> &Matrix, OperatorFunction<Field> &HermitianSolver)
: _Matrix(Matrix), _HermitianSolver(HermitianSolver) {};
void operator() (const Field &in, Field &out){
void operator() (const Field &in, Field &out){
Field src(in._grid);
Field src(in.Grid());
_Matrix.Mdag(in,src);
_HermitianSolver(src,out); // Mdag M out = Mdag in
_Matrix.Mdag(in,src);
_HermitianSolver(src,out); // Mdag M out = Mdag in
}
};
}
};
}
NAMESPACE_END(Grid);
#endif

2
Grid/algorithms/iterative/PowerMethod.h
View File

@ -14,7 +14,7 @@ template<class Field> class PowerMethod
RealD operator()(LinearOperatorBase<Field> &HermOp, const Field &src)
{
GridBase *grid = src._grid;
GridBase *grid = src.Grid();
// quickly get an idea of the largest eigenvalue to more properly normalize the residuum
RealD evalMaxApprox = 0.0;

128
Grid/algorithms/iterative/PrecConjugateResidual.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,97 +23,97 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_PREC_CONJUGATE_RESIDUAL_H
#define GRID_PREC_CONJUGATE_RESIDUAL_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
// single input vec, single output vec.
/////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
// single input vec, single output vec.
/////////////////////////////////////////////////////////////
template<class Field>
class PrecConjugateResidual : public OperatorFunction<Field> {
public:
RealD Tolerance;
Integer MaxIterations;
int verbose;
LinearFunction<Field> &Preconditioner;
template<class Field>
class PrecConjugateResidual : public OperatorFunction<Field> {
public:
RealD Tolerance;
Integer MaxIterations;
int verbose;
LinearFunction<Field> &Preconditioner;
PrecConjugateResidual(RealD tol,Integer maxit,LinearFunction<Field> &Prec) : Tolerance(tol), MaxIterations(maxit), Preconditioner(Prec)
{
verbose=1;
};
PrecConjugateResidual(RealD tol,Integer maxit,LinearFunction<Field> &Prec) : Tolerance(tol), MaxIterations(maxit), Preconditioner(Prec)
{
verbose=1;
};
void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
RealD a, b, c, d;
RealD cp, ssq,rsq;
RealD a, b, c, d;
RealD cp, ssq,rsq;
RealD rAr, rAAr, rArp;
RealD pAp, pAAp;
RealD rAr, rAAr, rArp;
RealD pAp, pAAp;
GridBase *grid = src._grid;
Field r(grid), p(grid), Ap(grid), Ar(grid), z(grid);
GridBase *grid = src.Grid();
Field r(grid), p(grid), Ap(grid), Ar(grid), z(grid);
psi=zero;
r = src;
Preconditioner(r,p);
psi=zero;
r = src;
Preconditioner(r,p);
Linop.HermOpAndNorm(p,Ap,pAp,pAAp);
Ar=Ap;
rAr=pAp;
rAAr=pAAp;
Linop.HermOpAndNorm(p,Ap,pAp,pAAp);
Ar=Ap;
rAr=pAp;
rAAr=pAAp;
cp =norm2(r);
ssq=norm2(src);
rsq=Tolerance*Tolerance*ssq;
cp =norm2(r);
ssq=norm2(src);
rsq=Tolerance*Tolerance*ssq;
if (verbose) std::cout<<GridLogMessage<<"PrecConjugateResidual: iteration " <<0<<" residual "<<cp<< " target"<< rsq<<std::endl;
if (verbose) std::cout<<GridLogMessage<<"PrecConjugateResidual: iteration " <<0<<" residual "<<cp<< " target"<< rsq<<std::endl;
for(int k=0;k<MaxIterations;k++){
for(int k=0;k<MaxIterations;k++){
Preconditioner(Ap,z);
RealD rq= real(innerProduct(Ap,z));
Preconditioner(Ap,z);
RealD rq= real(innerProduct(Ap,z));
a = rAr/rq;
a = rAr/rq;
axpy(psi,a,p,psi);
cp = axpy_norm(r,-a,z,r);
axpy(psi,a,p,psi);
cp = axpy_norm(r,-a,z,r);
rArp=rAr;
rArp=rAr;
Linop.HermOpAndNorm(r,Ar,rAr,rAAr);
Linop.HermOpAndNorm(r,Ar,rAr,rAAr);
b =rAr/rArp;
b =rAr/rArp;
axpy(p,b,p,r);
pAAp=axpy_norm(Ap,b,Ap,Ar);
axpy(p,b,p,r);
pAAp=axpy_norm(Ap,b,Ap,Ar);
if(verbose) std::cout<<GridLogMessage<<"PrecConjugateResidual: iteration " <<k<<" residual "<<cp<< " target"<< rsq<<std::endl;
if(cp<rsq) {
Linop.HermOp(psi,Ap);
axpy(r,-1.0,src,Ap);
RealD true_resid = norm2(r)/ssq;
std::cout<<GridLogMessage<<"PrecConjugateResidual: Converged on iteration " <<k
<< " computed residual "<<sqrt(cp/ssq)
<< " true residual "<<sqrt(true_resid)
<< " target " <<Tolerance <<std::endl;
return;
}
if(verbose) std::cout<<GridLogMessage<<"PrecConjugateResidual: iteration " <<k<<" residual "<<cp<< " target"<< rsq<<std::endl;
if(cp<rsq) {
Linop.HermOp(psi,Ap);
axpy(r,-1.0,src,Ap);
RealD true_resid = norm2(r)/ssq;
std::cout<<GridLogMessage<<"PrecConjugateResidual: Converged on iteration " <<k
<< " computed residual "<<sqrt(cp/ssq)
<< " true residual "<<sqrt(true_resid)
<< " target " <<Tolerance <<std::endl;
return;
}
std::cout<<GridLogMessage<<"PrecConjugateResidual did NOT converge"<<std::endl;
assert(0);
}
};
}
std::cout<<GridLogMessage<<"PrecConjugateResidual did NOT converge"<<std::endl;
assert(0);
}
};
NAMESPACE_END(Grid);
#endif

326
Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,8 +24,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_PREC_GCR_H
#define GRID_PREC_GCR_H
@ -36,206 +36,208 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
//NB. Likely not original reference since they are focussing on a preconditioner variant.
// but VPGCR was nicely written up in their paper
///////////////////////////////////////////////////////////////////////////////////////////////////////
namespace Grid {
NAMESPACE_BEGIN(Grid);
template<class Field>
class PrecGeneralisedConjugateResidual : public OperatorFunction<Field> {
public:
RealD Tolerance;
Integer MaxIterations;
int verbose;
int mmax;
int nstep;
int steps;
GridStopWatch PrecTimer;
GridStopWatch MatTimer;
GridStopWatch LinalgTimer;
template<class Field>
class PrecGeneralisedConjugateResidual : public OperatorFunction<Field> {
public:
using OperatorFunction<Field>::operator();
LinearFunction<Field> &Preconditioner;
RealD Tolerance;
Integer MaxIterations;
int verbose;
int mmax;
int nstep;
int steps;
GridStopWatch PrecTimer;
GridStopWatch MatTimer;
GridStopWatch LinalgTimer;
PrecGeneralisedConjugateResidual(RealD tol,Integer maxit,LinearFunction<Field> &Prec,int _mmax,int _nstep) :
Tolerance(tol),
MaxIterations(maxit),
Preconditioner(Prec),
mmax(_mmax),
nstep(_nstep)
{
verbose=1;
};
LinearFunction<Field> &Preconditioner;
void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
PrecGeneralisedConjugateResidual(RealD tol,Integer maxit,LinearFunction<Field> &Prec,int _mmax,int _nstep) :
Tolerance(tol),
MaxIterations(maxit),
Preconditioner(Prec),
mmax(_mmax),
nstep(_nstep)
{
verbose=1;
};
psi=zero;
RealD cp, ssq,rsq;
ssq=norm2(src);
rsq=Tolerance*Tolerance*ssq;
void operator() (LinearOperatorBase<Field> &Linop,const Field &src, Field &psi){
psi=Zero();
RealD cp, ssq,rsq;
ssq=norm2(src);
rsq=Tolerance*Tolerance*ssq;
Field r(src._grid);
Field r(src.Grid());
PrecTimer.Reset();
MatTimer.Reset();
LinalgTimer.Reset();
PrecTimer.Reset();
MatTimer.Reset();
LinalgTimer.Reset();
GridStopWatch SolverTimer;
SolverTimer.Start();
GridStopWatch SolverTimer;
SolverTimer.Start();
steps=0;
for(int k=0;k<MaxIterations;k++){
steps=0;
for(int k=0;k<MaxIterations;k++){
cp=GCRnStep(Linop,src,psi,rsq);
cp=GCRnStep(Linop,src,psi,rsq);
std::cout<<GridLogMessage<<"VPGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<std::endl;
std::cout<<GridLogMessage<<"VPGCR("<<mmax<<","<<nstep<<") "<< steps <<" steps cp = "<<cp<<std::endl;
if(cp<rsq) {
if(cp<rsq) {
SolverTimer.Stop();
SolverTimer.Stop();
Linop.HermOp(psi,r);
axpy(r,-1.0,src,r);
RealD tr = norm2(r);
std::cout<<GridLogMessage<<"PrecGeneralisedConjugateResidual: Converged on iteration " <<steps
<< " computed residual "<<sqrt(cp/ssq)
<< " true residual " <<sqrt(tr/ssq)
<< " target " <<Tolerance <<std::endl;
std::cout<<GridLogMessage<<"VPGCR Time elapsed: Total "<< SolverTimer.Elapsed() <<std::endl;
std::cout<<GridLogMessage<<"VPGCR Time elapsed: Precon "<< PrecTimer.Elapsed() <<std::endl;
std::cout<<GridLogMessage<<"VPGCR Time elapsed: Matrix "<< MatTimer.Elapsed() <<std::endl;
std::cout<<GridLogMessage<<"VPGCR Time elapsed: Linalg "<< LinalgTimer.Elapsed() <<std::endl;
return;
}
Linop.HermOp(psi,r);
axpy(r,-1.0,src,r);
RealD tr = norm2(r);
std::cout<<GridLogMessage<<"PrecGeneralisedConjugateResidual: Converged on iteration " <<steps
<< " computed residual "<<sqrt(cp/ssq)
<< " true residual " <<sqrt(tr/ssq)
<< " target " <<Tolerance <<std::endl;
std::cout<<GridLogMessage<<"VPGCR Time elapsed: Total "<< SolverTimer.Elapsed() <<std::endl;
std::cout<<GridLogMessage<<"VPGCR Time elapsed: Precon "<< PrecTimer.Elapsed() <<std::endl;
std::cout<<GridLogMessage<<"VPGCR Time elapsed: Matrix "<< MatTimer.Elapsed() <<std::endl;
std::cout<<GridLogMessage<<"VPGCR Time elapsed: Linalg "<< LinalgTimer.Elapsed() <<std::endl;
return;
}
std::cout<<GridLogMessage<<"Variable Preconditioned GCR did not converge"<<std::endl;
assert(0);
}
std::cout<<GridLogMessage<<"Variable Preconditioned GCR did not converge"<<std::endl;
assert(0);
}
RealD GCRnStep(LinearOperatorBase<Field> &Linop,const Field &src, Field &psi,RealD rsq){
RealD GCRnStep(LinearOperatorBase<Field> &Linop,const Field &src, Field &psi,RealD rsq){
RealD cp;
RealD a, b, c, d;
RealD zAz, zAAz;
RealD rAq, rq;
RealD cp;
RealD a, b;
RealD zAz, zAAz;
RealD rq;
GridBase *grid = src._grid;
GridBase *grid = src.Grid();
Field r(grid);
Field z(grid);
Field tmp(grid);
Field ttmp(grid);
Field Az(grid);
Field r(grid);
Field z(grid);
Field tmp(grid);
Field ttmp(grid);
Field Az(grid);
////////////////////////////////
// history for flexible orthog
////////////////////////////////
std::vector<Field> q(mmax,grid);
std::vector<Field> p(mmax,grid);
std::vector<RealD> qq(mmax);
////////////////////////////////
// history for flexible orthog
////////////////////////////////
std::vector<Field> q(mmax,grid);
std::vector<Field> p(mmax,grid);
std::vector<RealD> qq(mmax);
//////////////////////////////////
// initial guess x0 is taken as nonzero.
// r0=src-A x0 = src
//////////////////////////////////
MatTimer.Start();
Linop.HermOpAndNorm(psi,Az,zAz,zAAz);
MatTimer.Stop();
LinalgTimer.Start();
r=src-Az;
LinalgTimer.Stop();
//////////////////////////////////
// initial guess x0 is taken as nonzero.
// r0=src-A x0 = src
//////////////////////////////////
MatTimer.Start();
Linop.HermOpAndNorm(psi,Az,zAz,zAAz);
MatTimer.Stop();
LinalgTimer.Start();
r=src-Az;
LinalgTimer.Stop();
/////////////////////
// p = Prec(r)
/////////////////////
PrecTimer.Start();
Preconditioner(r,z);
PrecTimer.Stop();
/////////////////////
// p = Prec(r)
/////////////////////
PrecTimer.Start();
Preconditioner(r,z);
PrecTimer.Stop();
MatTimer.Start();
Linop.HermOp(z,tmp);
MatTimer.Stop();
MatTimer.Start();
Linop.HermOp(z,tmp);
MatTimer.Stop();
LinalgTimer.Start();
ttmp=tmp;
tmp=tmp-r;
LinalgTimer.Stop();
LinalgTimer.Start();
ttmp=tmp;
tmp=tmp-r;
LinalgTimer.Stop();
/*
/*
std::cout<<GridLogMessage<<r<<std::endl;
std::cout<<GridLogMessage<<z<<std::endl;
std::cout<<GridLogMessage<<ttmp<<std::endl;
std::cout<<GridLogMessage<<tmp<<std::endl;
*/
*/
MatTimer.Start();
Linop.HermOpAndNorm(z,Az,zAz,zAAz);
MatTimer.Stop();
LinalgTimer.Start();
//p[0],q[0],qq[0]
p[0]= z;
q[0]= Az;
qq[0]= zAAz;
cp =norm2(r);
LinalgTimer.Stop();
for(int k=0;k<nstep;k++){
steps++;
int kp = k+1;
int peri_k = k %mmax;
int peri_kp= kp%mmax;
LinalgTimer.Start();
rq= real(innerProduct(r,q[peri_k])); // what if rAr not real?
a = rq/qq[peri_k];
axpy(psi,a,p[peri_k],psi);
cp = axpy_norm(r,-a,q[peri_k],r);
LinalgTimer.Stop();
if((k==nstep-1)||(cp<rsq)){
return cp;
}
std::cout<<GridLogMessage<< " VPGCR_step["<<steps<<"] resid " <<sqrt(cp/rsq)<<std::endl;
PrecTimer.Start();
Preconditioner(r,z);// solve Az = r
PrecTimer.Stop();
MatTimer.Start();
Linop.HermOpAndNorm(z,Az,zAz,zAAz);
Linop.HermOpAndNorm(z,Az,zAz,zAAz);
Linop.HermOp(z,tmp);
MatTimer.Stop();
LinalgTimer.Start();
//p[0],q[0],qq[0]
p[0]= z;
q[0]= Az;
qq[0]= zAAz;
tmp=tmp-r;
std::cout<<GridLogMessage<< " Preconditioner resid " <<sqrt(norm2(tmp)/norm2(r))<<std::endl;
cp =norm2(r);
LinalgTimer.Stop();
q[peri_kp]=Az;
p[peri_kp]=z;
for(int k=0;k<nstep;k++){
int northog = ((kp)>(mmax-1))?(mmax-1):(kp); // if more than mmax done, we orthog all mmax history.
for(int back=0;back<northog;back++){
steps++;
int peri_back=(k-back)%mmax; assert((k-back)>=0);
int kp = k+1;
int peri_k = k %mmax;
int peri_kp= kp%mmax;
b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
p[peri_kp]=p[peri_kp]+b*p[peri_back];
q[peri_kp]=q[peri_kp]+b*q[peri_back];
LinalgTimer.Start();
rq= real(innerProduct(r,q[peri_k])); // what if rAr not real?
a = rq/qq[peri_k];
axpy(psi,a,p[peri_k],psi);
cp = axpy_norm(r,-a,q[peri_k],r);
LinalgTimer.Stop();
if((k==nstep-1)||(cp<rsq)){
return cp;
}
std::cout<<GridLogMessage<< " VPGCR_step["<<steps<<"] resid " <<sqrt(cp/rsq)<<std::endl;
PrecTimer.Start();
Preconditioner(r,z);// solve Az = r
PrecTimer.Stop();
MatTimer.Start();
Linop.HermOpAndNorm(z,Az,zAz,zAAz);
Linop.HermOp(z,tmp);
MatTimer.Stop();
LinalgTimer.Start();
tmp=tmp-r;
std::cout<<GridLogMessage<< " Preconditioner resid " <<sqrt(norm2(tmp)/norm2(r))<<std::endl;
q[peri_kp]=Az;
p[peri_kp]=z;
int northog = ((kp)>(mmax-1))?(mmax-1):(kp); // if more than mmax done, we orthog all mmax history.
for(int back=0;back<northog;back++){
int peri_back=(k-back)%mmax; assert((k-back)>=0);
b=-real(innerProduct(q[peri_back],Az))/qq[peri_back];
p[peri_kp]=p[peri_kp]+b*p[peri_back];
q[peri_kp]=q[peri_kp]+b*q[peri_back];
}
qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
LinalgTimer.Stop();
}
assert(0); // never reached
return cp;
qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
LinalgTimer.Stop();
}
};
}
assert(0); // never reached
return cp;
}
};
NAMESPACE_END(Grid);
#endif

56
Grid/algorithms/iterative/SchurRedBlack.h
View File

@ -297,9 +297,9 @@ namespace Grid {
/////////////////////////////////////////////////////
// src_o = (source_o - Moe MeeInv source_e)
/////////////////////////////////////////////////////
_Matrix.MooeeInv(src_e,tmp); assert( tmp.checkerboard ==Even);
_Matrix.Meooe (tmp,Mtmp); assert( Mtmp.checkerboard ==Odd);
tmp=src_o-Mtmp; assert( tmp.checkerboard ==Odd);
_Matrix.MooeeInv(src_e,tmp); assert( tmp.Checkerboard() ==Even);
_Matrix.Meooe (tmp,Mtmp); assert( Mtmp.Checkerboard() ==Odd);
tmp=src_o-Mtmp; assert( tmp.Checkerboard() ==Odd);
_Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
}
@ -317,17 +317,17 @@ namespace Grid {
///////////////////////////////////////////////////
// sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
///////////////////////////////////////////////////
_Matrix.Meooe(sol_o,tmp); assert( tmp.checkerboard ==Even);
src_e = src_e-tmp; assert( src_e.checkerboard ==Even);
_Matrix.MooeeInv(src_e,sol_e); assert( sol_e.checkerboard ==Even);
_Matrix.Meooe(sol_o,tmp); assert( tmp.Checkerboard() ==Even);
src_e = src_e-tmp; assert( src_e.Checkerboard() ==Even);
_Matrix.MooeeInv(src_e,sol_e); assert( sol_e.Checkerboard() ==Even);
setCheckerboard(sol,sol_e); assert( sol_e.checkerboard ==Even);
setCheckerboard(sol,sol_o); assert( sol_o.checkerboard ==Odd );
setCheckerboard(sol,sol_e); assert( sol_e.Checkerboard() ==Even);
setCheckerboard(sol,sol_o); assert( sol_o.Checkerboard() ==Odd );
}
virtual void RedBlackSolve (Matrix & _Matrix,const Field &src_o, Field &sol_o)
{
SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.checkerboard==Odd);
this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.Checkerboard()==Odd);
};
virtual void RedBlackSolve (Matrix & _Matrix,const std::vector<Field> &src_o, std::vector<Field> &sol_o)
{
@ -366,13 +366,13 @@ namespace Grid {
/////////////////////////////////////////////////////
// src_o = Mdag * (source_o - Moe MeeInv source_e)
/////////////////////////////////////////////////////
_Matrix.MooeeInv(src_e,tmp); assert( tmp.checkerboard ==Even);
_Matrix.Meooe (tmp,Mtmp); assert( Mtmp.checkerboard ==Odd);
tmp=src_o-Mtmp; assert( tmp.checkerboard ==Odd);
_Matrix.MooeeInv(src_e,tmp); assert( tmp.Checkerboard() ==Even);
_Matrix.Meooe (tmp,Mtmp); assert( Mtmp.Checkerboard() ==Odd);
tmp=src_o-Mtmp; assert( tmp.Checkerboard() ==Odd);
// get the right MpcDag
SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
_HermOpEO.MpcDag(tmp,src_o); assert(src_o.checkerboard ==Odd);
_HermOpEO.MpcDag(tmp,src_o); assert(src_o.Checkerboard() ==Odd);
}
virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@ -386,17 +386,17 @@ namespace Grid {
///////////////////////////////////////////////////
// sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
///////////////////////////////////////////////////
_Matrix.Meooe(sol_o,tmp); assert( tmp.checkerboard ==Even);
src_e_i = src_e-tmp; assert( src_e_i.checkerboard ==Even);
_Matrix.MooeeInv(src_e_i,sol_e); assert( sol_e.checkerboard ==Even);
_Matrix.Meooe(sol_o,tmp); assert( tmp.Checkerboard() ==Even);
src_e_i = src_e-tmp; assert( src_e_i.Checkerboard() ==Even);
_Matrix.MooeeInv(src_e_i,sol_e); assert( sol_e.Checkerboard() ==Even);
setCheckerboard(sol,sol_e); assert( sol_e.checkerboard ==Even);
setCheckerboard(sol,sol_o); assert( sol_o.checkerboard ==Odd );
setCheckerboard(sol,sol_e); assert( sol_e.Checkerboard() ==Even);
setCheckerboard(sol,sol_o); assert( sol_o.Checkerboard() ==Odd );
}
virtual void RedBlackSolve (Matrix & _Matrix,const Field &src_o, Field &sol_o)
{
SchurDiagMooeeOperator<Matrix,Field> _HermOpEO(_Matrix);
this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.checkerboard==Odd);
this->_HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.Checkerboard()==Odd);
};
virtual void RedBlackSolve (Matrix & _Matrix,const std::vector<Field> &src_o, std::vector<Field> &sol_o)
{
@ -437,12 +437,12 @@ namespace Grid {
/////////////////////////////////////////////////////
// src_o = Mdag * (source_o - Moe MeeInv source_e)
/////////////////////////////////////////////////////
_Matrix.MooeeInv(src_e,tmp); assert( tmp.checkerboard ==Even);
_Matrix.Meooe (tmp,Mtmp); assert( Mtmp.checkerboard ==Odd);
tmp=src_o-Mtmp; assert( tmp.checkerboard ==Odd);
_Matrix.MooeeInv(src_e,tmp); assert( tmp.Checkerboard() ==Even);
_Matrix.Meooe (tmp,Mtmp); assert( Mtmp.Checkerboard() ==Odd);
tmp=src_o-Mtmp; assert( tmp.Checkerboard() ==Odd);
// get the right MpcDag
_HermOpEO.MpcDag(tmp,src_o); assert(src_o.checkerboard ==Odd);
_HermOpEO.MpcDag(tmp,src_o); assert(src_o.Checkerboard() ==Odd);
}
virtual void RedBlackSolution(Matrix & _Matrix,const Field &sol_o, const Field &src_e,Field &sol)
@ -463,12 +463,12 @@ namespace Grid {
///////////////////////////////////////////////////
// sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
///////////////////////////////////////////////////
_Matrix.Meooe(sol_o_i,tmp); assert( tmp.checkerboard ==Even);
tmp = src_e-tmp; assert( src_e.checkerboard ==Even);
_Matrix.MooeeInv(tmp,sol_e); assert( sol_e.checkerboard ==Even);
_Matrix.Meooe(sol_o_i,tmp); assert( tmp.Checkerboard() ==Even);
tmp = src_e-tmp; assert( src_e.Checkerboard() ==Even);
_Matrix.MooeeInv(tmp,sol_e); assert( sol_e.Checkerboard() ==Even);
setCheckerboard(sol,sol_e); assert( sol_e.checkerboard ==Even);
setCheckerboard(sol,sol_o_i); assert( sol_o_i.checkerboard ==Odd );
setCheckerboard(sol,sol_e); assert( sol_e.Checkerboard() ==Even);
setCheckerboard(sol,sol_o_i); assert( sol_o_i.Checkerboard() ==Odd );
};
virtual void RedBlackSolve (Matrix & _Matrix,const Field &src_o, Field &sol_o)

26
Grid/allocator/AlignedAllocator.cc
View File

@ -1,11 +1,12 @@
#include <Grid/GridCore.h>
#include <fcntl.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
MemoryStats *MemoryProfiler::stats = nullptr;
bool MemoryProfiler::debug = false;
#ifdef POINTER_CACHE
int PointerCache::victim;
PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
@ -49,9 +50,9 @@ void *PointerCache::Insert(void *ptr,size_t bytes) {
void *PointerCache::Lookup(size_t bytes) {
if (bytes < 4096 ) return NULL;
if (bytes < 4096 ) return NULL;
#ifdef _OPENMP
#ifdef GRID_OMP
assert(omp_in_parallel()==0);
#endif
@ -63,7 +64,7 @@ void *PointerCache::Lookup(size_t bytes) {
}
return NULL;
}
#endif
void check_huge_pages(void *Buf,uint64_t BYTES)
{
@ -90,7 +91,7 @@ void check_huge_pages(void *Buf,uint64_t BYTES)
++n4ktotal;
if (pageaddr != baseaddr + j * page_size)
++nnothuge;
}
}
}
int rank = CartesianCommunicator::RankWorld();
printf("rank %d Allocated %d 4k pages, %d not in huge pages\n", rank, n4ktotal, nnothuge);
@ -106,20 +107,21 @@ std::string sizeString(const size_t bytes)
double count = bytes;
while (count >= 1024 && s < 7)
{
{
s++;
count /= 1024;
}
}
if (count - floor(count) == 0.0)
{
{
snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
}
}
else
{
{
snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
}
}
return std::string(buf);
}
}
NAMESPACE_END(Grid);

333
Grid/allocator/AlignedAllocator.h
View File

@ -24,8 +24,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_ALIGNED_ALLOCATOR_H
#define GRID_ALIGNED_ALLOCATOR_H
@ -40,89 +40,95 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <mm_malloc.h>
#endif
namespace Grid {
#define POINTER_CACHE
#define GRID_ALLOC_ALIGN (2*1024*1024)
class PointerCache {
private:
NAMESPACE_BEGIN(Grid);
static const int Ncache=8;
static int victim;
// Move control to configure.ac and Config.h?
#ifdef POINTER_CACHE
class PointerCache {
private:
typedef struct {
void *address;
size_t bytes;
int valid;
} PointerCacheEntry;
static const int Ncache=8;
static int victim;
typedef struct {
void *address;
size_t bytes;
int valid;
} PointerCacheEntry;
static PointerCacheEntry Entries[Ncache];
static PointerCacheEntry Entries[Ncache];
public:
public:
static void *Insert(void *ptr,size_t bytes) ;
static void *Lookup(size_t bytes) ;
static void *Insert(void *ptr,size_t bytes) ;
static void *Lookup(size_t bytes) ;
};
std::string sizeString(size_t bytes);
};
#endif
struct MemoryStats
{
size_t totalAllocated{0}, maxAllocated{0},
currentlyAllocated{0}, totalFreed{0};
};
std::string sizeString(size_t bytes);
struct MemoryStats
{
size_t totalAllocated{0}, maxAllocated{0},
currentlyAllocated{0}, totalFreed{0};
};
class MemoryProfiler
{
public:
static MemoryStats *stats;
static bool debug;
};
class MemoryProfiler
{
public:
static MemoryStats *stats;
static bool debug;
};
#define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
#define profilerDebugPrint \
if (MemoryProfiler::stats)\
{\
auto s = MemoryProfiler::stats;\
std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl;\
std::cout << GridLogDebug << "[Memory debug] total : " << memString(s->totalAllocated) \
<< std::endl;\
std::cout << GridLogDebug << "[Memory debug] max : " << memString(s->maxAllocated) \
<< std::endl;\
std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
<< std::endl;\
std::cout << GridLogDebug << "[Memory debug] freed : " << memString(s->totalFreed) \
<< std::endl;\
}
#define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
#define profilerDebugPrint \
if (MemoryProfiler::stats) \
{ \
auto s = MemoryProfiler::stats; \
std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl; \
std::cout << GridLogDebug << "[Memory debug] total : " << memString(s->totalAllocated) \
<< std::endl; \
std::cout << GridLogDebug << "[Memory debug] max : " << memString(s->maxAllocated) \
<< std::endl; \
std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
<< std::endl; \
std::cout << GridLogDebug << "[Memory debug] freed : " << memString(s->totalFreed) \
<< std::endl; \
}
#define profilerAllocate(bytes)\
if (MemoryProfiler::stats)\
{\
auto s = MemoryProfiler::stats;\
s->totalAllocated += (bytes);\
s->currentlyAllocated += (bytes);\
s->maxAllocated = std::max(s->maxAllocated, s->currentlyAllocated);\
}\
if (MemoryProfiler::debug)\
{\
std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl;\
profilerDebugPrint;\
}
#define profilerAllocate(bytes) \
if (MemoryProfiler::stats) \
{ \
auto s = MemoryProfiler::stats; \
s->totalAllocated += (bytes); \
s->currentlyAllocated += (bytes); \
s->maxAllocated = std::max(s->maxAllocated, s->currentlyAllocated); \
} \
if (MemoryProfiler::debug) \
{ \
std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl; \
profilerDebugPrint; \
}
#define profilerFree(bytes)\
if (MemoryProfiler::stats)\
{\
auto s = MemoryProfiler::stats;\
s->totalFreed += (bytes);\
s->currentlyAllocated -= (bytes);\
}\
if (MemoryProfiler::debug)\
{\
std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl;\
profilerDebugPrint;\
}
#define profilerFree(bytes) \
if (MemoryProfiler::stats) \
{ \
auto s = MemoryProfiler::stats; \
s->totalFreed += (bytes); \
s->currentlyAllocated -= (bytes); \
} \
if (MemoryProfiler::debug) \
{ \
std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl; \
profilerDebugPrint; \
}
void check_huge_pages(void *Buf,uint64_t BYTES);
void check_huge_pages(void *Buf,uint64_t BYTES);
////////////////////////////////////////////////////////////////////
// A lattice of something, but assume the something is SIMDized.
@ -152,29 +158,45 @@ public:
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
_Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
// if ( ptr != NULL )
// std::cout << "alignedAllocator "<<__n << " cache hit "<< std::hex << ptr <<std::dec <<std::endl;
//////////////////
// Hack 2MB align; could make option probably doesn't need configurability
//////////////////
//define GRID_ALLOC_ALIGN (128)
#define GRID_ALLOC_ALIGN (2*1024*1024)
#ifdef HAVE_MM_MALLOC_H
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
#ifdef POINTER_CACHE
_Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
#else
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
pointer ptr = nullptr;
#endif
// std::cout << "alignedAllocator " << std::hex << ptr <<std::dec <<std::endl;
// First touch optimise in threaded loop
uint8_t *cp = (uint8_t *)ptr;
#ifdef GRID_OMP
#pragma omp parallel for
#endif
for(size_type n=0;n<bytes;n+=4096){
cp[n]=0;
#ifdef GRID_NVCC
////////////////////////////////////
// Unified (managed) memory
////////////////////////////////////
if ( ptr == (_Tp *) NULL ) {
auto err = cudaMallocManaged((void **)&ptr,bytes);
if( err != cudaSuccess ) {
ptr = (_Tp *) NULL;
std::cerr << " cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
assert(0);
}
}
assert( ptr != (_Tp *)NULL);
#else
//////////////////////////////////////////////////////////////////////////////////////////
// 2MB align; could make option probably doesn't need configurability
//////////////////////////////////////////////////////////////////////////////////////////
#ifdef HAVE_MM_MALLOC_H
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) _mm_malloc(bytes,GRID_ALLOC_ALIGN);
#else
if ( ptr == (_Tp *) NULL ) ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,bytes);
#endif
assert( ptr != (_Tp *)NULL);
//////////////////////////////////////////////////
// First touch optimise in threaded loop
//////////////////////////////////////////////////
uint64_t *cp = (uint64_t *)ptr;
thread_for(n,bytes/sizeof(uint64_t), { // need only one touch per page
cp[n]=0;
});
#endif
return ptr;
}
@ -183,133 +205,40 @@ public:
profilerFree(bytes);
#ifdef POINTER_CACHE
pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
#else
pointer __freeme = __p;
#endif
#ifdef HAVE_MM_MALLOC_H
#ifdef GRID_NVCC
if ( __freeme ) cudaFree((void *)__freeme);
#else
#ifdef HAVE_MM_MALLOC_H
if ( __freeme ) _mm_free((void *)__freeme);
#else
#else
if ( __freeme ) free((void *)__freeme);
#endif
#endif
}
void construct(pointer __p, const _Tp& __val) { };
// FIXME: hack for the copy constructor, eventually it must be avoided
void construct(pointer __p, const _Tp& __val) { new((void *)__p) _Tp(__val); };
//void construct(pointer __p, const _Tp& __val) { };
void construct(pointer __p) { };
void destroy(pointer __p) { };
};
template<typename _Tp> inline bool operator==(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return true; }
template<typename _Tp> inline bool operator!=(const alignedAllocator<_Tp>&, const alignedAllocator<_Tp>&){ return false; }
//////////////////////////////////////////////////////////////////////////////////////////
// MPI3 : comms must use shm region
// SHMEM: comms must use symmetric heap
//////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_COMMS_SHMEM
extern "C" {
#include <mpp/shmem.h>
extern void * shmem_align(size_t, size_t);
extern void shmem_free(void *);
}
#define PARANOID_SYMMETRIC_HEAP
#endif
template<typename _Tp>
class commAllocator {
public:
typedef std::size_t size_type;
typedef std::ptrdiff_t difference_type;
typedef _Tp* pointer;
typedef const _Tp* const_pointer;
typedef _Tp& reference;
typedef const _Tp& const_reference;
typedef _Tp value_type;
template<typename _Tp1> struct rebind { typedef commAllocator<_Tp1> other; };
commAllocator() throw() { }
commAllocator(const commAllocator&) throw() { }
template<typename _Tp1> commAllocator(const commAllocator<_Tp1>&) throw() { }
~commAllocator() throw() { }
pointer address(reference __x) const { return &__x; }
size_type max_size() const throw() { return size_t(-1) / sizeof(_Tp); }
#ifdef GRID_COMMS_SHMEM
pointer allocate(size_type __n, const void* _p= 0)
{
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
#ifdef CRAY
_Tp *ptr = (_Tp *) shmem_align(bytes,64);
#else
_Tp *ptr = (_Tp *) shmem_align(64,bytes);
#endif
#ifdef PARANOID_SYMMETRIC_HEAP
static void * bcast;
static long psync[_SHMEM_REDUCE_SYNC_SIZE];
bcast = (void *) ptr;
shmem_broadcast32((void *)&bcast,(void *)&bcast,sizeof(void *)/4,0,0,0,shmem_n_pes(),psync);
if ( bcast != ptr ) {
std::printf("inconsistent alloc pe %d %lx %lx \n",shmem_my_pe(),bcast,ptr);std::fflush(stdout);
// BACKTRACEFILE();
exit(0);
}
assert( bcast == (void *) ptr);
#endif
return ptr;
}
void deallocate(pointer __p, size_type __n) {
size_type bytes = __n*sizeof(_Tp);
profilerFree(bytes);
shmem_free((void *)__p);
}
#else
pointer allocate(size_type __n, const void* _p= 0)
{
size_type bytes = __n*sizeof(_Tp);
profilerAllocate(bytes);
#ifdef HAVE_MM_MALLOC_H
_Tp * ptr = (_Tp *) _mm_malloc(bytes, GRID_ALLOC_ALIGN);
#else
_Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN, bytes);
#endif
uint8_t *cp = (uint8_t *)ptr;
if ( ptr ) {
// One touch per 4k page, static OMP loop to catch same loop order
#ifdef GRID_OMP
#pragma omp parallel for schedule(static)
#endif
for(size_type n=0;n<bytes;n+=4096){
cp[n]=0;
}
}
return ptr;
}
void deallocate(pointer __p, size_type __n) {
size_type bytes = __n*sizeof(_Tp);
profilerFree(bytes);
#ifdef HAVE_MM_MALLOC_H
_mm_free((void *)__p);
#else
free((void *)__p);
#endif
}
#endif
void construct(pointer __p, const _Tp& __val) { };
void construct(pointer __p) { };
void destroy(pointer __p) { };
};
template<typename _Tp> inline bool operator==(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return true; }
template<typename _Tp> inline bool operator!=(const commAllocator<_Tp>&, const commAllocator<_Tp>&){ return false; }
////////////////////////////////////////////////////////////////////////////////
// Template typedefs
////////////////////////////////////////////////////////////////////////////////
template<class T> using commAllocator = alignedAllocator<T>;
template<class T> using Vector = std::vector<T,alignedAllocator<T> >;
template<class T> using commVector = std::vector<T,commAllocator<T> >;
template<class T> using commVector = std::vector<T,alignedAllocator<T> >;
template<class T> using Matrix = std::vector<std::vector<T,alignedAllocator<T> > >;
}; // namespace Grid
NAMESPACE_END(Grid);
#endif

6
Grid/cartesian/Cartesian.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,8 +23,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CARTESIAN_H
#define GRID_CARTESIAN_H

452
Grid/cartesian/Cartesian_base.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -25,268 +25,266 @@
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CARTESIAN_BASE_H
#define GRID_CARTESIAN_BASE_H
NAMESPACE_BEGIN(Grid);
namespace Grid{
//////////////////////////////////////////////////////////////////////
// Commicator provides information on the processor grid
//////////////////////////////////////////////////////////////////////
// unsigned long _ndimension;
// std::vector<int> _processors; // processor grid
// int _processor; // linear processor rank
// std::vector<int> _processor_coor; // linear processor rank
//////////////////////////////////////////////////////////////////////
class GridBase : public CartesianCommunicator , public GridThread {
//////////////////////////////////////////////////////////////////////
// Commicator provides information on the processor grid
//////////////////////////////////////////////////////////////////////
// unsigned long _ndimension;
// Coordinate _processors; // processor grid
// int _processor; // linear processor rank
// Coordinate _processor_coor; // linear processor rank
//////////////////////////////////////////////////////////////////////
class GridBase : public CartesianCommunicator , public GridThread {
public:
int dummy;
// Give Lattice access
template<class object> friend class Lattice;
int dummy;
// Give Lattice access
template<class object> friend class Lattice;
GridBase(const std::vector<int> & processor_grid) : CartesianCommunicator(processor_grid) {};
GridBase(const std::vector<int> & processor_grid,
const CartesianCommunicator &parent,
int &split_rank)
: CartesianCommunicator(processor_grid,parent,split_rank) {};
GridBase(const std::vector<int> & processor_grid,
const CartesianCommunicator &parent)
: CartesianCommunicator(processor_grid,parent,dummy) {};
GridBase(const Coordinate & processor_grid) : CartesianCommunicator(processor_grid) {};
virtual ~GridBase() = default;
GridBase(const Coordinate & processor_grid,
const CartesianCommunicator &parent,
int &split_rank)
: CartesianCommunicator(processor_grid,parent,split_rank) {};
GridBase(const Coordinate & processor_grid,
const CartesianCommunicator &parent)
: CartesianCommunicator(processor_grid,parent,dummy) {};
virtual ~GridBase() = default;
// Physics Grid information.
std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.
std::vector<int> _fdimensions;// (full) Global dimensions of array prior to cb removal
std::vector<int> _gdimensions;// Global dimensions of array after cb removal
std::vector<int> _ldimensions;// local dimensions of array with processor images removed
std::vector<int> _rdimensions;// Reduced local dimensions with simd lane images and processor images removed
std::vector<int> _ostride; // Outer stride for each dimension
std::vector<int> _istride; // Inner stride i.e. within simd lane
int _osites; // _isites*_osites = product(dimensions).
int _isites;
int _fsites; // _isites*_osites = product(dimensions).
int _gsites;
std::vector<int> _slice_block;// subslice information
std::vector<int> _slice_stride;
std::vector<int> _slice_nblock;
// Physics Grid information.
Coordinate _simd_layout;// Which dimensions get relayed out over simd lanes.
Coordinate _fdimensions;// (full) Global dimensions of array prior to cb removal
Coordinate _gdimensions;// Global dimensions of array after cb removal
Coordinate _ldimensions;// local dimensions of array with processor images removed
Coordinate _rdimensions;// Reduced local dimensions with simd lane images and processor images removed
Coordinate _ostride; // Outer stride for each dimension
Coordinate _istride; // Inner stride i.e. within simd lane
int _osites; // _isites*_osites = product(dimensions).
int _isites;
int _fsites; // _isites*_osites = product(dimensions).
int _gsites;
Coordinate _slice_block;// subslice information
Coordinate _slice_stride;
Coordinate _slice_nblock;
std::vector<int> _lstart; // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
std::vector<int> _lend ; // local end of array in gcoors _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
Coordinate _lstart; // local start of array in gcoors _processor_coor[d]*_ldimensions[d]
Coordinate _lend ; // local end of array in gcoors _processor_coor[d]*_ldimensions[d]+_ldimensions_[d]-1
bool _isCheckerBoarded;
public:
////////////////////////////////////////////////////////////////
// Checkerboarding interface is virtual and overridden by
// GridCartesian / GridRedBlackCartesian
////////////////////////////////////////////////////////////////
virtual int CheckerBoarded(int dim)=0;
virtual int CheckerBoard(const std::vector<int> &site)=0;
virtual int CheckerBoardDestination(int source_cb,int shift,int dim)=0;
virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite)=0;
virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int cb)=0;
virtual int CheckerBoardFromOindex (int Oindex)=0;
virtual int CheckerBoardFromOindexTable (int Oindex)=0;
////////////////////////////////////////////////////////////////
// Checkerboarding interface is virtual and overridden by
// GridCartesian / GridRedBlackCartesian
////////////////////////////////////////////////////////////////
virtual int CheckerBoarded(int dim)=0;
virtual int CheckerBoard(const Coordinate &site)=0;
virtual int CheckerBoardDestination(int source_cb,int shift,int dim)=0;
virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite)=0;
virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int cb)=0;
virtual int CheckerBoardFromOindex (int Oindex)=0;
virtual int CheckerBoardFromOindexTable (int Oindex)=0;
//////////////////////////////////////////////////////////////////////////////////////////////
// Local layout calculations
//////////////////////////////////////////////////////////////////////////////////////////////
// These routines are key. Subdivide the linearised cartesian index into
// "inner" index identifying which simd lane of object<vFcomplex> is associated with coord
// "outer" index identifying which element of _odata in class "Lattice" is associated with coord.
//
// Compared to, say, Blitz++ we simply need to store BOTH an inner stride and an outer
// stride per dimension. The cost of evaluating the indexing information is doubled for an n-dimensional
// coordinate. Note, however, for data parallel operations the "inner" indexing cost is not paid and all
// lanes are operated upon simultaneously.
//////////////////////////////////////////////////////////////////////////////////////////////
// Local layout calculations
//////////////////////////////////////////////////////////////////////////////////////////////
// These routines are key. Subdivide the linearised cartesian index into
// "inner" index identifying which simd lane of object<vFcomplex> is associated with coord
// "outer" index identifying which element of _odata in class "Lattice" is associated with coord.
//
// Compared to, say, Blitz++ we simply need to store BOTH an inner stride and an outer
// stride per dimension. The cost of evaluating the indexing information is doubled for an n-dimensional
// coordinate. Note, however, for data parallel operations the "inner" indexing cost is not paid and all
// lanes are operated upon simultaneously.
virtual int oIndex(std::vector<int> &coor)
{
int idx=0;
// Works with either global or local coordinates
for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
return idx;
}
virtual int iIndex(std::vector<int> &lcoor)
{
int idx=0;
for(int d=0;d<_ndimension;d++) idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
return idx;
}
inline int oIndexReduced(std::vector<int> &ocoor)
{
int idx=0;
// ocoor is already reduced so can eliminate the modulo operation
// for fast indexing and inline the routine
for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*ocoor[d];
return idx;
}
inline void oCoorFromOindex (std::vector<int>& coor,int Oindex){
Lexicographic::CoorFromIndex(coor,Oindex,_rdimensions);
}
virtual int oIndex(Coordinate &coor)
{
int idx=0;
// Works with either global or local coordinates
for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*(coor[d]%_rdimensions[d]);
return idx;
}
virtual int iIndex(Coordinate &lcoor)
{
int idx=0;
for(int d=0;d<_ndimension;d++) idx+=_istride[d]*(lcoor[d]/_rdimensions[d]);
return idx;
}
inline int oIndexReduced(Coordinate &ocoor)
{
int idx=0;
// ocoor is already reduced so can eliminate the modulo operation
// for fast indexing and inline the routine
for(int d=0;d<_ndimension;d++) idx+=_ostride[d]*ocoor[d];
return idx;
}
inline void oCoorFromOindex (Coordinate& coor,int Oindex){
Lexicographic::CoorFromIndex(coor,Oindex,_rdimensions);
}
inline void InOutCoorToLocalCoor (std::vector<int> &ocoor, std::vector<int> &icoor, std::vector<int> &lcoor) {
lcoor.resize(_ndimension);
for (int d = 0; d < _ndimension; d++)
lcoor[d] = ocoor[d] + _rdimensions[d] * icoor[d];
}
inline void InOutCoorToLocalCoor (Coordinate &ocoor, Coordinate &icoor, Coordinate &lcoor) {
lcoor.resize(_ndimension);
for (int d = 0; d < _ndimension; d++)
lcoor[d] = ocoor[d] + _rdimensions[d] * icoor[d];
}
//////////////////////////////////////////////////////////
// SIMD lane addressing
//////////////////////////////////////////////////////////
inline void iCoorFromIindex(std::vector<int> &coor,int lane)
{
Lexicographic::CoorFromIndex(coor,lane,_simd_layout);
}
//////////////////////////////////////////////////////////
// SIMD lane addressing
//////////////////////////////////////////////////////////
inline void iCoorFromIindex(Coordinate &coor,int lane)
{
Lexicographic::CoorFromIndex(coor,lane,_simd_layout);
}
inline int PermuteDim(int dimension){
return _simd_layout[dimension]>1;
}
inline int PermuteType(int dimension){
int permute_type=0;
//
// FIXME:
//
// Best way to encode this would be to present a mask
// for which simd dimensions are rotated, and the rotation
// size. If there is only one simd dimension rotated, this is just
// a permute.
//
// Cases: PermuteType == 1,2,4,8
// Distance should be either 0,1,2..
//
if ( _simd_layout[dimension] > 2 ) {
for(int d=0;d<_ndimension;d++){
if ( d != dimension ) assert ( (_simd_layout[d]==1) );
}
permute_type = RotateBit; // How to specify distance; this is not just direction.
return permute_type;
}
for(int d=_ndimension-1;d>dimension;d--){
if (_simd_layout[d]>1 ) permute_type++;
inline int PermuteDim(int dimension){
return _simd_layout[dimension]>1;
}
inline int PermuteType(int dimension){
int permute_type=0;
//
// Best way to encode this would be to present a mask
// for which simd dimensions are rotated, and the rotation
// size. If there is only one simd dimension rotated, this is just
// a permute.
//
// Cases: PermuteType == 1,2,4,8
// Distance should be either 0,1,2..
//
if ( _simd_layout[dimension] > 2 ) {
for(int d=0;d<_ndimension;d++){
if ( d != dimension ) assert ( (_simd_layout[d]==1) );
}
permute_type = RotateBit; // How to specify distance; this is not just direction.
return permute_type;
}
////////////////////////////////////////////////////////////////
// Array sizing queries
////////////////////////////////////////////////////////////////
inline int iSites(void) const { return _isites; };
inline int Nsimd(void) const { return _isites; };// Synonymous with iSites
inline int oSites(void) const { return _osites; };
inline int lSites(void) const { return _isites*_osites; };
inline int gSites(void) const { return _isites*_osites*_Nprocessors; };
inline int Nd (void) const { return _ndimension;};
inline const std::vector<int> LocalStarts(void) { return _lstart; };
inline const std::vector<int> &FullDimensions(void) { return _fdimensions;};
inline const std::vector<int> &GlobalDimensions(void) { return _gdimensions;};
inline const std::vector<int> &LocalDimensions(void) { return _ldimensions;};
inline const std::vector<int> &VirtualLocalDimensions(void) { return _ldimensions;};
////////////////////////////////////////////////////////////////
// Utility to print the full decomposition details
////////////////////////////////////////////////////////////////
void show_decomposition(){
std::cout << GridLogMessage << "\tFull Dimensions : " << _fdimensions << std::endl;
std::cout << GridLogMessage << "\tSIMD layout : " << _simd_layout << std::endl;
std::cout << GridLogMessage << "\tGlobal Dimensions : " << _gdimensions << std::endl;
std::cout << GridLogMessage << "\tLocal Dimensions : " << _ldimensions << std::endl;
std::cout << GridLogMessage << "\tReduced Dimensions : " << _rdimensions << std::endl;
std::cout << GridLogMessage << "\tOuter strides : " << _ostride << std::endl;
std::cout << GridLogMessage << "\tInner strides : " << _istride << std::endl;
std::cout << GridLogMessage << "\tiSites : " << _isites << std::endl;
std::cout << GridLogMessage << "\toSites : " << _osites << std::endl;
std::cout << GridLogMessage << "\tlSites : " << lSites() << std::endl;
std::cout << GridLogMessage << "\tgSites : " << gSites() << std::endl;
std::cout << GridLogMessage << "\tNd : " << _ndimension << std::endl;
}
////////////////////////////////////////////////////////////////
// Global addressing
////////////////////////////////////////////////////////////////
void GlobalIndexToGlobalCoor(int gidx,std::vector<int> &gcoor){
assert(gidx< gSites());
Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
for(int d=_ndimension-1;d>dimension;d--){
if (_simd_layout[d]>1 ) permute_type++;
}
void LocalIndexToLocalCoor(int lidx,std::vector<int> &lcoor){
assert(lidx<lSites());
Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
return permute_type;
}
////////////////////////////////////////////////////////////////
// Array sizing queries
////////////////////////////////////////////////////////////////
inline int iSites(void) const { return _isites; };
inline int Nsimd(void) const { return _isites; };// Synonymous with iSites
inline int oSites(void) const { return _osites; };
inline int lSites(void) const { return _isites*_osites; };
inline int gSites(void) const { return _isites*_osites*_Nprocessors; };
inline int Nd (void) const { return _ndimension;};
inline const Coordinate LocalStarts(void) { return _lstart; };
inline const Coordinate &FullDimensions(void) { return _fdimensions;};
inline const Coordinate &GlobalDimensions(void) { return _gdimensions;};
inline const Coordinate &LocalDimensions(void) { return _ldimensions;};
inline const Coordinate &VirtualLocalDimensions(void) { return _ldimensions;};
////////////////////////////////////////////////////////////////
// Utility to print the full decomposition details
////////////////////////////////////////////////////////////////
void show_decomposition(){
std::cout << GridLogMessage << "\tFull Dimensions : " << _fdimensions << std::endl;
std::cout << GridLogMessage << "\tSIMD layout : " << _simd_layout << std::endl;
std::cout << GridLogMessage << "\tGlobal Dimensions : " << _gdimensions << std::endl;
std::cout << GridLogMessage << "\tLocal Dimensions : " << _ldimensions << std::endl;
std::cout << GridLogMessage << "\tReduced Dimensions : " << _rdimensions << std::endl;
std::cout << GridLogMessage << "\tOuter strides : " << _ostride << std::endl;
std::cout << GridLogMessage << "\tInner strides : " << _istride << std::endl;
std::cout << GridLogMessage << "\tiSites : " << _isites << std::endl;
std::cout << GridLogMessage << "\toSites : " << _osites << std::endl;
std::cout << GridLogMessage << "\tlSites : " << lSites() << std::endl;
std::cout << GridLogMessage << "\tgSites : " << gSites() << std::endl;
std::cout << GridLogMessage << "\tNd : " << _ndimension << std::endl;
}
////////////////////////////////////////////////////////////////
// Global addressing
////////////////////////////////////////////////////////////////
void GlobalIndexToGlobalCoor(int gidx,Coordinate &gcoor){
assert(gidx< gSites());
Lexicographic::CoorFromIndex(gcoor,gidx,_gdimensions);
}
void LocalIndexToLocalCoor(int lidx,Coordinate &lcoor){
assert(lidx<lSites());
Lexicographic::CoorFromIndex(lcoor,lidx,_ldimensions);
}
void GlobalCoorToGlobalIndex(const Coordinate & gcoor,int & gidx){
gidx=0;
int mult=1;
for(int mu=0;mu<_ndimension;mu++) {
gidx+=mult*gcoor[mu];
mult*=_gdimensions[mu];
}
void GlobalCoorToGlobalIndex(const std::vector<int> & gcoor,int & gidx){
gidx=0;
int mult=1;
for(int mu=0;mu<_ndimension;mu++) {
gidx+=mult*gcoor[mu];
mult*=_gdimensions[mu];
}
}
void GlobalCoorToProcessorCoorLocalCoor(Coordinate &pcoor,Coordinate &lcoor,const Coordinate &gcoor)
{
pcoor.resize(_ndimension);
lcoor.resize(_ndimension);
for(int mu=0;mu<_ndimension;mu++){
int _fld = _fdimensions[mu]/_processors[mu];
pcoor[mu] = gcoor[mu]/_fld;
lcoor[mu] = gcoor[mu]%_fld;
}
void GlobalCoorToProcessorCoorLocalCoor(std::vector<int> &pcoor,std::vector<int> &lcoor,const std::vector<int> &gcoor)
{
pcoor.resize(_ndimension);
lcoor.resize(_ndimension);
for(int mu=0;mu<_ndimension;mu++){
int _fld = _fdimensions[mu]/_processors[mu];
pcoor[mu] = gcoor[mu]/_fld;
lcoor[mu] = gcoor[mu]%_fld;
}
}
void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const std::vector<int> &gcoor)
{
std::vector<int> pcoor;
std::vector<int> lcoor;
GlobalCoorToProcessorCoorLocalCoor(pcoor,lcoor,gcoor);
rank = RankFromProcessorCoor(pcoor);
/*
std::vector<int> cblcoor(lcoor);
}
void GlobalCoorToRankIndex(int &rank, int &o_idx, int &i_idx ,const Coordinate &gcoor)
{
Coordinate pcoor;
Coordinate lcoor;
GlobalCoorToProcessorCoorLocalCoor(pcoor,lcoor,gcoor);
rank = RankFromProcessorCoor(pcoor);
/*
Coordinate cblcoor(lcoor);
for(int d=0;d<cblcoor.size();d++){
if( this->CheckerBoarded(d) ) {
cblcoor[d] = lcoor[d]/2;
}
if( this->CheckerBoarded(d) ) {
cblcoor[d] = lcoor[d]/2;
}
*/
i_idx= iIndex(lcoor);
o_idx= oIndex(lcoor);
}
}
*/
i_idx= iIndex(lcoor);
o_idx= oIndex(lcoor);
}
void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , std::vector<int> &gcoor)
{
gcoor.resize(_ndimension);
std::vector<int> coor(_ndimension);
void RankIndexToGlobalCoor(int rank, int o_idx, int i_idx , Coordinate &gcoor)
{
gcoor.resize(_ndimension);
Coordinate coor(_ndimension);
ProcessorCoorFromRank(rank,coor);
for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = _ldimensions[mu]*coor[mu];
ProcessorCoorFromRank(rank,coor);
for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = _ldimensions[mu]*coor[mu];
iCoorFromIindex(coor,i_idx);
for(int mu=0;mu<_ndimension;mu++) gcoor[mu] += _rdimensions[mu]*coor[mu];
iCoorFromIindex(coor,i_idx);
for(int mu=0;mu<_ndimension;mu++) gcoor[mu] += _rdimensions[mu]*coor[mu];
oCoorFromOindex (coor,o_idx);
for(int mu=0;mu<_ndimension;mu++) gcoor[mu] += coor[mu];
oCoorFromOindex (coor,o_idx);
for(int mu=0;mu<_ndimension;mu++) gcoor[mu] += coor[mu];
}
void RankIndexCbToFullGlobalCoor(int rank, int o_idx, int i_idx, int cb,Coordinate &fcoor)
{
RankIndexToGlobalCoor(rank,o_idx,i_idx ,fcoor);
if(CheckerBoarded(0)){
fcoor[0] = fcoor[0]*2+cb;
}
void RankIndexCbToFullGlobalCoor(int rank, int o_idx, int i_idx, int cb,std::vector<int> &fcoor)
{
RankIndexToGlobalCoor(rank,o_idx,i_idx ,fcoor);
if(CheckerBoarded(0)){
fcoor[0] = fcoor[0]*2+cb;
}
}
void ProcessorCoorLocalCoorToGlobalCoor(std::vector<int> &Pcoor,std::vector<int> &Lcoor,std::vector<int> &gcoor)
{
gcoor.resize(_ndimension);
for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = Pcoor[mu]*_ldimensions[mu]+Lcoor[mu];
}
}
void ProcessorCoorLocalCoorToGlobalCoor(Coordinate &Pcoor,Coordinate &Lcoor,Coordinate &gcoor)
{
gcoor.resize(_ndimension);
for(int mu=0;mu<_ndimension;mu++) gcoor[mu] = Pcoor[mu]*_ldimensions[mu]+Lcoor[mu];
}
};
}
NAMESPACE_END(Grid);
#endif

190
Grid/cartesian/Cartesian_full.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,97 +23,96 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CARTESIAN_FULL_H
#define GRID_CARTESIAN_FULL_H
namespace Grid{
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////////////
// Grid Support.
/////////////////////////////////////////////////////////////////////////////////////////
class GridCartesian: public GridBase {
public:
int dummy;
virtual int CheckerBoardFromOindexTable (int Oindex) {
return 0;
}
virtual int CheckerBoardFromOindex (int Oindex)
{
return 0;
}
virtual int CheckerBoarded(int dim){
return 0;
}
virtual int CheckerBoard(const std::vector<int> &site){
return 0;
}
virtual int CheckerBoardDestination(int cb,int shift,int dim){
return 0;
}
virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift, int ocb){
return shift;
}
virtual int CheckerBoardShift(int source_cb,int dim,int shift, int osite){
return shift;
}
/////////////////////////////////////////////////////////////////////////
// Constructor takes a parent grid and possibly subdivides communicator.
/////////////////////////////////////////////////////////////////////////
GridCartesian(const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid,
const GridCartesian &parent) : GridBase(processor_grid,parent,dummy)
{
Init(dimensions,simd_layout,processor_grid);
}
GridCartesian(const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid,
const GridCartesian &parent,int &split_rank) : GridBase(processor_grid,parent,split_rank)
{
Init(dimensions,simd_layout,processor_grid);
}
/////////////////////////////////////////////////////////////////////////
// Construct from comm world
/////////////////////////////////////////////////////////////////////////
GridCartesian(const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid) : GridBase(processor_grid)
{
Init(dimensions,simd_layout,processor_grid);
}
int dummy;
virtual int CheckerBoardFromOindexTable (int Oindex) {
return 0;
}
virtual int CheckerBoardFromOindex (int Oindex)
{
return 0;
}
virtual int CheckerBoarded(int dim){
return 0;
}
virtual int CheckerBoard(const Coordinate &site){
return 0;
}
virtual int CheckerBoardDestination(int cb,int shift,int dim){
return 0;
}
virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift, int ocb){
return shift;
}
virtual int CheckerBoardShift(int source_cb,int dim,int shift, int osite){
return shift;
}
/////////////////////////////////////////////////////////////////////////
// Constructor takes a parent grid and possibly subdivides communicator.
/////////////////////////////////////////////////////////////////////////
GridCartesian(const Coordinate &dimensions,
const Coordinate &simd_layout,
const Coordinate &processor_grid,
const GridCartesian &parent) : GridBase(processor_grid,parent,dummy)
{
Init(dimensions,simd_layout,processor_grid);
}
GridCartesian(const Coordinate &dimensions,
const Coordinate &simd_layout,
const Coordinate &processor_grid,
const GridCartesian &parent,int &split_rank) : GridBase(processor_grid,parent,split_rank)
{
Init(dimensions,simd_layout,processor_grid);
}
/////////////////////////////////////////////////////////////////////////
// Construct from comm world
/////////////////////////////////////////////////////////////////////////
GridCartesian(const Coordinate &dimensions,
const Coordinate &simd_layout,
const Coordinate &processor_grid) : GridBase(processor_grid)
{
Init(dimensions,simd_layout,processor_grid);
}
virtual ~GridCartesian() = default;
virtual ~GridCartesian() = default;
void Init(const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid)
{
///////////////////////
// Grid information
///////////////////////
void Init(const Coordinate &dimensions,
const Coordinate &simd_layout,
const Coordinate &processor_grid)
{
///////////////////////
// Grid information
///////////////////////
_isCheckerBoarded = false;
_ndimension = dimensions.size();
_ndimension = dimensions.size();
_fdimensions.resize(_ndimension);
_gdimensions.resize(_ndimension);
_ldimensions.resize(_ndimension);
_rdimensions.resize(_ndimension);
_simd_layout.resize(_ndimension);
_lstart.resize(_ndimension);
_lend.resize(_ndimension);
_fdimensions.resize(_ndimension);
_gdimensions.resize(_ndimension);
_ldimensions.resize(_ndimension);
_rdimensions.resize(_ndimension);
_simd_layout.resize(_ndimension);
_lstart.resize(_ndimension);
_lend.resize(_ndimension);
_ostride.resize(_ndimension);
_istride.resize(_ndimension);
_ostride.resize(_ndimension);
_istride.resize(_ndimension);
_fsites = _gsites = _osites = _isites = 1;
_fsites = _gsites = _osites = _isites = 1;
for (int d = 0; d < _ndimension; d++)
for (int d = 0; d < _ndimension; d++)
{
_fdimensions[d] = dimensions[d]; // Global dimensions
_gdimensions[d] = _fdimensions[d]; // Global dimensions
@ -136,30 +135,30 @@ public:
// Addressing support
if (d == 0)
{
_ostride[d] = 1;
_istride[d] = 1;
}
{
_ostride[d] = 1;
_istride[d] = 1;
}
else
{
_ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
_istride[d] = _istride[d - 1] * _simd_layout[d - 1];
}
{
_ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
_istride[d] = _istride[d - 1] * _simd_layout[d - 1];
}
}
///////////////////////
// subplane information
///////////////////////
_slice_block.resize(_ndimension);
_slice_stride.resize(_ndimension);
_slice_nblock.resize(_ndimension);
///////////////////////
// subplane information
///////////////////////
_slice_block.resize(_ndimension);
_slice_stride.resize(_ndimension);
_slice_nblock.resize(_ndimension);
int block = 1;
int nblock = 1;
for (int d = 0; d < _ndimension; d++)
nblock *= _rdimensions[d];
int block = 1;
int nblock = 1;
for (int d = 0; d < _ndimension; d++)
nblock *= _rdimensions[d];
for (int d = 0; d < _ndimension; d++)
for (int d = 0; d < _ndimension; d++)
{
nblock /= _rdimensions[d];
_slice_block[d] = block;
@ -167,8 +166,9 @@ public:
_slice_nblock[d] = nblock;
block = block * _rdimensions[d];
}
};
};
};
}
NAMESPACE_END(Grid);
#endif

373
Grid/cartesian/Cartesian_red_black.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,178 +24,147 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_CARTESIAN_RED_BLACK_H
#define GRID_CARTESIAN_RED_BLACK_H
NAMESPACE_BEGIN(Grid);
namespace Grid {
static const int CbRed =0;
static const int CbBlack=1;
static const int Even =CbRed;
static const int Odd =CbBlack;
static const int CbRed =0;
static const int CbBlack=1;
static const int Even =CbRed;
static const int Odd =CbBlack;
// Specialise this for red black grids storing half the data like a chess board.
class GridRedBlackCartesian : public GridBase
{
public:
std::vector<int> _checker_dim_mask;
int _checker_dim;
std::vector<int> _checker_board;
Coordinate _checker_dim_mask;
int _checker_dim;
std::vector<int> _checker_board;
virtual int CheckerBoarded(int dim){
if( dim==_checker_dim) return 1;
else return 0;
}
virtual int CheckerBoard(const std::vector<int> &site){
int linear=0;
assert(site.size()==_ndimension);
for(int d=0;d<_ndimension;d++){
if(_checker_dim_mask[d])
linear=linear+site[d];
}
return (linear&0x1);
virtual int CheckerBoarded(int dim){
if( dim==_checker_dim) return 1;
else return 0;
}
virtual int CheckerBoard(const Coordinate &site){
int linear=0;
assert(site.size()==_ndimension);
for(int d=0;d<_ndimension;d++){
if(_checker_dim_mask[d])
linear=linear+site[d];
}
return (linear&0x1);
}
// Depending on the cb of site, we toggle source cb.
// for block #b, element #e = (b, e)
// we need
virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int ocb){
if(dim != _checker_dim) return shift;
// Depending on the cb of site, we toggle source cb.
// for block #b, element #e = (b, e)
// we need
virtual int CheckerBoardShiftForCB(int source_cb,int dim,int shift,int ocb){
if(dim != _checker_dim) return shift;
int fulldim =_fdimensions[dim];
shift = (shift+fulldim)%fulldim;
int fulldim =_fdimensions[dim];
shift = (shift+fulldim)%fulldim;
// Probably faster with table lookup;
// or by looping over x,y,z and multiply rather than computing checkerboard.
// Probably faster with table lookup;
// or by looping over x,y,z and multiply rather than computing checkerboard.
if ( (source_cb+ocb)&1 ) {
return (shift)/2;
} else {
return (shift+1)/2;
}
if ( (source_cb+ocb)&1 ) {
return (shift)/2;
} else {
return (shift+1)/2;
}
virtual int CheckerBoardFromOindexTable (int Oindex) {
return _checker_board[Oindex];
}
virtual int CheckerBoardFromOindex (int Oindex)
{
std::vector<int> ocoor;
oCoorFromOindex(ocoor,Oindex);
return CheckerBoard(ocoor);
}
virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite){
}
virtual int CheckerBoardFromOindexTable (int Oindex) {
return _checker_board[Oindex];
}
virtual int CheckerBoardFromOindex (int Oindex)
{
Coordinate ocoor;
oCoorFromOindex(ocoor,Oindex);
return CheckerBoard(ocoor);
}
virtual int CheckerBoardShift(int source_cb,int dim,int shift,int osite){
if(dim != _checker_dim) return shift;
if(dim != _checker_dim) return shift;
int ocb=CheckerBoardFromOindex(osite);
int ocb=CheckerBoardFromOindex(osite);
return CheckerBoardShiftForCB(source_cb,dim,shift,ocb);
}
return CheckerBoardShiftForCB(source_cb,dim,shift,ocb);
}
virtual int CheckerBoardDestination(int source_cb,int shift,int dim){
if ( _checker_dim_mask[dim] ) {
// If _fdimensions[checker_dim] is odd, then shifting by 1 in other dims
// does NOT cause a parity hop.
int add=(dim==_checker_dim) ? 0 : _fdimensions[_checker_dim];
if ( (shift+add) &0x1) {
return 1-source_cb;
} else {
return source_cb;
}
virtual int CheckerBoardDestination(int source_cb,int shift,int dim){
if ( _checker_dim_mask[dim] ) {
// If _fdimensions[checker_dim] is odd, then shifting by 1 in other dims
// does NOT cause a parity hop.
int add=(dim==_checker_dim) ? 0 : _fdimensions[_checker_dim];
if ( (shift+add) &0x1) {
return 1-source_cb;
} else {
return source_cb;
}
};
} else {
return source_cb;
////////////////////////////////////////////////////////////
// Create Redblack from original grid; require full grid pointer ?
////////////////////////////////////////////////////////////
GridRedBlackCartesian(const GridBase *base) : GridBase(base->_processors,*base)
{
int dims = base->_ndimension;
std::vector<int> checker_dim_mask(dims,1);
int checker_dim = 0;
Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim);
};
////////////////////////////////////////////////////////////
// Create redblack from original grid, with non-trivial checker dim mask
////////////////////////////////////////////////////////////
GridRedBlackCartesian(const GridBase *base,
const std::vector<int> &checker_dim_mask,
int checker_dim
) : GridBase(base->_processors,*base)
{
Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim) ;
}
};
virtual ~GridRedBlackCartesian() = default;
#if 0
////////////////////////////////////////////////////////////
// Create redblack grid ;; deprecate these. Should not
// need direct creation of redblack without a full grid to base on
////////////////////////////////////////////////////////////
GridRedBlackCartesian(const GridBase *base,
const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid,
const std::vector<int> &checker_dim_mask,
int checker_dim
) : GridBase(processor_grid,*base)
{
Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
}
////////////////////////////////////////////////////////////
// Create Redblack from original grid; require full grid pointer ?
////////////////////////////////////////////////////////////
GridRedBlackCartesian(const GridBase *base) : GridBase(base->_processors,*base)
{
int dims = base->_ndimension;
Coordinate checker_dim_mask(dims,1);
int checker_dim = 0;
Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim);
};
////////////////////////////////////////////////////////////
// Create redblack grid
////////////////////////////////////////////////////////////
GridRedBlackCartesian(const GridBase *base,
const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid) : GridBase(processor_grid,*base)
{
std::vector<int> checker_dim_mask(dimensions.size(),1);
int checker_dim = 0;
Init(dimensions,simd_layout,processor_grid,checker_dim_mask,checker_dim);
}
#endif
////////////////////////////////////////////////////////////
// Create redblack from original grid, with non-trivial checker dim mask
////////////////////////////////////////////////////////////
GridRedBlackCartesian(const GridBase *base,
const Coordinate &checker_dim_mask,
int checker_dim
) : GridBase(base->_processors,*base)
{
Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim) ;
}
void Init(const std::vector<int> &dimensions,
const std::vector<int> &simd_layout,
const std::vector<int> &processor_grid,
const std::vector<int> &checker_dim_mask,
int checker_dim)
{
virtual ~GridRedBlackCartesian() = default;
void Init(const Coordinate &dimensions,
const Coordinate &simd_layout,
const Coordinate &processor_grid,
const Coordinate &checker_dim_mask,
int checker_dim)
{
_isCheckerBoarded = true;
_checker_dim = checker_dim;
assert(checker_dim_mask[checker_dim] == 1);
_ndimension = dimensions.size();
assert(checker_dim_mask.size() == _ndimension);
assert(processor_grid.size() == _ndimension);
assert(simd_layout.size() == _ndimension);
_checker_dim = checker_dim;
assert(checker_dim_mask[checker_dim] == 1);
_ndimension = dimensions.size();
assert(checker_dim_mask.size() == _ndimension);
assert(processor_grid.size() == _ndimension);
assert(simd_layout.size() == _ndimension);
_fdimensions.resize(_ndimension);
_gdimensions.resize(_ndimension);
_ldimensions.resize(_ndimension);
_rdimensions.resize(_ndimension);
_simd_layout.resize(_ndimension);
_lstart.resize(_ndimension);
_lend.resize(_ndimension);
_fdimensions.resize(_ndimension);
_gdimensions.resize(_ndimension);
_ldimensions.resize(_ndimension);
_rdimensions.resize(_ndimension);
_simd_layout.resize(_ndimension);
_lstart.resize(_ndimension);
_lend.resize(_ndimension);
_ostride.resize(_ndimension);
_istride.resize(_ndimension);
_ostride.resize(_ndimension);
_istride.resize(_ndimension);
_fsites = _gsites = _osites = _isites = 1;
_fsites = _gsites = _osites = _isites = 1;
_checker_dim_mask = checker_dim_mask;
_checker_dim_mask = checker_dim_mask;
for (int d = 0; d < _ndimension; d++)
for (int d = 0; d < _ndimension; d++)
{
_fdimensions[d] = dimensions[d];
_gdimensions[d] = _fdimensions[d];
@ -203,11 +172,11 @@ public:
_gsites = _gsites * _gdimensions[d];
if (d == _checker_dim)
{
assert((_gdimensions[d] & 0x1) == 0);
_gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
_gsites /= 2;
}
{
assert((_gdimensions[d] & 0x1) == 0);
_gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
_gsites /= 2;
}
_ldimensions[d] = _gdimensions[d] / _processors[d];
assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
_lstart[d] = _processor_coor[d] * _ldimensions[d];
@ -222,42 +191,42 @@ public:
// all elements of a simd vector must have same checkerboard.
// If Ls vectorised, this must still be the case; e.g. dwf rb5d
if (_simd_layout[d] > 1)
{
if (checker_dim_mask[d])
{
assert((_rdimensions[d] & 0x1) == 0);
}
}
{
if (checker_dim_mask[d])
{
assert((_rdimensions[d] & 0x1) == 0);
}
}
_osites *= _rdimensions[d];
_isites *= _simd_layout[d];
// Addressing support
if (d == 0)
{
_ostride[d] = 1;
_istride[d] = 1;
}
{
_ostride[d] = 1;
_istride[d] = 1;
}
else
{
_ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
_istride[d] = _istride[d - 1] * _simd_layout[d - 1];
}
{
_ostride[d] = _ostride[d - 1] * _rdimensions[d - 1];
_istride[d] = _istride[d - 1] * _simd_layout[d - 1];
}
}
////////////////////////////////////////////////////////////////////////////////////////////
// subplane information
////////////////////////////////////////////////////////////////////////////////////////////
_slice_block.resize(_ndimension);
_slice_stride.resize(_ndimension);
_slice_nblock.resize(_ndimension);
////////////////////////////////////////////////////////////////////////////////////////////
// subplane information
////////////////////////////////////////////////////////////////////////////////////////////
_slice_block.resize(_ndimension);
_slice_stride.resize(_ndimension);
_slice_nblock.resize(_ndimension);
int block = 1;
int nblock = 1;
for (int d = 0; d < _ndimension; d++)
nblock *= _rdimensions[d];
int block = 1;
int nblock = 1;
for (int d = 0; d < _ndimension; d++)
nblock *= _rdimensions[d];
for (int d = 0; d < _ndimension; d++)
for (int d = 0; d < _ndimension; d++)
{
nblock /= _rdimensions[d];
_slice_block[d] = block;
@ -266,55 +235,55 @@ public:
block = block * _rdimensions[d];
}
////////////////////////////////////////////////
// Create a checkerboard lookup table
////////////////////////////////////////////////
int rvol = 1;
for (int d = 0; d < _ndimension; d++)
////////////////////////////////////////////////
// Create a checkerboard lookup table
////////////////////////////////////////////////
int rvol = 1;
for (int d = 0; d < _ndimension; d++)
{
rvol = rvol * _rdimensions[d];
}
_checker_board.resize(rvol);
for (int osite = 0; osite < _osites; osite++)
_checker_board.resize(rvol);
for (int osite = 0; osite < _osites; osite++)
{
_checker_board[osite] = CheckerBoardFromOindex(osite);
}
};
};
protected:
virtual int oIndex(std::vector<int> &coor)
{
int idx = 0;
for (int d = 0; d < _ndimension; d++)
protected:
virtual int oIndex(Coordinate &coor)
{
int idx = 0;
for (int d = 0; d < _ndimension; d++)
{
if (d == _checker_dim)
{
idx += _ostride[d] * ((coor[d] / 2) % _rdimensions[d]);
}
{
idx += _ostride[d] * ((coor[d] / 2) % _rdimensions[d]);
}
else
{
idx += _ostride[d] * (coor[d] % _rdimensions[d]);
}
{
idx += _ostride[d] * (coor[d] % _rdimensions[d]);
}
}
return idx;
};
return idx;
};
virtual int iIndex(std::vector<int> &lcoor)
{
int idx = 0;
for (int d = 0; d < _ndimension; d++)
virtual int iIndex(Coordinate &lcoor)
{
int idx = 0;
for (int d = 0; d < _ndimension; d++)
{
if (d == _checker_dim)
{
idx += _istride[d] * (lcoor[d] / (2 * _rdimensions[d]));
}
{
idx += _istride[d] * (lcoor[d] / (2 * _rdimensions[d]));
}
else
{
idx += _istride[d] * (lcoor[d] / _rdimensions[d]);
}
{
idx += _istride[d] * (lcoor[d] / _rdimensions[d]);
}
}
return idx;
}
return idx;
}
};
}
NAMESPACE_END(Grid);
#endif

7
Grid/communicator/Communicator.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,11 +23,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_COMMUNICATOR_H
#define GRID_COMMUNICATOR_H
#include <Grid/util/Coordinate.h>
#include <Grid/communicator/SharedMemory.h>
#include <Grid/communicator/Communicator_base.h>

15
Grid/communicator/Communicator_base.cc
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,15 +23,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <fcntl.h>
#include <unistd.h>
#include <limits.h>
#include <sys/mman.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
@ -47,8 +47,8 @@ int CartesianCommunicator::Dimensions(void) { return
int CartesianCommunicator::IsBoss(void) { return _processor==0; };
int CartesianCommunicator::BossRank(void) { return 0; };
int CartesianCommunicator::ThisRank(void) { return _processor; };
const std::vector<int> & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
const std::vector<int> & CartesianCommunicator::ProcessorGrid(void) { return _processors; };
const Coordinate & CartesianCommunicator::ThisProcessorCoor(void) { return _processor_coor; };
const Coordinate & CartesianCommunicator::ProcessorGrid(void) { return _processors; };
int CartesianCommunicator::ProcessorCount(void) { return _Nprocessors; };
////////////////////////////////////////////////////////////////////////////////
@ -72,5 +72,6 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
GlobalSumVector((double *)c,2*N);
}
}
NAMESPACE_END(Grid);

41
Grid/communicator/Communicator_base.h
View File

@ -1,5 +1,5 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,8 +24,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_COMMUNICATOR_BASE_H
#define GRID_COMMUNICATOR_BASE_H
@ -34,7 +34,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
///////////////////////////////////
#include <Grid/communicator/SharedMemory.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
class CartesianCommunicator : public SharedMemory {
@ -52,9 +52,9 @@ public:
// Communicator should know nothing of the physics grid, only processor grid.
////////////////////////////////////////////
int _Nprocessors; // How many in all
std::vector<int> _processors; // Which dimensions get relayed out over processors lanes.
Coordinate _processors; // Which dimensions get relayed out over processors lanes.
int _processor; // linear processor rank
std::vector<int> _processor_coor; // linear processor coordinate
Coordinate _processor_coor; // linear processor coordinate
unsigned long _ndimension;
static Grid_MPI_Comm communicator_world;
Grid_MPI_Comm communicator;
@ -69,34 +69,34 @@ public:
// Constructors to sub-divide a parent communicator
// and default to comm world
////////////////////////////////////////////////
CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank);
CartesianCommunicator(const std::vector<int> &pdimensions_in);
CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank);
CartesianCommunicator(const Coordinate &pdimensions_in);
virtual ~CartesianCommunicator();
private:
private:
////////////////////////////////////////////////
// Private initialise from an MPI communicator
// Can use after an MPI_Comm_split, but hidden from user so private
////////////////////////////////////////////////
void InitFromMPICommunicator(const std::vector<int> &processors, Grid_MPI_Comm communicator_base);
public:
void InitFromMPICommunicator(const Coordinate &processors, Grid_MPI_Comm communicator_base);
public:
////////////////////////////////////////////////////////////////////////////////////////
// Wraps MPI_Cart routines, or implements equivalent on other impls
////////////////////////////////////////////////////////////////////////////////////////
void ShiftedRanks(int dim,int shift,int & source, int & dest);
int RankFromProcessorCoor(std::vector<int> &coor);
void ProcessorCoorFromRank(int rank,std::vector<int> &coor);
int RankFromProcessorCoor(Coordinate &coor);
void ProcessorCoorFromRank(int rank,Coordinate &coor);
int Dimensions(void) ;
int IsBoss(void) ;
int BossRank(void) ;
int ThisRank(void) ;
const std::vector<int> & ThisProcessorCoor(void) ;
const std::vector<int> & ProcessorGrid(void) ;
const Coordinate & ThisProcessorCoor(void) ;
const Coordinate & ProcessorGrid(void) ;
int ProcessorCount(void) ;
////////////////////////////////////////////////////////////////////////////////
@ -197,11 +197,12 @@ public:
void AllToAll(void *in,void *out,uint64_t words ,uint64_t bytes);
template<class obj> void Broadcast(int root,obj &data)
{
Broadcast(root,(void *)&data,sizeof(data));
};
{
Broadcast(root,(void *)&data,sizeof(data));
}
};
}
NAMESPACE_END(Grid);
#endif

74
Grid/communicator/Communicator_mpi3.cc
View File

@ -23,12 +23,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <Grid/communicator/SharedMemory.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
Grid_MPI_Comm CartesianCommunicator::communicator_world;
@ -44,10 +44,15 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
MPI_Initialized(&flag); // needed to coexist with other libs apparently
if ( !flag ) {
MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
//If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
if( (nCommThreads == 1 && provided == MPI_THREAD_SINGLE) ||
(nCommThreads > 1 && provided != MPI_THREAD_MULTIPLE) )
if( (nCommThreads == 1) && (provided == MPI_THREAD_SINGLE) ) {
assert(0);
}
if( (nCommThreads > 1) && (provided != MPI_THREAD_MULTIPLE) ) {
assert(0);
}
}
// Never clean up as done once.
@ -69,14 +74,14 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
assert(ierr==0);
}
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor)
{
int rank;
int ierr=MPI_Cart_rank (communicator, &coor[0], &rank);
assert(ierr==0);
return rank;
}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor)
{
coor.resize(_ndimension);
int ierr=MPI_Cart_coords (communicator, rank, _ndimension,&coor[0]);
@ -86,7 +91,7 @@ void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &c
////////////////////////////////////////////////////////////////////////////////////////////////////////
// Initialises from communicator_world
////////////////////////////////////////////////////////////////////////////////////////////////////////
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
{
MPI_Comm optimal_comm;
////////////////////////////////////////////////////
@ -105,12 +110,12 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
//////////////////////////////////
// Try to subdivide communicator
//////////////////////////////////
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
{
_ndimension = processors.size(); assert(_ndimension>=1);
int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
std::vector<int> parent_processor_coor(_ndimension,0);
std::vector<int> parent_processors (_ndimension,1);
Coordinate parent_processor_coor(_ndimension,0);
Coordinate parent_processors (_ndimension,1);
// Can make 5d grid from 4d etc...
int pad = _ndimension-parent_ndimension;
@ -133,9 +138,9 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
int Nchild = Nparent/childsize;
assert (childsize * Nchild == Nparent);
std::vector<int> ccoor(_ndimension); // coor within subcommunicator
std::vector<int> scoor(_ndimension); // coor of split within parent
std::vector<int> ssize(_ndimension); // coor of split within parent
Coordinate ccoor(_ndimension); // coor within subcommunicator
Coordinate scoor(_ndimension); // coor of split within parent
Coordinate ssize(_ndimension); // coor of split within parent
for(int d=0;d<_ndimension;d++){
ccoor[d] = parent_processor_coor[d] % processors[d];
@ -152,36 +157,6 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
MPI_Comm comm_split;
if ( Nchild > 1 ) {
if(0){
std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"] ";
for(int d=0;d<parent._ndimension;d++) std::cout << parent._processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" child grid["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << processors[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< parent._ndimension <<"] ";
for(int d=0;d<parent._ndimension;d++) std::cout << parent._processor_coor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new split "<< srank<<" scoor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << scoor[d] << " ";
std::cout<<std::endl;
std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"] ";
for(int d=0;d<processors.size();d++) std::cout << ccoor[d] << " ";
std::cout<<std::endl;
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Declare victory
//////////////////////////////////////////////////////////////////////////////////////////////////////
std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
<< Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
std::cout << " Split communicator " <<comm_split <<std::endl;
}
////////////////////////////////////////////////////////////////
// Split the communicator
////////////////////////////////////////////////////////////////
@ -203,7 +178,7 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
// Take the right SHM buffers
//////////////////////////////////////////////////////////////////////////////////////////////////////
SetCommunicator(comm_split);
///////////////////////////////////////////////
// Free the temp communicator
///////////////////////////////////////////////
@ -220,7 +195,7 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
}
}
void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
void CartesianCommunicator::InitFromMPICommunicator(const Coordinate &processors, MPI_Comm communicator_base)
{
////////////////////////////////////////////////////
// Creates communicator, and the communicator_halo
@ -237,7 +212,7 @@ void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &proc
_Nprocessors*=_processors[i];
}
std::vector<int> periodic(_ndimension,1);
Coordinate periodic(_ndimension,1);
MPI_Cart_create(communicator_base, _ndimension,&_processors[0],&periodic[0],0,&communicator);
MPI_Comm_rank(communicator,&_processor);
MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
@ -474,7 +449,7 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
void CartesianCommunicator::AllToAll(int dim,void *in,void *out,uint64_t words,uint64_t bytes)
{
std::vector<int> row(_ndimension,1);
Coordinate row(_ndimension,1);
assert(dim>=0 && dim<_ndimension);
// Split the communicator
@ -503,7 +478,6 @@ void CartesianCommunicator::AllToAll(void *in,void *out,uint64_t words,uint64_t
MPI_Type_free(&object);
}
NAMESPACE_END(Grid);
}

22
Grid/communicator/Communicator_none.cc
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,11 +23,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////////////////////////////////////////////
// Info that is setup once and indept of cartesian layout
@ -38,18 +38,18 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
{
GlobalSharedMemory::Init(communicator_world);
GlobalSharedMemory::SharedMemoryAllocate(
GlobalSharedMemory::MAX_MPI_SHM_BYTES,
GlobalSharedMemory::Hugepages);
GlobalSharedMemory::MAX_MPI_SHM_BYTES,
GlobalSharedMemory::Hugepages);
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
: CartesianCommunicator(processors)
{
srank=0;
SetCommunicator(communicator_world);
}
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
{
_processors = processors;
_ndimension = processors.size(); assert(_ndimension>=1);
@ -122,8 +122,8 @@ int CartesianCommunicator::RankWorld(void){return 0;}
void CartesianCommunicator::Barrier(void){}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor) { return 0;}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor){ coor = _processor_coor; }
int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) { return 0;}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){ coor = _processor_coor; }
void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
{
source =0;
@ -160,6 +160,6 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
void CartesianCommunicator::StencilBarrier(void){};
NAMESPACE_END(Grid);
}

8
Grid/communicator/SharedMemory.cc
View File

@ -28,10 +28,11 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
// static data
int GlobalSharedMemory::HPEhypercube = 1;
uint64_t GlobalSharedMemory::MAX_MPI_SHM_BYTES = 1024LL*1024LL*1024LL;
int GlobalSharedMemory::Hugepages = 0;
int GlobalSharedMemory::_ShmSetup;
@ -76,6 +77,7 @@ void *SharedMemory::ShmBufferMalloc(size_t bytes){
std::cout<< " Current value is " << (heap_size/(1024*1024)) <<std::endl;
assert(heap_bytes<heap_size);
}
//std::cerr << "ShmBufferMalloc "<<std::hex<< ptr<<" - "<<((uint64_t)ptr+bytes)<<std::dec<<std::endl;
return ptr;
}
void SharedMemory::ShmBufferFreeAll(void) {
@ -84,9 +86,9 @@ void SharedMemory::ShmBufferFreeAll(void) {
}
void *SharedMemory::ShmBufferSelf(void)
{
//std::cerr << "ShmBufferSelf "<<ShmRank<<" "<<std::hex<< ShmCommBufs[ShmRank] <<std::dec<<std::endl;
return ShmCommBufs[ShmRank];
}
NAMESPACE_END(Grid);
}

50
Grid/communicator/SharedMemory.h
View File

@ -25,18 +25,6 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
// TODO
// 1) move includes into SharedMemory.cc
//
// 2) split shared memory into a) optimal communicator creation from comm world
//
// b) shared memory buffers container
// -- static globally shared; init once
// -- per instance set of buffers.
//
#pragma once
#include <Grid/GridCore.h>
@ -57,26 +45,32 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <numaif.h>
#endif
namespace Grid {
NAMESPACE_BEGIN(Grid);
#if defined (GRID_COMMS_MPI3)
typedef MPI_Comm Grid_MPI_Comm;
typedef MPI_Request CommsRequest_t;
typedef MPI_Comm Grid_MPI_Comm;
typedef MPI_Request CommsRequest_t;
#else
typedef int CommsRequest_t;
typedef int Grid_MPI_Comm;
typedef int CommsRequest_t;
typedef int Grid_MPI_Comm;
#endif
class GlobalSharedMemory {
private:
private:
static const int MAXLOG2RANKSPERNODE = 16;
// Init once lock on the buffer allocation
static int _ShmSetup;
static int _ShmAlloc;
static uint64_t _ShmAllocBytes;
public:
public:
///////////////////////////////////////
// HPE 8600 hypercube optimisation
///////////////////////////////////////
static int HPEhypercube;
static int ShmSetup(void) { return _ShmSetup; }
static int ShmAlloc(void) { return _ShmAlloc; }
static uint64_t ShmAllocBytes(void) { return _ShmAllocBytes; }
@ -102,14 +96,16 @@ class GlobalSharedMemory {
// Create an optimal reordered communicator that makes MPI_Cart_create get it right
//////////////////////////////////////////////////////////////////////////////////////
static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
static void OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicatorHypercube(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicatorSharedMemory(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicator (const Coordinate &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicatorHypercube (const Coordinate &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm); // Turns MPI_COMM_WORLD into right layout for Cartesian
///////////////////////////////////////////////////
// Provide shared memory facilities off comm world
///////////////////////////////////////////////////
static void SharedMemoryAllocate(uint64_t bytes, int flags);
static void SharedMemoryFree(void);
static void SharedMemoryCopy(void *dest,const void *src,size_t bytes);
static void SharedMemoryZero(void *dest,size_t bytes);
};
@ -118,14 +114,14 @@ class GlobalSharedMemory {
//////////////////////////////
class SharedMemory
{
private:
private:
static const int MAXLOG2RANKSPERNODE = 16;
size_t heap_top;
size_t heap_bytes;
size_t heap_size;
protected:
protected:
Grid_MPI_Comm ShmComm; // for barriers
int ShmRank;
@ -133,7 +129,7 @@ class SharedMemory
std::vector<void *> ShmCommBufs;
std::vector<int> ShmRanks;// Mapping comm ranks to Shm ranks
public:
public:
SharedMemory() {};
~SharedMemory();
///////////////////////////////////////////////////////////////////////////////////////
@ -150,6 +146,7 @@ class SharedMemory
// Call on any instance
///////////////////////////////////////////////////
void SharedMemoryTest(void);
void *ShmBufferSelf(void);
void *ShmBuffer (int rank);
void *ShmBufferTranslate(int rank,void * local_p);
@ -164,4 +161,5 @@ class SharedMemory
};
}
NAMESPACE_END(Grid);

202
Grid/communicator/SharedMemoryMPI.cc
View File

@ -29,8 +29,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
#include <pwd.h>
namespace Grid {
#ifdef GRID_NVCC
#include <cuda_runtime_api.h>
#endif
NAMESPACE_BEGIN(Grid);
#define header "SharedMemoryMpi: "
/*Construct from an MPI communicator*/
void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
{
@ -46,6 +50,11 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&WorldShmComm);
MPI_Comm_rank(WorldShmComm ,&WorldShmRank);
MPI_Comm_size(WorldShmComm ,&WorldShmSize);
if ( WorldRank == 0) {
std::cout << header " World communicator of size " <<WorldSize << std::endl;
std::cout << header " Node communicator of size " <<WorldShmSize << std::endl;
}
// WorldShmComm, WorldShmSize, WorldShmRank
// WorldNodes
@ -130,7 +139,7 @@ int Log2Size(int TwoToPower,int MAXLOG2)
}
return log2size;
}
void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
{
//////////////////////////////////////////////////////////////////////////////
// Look and see if it looks like an HPE 8600 based on hostname conventions
@ -143,11 +152,10 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
gethostname(name,namelen);
int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
// if(nscan==3) OptimalCommunicatorHypercube(processors,optimal_comm);
// else OptimalCommunicatorSharedMemory(processors,optimal_comm);
OptimalCommunicatorSharedMemory(processors,optimal_comm);
if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
else OptimalCommunicatorSharedMemory(processors,optimal_comm);
}
void GlobalSharedMemory::OptimalCommunicatorHypercube(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
{
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
@ -189,9 +197,9 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const std::vector<int> &pr
}
std::string hname(name);
std::cout << "hostname "<<hname<<std::endl;
std::cout << "R " << R << " I " << I << " N "<< N
<< " hypercoor 0x"<<std::hex<<hypercoor<<std::dec<<std::endl;
// std::cout << "hostname "<<hname<<std::endl;
// std::cout << "R " << R << " I " << I << " N "<< N
// << " hypercoor 0x"<<std::hex<<hypercoor<<std::dec<<std::endl;
//////////////////////////////////////////////////////////////////
// broadcast node 0's base coordinate for this partition.
@ -214,7 +222,8 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const std::vector<int> &pr
////////////////////////////////////////////////////////////////
int ndimension = processors.size();
std::vector<int> processor_coor(ndimension);
std::vector<int> WorldDims = processors; std::vector<int> ShmDims (ndimension,1); std::vector<int> NodeDims (ndimension);
std::vector<int> WorldDims = processors.toVector();
std::vector<int> ShmDims (ndimension,1); std::vector<int> NodeDims (ndimension);
std::vector<int> ShmCoor (ndimension); std::vector<int> NodeCoor (ndimension); std::vector<int> WorldCoor(ndimension);
std::vector<int> HyperCoor(ndimension);
int dim = 0;
@ -222,7 +231,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const std::vector<int> &pr
while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
ShmDims[dim]*=2;
dim=(dim+1)%ndimension;
}
}
////////////////////////////////////////////////////////////////
// Establish torus of processes and nodes with sub-blockings
@ -241,7 +250,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const std::vector<int> &pr
HyperCoor[d]=hcoor & msk;
HyperCoor[d]=BinaryToGray(HyperCoor[d]); // Space filling curve magic
hcoor = hcoor >> bits;
}
}
////////////////////////////////////////////////////////////////
// Check processor counts match
////////////////////////////////////////////////////////////////
@ -270,7 +279,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const std::vector<int> &pr
int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
assert(ierr==0);
}
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
{
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
@ -283,9 +292,9 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const std::vector<int>
// in a maximally symmetrical way
////////////////////////////////////////////////////////////////
int ndimension = processors.size();
std::vector<int> processor_coor(ndimension);
std::vector<int> WorldDims = processors; std::vector<int> ShmDims (ndimension,1); std::vector<int> NodeDims (ndimension);
std::vector<int> ShmCoor (ndimension); std::vector<int> NodeCoor (ndimension); std::vector<int> WorldCoor(ndimension);
Coordinate processor_coor(ndimension);
Coordinate WorldDims = processors; Coordinate ShmDims(ndimension,1); Coordinate NodeDims (ndimension);
Coordinate ShmCoor(ndimension); Coordinate NodeCoor(ndimension); Coordinate WorldCoor(ndimension);
int dim = 0;
for(int l2=0;l2<log2size;l2++){
while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
@ -331,7 +340,7 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const std::vector<int>
#ifdef GRID_MPI3_SHMGET
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
std::cout << header "SharedMemoryAllocate "<< bytes<< " shmget implementation "<<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
@ -386,14 +395,101 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
_ShmAllocBytes = bytes;
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbfs mapping intended
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_NVCC
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
void * ShmCommBuf ;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// allocate the pointer array for shared windows for our group
//////////////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Barrier(WorldShmComm);
WorldShmCommBufs.resize(WorldShmSize);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// TODO/FIXME : NOT ALL NVLINK BOARDS have full Peer to peer connectivity.
// The annoyance is that they have partial peer 2 peer. This occurs on the 8 GPU blades.
// e.g. DGX1, supermicro board,
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
cudaSetDevice(WorldShmRank);
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
auto err = cudaMalloc(&ShmCommBuf, bytes);
if ( err != cudaSuccess) {
std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed for " << bytes<<" bytes " <<cudaGetErrorString(err)<< std::endl;
exit(EXIT_FAILURE);
}
if (ShmCommBuf == (void *)NULL ) {
std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
exit(EXIT_FAILURE);
}
if ( WorldRank == 0 ){
std::cout << header " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
}
SharedMemoryZero(ShmCommBuf,bytes);
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Loop over ranks/gpu's on our node
///////////////////////////////////////////////////////////////////////////////////////////////////////////
for(int r=0;r<WorldShmSize;r++){
//////////////////////////////////////////////////
// If it is me, pass around the IPC access key
//////////////////////////////////////////////////
cudaIpcMemHandle_t handle;
if ( r==WorldShmRank ) {
err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
if ( err != cudaSuccess) {
std::cerr << " SharedMemoryMPI.cc cudaIpcGetMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
exit(EXIT_FAILURE);
}
}
//////////////////////////////////////////////////
// Share this IPC handle across the Shm Comm
//////////////////////////////////////////////////
{
int ierr=MPI_Bcast(&handle,
sizeof(handle),
MPI_BYTE,
r,
WorldShmComm);
assert(ierr==0);
}
///////////////////////////////////////////////////////////////
// If I am not the source, overwrite thisBuf with remote buffer
///////////////////////////////////////////////////////////////
void * thisBuf = ShmCommBuf;
if ( r!=WorldShmRank ) {
err = cudaIpcOpenMemHandle(&thisBuf,handle,cudaIpcMemLazyEnablePeerAccess);
if ( err != cudaSuccess) {
std::cerr << " SharedMemoryMPI.cc cudaIpcOpenMemHandle failed for rank" << r <<" "<<cudaGetErrorString(err)<< std::endl;
exit(EXIT_FAILURE);
}
}
///////////////////////////////////////////////////////////////
// Save a copy of the device buffers
///////////////////////////////////////////////////////////////
WorldShmCommBufs[r] = thisBuf;
}
_ShmAllocBytes=bytes;
_ShmAlloc=1;
}
#else
#ifdef GRID_MPI3_SHMMMAP
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<< GRID_SHM_PATH <<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -430,7 +526,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
assert(((uint64_t)ptr&0x3F)==0);
close(fd);
WorldShmCommBufs[r] =ptr;
// std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
// std::cout << header "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
}
_ShmAlloc=1;
_ShmAllocBytes = bytes;
@ -440,7 +536,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
#ifdef GRID_MPI3_SHM_NONE
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
std::cout << header "SharedMemoryAllocate "<< bytes<< " MMAP anonymous implementation "<<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -487,7 +583,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
////////////////////////////////////////////////////////////////////////////////////////////
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
std::cout << header "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
MPI_Barrier(WorldShmComm);
@ -553,14 +649,31 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
_ShmAllocBytes = bytes;
}
#endif
#endif // End NVCC case for GPU device buffers
////////////////////////////////////////////////////////
// Global shared functionality finished
// Now move to per communicator functionality
////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////////////////
// Routines accessing shared memory should route through for GPU safety
/////////////////////////////////////////////////////////////////////////
void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
{
#ifdef GRID_NVCC
cudaMemset(dest,0,bytes);
#else
bzero(dest,bytes);
#endif
}
void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
{
#ifdef GRID_NVCC
cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
#else
bcopy(src,dest,bytes);
#endif
}
////////////////////////////////////////////////////////
// Global shared functionality finished
// Now move to per communicator functionality
////////////////////////////////////////////////////////
void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
{
int rank, size;
@ -588,7 +701,6 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
MPI_Allreduce(MPI_IN_PLACE,&wsr,1,MPI_UINT32_T,MPI_SUM,ShmComm);
ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[wsr];
// std::cout << "SetCommunicator ShmCommBufs ["<< r<< "] = "<< ShmCommBufs[r]<< " wsr = "<<wsr<<std::endl;
}
ShmBufferFreeAll();
@ -601,6 +713,8 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
std::vector<int> ranks(size); for(int r=0;r<size;r++) ranks[r]=r;
MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]);
SharedMemoryTest();
}
//////////////////////////////////////////////////////////////////
// On node barrier
@ -615,24 +729,26 @@ void SharedMemory::ShmBarrier(void)
void SharedMemory::SharedMemoryTest(void)
{
ShmBarrier();
uint64_t check[3];
uint64_t magic = 0x5A5A5A;
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
uint64_t * check = (uint64_t *) ShmCommBufs[r];
check[0] = GlobalSharedMemory::WorldNode;
check[1] = r;
check[2] = 0x5A5A5A;
for(uint64_t r=0;r<ShmSize;r++){
check[0]=GlobalSharedMemory::WorldNode;
check[1]=r;
check[2]=magic;
GlobalSharedMemory::SharedMemoryCopy( ShmCommBufs[r], check, 3*sizeof(uint64_t));
}
}
ShmBarrier();
for(int r=0;r<ShmSize;r++){
uint64_t * check = (uint64_t *) ShmCommBufs[r];
for(uint64_t r=0;r<ShmSize;r++){
ShmBarrier();
GlobalSharedMemory::SharedMemoryCopy(check,ShmCommBufs[r], 3*sizeof(uint64_t));
ShmBarrier();
assert(check[0]==GlobalSharedMemory::WorldNode);
assert(check[1]==r);
assert(check[2]==0x5A5A5A);
assert(check[2]==magic);
ShmBarrier();
}
ShmBarrier();
}
void *SharedMemory::ShmBuffer(int rank)
@ -646,7 +762,6 @@ void *SharedMemory::ShmBuffer(int rank)
}
void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
{
static int count =0;
int gpeer = ShmRanks[rank];
assert(gpeer!=ShmRank); // never send to self
if (gpeer == MPI_UNDEFINED){
@ -665,4 +780,5 @@ SharedMemory::~SharedMemory()
}
};
}
NAMESPACE_END(Grid);

15
Grid/communicator/SharedMemoryNone.cc
View File

@ -28,7 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
/*Construct from an MPI communicator*/
void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
@ -47,7 +47,7 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
_ShmSetup=1;
}
void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
{
optimal_comm = WorldComm;
}
@ -84,10 +84,10 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
_ShmAlloc=1;
};
////////////////////////////////////////////////////////
// Global shared functionality finished
// Now move to per communicator functionality
////////////////////////////////////////////////////////
////////////////////////////////////////////////////////
// Global shared functionality finished
// Now move to per communicator functionality
////////////////////////////////////////////////////////
void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
{
assert(GlobalSharedMemory::ShmAlloc()==1);
@ -125,4 +125,5 @@ void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
SharedMemory::~SharedMemory()
{};
}
NAMESPACE_END(Grid);

6
Grid/cshift/Cshift.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,8 +23,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef _GRID_CSHIFT_H_
#define _GRID_CSHIFT_H_

188
Grid/cshift/Cshift_common.h
View File

@ -25,10 +25,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef _GRID_CSHIFT_COMMON_H_
#define _GRID_CSHIFT_COMMON_H_
#pragma once
namespace Grid {
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////////////
// Gather for when there is no need to SIMD split
@ -36,20 +35,21 @@ namespace Grid {
template<class vobj> void
Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimension,int plane,int cbmask, int off=0)
{
int rd = rhs._grid->_rdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
if ( !rhs._grid->CheckerBoarded(dimension) ) {
if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
cbmask = 0x3;
}
int so=plane*rhs._grid->_ostride[dimension]; // base offset for start of plane
int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block[dimension];
int so=plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
int e1=rhs.Grid()->_slice_nblock[dimension];
int e2=rhs.Grid()->_slice_block[dimension];
int ent = 0;
static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
static Vector<std::pair<int,int> > table; table.resize(e1*e2);
int stride=rhs.Grid()->_slice_stride[dimension];
int stride=rhs._grid->_slice_stride[dimension];
auto rhs_v = rhs.View();
if ( cbmask == 0x3 ) {
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
@ -63,66 +63,68 @@ Gather_plane_simple (const Lattice<vobj> &rhs,commVector<vobj> &buffer,int dimen
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o = n*stride;
int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
if ( ocb &cbmask ) {
table[ent++]=std::pair<int,int> (off+bo++,so+o+b);
}
}
}
}
parallel_for(int i=0;i<ent;i++){
buffer[table[i].first]=rhs._odata[table[i].second];
}
thread_for(i,ent,{
buffer[table[i].first]=rhs_v[table[i].second];
});
}
///////////////////////////////////////////////////////////////////
// Gather for when there *is* need to SIMD split
///////////////////////////////////////////////////////////////////
template<class vobj> void
Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
Gather_plane_extract(const Lattice<vobj> &rhs,
ExtractPointerArray<typename vobj::scalar_object> pointers,
int dimension,int plane,int cbmask)
{
int rd = rhs._grid->_rdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
if ( !rhs._grid->CheckerBoarded(dimension) ) {
if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
cbmask = 0x3;
}
int so = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane
int so = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block[dimension];
int n1=rhs._grid->_slice_stride[dimension];
int e1=rhs.Grid()->_slice_nblock[dimension];
int e2=rhs.Grid()->_slice_block[dimension];
int n1=rhs.Grid()->_slice_stride[dimension];
auto rhs_v = rhs.View();
if ( cbmask ==0x3){
parallel_for_nest2(int n=0;n<e1;n++){
thread_for_collapse(2,n,e1,{
for(int b=0;b<e2;b++){
int o = n*n1;
int offset = b+n*e2;
vobj temp =rhs._odata[so+o+b];
vobj temp =rhs_v[so+o+b];
extract<vobj>(temp,pointers,offset);
}
}
});
} else {
// Case of SIMD split AND checker dim cannot currently be hit, except in
// Test_cshift_red_black code.
std::cout << " Dense packed buffer WARNING " <<std::endl;
parallel_for_nest2(int n=0;n<e1;n++){
thread_for_collapse(2,n,e1,{
for(int b=0;b<e2;b++){
int o=n*n1;
int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
int offset = b+n*e2;
if ( ocb & cbmask ) {
vobj temp =rhs._odata[so+o+b];
vobj temp =rhs_v[so+o+b];
extract<vobj>(temp,pointers,offset);
}
}
}
});
}
}
@ -131,17 +133,17 @@ Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename vobj::scalar_
//////////////////////////////////////////////////////
template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vobj> &buffer, int dimension,int plane,int cbmask)
{
int rd = rhs._grid->_rdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
if ( !rhs._grid->CheckerBoarded(dimension) ) {
if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
cbmask=0x3;
}
int so = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane
int so = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block[dimension];
int stride=rhs._grid->_slice_stride[dimension];
int e1=rhs.Grid()->_slice_nblock[dimension];
int e2=rhs.Grid()->_slice_block[dimension];
int stride=rhs.Grid()->_slice_stride[dimension];
static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
int ent =0;
@ -150,8 +152,8 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*rhs._grid->_slice_stride[dimension];
int bo =n*rhs._grid->_slice_block[dimension];
int o =n*rhs.Grid()->_slice_stride[dimension];
int bo =n*rhs.Grid()->_slice_block[dimension];
table[ent++] = std::pair<int,int>(so+o+b,bo+b);
}
}
@ -160,57 +162,60 @@ template<class vobj> void Scatter_plane_simple (Lattice<vobj> &rhs,commVector<vo
int bo=0;
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*rhs._grid->_slice_stride[dimension];
int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
int o =n*rhs.Grid()->_slice_stride[dimension];
int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
if ( ocb & cbmask ) {
table[ent++]=std::pair<int,int> (so+o+b,bo++);
}
}
}
}
parallel_for(int i=0;i<ent;i++){
rhs._odata[table[i].first]=buffer[table[i].second];
}
auto rhs_v = rhs.View();
thread_for(i,ent,{
rhs_v[table[i].first]=buffer[table[i].second];
});
}
//////////////////////////////////////////////////////
// Scatter for when there *is* need to SIMD split
//////////////////////////////////////////////////////
template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,std::vector<typename vobj::scalar_object *> pointers,int dimension,int plane,int cbmask)
template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerArray<typename vobj::scalar_object> pointers,int dimension,int plane,int cbmask)
{
int rd = rhs._grid->_rdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
if ( !rhs._grid->CheckerBoarded(dimension) ) {
if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
cbmask=0x3;
}
int so = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane
int so = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block[dimension];
int e1=rhs.Grid()->_slice_nblock[dimension];
int e2=rhs.Grid()->_slice_block[dimension];
if(cbmask ==0x3 ) {
parallel_for_nest2(int n=0;n<e1;n++){
auto rhs_v = rhs.View();
thread_for_collapse(2,n,e1,{
for(int b=0;b<e2;b++){
int o = n*rhs._grid->_slice_stride[dimension];
int offset = b+n*rhs._grid->_slice_block[dimension];
merge(rhs._odata[so+o+b],pointers,offset);
int o = n*rhs.Grid()->_slice_stride[dimension];
int offset = b+n*rhs.Grid()->_slice_block[dimension];
merge(rhs_v[so+o+b],pointers,offset);
}
}
});
} else {
// Case of SIMD split AND checker dim cannot currently be hit, except in
// Test_cshift_red_black code.
// std::cout << "Scatter_plane merge assert(0); think this is buggy FIXME "<< std::endl;// think this is buggy FIXME
std::cout<<" Unthreaded warning -- buffer is not densely packed ??"<<std::endl;
auto rhs_v = rhs.View();
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o = n*rhs._grid->_slice_stride[dimension];
int offset = b+n*rhs._grid->_slice_block[dimension];
int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);
int o = n*rhs.Grid()->_slice_stride[dimension];
int offset = b+n*rhs.Grid()->_slice_block[dimension];
int ocb=1<<rhs.Grid()->CheckerBoardFromOindex(o+b);
if ( ocb&cbmask ) {
merge(rhs._odata[so+o+b],pointers,offset);
merge(rhs_v[so+o+b],pointers,offset);
}
}
}
@ -222,18 +227,18 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,std::vector<typ
//////////////////////////////////////////////////////
template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask)
{
int rd = rhs._grid->_rdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
if ( !rhs._grid->CheckerBoarded(dimension) ) {
if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
cbmask=0x3;
}
int ro = rplane*rhs._grid->_ostride[dimension]; // base offset for start of plane
int lo = lplane*lhs._grid->_ostride[dimension]; // base offset for start of plane
int ro = rplane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
int lo = lplane*lhs.Grid()->_ostride[dimension]; // base offset for start of plane
int e1=rhs._grid->_slice_nblock[dimension]; // clearly loop invariant for icpc
int e2=rhs._grid->_slice_block[dimension];
int stride = rhs._grid->_slice_stride[dimension];
int e1=rhs.Grid()->_slice_nblock[dimension]; // clearly loop invariant for icpc
int e2=rhs.Grid()->_slice_block[dimension];
int stride = rhs.Grid()->_slice_stride[dimension];
static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
int ent=0;
@ -248,7 +253,7 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*stride+b;
int ocb=1<<lhs._grid->CheckerBoardFromOindex(o);
int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o);
if ( ocb&cbmask ) {
table[ent++] = std::pair<int,int>(lo+o,ro+o);
}
@ -256,32 +261,33 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
}
}
parallel_for(int i=0;i<ent;i++){
lhs._odata[table[i].first]=rhs._odata[table[i].second];
}
auto rhs_v = rhs.View();
auto lhs_v = lhs.View();
thread_for(i,ent,{
lhs_v[table[i].first]=rhs_v[table[i].second];
});
}
template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vobj> &rhs, int dimension,int lplane,int rplane,int cbmask,int permute_type)
{
int rd = rhs._grid->_rdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
if ( !rhs._grid->CheckerBoarded(dimension) ) {
if ( !rhs.Grid()->CheckerBoarded(dimension) ) {
cbmask=0x3;
}
int ro = rplane*rhs._grid->_ostride[dimension]; // base offset for start of plane
int lo = lplane*lhs._grid->_ostride[dimension]; // base offset for start of plane
int ro = rplane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
int lo = lplane*lhs.Grid()->_ostride[dimension]; // base offset for start of plane
int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block [dimension];
int stride = rhs._grid->_slice_stride[dimension];
int e1=rhs.Grid()->_slice_nblock[dimension];
int e2=rhs.Grid()->_slice_block [dimension];
int stride = rhs.Grid()->_slice_stride[dimension];
static std::vector<std::pair<int,int> > table; table.resize(e1*e2);
int ent=0;
double t_tab,t_perm;
if ( cbmask == 0x3 ) {
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
@ -292,14 +298,16 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o =n*stride;
int ocb=1<<lhs._grid->CheckerBoardFromOindex(o+b);
int ocb=1<<lhs.Grid()->CheckerBoardFromOindex(o+b);
if ( ocb&cbmask ) table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
}}
}
parallel_for(int i=0;i<ent;i++){
permute(lhs._odata[table[i].first],rhs._odata[table[i].second],permute_type);
}
auto rhs_v = rhs.View();
auto lhs_v = lhs.View();
thread_for(i,ent,{
permute(lhs_v[table[i].first],rhs_v[table[i].second],permute_type);
});
}
//////////////////////////////////////////////////////
@ -309,11 +317,9 @@ template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &r
{
int sshift[2];
sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
double t_local;
if ( sshift[0] == sshift[1] ) {
Cshift_local(ret,rhs,dimension,shift,0x3);
} else {
@ -324,7 +330,7 @@ template<class vobj> void Cshift_local(Lattice<vobj>& ret,const Lattice<vobj> &r
template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
{
GridBase *grid = rhs._grid;
GridBase *grid = rhs.Grid();
int fd = grid->_fdimensions[dimension];
int rd = grid->_rdimensions[dimension];
int ld = grid->_ldimensions[dimension];
@ -335,18 +341,18 @@ template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &r
shift = (shift+fd)%fd;
// the permute type
ret.checkerboard = grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
ret.Checkerboard() = grid->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
int permute_dim =grid->PermuteDim(dimension);
int permute_type=grid->PermuteType(dimension);
int permute_type_dist;
for(int x=0;x<rd;x++){
int o = 0;
// int o = 0;
int bo = x * grid->_ostride[dimension];
int cb= (cbmask==0x2)? Odd : Even;
int sshift = grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
int sshift = grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
int sx = (x+sshift)%rd;
// wrap is whether sshift > rd.
@ -387,5 +393,5 @@ template<class vobj> void Cshift_local(Lattice<vobj> &ret,const Lattice<vobj> &r
}
}
}
#endif
NAMESPACE_END(Grid);

66
Grid/cshift/Cshift_mpi.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,33 +24,33 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef _GRID_CSHIFT_MPI_H_
#define _GRID_CSHIFT_MPI_H_
namespace Grid {
NAMESPACE_BEGIN(Grid);
template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
{
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
Lattice<vobj> ret(rhs._grid);
Lattice<vobj> ret(rhs.Grid());
int fd = rhs._grid->_fdimensions[dimension];
int rd = rhs._grid->_rdimensions[dimension];
int fd = rhs.Grid()->_fdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
// Map to always positive shift modulo global full dimension.
shift = (shift+fd)%fd;
ret.checkerboard = rhs._grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
// the permute type
int simd_layout = rhs._grid->_simd_layout[dimension];
int comm_dim = rhs._grid->_processors[dimension] >1 ;
int splice_dim = rhs._grid->_simd_layout[dimension]>1 && (comm_dim);
int simd_layout = rhs.Grid()->_simd_layout[dimension];
int comm_dim = rhs.Grid()->_processors[dimension] >1 ;
int splice_dim = rhs.Grid()->_simd_layout[dimension]>1 && (comm_dim);
if ( !comm_dim ) {
@ -70,10 +70,10 @@ template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &r
{
int sshift[2];
sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
// std::cout << "Cshift_comms dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
// std::cout << "Cshift_comms dim "<<dimension<<"cb "<<rhs.Checkerboard()<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
if ( sshift[0] == sshift[1] ) {
// std::cout << "Single pass Cshift_comms" <<std::endl;
Cshift_comms(ret,rhs,dimension,shift,0x3);
@ -88,8 +88,8 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj>& ret,const Lattice<vob
{
int sshift[2];
sshift[0] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Even);
sshift[1] = rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,Odd);
sshift[0] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Even);
sshift[1] = rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,Odd);
//std::cout << "Cshift_comms_simd dim "<<dimension<<"cb "<<rhs.checkerboard<<"shift "<<shift<<" sshift " << sshift[0]<<" "<<sshift[1]<<std::endl;
if ( sshift[0] == sshift[1] ) {
@ -107,25 +107,25 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
GridBase *grid=rhs._grid;
Lattice<vobj> temp(rhs._grid);
GridBase *grid=rhs.Grid();
Lattice<vobj> temp(rhs.Grid());
int fd = rhs._grid->_fdimensions[dimension];
int rd = rhs._grid->_rdimensions[dimension];
int pd = rhs._grid->_processors[dimension];
int simd_layout = rhs._grid->_simd_layout[dimension];
int comm_dim = rhs._grid->_processors[dimension] >1 ;
int fd = rhs.Grid()->_fdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
int pd = rhs.Grid()->_processors[dimension];
int simd_layout = rhs.Grid()->_simd_layout[dimension];
int comm_dim = rhs.Grid()->_processors[dimension] >1 ;
assert(simd_layout==1);
assert(comm_dim==1);
assert(shift>=0);
assert(shift<fd);
int buffer_size = rhs._grid->_slice_nblock[dimension]*rhs._grid->_slice_block[dimension];
int buffer_size = rhs.Grid()->_slice_nblock[dimension]*rhs.Grid()->_slice_block[dimension];
commVector<vobj> send_buf(buffer_size);
commVector<vobj> recv_buf(buffer_size);
int cb= (cbmask==0x2)? Odd : Even;
int sshift= rhs._grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
int sshift= rhs.Grid()->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
for(int x=0;x<rd;x++){
@ -145,7 +145,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
int rank = grid->_processor;
// int rank = grid->_processor;
int recv_from_rank;
int xmit_to_rank;
grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
@ -165,7 +165,7 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vobj> &rhs,int dimension,int shift,int cbmask)
{
GridBase *grid=rhs._grid;
GridBase *grid=rhs.Grid();
const int Nsimd = grid->Nsimd();
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_object scalar_object;
@ -193,21 +193,21 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
// Simd direction uses an extract/merge pair
///////////////////////////////////////////////
int buffer_size = grid->_slice_nblock[dimension]*grid->_slice_block[dimension];
int words = sizeof(vobj)/sizeof(vector_type);
// int words = sizeof(vobj)/sizeof(vector_type);
std::vector<commVector<scalar_object> > send_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
std::vector<commVector<scalar_object> > recv_buf_extract(Nsimd,commVector<scalar_object>(buffer_size) );
int bytes = buffer_size*sizeof(scalar_object);
std::vector<scalar_object *> pointers(Nsimd); //
std::vector<scalar_object *> rpointers(Nsimd); // received pointers
ExtractPointerArray<scalar_object> pointers(Nsimd); //
ExtractPointerArray<scalar_object> rpointers(Nsimd); // received pointers
///////////////////////////////////////////
// Work out what to send where
///////////////////////////////////////////
int cb = (cbmask==0x2)? Odd : Even;
int sshift= grid->CheckerBoardShiftForCB(rhs.checkerboard,dimension,shift,cb);
int sshift= grid->CheckerBoardShiftForCB(rhs.Checkerboard(),dimension,shift,cb);
// loop over outer coord planes orthog to dim
for(int x=0;x<rd;x++){
@ -257,6 +257,8 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
Scatter_plane_merge(ret,rpointers,dimension,x,cbmask);
}
}
}
NAMESPACE_END(Grid);
#endif

15
Grid/cshift/Cshift_none.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,17 +23,18 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef _GRID_CSHIFT_NONE_H_
#define _GRID_CSHIFT_NONE_H_
namespace Grid {
NAMESPACE_BEGIN(Grid);
template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension,int shift)
{
Lattice<vobj> ret(rhs._grid);
ret.checkerboard = rhs._grid->CheckerBoardDestination(rhs.checkerboard,shift,dimension);
Lattice<vobj> ret(rhs.Grid());
ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
Cshift_local(ret,rhs,dimension,shift);
return ret;
}
}
NAMESPACE_END(Grid);
#endif

2
Grid/json/json.hpp
View File

@ -1,3 +1,4 @@
#ifndef __NVCC__
/*
__ _____ _____ _____
__| | __| | | | JSON for Modern C++
@ -18918,3 +18919,4 @@ inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std
#endif
#endif

21
Grid/lattice/Lattice.h
View File

@ -25,9 +25,22 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_H
#define GRID_LATTICE_H
#pragma once
#include <Grid/lattice/Lattice_base.h>
#include <Grid/lattice/Lattice_conformable.h>
#include <Grid/lattice/Lattice_ET.h>
#include <Grid/lattice/Lattice_arith.h>
#include <Grid/lattice/Lattice_trace.h>
#include <Grid/lattice/Lattice_transpose.h>
#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_comparison_utils.h>
#include <Grid/lattice/Lattice_comparison.h>
#include <Grid/lattice/Lattice_coordinate.h>
//#include <Grid/lattice/Lattice_where.h>
#include <Grid/lattice/Lattice_rng.h>
#include <Grid/lattice/Lattice_unary.h>
#include <Grid/lattice/Lattice_transfer.h>
#endif

401
Grid/lattice/Lattice_ET.h
View File

@ -27,7 +27,7 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
/* END LEGAL */
#ifndef GRID_LATTICE_ET_H
#define GRID_LATTICE_ET_H
@ -36,13 +36,13 @@ directory
#include <typeinfo>
#include <vector>
namespace Grid {
NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////
// Predicated where support
////////////////////////////////////////////////////
template <class iobj, class vobj, class robj>
inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
accelerator_inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
const robj &iffalse) {
typename std::remove_const<vobj>::type ret;
@ -51,11 +51,10 @@ inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
typedef typename vobj::vector_type vector_type;
const int Nsimd = vobj::vector_type::Nsimd();
const int words = sizeof(vobj) / sizeof(vector_type);
std::vector<Integer> mask(Nsimd);
std::vector<scalar_object> truevals(Nsimd);
std::vector<scalar_object> falsevals(Nsimd);
ExtractBuffer<Integer> mask(Nsimd);
ExtractBuffer<scalar_object> truevals(Nsimd);
ExtractBuffer<scalar_object> falsevals(Nsimd);
extract(iftrue, truevals);
extract(iffalse, falsevals);
@ -69,158 +68,148 @@ inline vobj predicatedWhere(const iobj &predicate, const vobj &iftrue,
return ret;
}
////////////////////////////////////////////
// recursive evaluation of expressions; Could
// switch to generic approach with variadics, a la
// Antonin's Lat Sim but the repack to variadic with popped
// from tuple is hideous; C++14 introduces std::make_index_sequence for this
////////////////////////////////////////////
// leaf eval of lattice ; should enable if protect using traits
template <typename T>
using is_lattice = std::is_base_of<LatticeBase, T>;
template <typename T>
using is_lattice_expr = std::is_base_of<LatticeExpressionBase, T>;
template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
/////////////////////////////////////////////////////
//Specialization of getVectorType for lattices
/////////////////////////////////////////////////////
template<typename T>
struct getVectorType<Lattice<T> >{
typedef typename Lattice<T>::vector_object type;
};
template<class sobj>
inline sobj eval(const unsigned int ss, const sobj &arg)
////////////////////////////////////////////
//-- recursive evaluation of expressions; --
// handle leaves of syntax tree
///////////////////////////////////////////////////
template<class sobj> accelerator_inline
sobj eval(const uint64_t ss, const sobj &arg)
{
return arg;
}
template <class lobj>
inline const lobj &eval(const unsigned int ss, const Lattice<lobj> &arg) {
return arg._odata[ss];
template <class lobj> accelerator_inline
const lobj & eval(const uint64_t ss, const LatticeView<lobj> &arg)
{
return arg[ss];
}
template <class lobj> accelerator_inline
const lobj & eval(const uint64_t ss, const Lattice<lobj> &arg)
{
auto view = arg.View();
return view[ss];
}
// handle nodes in syntax tree
template <typename Op, typename T1>
auto inline eval(
const unsigned int ss,
const LatticeUnaryExpression<Op, T1> &expr) // eval one operand
-> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)))) {
return expr.first.func(eval(ss, std::get<0>(expr.second)));
///////////////////////////////////////////////////
// handle nodes in syntax tree- eval one operand
///////////////////////////////////////////////////
template <typename Op, typename T1> accelerator_inline
auto eval(const uint64_t ss, const LatticeUnaryExpression<Op, T1> &expr)
-> decltype(expr.op.func( eval(ss, expr.arg1)))
{
return expr.op.func( eval(ss, expr.arg1) );
}
template <typename Op, typename T1, typename T2>
auto inline eval(
const unsigned int ss,
const LatticeBinaryExpression<Op, T1, T2> &expr) // eval two operands
-> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)),
eval(ss, std::get<1>(expr.second)))) {
return expr.first.func(eval(ss, std::get<0>(expr.second)),
eval(ss, std::get<1>(expr.second)));
///////////////////////
// eval two operands
///////////////////////
template <typename Op, typename T1, typename T2> accelerator_inline
auto eval(const uint64_t ss, const LatticeBinaryExpression<Op, T1, T2> &expr)
-> decltype(expr.op.func( eval(ss,expr.arg1),eval(ss,expr.arg2)))
{
return expr.op.func( eval(ss,expr.arg1), eval(ss,expr.arg2) );
}
template <typename Op, typename T1, typename T2, typename T3>
auto inline eval(const unsigned int ss,
const LatticeTrinaryExpression<Op, T1, T2, T3>
&expr) // eval three operands
-> decltype(expr.first.func(eval(ss, std::get<0>(expr.second)),
eval(ss, std::get<1>(expr.second)),
eval(ss, std::get<2>(expr.second)))) {
return expr.first.func(eval(ss, std::get<0>(expr.second)),
eval(ss, std::get<1>(expr.second)),
eval(ss, std::get<2>(expr.second)));
///////////////////////
// eval three operands
///////////////////////
template <typename Op, typename T1, typename T2, typename T3> accelerator_inline
auto eval(const uint64_t ss, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-> decltype(expr.op.func(eval(ss, expr.arg1), eval(ss, expr.arg2), eval(ss, expr.arg3)))
{
return expr.op.func(eval(ss, expr.arg1), eval(ss, expr.arg2), eval(ss, expr.arg3));
}
//////////////////////////////////////////////////////////////////////////
// Obtain the grid from an expression, ensuring conformable. This must follow a
// tree recursion
// tree recursion; must retain grid pointer in the LatticeView class which sucks
// Use a different method, and make it void *.
// Perhaps a conformable method.
//////////////////////////////////////////////////////////////////////////
template <class T1,
typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
inline void GridFromExpression(GridBase *&grid, const T1 &lat) // Lattice leaf
template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
accelerator_inline void GridFromExpression(GridBase *&grid, const T1 &lat) // Lattice leaf
{
if (grid) {
conformable(grid, lat._grid);
}
grid = lat._grid;
lat.Conformable(grid);
}
template <class T1,
typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
inline void GridFromExpression(GridBase *&grid,
const T1 &notlat) // non-lattice leaf
template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
accelerator_inline
void GridFromExpression(GridBase *&grid,const T1 &notlat) // non-lattice leaf
{}
template <typename Op, typename T1>
inline void GridFromExpression(GridBase *&grid,
const LatticeUnaryExpression<Op, T1> &expr) {
GridFromExpression(grid, std::get<0>(expr.second)); // recurse
accelerator_inline
void GridFromExpression(GridBase *&grid,const LatticeUnaryExpression<Op, T1> &expr)
{
GridFromExpression(grid, expr.arg1); // recurse
}
template <typename Op, typename T1, typename T2>
inline void GridFromExpression(
GridBase *&grid, const LatticeBinaryExpression<Op, T1, T2> &expr) {
GridFromExpression(grid, std::get<0>(expr.second)); // recurse
GridFromExpression(grid, std::get<1>(expr.second));
accelerator_inline
void GridFromExpression(GridBase *&grid, const LatticeBinaryExpression<Op, T1, T2> &expr)
{
GridFromExpression(grid, expr.arg1); // recurse
GridFromExpression(grid, expr.arg2);
}
template <typename Op, typename T1, typename T2, typename T3>
inline void GridFromExpression(
GridBase *&grid, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) {
GridFromExpression(grid, std::get<0>(expr.second)); // recurse
GridFromExpression(grid, std::get<1>(expr.second));
GridFromExpression(grid, std::get<2>(expr.second));
accelerator_inline
void GridFromExpression(GridBase *&grid, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
{
GridFromExpression(grid, expr.arg1); // recurse
GridFromExpression(grid, expr.arg2); // recurse
GridFromExpression(grid, expr.arg3); // recurse
}
//////////////////////////////////////////////////////////////////////////
// Obtain the CB from an expression, ensuring conformable. This must follow a
// tree recursion
//////////////////////////////////////////////////////////////////////////
template <class T1,
typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
template <class T1,typename std::enable_if<is_lattice<T1>::value, T1>::type * = nullptr>
inline void CBFromExpression(int &cb, const T1 &lat) // Lattice leaf
{
if ((cb == Odd) || (cb == Even)) {
assert(cb == lat.checkerboard);
assert(cb == lat.Checkerboard());
}
cb = lat.checkerboard;
// std::cout<<GridLogMessage<<"Lattice leaf cb "<<cb<<std::endl;
cb = lat.Checkerboard();
}
template <class T1,
typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
template <class T1,typename std::enable_if<!is_lattice<T1>::value, T1>::type * = nullptr>
inline void CBFromExpression(int &cb, const T1 &notlat) // non-lattice leaf
{
// std::cout<<GridLogMessage<<"Non lattice leaf cb"<<cb<<std::endl;
}
template <typename Op, typename T1>
inline void CBFromExpression(int &cb,
const LatticeUnaryExpression<Op, T1> &expr) {
CBFromExpression(cb, std::get<0>(expr.second)); // recurse
// std::cout<<GridLogMessage<<"Unary node cb "<<cb<<std::endl;
}
template <typename Op, typename T1, typename T2>
inline void CBFromExpression(int &cb,
const LatticeBinaryExpression<Op, T1, T2> &expr) {
CBFromExpression(cb, std::get<0>(expr.second)); // recurse
CBFromExpression(cb, std::get<1>(expr.second));
// std::cout<<GridLogMessage<<"Binary node cb "<<cb<<std::endl;
template <typename Op, typename T1> inline
void CBFromExpression(int &cb,const LatticeUnaryExpression<Op, T1> &expr)
{
CBFromExpression(cb, expr.arg1); // recurse AST
}
template <typename Op, typename T1, typename T2> inline
void CBFromExpression(int &cb,const LatticeBinaryExpression<Op, T1, T2> &expr)
{
CBFromExpression(cb, expr.arg1); // recurse AST
CBFromExpression(cb, expr.arg2); // recurse AST
}
template <typename Op, typename T1, typename T2, typename T3>
inline void CBFromExpression(
int &cb, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr) {
CBFromExpression(cb, std::get<0>(expr.second)); // recurse
CBFromExpression(cb, std::get<1>(expr.second));
CBFromExpression(cb, std::get<2>(expr.second));
// std::cout<<GridLogMessage<<"Trinary node cb "<<cb<<std::endl;
inline void CBFromExpression(int &cb, const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
{
CBFromExpression(cb, expr.arg1); // recurse AST
CBFromExpression(cb, expr.arg2); // recurse AST
CBFromExpression(cb, expr.arg3); // recurse AST
}
////////////////////////////////////////////
// Unary operators and funcs
////////////////////////////////////////////
#define GridUnopClass(name, ret) \
template <class arg> \
struct name { \
static auto inline func(const arg a) -> decltype(ret) { return ret; } \
#define GridUnopClass(name, ret) \
template <class arg> \
struct name { \
static auto accelerator_inline func(const arg a) -> decltype(ret) { return ret; } \
};
GridUnopClass(UnarySub, -a);
@ -250,19 +239,21 @@ GridUnopClass(UnaryExp, exp(a));
////////////////////////////////////////////
// Binary operators
////////////////////////////////////////////
#define GridBinOpClass(name, combination) \
template <class left, class right> \
struct name { \
static auto inline func(const left &lhs, const right &rhs) \
-> decltype(combination) const { \
return combination; \
} \
}
#define GridBinOpClass(name, combination) \
template <class left, class right> \
struct name { \
static auto accelerator_inline \
func(const left &lhs, const right &rhs) \
-> decltype(combination) const \
{ \
return combination; \
} \
};
GridBinOpClass(BinaryAdd, lhs + rhs);
GridBinOpClass(BinarySub, lhs - rhs);
GridBinOpClass(BinaryMul, lhs *rhs);
GridBinOpClass(BinaryDiv, lhs /rhs);
GridBinOpClass(BinaryAnd, lhs &rhs);
GridBinOpClass(BinaryOr, lhs | rhs);
GridBinOpClass(BinaryAndAnd, lhs &&rhs);
@ -271,92 +262,71 @@ GridBinOpClass(BinaryOrOr, lhs || rhs);
////////////////////////////////////////////////////
// Trinary conditional op
////////////////////////////////////////////////////
#define GridTrinOpClass(name, combination) \
template <class predicate, class left, class right> \
struct name { \
static auto inline func(const predicate &pred, const left &lhs, \
const right &rhs) -> decltype(combination) const { \
return combination; \
} \
}
#define GridTrinOpClass(name, combination) \
template <class predicate, class left, class right> \
struct name { \
static auto accelerator_inline \
func(const predicate &pred, const left &lhs, const right &rhs) \
-> decltype(combination) const \
{ \
return combination; \
} \
};
GridTrinOpClass(
TrinaryWhere,
(predicatedWhere<predicate, typename std::remove_reference<left>::type,
typename std::remove_reference<right>::type>(pred, lhs,
rhs)));
GridTrinOpClass(TrinaryWhere,
(predicatedWhere<predicate,
typename std::remove_reference<left>::type,
typename std::remove_reference<right>::type>(pred, lhs,rhs)));
////////////////////////////////////////////
// Operator syntactical glue
////////////////////////////////////////////
#define GRID_UNOP(name) name<decltype(eval(0, arg))>
#define GRID_BINOP(name) name<decltype(eval(0, lhs)), decltype(eval(0, rhs))>
#define GRID_TRINOP(name) \
name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
#define GRID_UNOP(name) name<decltype(eval(0, arg))>
#define GRID_BINOP(name) name<decltype(eval(0, lhs)), decltype(eval(0, rhs))>
#define GRID_TRINOP(name) name<decltype(eval(0, pred)), decltype(eval(0, lhs)), decltype(eval(0, rhs))>
#define GRID_DEF_UNOP(op, name) \
template <typename T1, \
typename std::enable_if<is_lattice<T1>::value || \
is_lattice_expr<T1>::value, \
T1>::type * = nullptr> \
inline auto op(const T1 &arg) \
->decltype(LatticeUnaryExpression<GRID_UNOP(name), const T1 &>( \
std::make_pair(GRID_UNOP(name)(), std::forward_as_tuple(arg)))) { \
return LatticeUnaryExpression<GRID_UNOP(name), const T1 &>( \
std::make_pair(GRID_UNOP(name)(), std::forward_as_tuple(arg))); \
#define GRID_DEF_UNOP(op, name) \
template <typename T1, typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
inline auto op(const T1 &arg) ->decltype(LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg)) \
{ \
return LatticeUnaryExpression<GRID_UNOP(name),T1>(GRID_UNOP(name)(), arg); \
}
#define GRID_BINOP_LEFT(op, name) \
template <typename T1, typename T2, \
typename std::enable_if<is_lattice<T1>::value || \
is_lattice_expr<T1>::value, \
T1>::type * = nullptr> \
inline auto op(const T1 &lhs, const T2 &rhs) \
->decltype( \
LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
std::make_pair(GRID_BINOP(name)(), \
std::forward_as_tuple(lhs, rhs)))) { \
return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
std::make_pair(GRID_BINOP(name)(), std::forward_as_tuple(lhs, rhs))); \
#define GRID_BINOP_LEFT(op, name) \
template <typename T1, typename T2, \
typename std::enable_if<is_lattice<T1>::value||is_lattice_expr<T1>::value,T1>::type * = nullptr> \
inline auto op(const T1 &lhs, const T2 &rhs) \
->decltype(LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs,rhs)) \
{ \
return LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs,rhs);\
}
#define GRID_BINOP_RIGHT(op, name) \
template <typename T1, typename T2, \
typename std::enable_if<!is_lattice<T1>::value && \
!is_lattice_expr<T1>::value, \
T1>::type * = nullptr, \
typename std::enable_if<is_lattice<T2>::value || \
is_lattice_expr<T2>::value, \
T2>::type * = nullptr> \
inline auto op(const T1 &lhs, const T2 &rhs) \
->decltype( \
LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
std::make_pair(GRID_BINOP(name)(), \
std::forward_as_tuple(lhs, rhs)))) { \
return LatticeBinaryExpression<GRID_BINOP(name), const T1 &, const T2 &>( \
std::make_pair(GRID_BINOP(name)(), std::forward_as_tuple(lhs, rhs))); \
#define GRID_BINOP_RIGHT(op, name) \
template <typename T1, typename T2, \
typename std::enable_if<!is_lattice<T1>::value&&!is_lattice_expr<T1>::value,T1>::type * = nullptr, \
typename std::enable_if< is_lattice<T2>::value|| is_lattice_expr<T2>::value,T2>::type * = nullptr> \
inline auto op(const T1 &lhs, const T2 &rhs) \
->decltype(LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs, rhs)) \
{ \
return LatticeBinaryExpression<GRID_BINOP(name),T1,T2>(GRID_BINOP(name)(),lhs, rhs); \
}
#define GRID_DEF_BINOP(op, name) \
GRID_BINOP_LEFT(op, name); \
#define GRID_DEF_BINOP(op, name) \
GRID_BINOP_LEFT(op, name); \
GRID_BINOP_RIGHT(op, name);
#define GRID_DEF_TRINOP(op, name) \
template <typename T1, typename T2, typename T3> \
inline auto op(const T1 &pred, const T2 &lhs, const T3 &rhs) \
->decltype( \
LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &, \
const T3 &>(std::make_pair( \
GRID_TRINOP(name)(), std::forward_as_tuple(pred, lhs, rhs)))) { \
return LatticeTrinaryExpression<GRID_TRINOP(name), const T1 &, const T2 &, \
const T3 &>(std::make_pair( \
GRID_TRINOP(name)(), std::forward_as_tuple(pred, lhs, rhs))); \
#define GRID_DEF_TRINOP(op, name) \
template <typename T1, typename T2, typename T3> \
inline auto op(const T1 &pred, const T2 &lhs, const T3 &rhs) \
->decltype(LatticeTrinaryExpression<GRID_TRINOP(name),T1,T2,T3>(GRID_TRINOP(name)(),pred, lhs, rhs)) \
{ \
return LatticeTrinaryExpression<GRID_TRINOP(name),T1,T2,T3>(GRID_TRINOP(name)(),pred, lhs, rhs); \
}
////////////////////////
// Operator definitions
////////////////////////
GRID_DEF_UNOP(operator-, UnarySub);
GRID_DEF_UNOP(Not, UnaryNot);
GRID_DEF_UNOP(operator!, UnaryNot);
@ -400,29 +370,27 @@ GRID_DEF_TRINOP(where, TrinaryWhere);
/////////////////////////////////////////////////////////////
template <class Op, class T1>
auto closure(const LatticeUnaryExpression<Op, T1> &expr)
-> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second))))> {
Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second))))> ret(
expr);
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1)))>
{
Lattice<decltype(expr.op.func(eval(0, expr.arg1)))> ret(expr);
return ret;
}
template <class Op, class T1, class T2>
auto closure(const LatticeBinaryExpression<Op, T1, T2> &expr)
-> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
eval(0, std::get<1>(expr.second))))> {
Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
eval(0, std::get<1>(expr.second))))>
ret(expr);
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))>
{
Lattice<decltype(expr.op.func(eval(0, expr.arg1),eval(0, expr.arg2)))> ret(expr);
return ret;
}
template <class Op, class T1, class T2, class T3>
auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
-> Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
eval(0, std::get<1>(expr.second)),
eval(0, std::get<2>(expr.second))))> {
Lattice<decltype(expr.first.func(eval(0, std::get<0>(expr.second)),
eval(0, std::get<1>(expr.second)),
eval(0, std::get<2>(expr.second))))>
ret(expr);
-> Lattice<decltype(expr.op.func(eval(0, expr.arg1),
eval(0, expr.arg2),
eval(0, expr.arg3)))>
{
Lattice<decltype(expr.op.func(eval(0, expr.arg1),
eval(0, expr.arg2),
eval(0, expr.arg3)))> ret(expr);
return ret;
}
@ -433,34 +401,7 @@ auto closure(const LatticeTrinaryExpression<Op, T1, T2, T3> &expr)
#undef GRID_DEF_UNOP
#undef GRID_DEF_BINOP
#undef GRID_DEF_TRINOP
}
#if 0
using namespace Grid;
int main(int argc,char **argv){
Lattice<double> v1(16);
Lattice<double> v2(16);
Lattice<double> v3(16);
BinaryAdd<double,double> tmp;
LatticeBinaryExpression<BinaryAdd<double,double>,Lattice<double> &,Lattice<double> &>
expr(std::make_pair(tmp,
std::forward_as_tuple(v1,v2)));
tmp.func(eval(0,v1),eval(0,v2));
auto var = v1+v2;
std::cout<<GridLogMessage<<typeid(var).name()<<std::endl;
v3=v1+v2;
v3=v1+v2+v1*v2;
};
void testit(Lattice<double> &v1,Lattice<double> &v2,Lattice<double> &v3)
{
v3=v1+v2+v1*v2;
}
#endif
NAMESPACE_END(Grid);
#endif

452
Grid/lattice/Lattice_arith.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,233 +23,235 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_ARITH_H
#define GRID_LATTICE_ARITH_H
namespace Grid {
//////////////////////////////////////////////////////////////////////////////////////////////////////
// avoid copy back routines for mult, mac, sub, add
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> strong_inline
void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.checkerboard = lhs.checkerboard;
conformable(ret,rhs);
conformable(lhs,rhs);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
mult(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
vstream(ret._odata[ss],tmp);
#else
mult(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
#endif
}
}
template<class obj1,class obj2,class obj3> strong_inline
void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.checkerboard = lhs.checkerboard;
conformable(ret,rhs);
conformable(lhs,rhs);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
mac(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
vstream(ret._odata[ss],tmp);
#else
mac(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
#endif
}
}
template<class obj1,class obj2,class obj3> strong_inline
void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.checkerboard = lhs.checkerboard;
conformable(ret,rhs);
conformable(lhs,rhs);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
sub(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
vstream(ret._odata[ss],tmp);
#else
sub(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
#endif
}
}
template<class obj1,class obj2,class obj3> strong_inline
void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.checkerboard = lhs.checkerboard;
conformable(ret,rhs);
conformable(lhs,rhs);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
add(&tmp,&lhs._odata[ss],&rhs._odata[ss]);
vstream(ret._odata[ss],tmp);
#else
add(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
#endif
}
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// avoid copy back routines for mult, mac, sub, add
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> strong_inline
void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.checkerboard = lhs.checkerboard;
conformable(lhs,ret);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
obj1 tmp;
mult(&tmp,&lhs._odata[ss],&rhs);
vstream(ret._odata[ss],tmp);
}
}
template<class obj1,class obj2,class obj3> strong_inline
void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.checkerboard = lhs.checkerboard;
conformable(ret,lhs);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
obj1 tmp;
mac(&tmp,&lhs._odata[ss],&rhs);
vstream(ret._odata[ss],tmp);
}
}
template<class obj1,class obj2,class obj3> strong_inline
void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.checkerboard = lhs.checkerboard;
conformable(ret,lhs);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
sub(&tmp,&lhs._odata[ss],&rhs);
vstream(ret._odata[ss],tmp);
#else
sub(&ret._odata[ss],&lhs._odata[ss],&rhs);
#endif
}
}
template<class obj1,class obj2,class obj3> strong_inline
void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.checkerboard = lhs.checkerboard;
conformable(lhs,ret);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
add(&tmp,&lhs._odata[ss],&rhs);
vstream(ret._odata[ss],tmp);
#else
add(&ret._odata[ss],&lhs._odata[ss],&rhs);
#endif
}
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// avoid copy back routines for mult, mac, sub, add
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> strong_inline
void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.checkerboard = rhs.checkerboard;
conformable(ret,rhs);
parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
mult(&tmp,&lhs,&rhs._odata[ss]);
vstream(ret._odata[ss],tmp);
#else
mult(&ret._odata[ss],&lhs,&rhs._odata[ss]);
#endif
}
}
template<class obj1,class obj2,class obj3> strong_inline
void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.checkerboard = rhs.checkerboard;
conformable(ret,rhs);
parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
mac(&tmp,&lhs,&rhs._odata[ss]);
vstream(ret._odata[ss],tmp);
#else
mac(&ret._odata[ss],&lhs,&rhs._odata[ss]);
#endif
}
}
template<class obj1,class obj2,class obj3> strong_inline
void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.checkerboard = rhs.checkerboard;
conformable(ret,rhs);
parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
sub(&tmp,&lhs,&rhs._odata[ss]);
vstream(ret._odata[ss],tmp);
#else
sub(&ret._odata[ss],&lhs,&rhs._odata[ss]);
#endif
}
}
template<class obj1,class obj2,class obj3> strong_inline
void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.checkerboard = rhs.checkerboard;
conformable(ret,rhs);
parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
#ifdef STREAMING_STORES
obj1 tmp;
add(&tmp,&lhs,&rhs._odata[ss]);
vstream(ret._odata[ss],tmp);
#else
add(&ret._odata[ss],&lhs,&rhs._odata[ss]);
#endif
}
}
template<class sobj,class vobj> strong_inline
void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
ret.checkerboard = x.checkerboard;
conformable(ret,x);
conformable(x,y);
parallel_for(int ss=0;ss<x._grid->oSites();ss++){
#ifdef STREAMING_STORES
vobj tmp = a*x._odata[ss]+y._odata[ss];
vstream(ret._odata[ss],tmp);
#else
ret._odata[ss]=a*x._odata[ss]+y._odata[ss];
#endif
}
}
template<class sobj,class vobj> strong_inline
void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
ret.checkerboard = x.checkerboard;
conformable(ret,x);
conformable(x,y);
parallel_for(int ss=0;ss<x._grid->oSites();ss++){
#ifdef STREAMING_STORES
vobj tmp = a*x._odata[ss]+b*y._odata[ss];
vstream(ret._odata[ss],tmp);
#else
ret._odata[ss]=a*x._odata[ss]+b*y._odata[ss];
#endif
}
}
template<class sobj,class vobj> strong_inline
RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
return axpy_norm_fast(ret,a,x,y);
}
template<class sobj,class vobj> strong_inline
RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
return axpby_norm_fast(ret,a,b,x,y);
}
NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////////////////////////////////////////////////
// avoid copy back routines for mult, mac, sub, add
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = lhs.Checkerboard();
auto ret_v = ret.View();
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
conformable(ret,rhs);
conformable(lhs,rhs);
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t = lhs_v(ss);
auto rhs_t = rhs_v(ss);
mult(&tmp,&lhs_t,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
auto rhs_t=rhs_v(ss);
mac(&tmp,&lhs_t,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
auto rhs_t=rhs_v(ss);
sub(&tmp,&lhs_t,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,rhs);
conformable(lhs,rhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
auto rhs_t=rhs_v(ss);
add(&tmp,&lhs_t,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// avoid copy back routines for mult, mac, sub, add
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(lhs,ret);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
mult(&tmp,&lhs_v(ss),&rhs);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void mac(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,lhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
mac(&tmp,&lhs_t,&rhs);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void sub(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(ret,lhs);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
sub(&tmp,&lhs_t,&rhs);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void add(Lattice<obj1> &ret,const Lattice<obj2> &lhs,const obj3 &rhs){
ret.Checkerboard() = lhs.Checkerboard();
conformable(lhs,ret);
auto ret_v = ret.View();
auto lhs_v = lhs.View();
accelerator_for(ss,lhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto lhs_t=lhs_v(ss);
add(&tmp,&lhs_t,&rhs);
coalescedWrite(ret_v[ss],tmp);
});
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
// avoid copy back routines for mult, mac, sub, add
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class obj1,class obj2,class obj3> inline
void mult(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
auto ret_v = ret.View();
auto rhs_v = lhs.View();
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto rhs_t=rhs_v(ss);
mult(&tmp,&lhs,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void mac(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
auto ret_v = ret.View();
auto rhs_v = lhs.View();
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto rhs_t=rhs_v(ss);
mac(&tmp,&lhs,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void sub(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
auto ret_v = ret.View();
auto rhs_v = lhs.View();
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto rhs_t=rhs_v(ss);
sub(&tmp,&lhs,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class obj1,class obj2,class obj3> inline
void add(Lattice<obj1> &ret,const obj2 &lhs,const Lattice<obj3> &rhs){
ret.Checkerboard() = rhs.Checkerboard();
conformable(ret,rhs);
auto ret_v = ret.View();
auto rhs_v = lhs.View();
accelerator_for(ss,rhs_v.size(),obj1::Nsimd(),{
decltype(coalescedRead(obj1())) tmp;
auto rhs_t=rhs_v(ss);
add(&tmp,&lhs,&rhs_t);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class sobj,class vobj> inline
void axpy(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y){
ret.Checkerboard() = x.Checkerboard();
conformable(ret,x);
conformable(x,y);
auto ret_v = ret.View();
auto x_v = x.View();
auto y_v = y.View();
accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
auto tmp = a*x_v(ss)+y_v(ss);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class sobj,class vobj> inline
void axpby(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y){
ret.Checkerboard() = x.Checkerboard();
conformable(ret,x);
conformable(x,y);
auto ret_v = ret.View();
auto x_v = x.View();
auto y_v = y.View();
accelerator_for(ss,x_v.size(),vobj::Nsimd(),{
auto tmp = a*x_v(ss)+b*y_v(ss);
coalescedWrite(ret_v[ss],tmp);
});
}
template<class sobj,class vobj> inline
RealD axpy_norm(Lattice<vobj> &ret,sobj a,const Lattice<vobj> &x,const Lattice<vobj> &y)
{
return axpy_norm_fast(ret,a,x,y);
}
template<class sobj,class vobj> inline
RealD axpby_norm(Lattice<vobj> &ret,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
{
return axpby_norm_fast(ret,a,b,x,y);
}
NAMESPACE_END(Grid);
#endif

595
Grid/lattice/Lattice_base.h
View File

@ -27,349 +27,440 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_BASE_H
#define GRID_LATTICE_BASE_H
/* END LEGAL */
#pragma once
#define STREAMING_STORES
namespace Grid {
// TODO:
// mac,real,imag
// Functionality:
// -=,+=,*=,()
// add,+,sub,-,mult,mac,*
// adj,conjugate
// real,imag
// transpose,transposeIndex
// trace,traceIndex
// peekIndex
// innerProduct,outerProduct,
// localNorm2
// localInnerProduct
NAMESPACE_BEGIN(Grid);
extern int GridCshiftPermuteMap[4][16];
////////////////////////////////////////////////
// Basic expressions used in Expression Template
////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////
// Base class which can be used by traits to pick up behaviour
///////////////////////////////////////////////////////////////////
class LatticeBase {};
class LatticeBase
/////////////////////////////////////////////////////////////////////////////////////////
// Conformable checks; same instance of Grid required
/////////////////////////////////////////////////////////////////////////////////////////
void accelerator_inline conformable(GridBase *lhs,GridBase *rhs)
{
public:
virtual ~LatticeBase(void) = default;
GridBase *_grid;
};
class LatticeExpressionBase {};
template <typename Op, typename T1>
class LatticeUnaryExpression : public std::pair<Op,std::tuple<T1> > , public LatticeExpressionBase {
public:
LatticeUnaryExpression(const std::pair<Op,std::tuple<T1> > &arg): std::pair<Op,std::tuple<T1> >(arg) {};
};
template <typename Op, typename T1, typename T2>
class LatticeBinaryExpression : public std::pair<Op,std::tuple<T1,T2> > , public LatticeExpressionBase {
public:
LatticeBinaryExpression(const std::pair<Op,std::tuple<T1,T2> > &arg): std::pair<Op,std::tuple<T1,T2> >(arg) {};
};
template <typename Op, typename T1, typename T2, typename T3>
class LatticeTrinaryExpression :public std::pair<Op,std::tuple<T1,T2,T3> >, public LatticeExpressionBase {
public:
LatticeTrinaryExpression(const std::pair<Op,std::tuple<T1,T2,T3> > &arg): std::pair<Op,std::tuple<T1,T2,T3> >(arg) {};
};
void inline conformable(GridBase *lhs,GridBase *rhs)
{
assert((lhs == rhs) && " conformable check pointers mismatch ");
assert(lhs == rhs);
}
template<class vobj>
class Lattice : public LatticeBase
////////////////////////////////////////////////////////////////////////////
// Minimal base class containing only data valid to access from accelerator
// _odata will be a managed pointer in CUDA
////////////////////////////////////////////////////////////////////////////
// Force access to lattice through a view object.
// prevents writing of code that will not offload to GPU, but perhaps annoyingly
// strict since host could could in principle direct access through the lattice object
// Need to decide programming model.
#define LATTICE_VIEW_STRICT
template<class vobj> class LatticeAccelerator : public LatticeBase
{
protected:
GridBase *_grid;
int checkerboard;
vobj *_odata; // A managed pointer
uint64_t _odata_size;
public:
accelerator_inline LatticeAccelerator() : checkerboard(0), _odata(nullptr), _odata_size(0), _grid(nullptr) { };
accelerator_inline uint64_t oSites(void) const { return _odata_size; };
accelerator_inline int Checkerboard(void) const { return checkerboard; };
accelerator_inline int &Checkerboard(void) { return this->checkerboard; }; // can assign checkerboard on a container, not a view
accelerator_inline void Conformable(GridBase * &grid) const
{
if (grid) conformable(grid, _grid);
else grid = _grid;
};
};
/////////////////////////////////////////////////////////////////////////////////////////
// A View class which provides accessor to the data.
// This will be safe to call from accelerator_for and is trivially copy constructible
// The copy constructor for this will need to be used by device lambda functions
/////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
class LatticeView : public LatticeAccelerator<vobj>
{
public:
int checkerboard;
Vector<vobj> _odata;
// to pthread need a computable loop where loop induction is not required
int begin(void) { return 0;};
int end(void) { return _odata.size(); }
vobj & operator[](int i) { return _odata[i]; };
const vobj & operator[](int i) const { return _odata[i]; };
// Rvalue
#ifdef __CUDA_ARCH__
accelerator_inline const typename vobj::scalar_object operator()(size_t i) const { return coalescedRead(this->_odata[i]); }
#else
accelerator_inline const vobj & operator()(size_t i) const { return this->_odata[i]; }
#endif
accelerator_inline const vobj & operator[](size_t i) const { return this->_odata[i]; };
accelerator_inline vobj & operator[](size_t i) { return this->_odata[i]; };
accelerator_inline uint64_t begin(void) const { return 0;};
accelerator_inline uint64_t end(void) const { return this->_odata_size; };
accelerator_inline uint64_t size(void) const { return this->_odata_size; };
LatticeView(const LatticeAccelerator<vobj> &refer_to_me) : LatticeAccelerator<vobj> (refer_to_me)
{
}
};
/////////////////////////////////////////////////////////////////////////////////////////
// Lattice expression types used by ET to assemble the AST
//
// Need to be able to detect code paths according to the whether a lattice object or not
// so introduce some trait type things
/////////////////////////////////////////////////////////////////////////////////////////
class LatticeExpressionBase {};
template <typename T> using is_lattice = std::is_base_of<LatticeBase, T>;
template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
template<class T, bool isLattice> struct ViewMapBase { typedef T Type; };
template<class T> struct ViewMapBase<T,true> { typedef LatticeView<typename T::vector_object> Type; };
template<class T> using ViewMap = ViewMapBase<T,std::is_base_of<LatticeBase, T>::value >;
template <typename Op, typename _T1>
class LatticeUnaryExpression : public LatticeExpressionBase
{
public:
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef vobj vector_object;
typedef typename ViewMap<_T1>::Type T1;
Op op;
T1 arg1;
LatticeUnaryExpression(Op _op,const _T1 &_arg1) : op(_op), arg1(_arg1) {};
};
template <typename Op, typename _T1, typename _T2>
class LatticeBinaryExpression : public LatticeExpressionBase
{
public:
typedef typename ViewMap<_T1>::Type T1;
typedef typename ViewMap<_T2>::Type T2;
Op op;
T1 arg1;
T2 arg2;
LatticeBinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2) : op(_op), arg1(_arg1), arg2(_arg2) {};
};
template <typename Op, typename _T1, typename _T2, typename _T3>
class LatticeTrinaryExpression : public LatticeExpressionBase
{
public:
typedef typename ViewMap<_T1>::Type T1;
typedef typename ViewMap<_T2>::Type T2;
typedef typename ViewMap<_T3>::Type T3;
Op op;
T1 arg1;
T2 arg2;
T3 arg3;
LatticeTrinaryExpression(Op _op,const _T1 &_arg1,const _T2 &_arg2,const _T3 &_arg3) : op(_op), arg1(_arg1), arg2(_arg2), arg3(_arg3) {};
};
/////////////////////////////////////////////////////////////////////////////////////////
// The real lattice class, with normal copy and assignment semantics.
// This contains extra (host resident) grid pointer data that may be accessed by host code
/////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
class Lattice : public LatticeAccelerator<vobj>
{
public:
GridBase *Grid(void) const { return this->_grid; }
///////////////////////////////////////////////////
// Member types
///////////////////////////////////////////////////
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef vobj vector_object;
private:
void dealloc(void)
{
alignedAllocator<vobj> alloc;
if( this->_odata_size ) {
alloc.deallocate(this->_odata,this->_odata_size);
this->_odata=nullptr;
this->_odata_size=0;
}
}
void resize(uint64_t size)
{
alignedAllocator<vobj> alloc;
if ( this->_odata_size != size ) {
dealloc();
}
this->_odata_size = size;
if ( size )
this->_odata = alloc.allocate(this->_odata_size);
else
this->_odata = nullptr;
}
public:
/////////////////////////////////////////////////////////////////////////////////
// Return a view object that may be dereferenced in site loops.
// The view is trivially copy constructible and may be copied to an accelerator device
// in device lambdas
/////////////////////////////////////////////////////////////////////////////////
LatticeView<vobj> View (void) const
{
LatticeView<vobj> accessor(*( (LatticeAccelerator<vobj> *) this));
return accessor;
}
~Lattice() {
if ( this->_odata_size ) {
dealloc();
}
}
////////////////////////////////////////////////////////////////////////////////
// Expression Template closure support
////////////////////////////////////////////////////////////////////////////////
template <typename Op, typename T1> strong_inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
template <typename Op, typename T1> inline Lattice<vobj> & operator=(const LatticeUnaryExpression<Op,T1> &expr)
{
GridBase *egrid(nullptr);
GridFromExpression(egrid,expr);
assert(egrid!=nullptr);
conformable(_grid,egrid);
conformable(this->_grid,egrid);
int cb=-1;
CBFromExpression(cb,expr);
assert( (cb==Odd) || (cb==Even));
checkerboard=cb;
this->checkerboard=cb;
parallel_for(int ss=0;ss<_grid->oSites();ss++){
#ifdef STREAMING_STORES
vobj tmp = eval(ss,expr);
vstream(_odata[ss] ,tmp);
#else
_odata[ss]=eval(ss,expr);
#endif
}
auto me = View();
accelerator_for(ss,me.size(),1,{
auto tmp = eval(ss,expr);
vstream(me[ss],tmp);
});
return *this;
}
template <typename Op, typename T1,typename T2> strong_inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
template <typename Op, typename T1,typename T2> inline Lattice<vobj> & operator=(const LatticeBinaryExpression<Op,T1,T2> &expr)
{
GridBase *egrid(nullptr);
GridFromExpression(egrid,expr);
assert(egrid!=nullptr);
conformable(_grid,egrid);
conformable(this->_grid,egrid);
int cb=-1;
CBFromExpression(cb,expr);
assert( (cb==Odd) || (cb==Even));
checkerboard=cb;
this->checkerboard=cb;
parallel_for(int ss=0;ss<_grid->oSites();ss++){
#ifdef STREAMING_STORES
vobj tmp = eval(ss,expr);
vstream(_odata[ss] ,tmp);
#else
_odata[ss]=eval(ss,expr);
#endif
}
auto me = View();
accelerator_for(ss,me.size(),1,{
auto tmp = eval(ss,expr);
vstream(me[ss],tmp);
});
return *this;
}
template <typename Op, typename T1,typename T2,typename T3> strong_inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
template <typename Op, typename T1,typename T2,typename T3> inline Lattice<vobj> & operator=(const LatticeTrinaryExpression<Op,T1,T2,T3> &expr)
{
GridBase *egrid(nullptr);
GridFromExpression(egrid,expr);
assert(egrid!=nullptr);
conformable(_grid,egrid);
conformable(this->_grid,egrid);
int cb=-1;
CBFromExpression(cb,expr);
assert( (cb==Odd) || (cb==Even));
checkerboard=cb;
parallel_for(int ss=0;ss<_grid->oSites();ss++){
#ifdef STREAMING_STORES
//vobj tmp = eval(ss,expr);
vstream(_odata[ss] ,eval(ss,expr));
#else
_odata[ss] = eval(ss,expr);
#endif
}
this->checkerboard=cb;
auto me = View();
accelerator_for(ss,me.size(),1,{
auto tmp = eval(ss,expr);
vstream(me[ss],tmp);
});
return *this;
}
//GridFromExpression is tricky to do
template<class Op,class T1>
Lattice(const LatticeUnaryExpression<Op,T1> & expr) {
_grid = nullptr;
GridFromExpression(_grid,expr);
assert(_grid!=nullptr);
Lattice(const LatticeUnaryExpression<Op,T1> & expr) {
this->_grid = nullptr;
GridFromExpression(this->_grid,expr);
assert(this->_grid!=nullptr);
int cb=-1;
CBFromExpression(cb,expr);
assert( (cb==Odd) || (cb==Even));
checkerboard=cb;
this->checkerboard=cb;
_odata.resize(_grid->oSites());
parallel_for(int ss=0;ss<_grid->oSites();ss++){
#ifdef STREAMING_STORES
vobj tmp = eval(ss,expr);
vstream(_odata[ss] ,tmp);
#else
_odata[ss]=eval(ss,expr);
#endif
}
};
resize(this->_grid->oSites());
*this = expr;
}
template<class Op,class T1, class T2>
Lattice(const LatticeBinaryExpression<Op,T1,T2> & expr) {
_grid = nullptr;
GridFromExpression(_grid,expr);
assert(_grid!=nullptr);
this->_grid = nullptr;
GridFromExpression(this->_grid,expr);
assert(this->_grid!=nullptr);
int cb=-1;
CBFromExpression(cb,expr);
assert( (cb==Odd) || (cb==Even));
checkerboard=cb;
this->checkerboard=cb;
_odata.resize(_grid->oSites());
parallel_for(int ss=0;ss<_grid->oSites();ss++){
#ifdef STREAMING_STORES
vobj tmp = eval(ss,expr);
vstream(_odata[ss] ,tmp);
#else
_odata[ss]=eval(ss,expr);
#endif
}
};
resize(this->_grid->oSites());
*this = expr;
}
template<class Op,class T1, class T2, class T3>
Lattice(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr) {
_grid = nullptr;
GridFromExpression(_grid,expr);
assert(_grid!=nullptr);
this->_grid = nullptr;
GridFromExpression(this->_grid,expr);
assert(this->_grid!=nullptr);
int cb=-1;
CBFromExpression(cb,expr);
assert( (cb==Odd) || (cb==Even));
checkerboard=cb;
this->checkerboard=cb;
_odata.resize(_grid->oSites());
parallel_for(int ss=0;ss<_grid->oSites();ss++){
vstream(_odata[ss] ,eval(ss,expr));
}
};
resize(this->_grid->oSites());
*this = expr;
}
template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
auto me = View();
thread_for(ss,me.size(),{
me[ss] = r;
});
return *this;
}
//////////////////////////////////////////////////////////////////
// Constructor requires "grid" passed.
// what about a default grid?
//////////////////////////////////////////////////////////////////
Lattice(GridBase *grid) : _odata(grid->oSites()) {
_grid = grid;
// _odata.reserve(_grid->oSites());
// _odata.resize(_grid->oSites());
// std::cout << "Constructing lattice object with Grid pointer "<<_grid<<std::endl;
assert((((uint64_t)&_odata[0])&0xF) ==0);
checkerboard=0;
// Follow rule of five, with Constructor requires "grid" passed
// to user defined constructor
///////////////////////////////////////////
// user defined constructor
///////////////////////////////////////////
Lattice(GridBase *grid) {
this->_grid = grid;
resize(this->_grid->oSites());
assert((((uint64_t)&this->_odata[0])&0xF) ==0);
this->checkerboard=0;
}
Lattice(const Lattice& r){ // copy constructor
_grid = r._grid;
checkerboard = r.checkerboard;
_odata.resize(_grid->oSites());// essential
parallel_for(int ss=0;ss<_grid->oSites();ss++){
_odata[ss]=r._odata[ss];
}
}
Lattice(Lattice&& r){ // move constructor
_grid = r._grid;
checkerboard = r.checkerboard;
_odata=std::move(r._odata);
}
inline Lattice<vobj> & operator = (Lattice<vobj> && r)
{
_grid = r._grid;
checkerboard = r.checkerboard;
_odata =std::move(r._odata);
return *this;
}
inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
_grid = r._grid;
checkerboard = r.checkerboard;
_odata.resize(_grid->oSites());// essential
parallel_for(int ss=0;ss<_grid->oSites();ss++){
_odata[ss]=r._odata[ss];
}
return *this;
}
template<class robj> strong_inline Lattice<vobj> & operator = (const Lattice<robj> & r){
this->checkerboard = r.checkerboard;
conformable(*this,r);
parallel_for(int ss=0;ss<_grid->oSites();ss++){
this->_odata[ss]=r._odata[ss];
}
return *this;
}
virtual ~Lattice(void) = default;
// virtual ~Lattice(void) = default;
void reset(GridBase* grid) {
if (_grid != grid) {
_grid = grid;
_odata.resize(grid->oSites());
checkerboard = 0;
if (this->_grid != grid) {
this->_grid = grid;
this->_odata.resize(grid->oSites());
this->checkerboard = 0;
}
}
template<class sobj> strong_inline Lattice<vobj> & operator = (const sobj & r){
parallel_for(int ss=0;ss<_grid->oSites();ss++){
this->_odata[ss]=r;
}
///////////////////////////////////////////
// copy constructor
///////////////////////////////////////////
Lattice(const Lattice& r){
// std::cout << "Lattice constructor(const Lattice &) "<<this<<std::endl;
this->_grid = r.Grid();
resize(this->_grid->oSites());
*this = r;
}
///////////////////////////////////////////
// move constructor
///////////////////////////////////////////
Lattice(Lattice && r){
this->_grid = r.Grid();
this->_odata = r._odata;
this->_odata_size = r._odata_size;
this->checkerboard= r.Checkerboard();
r._odata = nullptr;
r._odata_size = 0;
}
///////////////////////////////////////////
// assignment template
///////////////////////////////////////////
template<class robj> inline Lattice<vobj> & operator = (const Lattice<robj> & r){
typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
conformable(*this,r);
this->checkerboard = r.Checkerboard();
auto me = View();
auto him= r.View();
accelerator_for(ss,me.size(),vobj::Nsimd(),{
coalescedWrite(me[ss],him(ss));
});
return *this;
}
///////////////////////////////////////////
// Copy assignment
///////////////////////////////////////////
inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
this->checkerboard = r.Checkerboard();
conformable(*this,r);
auto me = View();
auto him= r.View();
accelerator_for(ss,me.size(),vobj::Nsimd(),{
coalescedWrite(me[ss],him(ss));
});
return *this;
}
///////////////////////////////////////////
// Move assignment possible if same type
///////////////////////////////////////////
inline Lattice<vobj> & operator = (Lattice<vobj> && r){
resize(0); // deletes if appropriate
this->_grid = r.Grid();
this->_odata = r._odata;
this->_odata_size = r._odata_size;
this->checkerboard= r.Checkerboard();
r._odata = nullptr;
r._odata_size = 0;
return *this;
}
/////////////////////////////////////////////////////////////////////////////
// *=,+=,-= operators inherit behvour from correspond */+/- operation
template<class T> strong_inline Lattice<vobj> &operator *=(const T &r) {
/////////////////////////////////////////////////////////////////////////////
template<class T> inline Lattice<vobj> &operator *=(const T &r) {
*this = (*this)*r;
return *this;
}
template<class T> strong_inline Lattice<vobj> &operator -=(const T &r) {
template<class T> inline Lattice<vobj> &operator -=(const T &r) {
*this = (*this)-r;
return *this;
}
template<class T> strong_inline Lattice<vobj> &operator +=(const T &r) {
template<class T> inline Lattice<vobj> &operator +=(const T &r) {
*this = (*this)+r;
return *this;
}
}; // class Lattice
template<class vobj> std::ostream& operator<< (std::ostream& stream, const Lattice<vobj> &o){
std::vector<int> gcoor;
typedef typename vobj::scalar_object sobj;
sobj ss;
for(int g=0;g<o._grid->_gsites;g++){
o._grid->GlobalIndexToGlobalCoor(g,gcoor);
peekSite(ss,o,gcoor);
stream<<"[";
for(int d=0;d<gcoor.size();d++){
stream<<gcoor[d];
if(d!=gcoor.size()-1) stream<<",";
}
stream<<"]\t";
stream<<ss<<std::endl;
}
return stream;
friend inline void swap(Lattice &l, Lattice &r) {
conformable(l,r);
LatticeAccelerator<vobj> tmp;
LatticeAccelerator<vobj> *lp = (LatticeAccelerator<vobj> *)&l;
LatticeAccelerator<vobj> *rp = (LatticeAccelerator<vobj> *)&r;
tmp = *lp; *lp=*rp; *rp=tmp;
}
}; // class Lattice
template<class vobj> std::ostream& operator<< (std::ostream& stream, const Lattice<vobj> &o){
typedef typename vobj::scalar_object sobj;
for(int g=0;g<o.Grid()->_gsites;g++){
Coordinate gcoor;
o.Grid()->GlobalIndexToGlobalCoor(g,gcoor);
sobj ss;
peekSite(ss,o,gcoor);
stream<<"[";
for(int d=0;d<gcoor.size();d++){
stream<<gcoor[d];
if(d!=gcoor.size()-1) stream<<",";
}
stream<<"]\t";
stream<<ss<<std::endl;
}
return stream;
}
NAMESPACE_END(Grid);
#include "Lattice_conformable.h"
#define GRID_LATTICE_EXPRESSION_TEMPLATES
#ifdef GRID_LATTICE_EXPRESSION_TEMPLATES
#include "Lattice_ET.h"
#else
#include "Lattice_overload.h"
#endif
#include "Lattice_arith.h"
#include "Lattice_trace.h"
#include "Lattice_transpose.h"
#include "Lattice_local.h"
#include "Lattice_reduction.h"
#include "Lattice_peekpoke.h"
#include "Lattice_reality.h"
#include "Lattice_comparison_utils.h"
#include "Lattice_comparison.h"
#include "Lattice_coordinate.h"
#include "Lattice_where.h"
#include "Lattice_rng.h"
#include "Lattice_unary.h"
#include "Lattice_transfer.h"
#endif

310
Grid/lattice/Lattice_comparison.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,146 +24,184 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_COMPARISON_H
#define GRID_LATTICE_COMPARISON_H
namespace Grid {
NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////////////////////
// relational operators
//
// Support <,>,<=,>=,==,!=
//
//Query supporting bitwise &, |, ^, !
//Query supporting logical &&, ||,
//////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////
// relational operators
//
// Support <,>,<=,>=,==,!=
//
//Query supporting bitwise &, |, ^, !
//Query supporting logical &&, ||,
//////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////
// compare lattice to lattice
//////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj>
inline Lattice<vInteger> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
{
Lattice<vInteger> ret(rhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
ret._odata[ss]=op(lhs._odata[ss],rhs._odata[ss]);
}
return ret;
typedef iScalar<vInteger> vPredicate ;
/*
template <class iobj, class vobj, class robj> accelerator_inline
vobj predicatedWhere(const iobj &predicate, const vobj &iftrue, const robj &iffalse)
{
typename std::remove_const<vobj>::type ret;
typedef typename vobj::scalar_object scalar_object;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
const int Nsimd = vobj::vector_type::Nsimd();
ExtractBuffer<Integer> mask(Nsimd);
ExtractBuffer<scalar_object> truevals(Nsimd);
ExtractBuffer<scalar_object> falsevals(Nsimd);
extract(iftrue, truevals);
extract(iffalse, falsevals);
extract<vInteger, Integer>(TensorRemove(predicate), mask);
for (int s = 0; s < Nsimd; s++) {
if (mask[s]) falsevals[s] = truevals[s];
}
//////////////////////////////////////////////////////////////////////////
// compare lattice to scalar
//////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj>
inline Lattice<vInteger> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
{
Lattice<vInteger> ret(lhs._grid);
parallel_for(int ss=0;ss<lhs._grid->oSites(); ss++){
ret._odata[ss]=op(lhs._odata[ss],rhs);
}
return ret;
}
//////////////////////////////////////////////////////////////////////////
// compare scalar to lattice
//////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj>
inline Lattice<vInteger> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
{
Lattice<vInteger> ret(rhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
ret._odata[ss]=op(lhs._odata[ss],rhs);
}
return ret;
}
//////////////////////////////////////////////////////////////////////////
// Map to functors
//////////////////////////////////////////////////////////////////////////
// Less than
template<class lobj,class robj>
inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vlt<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vlt<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vlt<lobj,robj>(),lhs,rhs);
}
// Less than equal
template<class lobj,class robj>
inline Lattice<vInteger> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vle<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vle<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vle<lobj,robj>(),lhs,rhs);
}
// Greater than
template<class lobj,class robj>
inline Lattice<vInteger> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vgt<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vgt<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vgt<lobj,robj>(),lhs,rhs);
}
// Greater than equal
template<class lobj,class robj>
inline Lattice<vInteger> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vge<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vge<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vge<lobj,robj>(),lhs,rhs);
}
// equal
template<class lobj,class robj>
inline Lattice<vInteger> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(veq<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(veq<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(veq<lobj,robj>(),lhs,rhs);
}
// not equal
template<class lobj,class robj>
inline Lattice<vInteger> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vne<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vne<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vInteger> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vne<lobj,robj>(),lhs,rhs);
}
merge(ret, falsevals);
return ret;
}
*/
//////////////////////////////////////////////////////////////////////////
// compare lattice to lattice
//////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> LLComparison(vfunctor op,const Lattice<lobj> &lhs,const Lattice<robj> &rhs)
{
Lattice<vPredicate> ret(rhs.Grid());
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
auto ret_v = ret.View();
thread_for( ss, rhs_v.size(), {
ret_v[ss]=op(lhs_v[ss],rhs_v[ss]);
});
return ret;
}
//////////////////////////////////////////////////////////////////////////
// compare lattice to scalar
//////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> LSComparison(vfunctor op,const Lattice<lobj> &lhs,const robj &rhs)
{
Lattice<vPredicate> ret(lhs.Grid());
auto lhs_v = lhs.View();
auto ret_v = ret.View();
thread_for( ss, lhs_v.size(), {
ret_v[ss]=op(lhs_v[ss],rhs);
});
return ret;
}
//////////////////////////////////////////////////////////////////////////
// compare scalar to lattice
//////////////////////////////////////////////////////////////////////////
template<class vfunctor,class lobj,class robj>
inline Lattice<vPredicate> SLComparison(vfunctor op,const lobj &lhs,const Lattice<robj> &rhs)
{
Lattice<vPredicate> ret(rhs.Grid());
auto rhs_v = rhs.View();
auto ret_v = ret.View();
thread_for( ss, rhs_v.size(), {
ret_v[ss]=op(lhs,rhs_v[ss]);
});
return ret;
}
//////////////////////////////////////////////////////////////////////////
// Map to functors
//////////////////////////////////////////////////////////////////////////
// Less than
template<class lobj,class robj>
inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vlt<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator < (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vlt<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator < (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vlt<lobj,robj>(),lhs,rhs);
}
// Less than equal
template<class lobj,class robj>
inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vle<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator <= (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vle<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator <= (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vle<lobj,robj>(),lhs,rhs);
}
// Greater than
template<class lobj,class robj>
inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vgt<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator > (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vgt<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator > (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vgt<lobj,robj>(),lhs,rhs);
}
// Greater than equal
template<class lobj,class robj>
inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vge<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator >= (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vge<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator >= (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vge<lobj,robj>(),lhs,rhs);
}
// equal
template<class lobj,class robj>
inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(veq<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator == (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(veq<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator == (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(veq<lobj,robj>(),lhs,rhs);
}
// not equal
template<class lobj,class robj>
inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const Lattice<robj> & rhs) {
return LLComparison(vne<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator != (const Lattice<lobj> & lhs, const robj & rhs) {
return LSComparison(vne<lobj,robj>(),lhs,rhs);
}
template<class lobj,class robj>
inline Lattice<vPredicate> operator != (const lobj & lhs, const Lattice<robj> & rhs) {
return SLComparison(vne<lobj,robj>(),lhs,rhs);
}
NAMESPACE_END(Grid);
#endif

66
Grid/lattice/Lattice_comparison_utils.h
View File

@ -26,10 +26,10 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_COMPARISON_H
#define GRID_COMPARISON_H
namespace Grid {
#pragma once
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////
// This implementation is a bit poor.
@ -44,42 +44,42 @@ namespace Grid {
//
template<class lobj,class robj> class veq {
public:
vInteger operator()(const lobj &lhs, const robj &rhs)
accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) == (rhs);
}
};
template<class lobj,class robj> class vne {
public:
vInteger operator()(const lobj &lhs, const robj &rhs)
accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) != (rhs);
}
};
template<class lobj,class robj> class vlt {
public:
vInteger operator()(const lobj &lhs, const robj &rhs)
accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) < (rhs);
}
};
template<class lobj,class robj> class vle {
public:
vInteger operator()(const lobj &lhs, const robj &rhs)
accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) <= (rhs);
}
};
template<class lobj,class robj> class vgt {
public:
vInteger operator()(const lobj &lhs, const robj &rhs)
accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) > (rhs);
}
};
template<class lobj,class robj> class vge {
public:
vInteger operator()(const lobj &lhs, const robj &rhs)
accelerator vInteger operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) >= (rhs);
}
@ -88,42 +88,42 @@ namespace Grid {
// Generic list of functors
template<class lobj,class robj> class seq {
public:
Integer operator()(const lobj &lhs, const robj &rhs)
accelerator Integer operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) == (rhs);
}
};
template<class lobj,class robj> class sne {
public:
Integer operator()(const lobj &lhs, const robj &rhs)
accelerator Integer operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) != (rhs);
}
};
template<class lobj,class robj> class slt {
public:
Integer operator()(const lobj &lhs, const robj &rhs)
accelerator Integer operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) < (rhs);
}
};
template<class lobj,class robj> class sle {
public:
Integer operator()(const lobj &lhs, const robj &rhs)
accelerator Integer operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) <= (rhs);
}
};
template<class lobj,class robj> class sgt {
public:
Integer operator()(const lobj &lhs, const robj &rhs)
accelerator Integer operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) > (rhs);
}
};
template<class lobj,class robj> class sge {
public:
Integer operator()(const lobj &lhs, const robj &rhs)
accelerator Integer operator()(const lobj &lhs, const robj &rhs)
{
return (lhs) >= (rhs);
}
@ -133,12 +133,12 @@ namespace Grid {
// Integer and real get extra relational functions.
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0>
inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const vsimd & rhs)
{
typedef typename vsimd::scalar_type scalar;
std::vector<scalar> vlhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
std::vector<scalar> vrhs(vsimd::Nsimd());
std::vector<Integer> vpred(vsimd::Nsimd());
ExtractBuffer<scalar> vlhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
ExtractBuffer<scalar> vrhs(vsimd::Nsimd());
ExtractBuffer<Integer> vpred(vsimd::Nsimd());
vInteger ret;
extract<vsimd,scalar>(lhs,vlhs);
extract<vsimd,scalar>(rhs,vrhs);
@ -150,11 +150,11 @@ namespace Grid {
}
template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0>
inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
accelerator_inline vInteger Comparison(sfunctor sop,const vsimd & lhs, const typename vsimd::scalar_type & rhs)
{
typedef typename vsimd::scalar_type scalar;
std::vector<scalar> vlhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
std::vector<Integer> vpred(vsimd::Nsimd());
ExtractBuffer<scalar> vlhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
ExtractBuffer<Integer> vpred(vsimd::Nsimd());
vInteger ret;
extract<vsimd,scalar>(lhs,vlhs);
for(int s=0;s<vsimd::Nsimd();s++){
@ -165,11 +165,11 @@ namespace Grid {
}
template<class sfunctor, class vsimd,IfNotComplex<vsimd> = 0>
inline vInteger Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
accelerator_inline vInteger Comparison(sfunctor sop,const typename vsimd::scalar_type & lhs, const vsimd & rhs)
{
typedef typename vsimd::scalar_type scalar;
std::vector<scalar> vrhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
std::vector<Integer> vpred(vsimd::Nsimd());
ExtractBuffer<scalar> vrhs(vsimd::Nsimd()); // Use functors to reduce this to single implementation
ExtractBuffer<Integer> vpred(vsimd::Nsimd());
vInteger ret;
extract<vsimd,scalar>(rhs,vrhs);
for(int s=0;s<vsimd::Nsimd();s++){
@ -181,30 +181,30 @@ namespace Grid {
#define DECLARE_RELATIONAL_EQ(op,functor) \
template<class vsimd,IfSimd<vsimd> = 0>\
inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\
accelerator_inline vInteger operator op (const vsimd & lhs, const vsimd & rhs)\
{\
typedef typename vsimd::scalar_type scalar;\
return Comparison(functor<scalar,scalar>(),lhs,rhs);\
}\
template<class vsimd,IfSimd<vsimd> = 0>\
inline vInteger operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
accelerator_inline vInteger operator op (const vsimd & lhs, const typename vsimd::scalar_type & rhs) \
{\
typedef typename vsimd::scalar_type scalar;\
return Comparison(functor<scalar,scalar>(),lhs,rhs);\
}\
template<class vsimd,IfSimd<vsimd> = 0>\
inline vInteger operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
accelerator_inline vInteger operator op (const typename vsimd::scalar_type & lhs, const vsimd & rhs) \
{\
typedef typename vsimd::scalar_type scalar;\
return Comparison(functor<scalar,scalar>(),lhs,rhs);\
}\
template<class vsimd>\
inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const typename vsimd::scalar_type &rhs) \
{ \
return lhs._internal op rhs; \
} \
template<class vsimd>\
inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
accelerator_inline vInteger operator op(const typename vsimd::scalar_type &lhs,const iScalar<vsimd> &rhs) \
{ \
return lhs op rhs._internal; \
} \
@ -212,7 +212,7 @@ namespace Grid {
#define DECLARE_RELATIONAL(op,functor) \
DECLARE_RELATIONAL_EQ(op,functor) \
template<class vsimd>\
inline vInteger operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
accelerator_inline vInteger operator op(const iScalar<vsimd> &lhs,const iScalar<vsimd> &rhs)\
{ \
return lhs._internal op rhs._internal; \
}
@ -226,7 +226,7 @@ DECLARE_RELATIONAL(!=,sne);
#undef DECLARE_RELATIONAL
}
NAMESPACE_END(Grid);
#endif

20
Grid/lattice/Lattice_conformable.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,18 +23,18 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_CONFORMABLE_H
#define GRID_LATTICE_CONFORMABLE_H
namespace Grid {
template<class obj1,class obj2> void conformable(const Lattice<obj1> &lhs,const Lattice<obj2> &rhs)
{
assert(lhs._grid == rhs._grid);
assert(lhs.checkerboard == rhs.checkerboard);
}
NAMESPACE_BEGIN(Grid);
template<class obj1,class obj2> void conformable(const Lattice<obj1> &lhs,const Lattice<obj2> &rhs)
{
assert(lhs.Grid() == rhs.Grid());
assert(lhs.Checkerboard() == rhs.Checkerboard());
}
NAMESPACE_END(Grid);
#endif

68
Grid/lattice/Lattice_coordinate.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,34 +23,52 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_COORDINATE_H
#define GRID_LATTICE_COORDINATE_H
*************************************************************************************/
/* END LEGAL */
#pragma once
namespace Grid {
NAMESPACE_BEGIN(Grid);
template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
{
typedef typename iobj::scalar_type scalar_type;
typedef typename iobj::vector_type vector_type;
template<class iobj> inline void LatticeCoordinate(Lattice<iobj> &l,int mu)
{
typedef typename iobj::scalar_type scalar_type;
typedef typename iobj::vector_type vector_type;
GridBase *grid = l._grid;
int Nsimd = grid->iSites();
GridBase *grid = l.Grid();
int Nsimd = grid->iSites();
std::vector<int> gcoor;
std::vector<scalar_type> mergebuf(Nsimd);
Coordinate gcoor;
ExtractBuffer<scalar_type> mergebuf(Nsimd);
vector_type vI;
for(int o=0;o<grid->oSites();o++){
for(int i=0;i<grid->iSites();i++){
grid->RankIndexToGlobalCoor(grid->ThisRank(),o,i,gcoor);
mergebuf[i]=(Integer)gcoor[mu];
}
merge<vector_type,scalar_type>(vI,mergebuf);
l._odata[o]=vI;
vector_type vI;
auto l_v = l.View();
for(int o=0;o<grid->oSites();o++){
for(int i=0;i<grid->iSites();i++){
grid->RankIndexToGlobalCoor(grid->ThisRank(),o,i,gcoor);
mergebuf[i]=(Integer)gcoor[mu];
}
merge<vector_type,scalar_type>(vI,mergebuf);
l_v[o]=vI;
}
};
// LatticeCoordinate();
// FIXME for debug; deprecate this; made obscelete by
template<class vobj> void lex_sites(Lattice<vobj> &l){
auto l_v = l.View();
Real *v_ptr = (Real *)&l_v[0];
size_t o_len = l.Grid()->oSites();
size_t v_len = sizeof(vobj)/sizeof(vRealF);
size_t vec_len = vRealF::Nsimd();
for(int i=0;i<o_len;i++){
for(int j=0;j<v_len;j++){
for(int vv=0;vv<vec_len;vv+=2){
v_ptr[i*v_len*vec_len+j*vec_len+vv ]= i+vv*500;
v_ptr[i*v_len*vec_len+j*vec_len+vv+1]= i+vv*500;
}
};
}}
}
#endif
NAMESPACE_END(Grid);

92
Grid/lattice/Lattice_local.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,8 +23,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_LOCALREDUCTION_H
#define GRID_LATTICE_LOCALREDUCTION_H
@ -32,44 +32,56 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
// localInner, localNorm, outerProduct
///////////////////////////////////////////////
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////
// Non site, reduced locally reduced routines
/////////////////////////////////////////////////////
/////////////////////////////////////////////////////
// Non site, reduced locally reduced routines
/////////////////////////////////////////////////////
// localNorm2,
template<class vobj>
inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
{
Lattice<typename vobj::tensor_reduced> ret(rhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
ret._odata[ss]=innerProduct(rhs._odata[ss],rhs._odata[ss]);
}
return ret;
}
// localInnerProduct
template<class vobj>
inline auto localInnerProduct (const Lattice<vobj> &lhs,const Lattice<vobj> &rhs) -> Lattice<typename vobj::tensor_reduced>
{
Lattice<typename vobj::tensor_reduced> ret(rhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
ret._odata[ss]=innerProduct(lhs._odata[ss],rhs._odata[ss]);
}
return ret;
}
// outerProduct Scalar x Scalar -> Scalar
// Vector x Vector -> Matrix
template<class ll,class rr>
inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Lattice<decltype(outerProduct(lhs._odata[0],rhs._odata[0]))>
{
Lattice<decltype(outerProduct(lhs._odata[0],rhs._odata[0]))> ret(rhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
ret._odata[ss]=outerProduct(lhs._odata[ss],rhs._odata[ss]);
}
return ret;
}
// localNorm2,
template<class vobj>
inline auto localNorm2 (const Lattice<vobj> &rhs)-> Lattice<typename vobj::tensor_reduced>
{
Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
auto rhs_v = rhs.View();
auto ret_v = ret.View();
accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss],innerProduct(rhs_v(ss),rhs_v(ss)));
});
return ret;
}
// localInnerProduct
template<class vobj>
inline auto localInnerProduct (const Lattice<vobj> &lhs,const Lattice<vobj> &rhs) -> Lattice<typename vobj::tensor_reduced>
{
Lattice<typename vobj::tensor_reduced> ret(rhs.Grid());
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
auto ret_v = ret.View();
accelerator_for(ss,rhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss],innerProduct(lhs_v(ss),rhs_v(ss)));
});
return ret;
}
// outerProduct Scalar x Scalar -> Scalar
// Vector x Vector -> Matrix
template<class ll,class rr>
inline auto outerProduct (const Lattice<ll> &lhs,const Lattice<rr> &rhs) -> Lattice<decltype(outerProduct(ll(),rr()))>
{
typedef decltype(coalescedRead(ll())) sll;
typedef decltype(coalescedRead(rr())) srr;
Lattice<decltype(outerProduct(ll(),rr()))> ret(rhs.Grid());
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
auto ret_v = ret.View();
accelerator_for(ss,rhs_v.size(),1,{
// FIXME had issues with scalar version of outer
// Use vector [] operator and don't read coalesce this loop
ret_v[ss]=outerProduct(lhs_v[ss],rhs_v[ss]);
});
return ret;
}
NAMESPACE_END(Grid);
#endif

202
Grid/lattice/Lattice_matrix_reduction.h Normal file
View File

@ -0,0 +1,202 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/lattice/Lattice_reduction.h
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>
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 <Grid/Grid_Eigen_Dense.h>
#ifdef GRID_WARN_SUBOPTIMAL
#warning "Optimisation alert all these reduction loops are NOT threaded "
#endif
NAMESPACE_BEGIN(Grid);
template<class vobj>
static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X.Grid()->GlobalDimensions()[Orthog];
GridBase *FullGrid = X.Grid();
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
// Lattice<vobj> Xslice(SliceGrid);
// Lattice<vobj> Rslice(SliceGrid);
assert( FullGrid->_simd_layout[Orthog]==1);
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
int stride=FullGrid->_slice_stride[Orthog];
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
auto X_v = X.View();
auto Y_v = Y.View();
auto R_v = R.View();
thread_region
{
std::vector<vobj> s_x(Nblock);
thread_loop_collapse2( (int n=0;n<nblock;n++),{
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
s_x[i] = X_v[o+i*ostride];
}
vobj dot;
for(int i=0;i<Nblock;i++){
dot = Y_v[o+i*ostride];
for(int j=0;j<Nblock;j++){
dot = dot + s_x[j]*(scale*aa(j,i));
}
R_v[o+i*ostride]=dot;
}
}});
}
};
template<class vobj>
static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,int Orthog,RealD scale=1.0)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X.Grid()->GlobalDimensions()[Orthog];
GridBase *FullGrid = X.Grid();
assert( FullGrid->_simd_layout[Orthog]==1);
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
int stride=FullGrid->_slice_stride[Orthog];
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
auto X_v = X.View();
auto R_v = R.View();
thread_region
{
std::vector<vobj> s_x(Nblock);
thread_loop_collapse2( (int n=0;n<nblock;n++),{
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
s_x[i] = X_v[o+i*ostride];
}
vobj dot;
for(int i=0;i<Nblock;i++){
dot = s_x[0]*(scale*aa(0,i));
for(int j=1;j<Nblock;j++){
dot = dot + s_x[j]*(scale*aa(j,i));
}
R_v[o+i*ostride]=dot;
}
}});
}
};
template<class vobj>
static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
GridBase *FullGrid = lhs.Grid();
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
int Nblock = FullGrid->GlobalDimensions()[Orthog];
// Lattice<vobj> Lslice(SliceGrid);
// Lattice<vobj> Rslice(SliceGrid);
mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
assert( FullGrid->_simd_layout[Orthog]==1);
// int nh = FullGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
// int nl = nh-1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
int stride=FullGrid->_slice_stride[Orthog];
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
typedef typename vobj::vector_typeD vector_typeD;
auto lhs_v = lhs.View();
auto rhs_v = rhs.View();
thread_region {
std::vector<vobj> Left(Nblock);
std::vector<vobj> Right(Nblock);
Eigen::MatrixXcd mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
thread_loop_collapse2((int n=0;n<nblock;n++),{
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
Left [i] = lhs_v[o+i*ostride];
Right[i] = rhs_v[o+i*ostride];
}
for(int i=0;i<Nblock;i++){
for(int j=0;j<Nblock;j++){
auto tmp = innerProduct(Left[i],Right[j]);
auto rtmp = TensorRemove(tmp);
ComplexD z = Reduce(rtmp);
mat_thread(i,j) += std::complex<double>(real(z),imag(z));
}}
}});
thread_critical {
mat += mat_thread;
}
}
for(int i=0;i<Nblock;i++){
for(int j=0;j<Nblock;j++){
ComplexD sum = mat(i,j);
FullGrid->GlobalSum(sum);
mat(i,j)=sum;
}}
return;
}
NAMESPACE_END(Grid);

138
Grid/lattice/Lattice_overload.h
View File

@ -1,138 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/lattice/Lattice_overload.h
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_OVERLOAD_H
#define GRID_LATTICE_OVERLOAD_H
namespace Grid {
//////////////////////////////////////////////////////////////////////////////////////////////////////
// unary negation
//////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
inline Lattice<vobj> operator -(const Lattice<vobj> &r)
{
Lattice<vobj> ret(r._grid);
parallel_for(int ss=0;ss<r._grid->oSites();ss++){
vstream(ret._odata[ss], -r._odata[ss]);
}
return ret;
}
/////////////////////////////////////////////////////////////////////////////////////
// Lattice BinOp Lattice,
//NB mult performs conformable check. Do not reapply here for performance.
/////////////////////////////////////////////////////////////////////////////////////
template<class left,class right>
inline auto operator * (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]*rhs._odata[0])>
{
Lattice<decltype(lhs._odata[0]*rhs._odata[0])> ret(rhs._grid);
mult(ret,lhs,rhs);
return ret;
}
template<class left,class right>
inline auto operator + (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]+rhs._odata[0])>
{
Lattice<decltype(lhs._odata[0]+rhs._odata[0])> ret(rhs._grid);
add(ret,lhs,rhs);
return ret;
}
template<class left,class right>
inline auto operator - (const Lattice<left> &lhs,const Lattice<right> &rhs)-> Lattice<decltype(lhs._odata[0]-rhs._odata[0])>
{
Lattice<decltype(lhs._odata[0]-rhs._odata[0])> ret(rhs._grid);
sub(ret,lhs,rhs);
return ret;
}
// Scalar BinOp Lattice ;generate return type
template<class left,class right>
inline auto operator * (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs*rhs._odata[0])>
{
Lattice<decltype(lhs*rhs._odata[0])> ret(rhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
decltype(lhs*rhs._odata[0]) tmp=lhs*rhs._odata[ss];
vstream(ret._odata[ss],tmp);
// ret._odata[ss]=lhs*rhs._odata[ss];
}
return ret;
}
template<class left,class right>
inline auto operator + (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs+rhs._odata[0])>
{
Lattice<decltype(lhs+rhs._odata[0])> ret(rhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
decltype(lhs+rhs._odata[0]) tmp =lhs-rhs._odata[ss];
vstream(ret._odata[ss],tmp);
// ret._odata[ss]=lhs+rhs._odata[ss];
}
return ret;
}
template<class left,class right>
inline auto operator - (const left &lhs,const Lattice<right> &rhs) -> Lattice<decltype(lhs-rhs._odata[0])>
{
Lattice<decltype(lhs-rhs._odata[0])> ret(rhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
decltype(lhs-rhs._odata[0]) tmp=lhs-rhs._odata[ss];
vstream(ret._odata[ss],tmp);
}
return ret;
}
template<class left,class right>
inline auto operator * (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]*rhs)>
{
Lattice<decltype(lhs._odata[0]*rhs)> ret(lhs._grid);
parallel_for(int ss=0;ss<lhs._grid->oSites(); ss++){
decltype(lhs._odata[0]*rhs) tmp =lhs._odata[ss]*rhs;
vstream(ret._odata[ss],tmp);
// ret._odata[ss]=lhs._odata[ss]*rhs;
}
return ret;
}
template<class left,class right>
inline auto operator + (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]+rhs)>
{
Lattice<decltype(lhs._odata[0]+rhs)> ret(lhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
decltype(lhs._odata[0]+rhs) tmp=lhs._odata[ss]+rhs;
vstream(ret._odata[ss],tmp);
// ret._odata[ss]=lhs._odata[ss]+rhs;
}
return ret;
}
template<class left,class right>
inline auto operator - (const Lattice<left> &lhs,const right &rhs) -> Lattice<decltype(lhs._odata[0]-rhs)>
{
Lattice<decltype(lhs._odata[0]-rhs)> ret(lhs._grid);
parallel_for(int ss=0;ss<rhs._grid->oSites(); ss++){
decltype(lhs._odata[0]-rhs) tmp=lhs._odata[ss]-rhs;
vstream(ret._odata[ss],tmp);
// ret._odata[ss]=lhs._odata[ss]-rhs;
}
return ret;
}
}
#endif

344
Grid/lattice/Lattice_peekpoke.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -25,8 +25,8 @@ Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_PEEK_H
#define GRID_LATTICE_PEEK_H
@ -34,172 +34,184 @@ Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
// Peeking and poking around
///////////////////////////////////////////////
namespace Grid {
////////////////////////////////////////////////////////////////////////////////////////////////////
// Peek internal indices of a Lattice object
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Index,class vobj>
auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Index>(lhs._odata[0],i))>
{
Lattice<decltype(peekIndex<Index>(lhs._odata[0],i))> ret(lhs._grid);
ret.checkerboard=lhs.checkerboard;
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i);
}
return ret;
};
template<int Index,class vobj>
auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekIndex<Index>(lhs._odata[0],i,j))>
{
Lattice<decltype(peekIndex<Index>(lhs._odata[0],i,j))> ret(lhs._grid);
ret.checkerboard=lhs.checkerboard;
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
ret._odata[ss] = peekIndex<Index>(lhs._odata[ss],i,j);
}
return ret;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
// Poke internal indices of a Lattice object
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Index,class vobj>
void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(lhs._odata[0],0))> & rhs,int i)
{
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i);
}
}
template<int Index,class vobj>
void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(lhs._odata[0],0,0))> & rhs,int i,int j)
{
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
pokeIndex<Index>(lhs._odata[ss],rhs._odata[ss],i,j);
}
}
//////////////////////////////////////////////////////
// Poke a scalar object into the SIMD array
//////////////////////////////////////////////////////
template<class vobj,class sobj>
void pokeSite(const sobj &s,Lattice<vobj> &l,const std::vector<int> &site){
GridBase *grid=l._grid;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.checkerboard== l._grid->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
int rank,odx,idx;
// Optional to broadcast from node 0.
grid->GlobalCoorToRankIndex(rank,odx,idx,site);
grid->Broadcast(grid->BossRank(),s);
std::vector<sobj> buf(Nsimd);
// extract-modify-merge cycle is easiest way and this is not perf critical
if ( rank == grid->ThisRank() ) {
extract(l._odata[odx],buf);
buf[idx] = s;
merge(l._odata[odx],buf);
}
return;
};
NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////
// Peek a scalar object from the SIMD array
//////////////////////////////////////////////////////////
template<class vobj,class sobj>
void peekSite(sobj &s,const Lattice<vobj> &l,const std::vector<int> &site){
GridBase *grid=l._grid;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.checkerboard == l._grid->CheckerBoard(site));
int rank,odx,idx;
grid->GlobalCoorToRankIndex(rank,odx,idx,site);
std::vector<sobj> buf(Nsimd);
extract(l._odata[odx],buf);
s = buf[idx];
grid->Broadcast(rank,s);
return;
};
//////////////////////////////////////////////////////////
// Peek a scalar object from the SIMD array
//////////////////////////////////////////////////////////
template<class vobj,class sobj>
void peekLocalSite(sobj &s,const Lattice<vobj> &l,std::vector<int> &site){
GridBase *grid = l._grid;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.checkerboard== l._grid->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
static const int words=sizeof(vobj)/sizeof(vector_type);
int odx,idx;
idx= grid->iIndex(site);
odx= grid->oIndex(site);
scalar_type * vp = (scalar_type *)&l._odata[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
pt[w] = vp[idx+w*Nsimd];
}
return;
};
template<class vobj,class sobj>
void pokeLocalSite(const sobj &s,Lattice<vobj> &l,std::vector<int> &site){
GridBase *grid=l._grid;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.checkerboard== l._grid->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
static const int words=sizeof(vobj)/sizeof(vector_type);
int odx,idx;
idx= grid->iIndex(site);
odx= grid->oIndex(site);
scalar_type * vp = (scalar_type *)&l._odata[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
vp[idx+w*Nsimd] = pt[w];
}
return;
};
// FIXME accelerator_loop and accelerator_inline these
////////////////////////////////////////////////////////////////////////////////////////////////////
// Peek internal indices of a Lattice object
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Index,class vobj>
auto PeekIndex(const Lattice<vobj> &lhs,int i) -> Lattice<decltype(peekIndex<Index>(vobj(),i))>
{
Lattice<decltype(peekIndex<Index>(vobj(),i))> ret(lhs.Grid());
ret.Checkerboard()=lhs.Checkerboard();
auto ret_v = ret.View();
auto lhs_v = lhs.View();
thread_for( ss, lhs_v.size(), {
ret_v[ss] = peekIndex<Index>(lhs_v[ss],i);
});
return ret;
};
template<int Index,class vobj>
auto PeekIndex(const Lattice<vobj> &lhs,int i,int j) -> Lattice<decltype(peekIndex<Index>(vobj(),i,j))>
{
Lattice<decltype(peekIndex<Index>(vobj(),i,j))> ret(lhs.Grid());
ret.Checkerboard()=lhs.Checkerboard();
auto ret_v = ret.View();
auto lhs_v = lhs.View();
thread_for( ss, lhs_v.size(), {
ret_v[ss] = peekIndex<Index>(lhs_v[ss],i,j);
});
return ret;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
// Poke internal indices of a Lattice object
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Index,class vobj>
void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0))> & rhs,int i)
{
auto rhs_v = rhs.View();
auto lhs_v = lhs.View();
thread_for( ss, lhs_v.size(), {
pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i);
});
}
template<int Index,class vobj>
void PokeIndex(Lattice<vobj> &lhs,const Lattice<decltype(peekIndex<Index>(vobj(),0,0))> & rhs,int i,int j)
{
auto rhs_v = rhs.View();
auto lhs_v = lhs.View();
thread_for( ss, lhs_v.size(), {
pokeIndex<Index>(lhs_v[ss],rhs_v[ss],i,j);
});
}
//////////////////////////////////////////////////////
// Poke a scalar object into the SIMD array
//////////////////////////////////////////////////////
template<class vobj,class sobj>
void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
GridBase *grid=l.Grid();
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
int rank,odx,idx;
// Optional to broadcast from node 0.
grid->GlobalCoorToRankIndex(rank,odx,idx,site);
grid->Broadcast(grid->BossRank(),s);
// extract-modify-merge cycle is easiest way and this is not perf critical
ExtractBuffer<sobj> buf(Nsimd);
auto l_v = l.View();
if ( rank == grid->ThisRank() ) {
extract(l_v[odx],buf);
buf[idx] = s;
merge(l_v[odx],buf);
}
return;
};
//////////////////////////////////////////////////////////
// Peek a scalar object from the SIMD array
//////////////////////////////////////////////////////////
template<class vobj,class sobj>
void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
GridBase *grid=l.Grid();
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard() == l.Grid()->CheckerBoard(site));
int rank,odx,idx;
grid->GlobalCoorToRankIndex(rank,odx,idx,site);
ExtractBuffer<sobj> buf(Nsimd);
auto l_v = l.View();
extract(l_v[odx],buf);
s = buf[idx];
grid->Broadcast(rank,s);
return;
};
//////////////////////////////////////////////////////////
// Peek a scalar object from the SIMD array
//////////////////////////////////////////////////////////
template<class vobj,class sobj>
void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
GridBase *grid = l.Grid();
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
static const int words=sizeof(vobj)/sizeof(vector_type);
int odx,idx;
idx= grid->iIndex(site);
odx= grid->oIndex(site);
auto l_v = l.View();
scalar_type * vp = (scalar_type *)&l_v[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
pt[w] = vp[idx+w*Nsimd];
}
return;
};
template<class vobj,class sobj>
void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
GridBase *grid=l.Grid();
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
static const int words=sizeof(vobj)/sizeof(vector_type);
int odx,idx;
idx= grid->iIndex(site);
odx= grid->oIndex(site);
auto l_v = l.View();
scalar_type * vp = (scalar_type *)&l_v[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
vp[idx+w*Nsimd] = pt[w];
}
return;
};
NAMESPACE_END(Grid);
#endif

44
Grid/lattice/Lattice_reality.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -25,8 +25,8 @@ Author: neo <cossu@post.kek.jp>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_REALITY_H
#define GRID_LATTICE_REALITY_H
@ -36,22 +36,28 @@ Author: neo <cossu@post.kek.jp>
// The choice of burying complex in the SIMD
// is making the use of "real" and "imag" very cumbersome
namespace Grid {
NAMESPACE_BEGIN(Grid);
template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs._grid);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
ret._odata[ss] = adj(lhs._odata[ss]);
}
return ret;
};
template<class vobj> inline Lattice<vobj> adj(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
auto lhs_v = lhs.View();
auto ret_v = ret.View();
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
coalescedWrite(ret_v[ss], adj(lhs_v(ss)));
});
return ret;
};
template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
auto lhs_v = lhs.View();
auto ret_v = ret.View();
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
coalescedWrite( ret_v[ss] , conjugate(lhs_v(ss)));
});
return ret;
};
NAMESPACE_END(Grid);
template<class vobj> inline Lattice<vobj> conjugate(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs._grid);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
ret._odata[ss] = conjugate(lhs._odata[ss]);
}
return ret;
};
}
#endif

418
Grid/lattice/Lattice_reduction.h
View File

@ -19,22 +19,76 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_REDUCTION_H
#define GRID_LATTICE_REDUCTION_H
#pragma once
#include <Grid/Grid_Eigen_Dense.h>
namespace Grid {
#ifdef GRID_WARN_SUBOPTIMAL
#warning "Optimisation alert all these reduction loops are NOT threaded "
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// Deterministic Reduction operations
////////////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_NVCC
#include <Grid/lattice/Lattice_reduction_gpu.h>
#endif
NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////
// FIXME this should promote to double and accumulate
//////////////////////////////////////////////////////
template<class vobj>
inline typename vobj::scalar_object sum_cpu(const vobj *arg, Integer osites)
{
typedef typename vobj::scalar_object sobj;
const int Nsimd = vobj::Nsimd();
const int 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)
{
#ifdef GRID_NVCC
return sum_gpu(arg,osites);
#else
return sum_cpu(arg,osites);
#endif
}
template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
{
auto arg_v = arg.View();
Integer osites = arg.Grid()->oSites();
auto ssum= sum(&arg_v[0],osites);
arg.Grid()->GlobalSum(ssum);
return ssum;
}
////////////////////////////////////////////////////////////////////////////////////////////////////
// Deterministic Reduction operations
////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
auto nrm = innerProduct(arg,arg);
return std::real(nrm);
ComplexD nrm = innerProduct(arg,arg);
return real(nrm);
}
// Double inner product
@ -43,32 +97,49 @@ inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &righ
{
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_typeD vector_type;
GridBase *grid = left._grid;
const int pad = 8;
ComplexD inner;
Vector<ComplexD> sumarray(grid->SumArraySize()*pad);
parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
decltype(innerProductD(left._odata[0],right._odata[0])) vinner=zero; // private to thread; sub summation
for(int ss=myoff;ss<mywork+myoff; ss++){
vinner = vinner + innerProductD(left._odata[ss],right._odata[ss]);
}
// All threads sum across SIMD; reduce serial work at end
// one write per cacheline with streaming store
ComplexD tmp = Reduce(TensorRemove(vinner)) ;
vstream(sumarray[thr*pad],tmp);
}
ComplexD nrm;
inner=0.0;
for(int i=0;i<grid->SumArraySize();i++){
inner = inner+sumarray[i*pad];
}
right._grid->GlobalSum(inner);
return inner;
GridBase *grid = left.Grid();
// Might make all code paths go this way.
auto left_v = left.View();
auto right_v=right.View();
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites = grid->oSites();
#ifdef GRID_NVCC
// GPU - SIMT lane compliance...
typedef decltype(innerProduct(left_v[0],right_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
accelerator_for( ss, sites, nsimd,{
auto x_l = left_v(ss);
auto y_l = right_v(ss);
coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
})
// This is in single precision and fails some tests
// Need a sumD that sums in double
nrm = TensorRemove(sumD_gpu(inner_tmp_v,sites));
#else
// CPU
typedef decltype(innerProductD(left_v[0],right_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
accelerator_for( ss, sites, nsimd,{
auto x_l = left_v[ss];
auto y_l = right_v[ss];
inner_tmp_v[ss]=innerProductD(x_l,y_l);
})
nrm = TensorRemove(sum(inner_tmp_v,sites));
#endif
grid->GlobalSum(nrm);
return nrm;
}
/////////////////////////
@ -86,8 +157,7 @@ axpy_norm_fast(Lattice<vobj> &z,sobj a,const Lattice<vobj> &x,const Lattice<vobj
template<class sobj,class vobj> strong_inline RealD
axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y)
{
const int pad = 8;
z.checkerboard = x.checkerboard;
z.Checkerboard() = x.Checkerboard();
conformable(z,x);
conformable(x,y);
@ -95,43 +165,57 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
typedef typename vobj::vector_typeD vector_type;
RealD nrm;
GridBase *grid = x._grid;
GridBase *grid = x.Grid();
auto x_v=x.View();
auto y_v=y.View();
auto z_v=z.View();
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites = grid->oSites();
Vector<RealD> sumarray(grid->SumArraySize()*pad);
#ifdef GRID_NVCC
// GPU
typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
accelerator_for( ss, sites, nsimd,{
auto tmp = a*x_v(ss)+b*y_v(ss);
coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
coalescedWrite(z_v[ss],tmp);
});
nrm = real(TensorRemove(sumD_gpu(inner_tmp_v,sites)));
#else
// CPU
typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(x._grid->oSites(),thr,mywork,myoff);
// private to thread; sub summation
decltype(innerProductD(z._odata[0],z._odata[0])) vnrm=zero;
for(int ss=myoff;ss<mywork+myoff; ss++){
vobj tmp = a*x._odata[ss]+b*y._odata[ss];
vnrm = vnrm + innerProductD(tmp,tmp);
vstream(z._odata[ss],tmp);
}
vstream(sumarray[thr*pad],real(Reduce(TensorRemove(vnrm)))) ;
}
nrm = 0.0; // sum across threads; linear in thread count but fast
for(int i=0;i<grid->SumArraySize();i++){
nrm = nrm+sumarray[i*pad];
}
z._grid->GlobalSum(nrm);
accelerator_for( ss, sites, nsimd,{
auto tmp = a*x_v(ss)+b*y_v(ss);
inner_tmp_v[ss]=innerProductD(tmp,tmp);
z_v[ss]=tmp;
});
// Already promoted to double
nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
#endif
grid->GlobalSum(nrm);
return nrm;
}
template<class Op,class T1>
inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second))))::scalar_object
->typename decltype(expr.op.func(eval(0,expr.arg1)))::scalar_object
{
return sum(closure(expr));
}
template<class Op,class T1,class T2>
inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),eval(0,std::get<1>(expr.second))))::scalar_object
->typename decltype(expr.op.func(eval(0,expr.arg1),eval(0,expr.arg2)))::scalar_object
{
return sum(closure(expr));
}
@ -139,54 +223,14 @@ inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
template<class Op,class T1,class T2,class T3>
inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),
eval(0,std::get<1>(expr.second)),
eval(0,std::get<2>(expr.second))
->typename decltype(expr.op.func(eval(0,expr.arg1),
eval(0,expr.arg2),
eval(0,expr.arg3)
))::scalar_object
{
return sum(closure(expr));
}
template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
{
GridBase *grid=arg._grid;
int Nsimd = grid->Nsimd();
std::vector<vobj,alignedAllocator<vobj> > sumarray(grid->SumArraySize());
for(int i=0;i<grid->SumArraySize();i++){
sumarray[i]=zero;
}
parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(grid->oSites(),thr,mywork,myoff);
vobj vvsum=zero;
for(int ss=myoff;ss<mywork+myoff; ss++){
vvsum = vvsum + arg._odata[ss];
}
sumarray[thr]=vvsum;
}
vobj vsum=zero; // sum across threads
for(int i=0;i<grid->SumArraySize();i++){
vsum = vsum+sumarray[i];
}
typedef typename vobj::scalar_object sobj;
sobj ssum=zero;
std::vector<sobj> buf(Nsimd);
extract(vsum,buf);
for(int i=0;i<Nsimd;i++) ssum = ssum + buf[i];
arg._grid->GlobalSum(ssum);
return ssum;
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
// sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -199,7 +243,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
// But easily avoided by using double precision fields
///////////////////////////////////////////////////////
typedef typename vobj::scalar_object sobj;
GridBase *grid = Data._grid;
GridBase *grid = Data.Grid();
assert(grid!=NULL);
const int Nd = grid->_ndimension;
@ -212,13 +256,13 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
int ld=grid->_ldimensions[orthogdim];
int rd=grid->_rdimensions[orthogdim];
std::vector<vobj,alignedAllocator<vobj> > lvSum(rd); // will locally sum vectors first
std::vector<sobj> lsSum(ld,zero); // sum across these down to scalars
std::vector<sobj> extracted(Nsimd); // splitting the SIMD
Vector<vobj> lvSum(rd); // will locally sum vectors first
Vector<sobj> lsSum(ld,Zero()); // sum across these down to scalars
ExtractBuffer<sobj> extracted(Nsimd); // splitting the SIMD
result.resize(fd); // And then global sum to return the same vector to every node
for(int r=0;r<rd;r++){
lvSum[r]=zero;
lvSum[r]=Zero();
}
int e1= grid->_slice_nblock[orthogdim];
@ -227,20 +271,19 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
// sum over reduced dimension planes, breaking out orthog dir
// Parallel over orthog direction
parallel_for(int r=0;r<rd;r++){
auto Data_v=Data.View();
thread_for( r,rd, {
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
lvSum[r]=lvSum[r]+Data._odata[ss];
lvSum[r]=lvSum[r]+Data_v[ss];
}
}
}
});
// Sum across simd lanes in the plane, breaking out orthog dir.
std::vector<int> icoor(Nd);
Coordinate icoor(Nd);
for(int rt=0;rt<rd;rt++){
@ -265,7 +308,7 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
if ( pt == grid->_processor_coor[orthogdim] ) {
gsum=lsSum[lt];
} else {
gsum=zero;
gsum=Zero();
}
grid->GlobalSum(gsum);
@ -292,9 +335,9 @@ static void localSliceInnerProductVector(std::vector<ComplexD> &result, const La
// std::cout << GridLogMessage << "Start prep" << std::endl;
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
GridBase *grid = lhs._grid;
GridBase *grid = lhs.Grid();
assert(grid!=NULL);
conformable(grid,rhs._grid);
conformable(grid,rhs.Grid());
const int Nd = grid->_ndimension;
const int Nsimd = grid->Nsimd();
@ -307,14 +350,14 @@ static void localSliceInnerProductVector(std::vector<ComplexD> &result, const La
int rd=grid->_rdimensions[orthogdim];
// std::cout << GridLogMessage << "Start alloc" << std::endl;
std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd); // will locally sum vectors first
Vector<vector_type> lvSum(rd); // will locally sum vectors first
lsSum.resize(ld,scalar_type(0.0)); // sum across these down to scalars
std::vector<iScalar<scalar_type>> extracted(Nsimd); // splitting the SIMD
ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd); // splitting the SIMD
// std::cout << GridLogMessage << "End alloc" << std::endl;
result.resize(fd); // And then global sum to return the same vector to every node for IO to file
for(int r=0;r<rd;r++){
lvSum[r]=zero;
lvSum[r]=Zero();
}
int e1= grid->_slice_nblock[orthogdim];
@ -323,23 +366,24 @@ static void localSliceInnerProductVector(std::vector<ComplexD> &result, const La
// std::cout << GridLogMessage << "End prep" << std::endl;
// std::cout << GridLogMessage << "Start parallel inner product, _rd = " << rd << std::endl;
vector_type vv;
parallel_for(int r=0;r<rd;r++)
{
auto l_v=lhs.View();
auto r_v=rhs.View();
thread_for( r,rd,{
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss = so + n * stride + b;
vv = TensorRemove(innerProduct(lhs._odata[ss], rhs._odata[ss]));
vv = TensorRemove(innerProduct(l_v[ss], r_v[ss]));
lvSum[r] = lvSum[r] + vv;
}
}
}
});
// std::cout << GridLogMessage << "End parallel inner product" << std::endl;
// Sum across simd lanes in the plane, breaking out orthog dir.
std::vector<int> icoor(Nd);
Coordinate icoor(Nd);
for(int rt=0;rt<rd;rt++){
iScalar<vector_type> temp;
@ -362,7 +406,7 @@ template <class vobj>
static void globalSliceInnerProductVector(std::vector<ComplexD> &result, const Lattice<vobj> &lhs, std::vector<typename vobj::scalar_type> &lsSum, int orthogdim)
{
typedef typename vobj::scalar_type scalar_type;
GridBase *grid = lhs._grid;
GridBase *grid = lhs.Grid();
int fd = result.size();
int ld = lsSum.size();
// sum over nodes.
@ -388,9 +432,9 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
{
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
GridBase *grid = lhs._grid;
GridBase *grid = lhs.Grid();
assert(grid!=NULL);
conformable(grid,rhs._grid);
conformable(grid,rhs.Grid());
const int Nd = grid->_ndimension;
const int Nsimd = grid->Nsimd();
@ -402,34 +446,36 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
int ld=grid->_ldimensions[orthogdim];
int rd=grid->_rdimensions[orthogdim];
std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd); // will locally sum vectors first
std::vector<scalar_type > lsSum(ld,scalar_type(0.0)); // sum across these down to scalars
std::vector<iScalar<scalar_type> > extracted(Nsimd); // splitting the SIMD
Vector<vector_type> lvSum(rd); // will locally sum vectors first
Vector<scalar_type > lsSum(ld,scalar_type(0.0)); // sum across these down to scalars
ExtractBuffer<iScalar<scalar_type> > extracted(Nsimd); // splitting the SIMD
result.resize(fd); // And then global sum to return the same vector to every node for IO to file
for(int r=0;r<rd;r++){
lvSum[r]=zero;
lvSum[r]=Zero();
}
int e1= grid->_slice_nblock[orthogdim];
int e2= grid->_slice_block [orthogdim];
int stride=grid->_slice_stride[orthogdim];
parallel_for(int r=0;r<rd;r++){
auto lhv=lhs.View();
auto rhv=rhs.View();
thread_for( r,rd,{
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
vector_type vv = TensorRemove(innerProduct(lhs._odata[ss],rhs._odata[ss]));
vector_type vv = TensorRemove(innerProduct(lhv[ss],rhv[ss]));
lvSum[r]=lvSum[r]+vv;
}
}
}
});
// Sum across simd lanes in the plane, breaking out orthog dir.
std::vector<int> icoor(Nd);
Coordinate icoor(Nd);
for(int rt=0;rt<rd;rt++){
iScalar<vector_type> temp;
@ -470,7 +516,7 @@ static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Ortho
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = rhs._grid->GlobalDimensions()[Orthog];
int Nblock = rhs.Grid()->GlobalDimensions()[Orthog];
std::vector<ComplexD> ip(Nblock);
sn.resize(Nblock);
@ -492,7 +538,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
scalar_type zscale(scale);
GridBase *grid = X._grid;
GridBase *grid = X.Grid();
int Nsimd =grid->Nsimd();
int Nblock =grid->GlobalDimensions()[orthogdim];
@ -505,8 +551,7 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
int e2 =grid->_slice_block [orthogdim];
int stride =grid->_slice_stride[orthogdim];
std::vector<int> icoor;
Coordinate icoor;
for(int r=0;r<rd;r++){
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
@ -522,12 +567,15 @@ static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice
tensor_reduced at; at=av;
parallel_for_nest2(int n=0;n<e1;n++){
auto Rv=R.View();
auto Xv=X.View();
auto Yv=Y.View();
thread_for_collapse(2, n, e1, {
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
R._odata[ss] = at*X._odata[ss]+Y._odata[ss];
Rv[ss] = at*Xv[ss]+Yv[ss];
}
}
});
}
};
@ -559,18 +607,18 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X._grid->GlobalDimensions()[Orthog];
int Nblock = X.Grid()->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *FullGrid = X.Grid();
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
// Lattice<vobj> Xslice(SliceGrid);
// Lattice<vobj> Rslice(SliceGrid);
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
// int nh = FullGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl = nh-1;
// int nl = nh-1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
@ -578,28 +626,31 @@ static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
#pragma omp parallel
{
std::vector<vobj> s_x(Nblock);
#pragma omp for collapse(2)
for(int n=0;n<nblock;n++){
for(int b=0;b<block;b++){
auto X_v=X.View();
auto Y_v=Y.View();
auto R_v=R.View();
thread_region
{
Vector<vobj> s_x(Nblock);
thread_for_collapse_in_region(2, n,nblock, {
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
s_x[i] = X[o+i*ostride];
s_x[i] = X_v[o+i*ostride];
}
vobj dot;
for(int i=0;i<Nblock;i++){
dot = Y[o+i*ostride];
dot = Y_v[o+i*ostride];
for(int j=0;j<Nblock;j++){
dot = dot + s_x[j]*(scale*aa(j,i));
}
R[o+i*ostride]=dot;
R_v[o+i*ostride]=dot;
}
}}
}});
}
};
@ -610,17 +661,17 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X._grid->GlobalDimensions()[Orthog];
int Nblock = X.Grid()->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *FullGrid = X.Grid();
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
// Lattice<vobj> Xslice(SliceGrid);
// Lattice<vobj> Rslice(SliceGrid);
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
// int nh = FullGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl=1;
// int nl=1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
@ -628,17 +679,19 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
int block =FullGrid->_slice_block [Orthog];
int nblock=FullGrid->_slice_nblock[Orthog];
int ostride=FullGrid->_ostride[Orthog];
#pragma omp parallel
auto R_v = R.View();
auto X_v = X.View();
thread_region
{
std::vector<vobj> s_x(Nblock);
#pragma omp for collapse(2)
for(int n=0;n<nblock;n++){
thread_for_collapse_in_region( 2 ,n,nblock,{
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
s_x[i] = X[o+i*ostride];
s_x[i] = X_v[o+i*ostride];
}
vobj dot;
@ -647,11 +700,10 @@ static void sliceMulMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<
for(int j=1;j<Nblock;j++){
dot = dot + s_x[j]*(scale*aa(j,i));
}
R[o+i*ostride]=dot;
R_v[o+i*ostride]=dot;
}
}}
}});
}
};
@ -662,7 +714,7 @@ static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj>
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
GridBase *FullGrid = lhs._grid;
GridBase *FullGrid = lhs.Grid();
// GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
int Nblock = FullGrid->GlobalDimensions()[Orthog];
@ -673,9 +725,9 @@ static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj>
mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
assert( FullGrid->_simd_layout[Orthog]==1);
int nh = FullGrid->_ndimension;
// int nh = FullGrid->_ndimension;
// int nl = SliceGrid->_ndimension;
int nl = nh-1;
// int nl = nh-1;
//FIXME package in a convenient iterator
//Should loop over a plane orthogonal to direction "Orthog"
@ -686,31 +738,33 @@ static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj>
typedef typename vobj::vector_typeD vector_typeD;
#pragma omp parallel
auto lhs_v=lhs.View();
auto rhs_v=rhs.View();
thread_region
{
std::vector<vobj> Left(Nblock);
std::vector<vobj> Right(Nblock);
Eigen::MatrixXcd mat_thread = Eigen::MatrixXcd::Zero(Nblock,Nblock);
#pragma omp for collapse(2)
for(int n=0;n<nblock;n++){
thread_for_collapse_in_region( 2, n,nblock,{
for(int b=0;b<block;b++){
int o = n*stride + b;
for(int i=0;i<Nblock;i++){
Left [i] = lhs[o+i*ostride];
Right[i] = rhs[o+i*ostride];
Left [i] = lhs_v[o+i*ostride];
Right[i] = rhs_v[o+i*ostride];
}
for(int i=0;i<Nblock;i++){
for(int j=0;j<Nblock;j++){
auto tmp = innerProduct(Left[i],Right[j]);
auto rtmp = TensorRemove(tmp);
mat_thread(i,j) += Reduce(rtmp);
auto red = Reduce(rtmp);
mat_thread(i,j) += std::complex<double>(real(red),imag(red));
}}
}}
#pragma omp critical
}});
thread_critical
{
mat += mat_thread;
}
@ -726,8 +780,8 @@ static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj>
return;
}
} /*END NAMESPACE GRID*/
#endif
NAMESPACE_END(Grid);

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

@ -0,0 +1,226 @@
NAMESPACE_BEGIN(Grid);
#define WARP_SIZE 32
extern cudaDeviceProp *gpu_props;
__device__ unsigned int retirementCount = 0;
template <class Iterator>
unsigned int nextPow2(Iterator x) {
--x;
x |= x >> 1;
x |= x >> 2;
x |= x >> 4;
x |= x >> 8;
x |= x >> 16;
return ++x;
}
template <class Iterator>
void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
int device;
cudaGetDevice(&device);
Iterator warpSize = gpu_props[device].warpSize;
Iterator sharedMemPerBlock = gpu_props[device].sharedMemPerBlock;
Iterator maxThreadsPerBlock = gpu_props[device].maxThreadsPerBlock;
Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
std::cout << GridLogDebug << "GPU has:" << std::endl;
std::cout << GridLogDebug << "\twarpSize = " << warpSize << std::endl;
std::cout << GridLogDebug << "\tsharedMemPerBlock = " << sharedMemPerBlock << std::endl;
std::cout << GridLogDebug << "\tmaxThreadsPerBlock = " << maxThreadsPerBlock << std::endl;
std::cout << GridLogDebug << "\tmaxThreadsPerBlock = " << warpSize << std::endl;
std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
if (warpSize != WARP_SIZE) {
std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
exit(EXIT_FAILURE);
}
// let the number of threads in a block be a multiple of 2, starting from warpSize
threads = warpSize;
while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
// keep all the streaming multiprocessors busy
blocks = nextPow2(multiProcessorCount);
}
template <class sobj, class Iterator>
__device__ void reduceBlock(volatile sobj *sdata, sobj mySum, const Iterator tid) {
Iterator blockSize = blockDim.x;
// cannot use overloaded operators for sobj as they are not volatile-qualified
memcpy((void *)&sdata[tid], (void *)&mySum, sizeof(sobj));
__syncwarp();
const Iterator VEC = WARP_SIZE;
const Iterator vid = tid & (VEC-1);
sobj beta, temp;
memcpy((void *)&beta, (void *)&mySum, sizeof(sobj));
for (int i = VEC/2; i > 0; i>>=1) {
if (vid < i) {
memcpy((void *)&temp, (void *)&sdata[tid+i], sizeof(sobj));
beta += temp;
memcpy((void *)&sdata[tid], (void *)&beta, sizeof(sobj));
}
__syncwarp();
}
__syncthreads();
if (threadIdx.x == 0) {
beta = Zero();
for (Iterator i = 0; i < blockSize; i += VEC) {
memcpy((void *)&temp, (void *)&sdata[i], sizeof(sobj));
beta += temp;
}
memcpy((void *)&sdata[0], (void *)&beta, sizeof(sobj));
}
__syncthreads();
}
template <class vobj, class sobj, class Iterator>
__device__ void reduceBlocks(const vobj *g_idata, sobj *g_odata, Iterator n)
{
constexpr Iterator nsimd = vobj::Nsimd();
Iterator blockSize = blockDim.x;
// force shared memory alignment
extern __shared__ __align__(COALESCE_GRANULARITY) unsigned char shmem_pointer[];
// it's not possible to have two extern __shared__ arrays with same name
// but different types in different scopes -- need to cast each time
sobj *sdata = (sobj *)shmem_pointer;
// first level of reduction,
// each thread writes result in mySum
Iterator tid = threadIdx.x;
Iterator i = blockIdx.x*(blockSize*2) + threadIdx.x;
Iterator gridSize = blockSize*2*gridDim.x;
sobj mySum = Zero();
while (i < n) {
Iterator lane = i % nsimd;
Iterator ss = i / nsimd;
auto tmp = extractLane(lane,g_idata[ss]);
sobj tmpD;
tmpD=tmp;
mySum +=tmpD;
if (i + blockSize < n) {
lane = (i+blockSize) % nsimd;
ss = (i+blockSize) / nsimd;
tmp = extractLane(lane,g_idata[ss]);
tmpD = tmp;
mySum += tmpD;
}
i += gridSize;
}
// copy mySum to shared memory and perform
// reduction for all threads in this block
reduceBlock(sdata, mySum, tid);
if (tid == 0) g_odata[blockIdx.x] = sdata[0];
}
template <class vobj, class sobj,class Iterator>
__global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
Iterator blockSize = blockDim.x;
// perform reduction for this block and
// write result to global memory buffer
reduceBlocks(lat, buffer, n);
if (gridDim.x > 1) {
const Iterator tid = threadIdx.x;
__shared__ bool amLast;
// force shared memory alignment
extern __shared__ __align__(COALESCE_GRANULARITY) unsigned char shmem_pointer[];
// it's not possible to have two extern __shared__ arrays with same name
// but different types in different scopes -- need to cast each time
sobj *smem = (sobj *)shmem_pointer;
// wait until all outstanding memory instructions in this thread are finished
__threadfence();
if (tid==0) {
unsigned int ticket = atomicInc(&retirementCount, gridDim.x);
// true if this block is the last block to be done
amLast = (ticket == gridDim.x-1);
}
// each thread must read the correct value of amLast
__syncthreads();
if (amLast) {
// reduce buffer[0], ..., buffer[gridDim.x-1]
Iterator i = tid;
sobj mySum = Zero();
while (i < gridDim.x) {
mySum += buffer[i];
i += blockSize;
}
reduceBlock(smem, mySum, tid);
if (tid==0) {
buffer[0] = smem[0];
// reset count variable
retirementCount = 0;
}
}
}
}
/////////////////////////////////////////////////////////////////////////////////////////////////////////
// Possibly promote to double and sum
/////////////////////////////////////////////////////////////////////////////////////////////////////////
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
{
typedef typename vobj::scalar_objectD sobj;
typedef decltype(lat) Iterator;
Integer nsimd= vobj::Nsimd();
Integer size = osites*nsimd;
Integer numThreads, numBlocks;
getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
Integer smemSize = numThreads * sizeof(sobj);
Vector<sobj> buffer(numBlocks);
sobj *buffer_v = &buffer[0];
reduceKernel<<< numBlocks, numThreads, smemSize >>>(lat, buffer_v, size);
cudaDeviceSynchronize();
cudaError err = cudaGetLastError();
if ( cudaSuccess != err ) {
printf("Cuda error %s\n",cudaGetErrorString( err ));
exit(0);
}
auto result = buffer_v[0];
return result;
}
/////////////////////////////////////////////////////////////////////////////////////////////////////////
// Return as same precision as input performing reduction in double precision though
/////////////////////////////////////////////////////////////////////////////////////////////////////////
template <class vobj>
inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
{
typedef typename vobj::scalar_object sobj;
sobj result;
result = sumD_gpu(lat,osites);
return result;
}
NAMESPACE_END(Grid);

733
Grid/lattice/Lattice_rng.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,8 +24,8 @@
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_RNG_H
#define GRID_LATTICE_RNG_H
@ -41,282 +41,289 @@
#undef RNG_FAST_DISCARD
#endif
namespace Grid {
NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////////
// Allow the RNG state to be less dense than the fine grid
//////////////////////////////////////////////////////////////
inline int RNGfillable(GridBase *coarse,GridBase *fine)
{
//////////////////////////////////////////////////////////////
// Allow the RNG state to be less dense than the fine grid
//////////////////////////////////////////////////////////////
inline int RNGfillable(GridBase *coarse,GridBase *fine)
{
int rngdims = coarse->_ndimension;
int rngdims = coarse->_ndimension;
// trivially extended in higher dims, with locality guaranteeing RNG state is local to node
int lowerdims = fine->_ndimension - coarse->_ndimension;
assert(lowerdims >= 0);
for(int d=0;d<lowerdims;d++){
assert(fine->_simd_layout[d]==1);
assert(fine->_processors[d]==1);
}
int multiplicity=1;
for(int d=0;d<lowerdims;d++){
multiplicity=multiplicity*fine->_rdimensions[d];
}
// local and global volumes subdivide cleanly after SIMDization
for(int d=0;d<rngdims;d++){
int fd= d+lowerdims;
assert(coarse->_processors[d] == fine->_processors[fd]);
assert(coarse->_simd_layout[d] == fine->_simd_layout[fd]);
assert(((fine->_rdimensions[fd] / coarse->_rdimensions[d])* coarse->_rdimensions[d])==fine->_rdimensions[fd]);
multiplicity = multiplicity *fine->_rdimensions[fd] / coarse->_rdimensions[d];
}
return multiplicity;
// trivially extended in higher dims, with locality guaranteeing RNG state is local to node
int lowerdims = fine->_ndimension - coarse->_ndimension;
assert(lowerdims >= 0);
for(int d=0;d<lowerdims;d++){
assert(fine->_simd_layout[d]==1);
assert(fine->_processors[d]==1);
}
int multiplicity=1;
for(int d=0;d<lowerdims;d++){
multiplicity=multiplicity*fine->_rdimensions[d];
}
// local and global volumes subdivide cleanly after SIMDization
for(int d=0;d<rngdims;d++){
int fd= d+lowerdims;
assert(coarse->_processors[d] == fine->_processors[fd]);
assert(coarse->_simd_layout[d] == fine->_simd_layout[fd]);
assert(((fine->_rdimensions[fd] / coarse->_rdimensions[d])* coarse->_rdimensions[d])==fine->_rdimensions[fd]);
multiplicity = multiplicity *fine->_rdimensions[fd] / coarse->_rdimensions[d];
}
return multiplicity;
}
// merge of April 11 2017
// this function is necessary for the LS vectorised field
inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
{
int rngdims = coarse->_ndimension;
// this function is necessary for the LS vectorised field
inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
{
int rngdims = coarse->_ndimension;
// trivially extended in higher dims, with locality guaranteeing RNG state is local to node
int lowerdims = fine->_ndimension - coarse->_ndimension; assert(lowerdims >= 0);
// assumes that the higher dimensions are not using more processors
// all further divisions are local
for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
// trivially extended in higher dims, with locality guaranteeing RNG state is local to node
int lowerdims = fine->_ndimension - coarse->_ndimension; assert(lowerdims >= 0);
// assumes that the higher dimensions are not using more processors
// all further divisions are local
for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
// then divide the number of local sites
// check that the total number of sims agree, meanse the iSites are the same
assert(fine->Nsimd() == coarse->Nsimd());
// then divide the number of local sites
// check that the total number of sims agree, meanse the iSites are the same
assert(fine->Nsimd() == coarse->Nsimd());
// check that the two grids divide cleanly
assert( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
// check that the two grids divide cleanly
assert( (fine->lSites() / coarse->lSites() ) * coarse->lSites() == fine->lSites() );
return fine->lSites() / coarse->lSites();
}
return fine->lSites() / coarse->lSites();
}
// real scalars are one component
template<class scalar,class distribution,class generator>
void fillScalar(scalar &s,distribution &dist,generator & gen)
{
s=dist(gen);
}
template<class distribution,class generator>
void fillScalar(ComplexF &s,distribution &dist, generator &gen)
{
s=ComplexF(dist(gen),dist(gen));
}
template<class distribution,class generator>
void fillScalar(ComplexD &s,distribution &dist,generator &gen)
{
s=ComplexD(dist(gen),dist(gen));
}
// real scalars are one component
template<class scalar,class distribution,class generator>
void fillScalar(scalar &s,distribution &dist,generator & gen)
{
s=dist(gen);
}
template<class distribution,class generator>
void fillScalar(ComplexF &s,distribution &dist, generator &gen)
{
// s=ComplexF(dist(gen),dist(gen));
s.real(dist(gen));
s.imag(dist(gen));
}
template<class distribution,class generator>
void fillScalar(ComplexD &s,distribution &dist,generator &gen)
{
// s=ComplexD(dist(gen),dist(gen));
s.real(dist(gen));
s.imag(dist(gen));
}
class GridRNGbase {
public:
// One generator per site.
// Uniform and Gaussian distributions from these generators.
class GridRNGbase {
public:
// One generator per site.
// Uniform and Gaussian distributions from these generators.
#ifdef RNG_RANLUX
typedef std::ranlux48 RngEngine;
typedef uint64_t RngStateType;
static const int RngStateCount = 15;
typedef std::ranlux48 RngEngine;
typedef uint64_t RngStateType;
static const int RngStateCount = 15;
#endif
#ifdef RNG_MT19937
typedef std::mt19937 RngEngine;
typedef uint32_t RngStateType;
static const int RngStateCount = std::mt19937::state_size;
typedef std::mt19937 RngEngine;
typedef uint32_t RngStateType;
static const int RngStateCount = std::mt19937::state_size;
#endif
#ifdef RNG_SITMO
typedef sitmo::prng_engine RngEngine;
typedef uint64_t RngStateType;
static const int RngStateCount = 13;
typedef sitmo::prng_engine RngEngine;
typedef uint64_t RngStateType;
static const int RngStateCount = 13;
#endif
std::vector<RngEngine> _generators;
std::vector<std::uniform_real_distribution<RealD> > _uniform;
std::vector<std::normal_distribution<RealD> > _gaussian;
std::vector<std::discrete_distribution<int32_t> > _bernoulli;
std::vector<std::uniform_int_distribution<uint32_t> > _uid;
std::vector<RngEngine> _generators;
std::vector<std::uniform_real_distribution<RealD> > _uniform;
std::vector<std::normal_distribution<RealD> > _gaussian;
std::vector<std::discrete_distribution<int32_t> > _bernoulli;
std::vector<std::uniform_int_distribution<uint32_t> > _uid;
///////////////////////
// support for parallel init
///////////////////////
///////////////////////
// support for parallel init
///////////////////////
#ifdef RNG_FAST_DISCARD
static void Skip(RngEngine &eng,uint64_t site)
{
/////////////////////////////////////////////////////////////////////////////////////
// Skip by 2^40 elements between successive lattice sites
// This goes by 10^12.
// Consider quenched updating; likely never exceeding rate of 1000 sweeps
// per second on any machine. This gives us of order 10^9 seconds, or 100 years
// skip ahead.
// For HMC unlikely to go at faster than a solve per second, and
// tens of seconds per trajectory so this is clean in all reasonable cases,
// and margin of safety is orders of magnitude.
// We could hack Sitmo to skip in the higher order words of state if necessary
static void Skip(RngEngine &eng,uint64_t site)
{
/////////////////////////////////////////////////////////////////////////////////////
// Skip by 2^40 elements between successive lattice sites
// This goes by 10^12.
// Consider quenched updating; likely never exceeding rate of 1000 sweeps
// per second on any machine. This gives us of order 10^9 seconds, or 100 years
// skip ahead.
// For HMC unlikely to go at faster than a solve per second, and
// tens of seconds per trajectory so this is clean in all reasonable cases,
// and margin of safety is orders of magnitude.
// We could hack Sitmo to skip in the higher order words of state if necessary
//
// Replace with 2^30 ; avoid problem on large volumes
//
/////////////////////////////////////////////////////////////////////////////////////
// uint64_t skip = site+1; // Old init Skipped then drew. Checked compat with faster init
const int shift = 30;
/////////////////////////////////////////////////////////////////////////////////////
// uint64_t skip = site+1; // Old init Skipped then drew. Checked compat with faster init
const int shift = 30;
uint64_t skip = site;
////////////////////////////////////////////////////////////////////
// Weird compiler bug in Intel 2018.1 under O3 was generating 32bit and not 64 bit left shift.
////////////////////////////////////////////////////////////////////
volatile uint64_t skip = site;
skip = skip<<shift;
skip = skip<<shift;
assert((skip >> shift)==site); // check for overflow
assert((skip >> shift)==site); // check for overflow
eng.discard(skip);
// std::cout << " Engine " <<site << " state " <<eng<<std::endl;
}
eng.discard(skip);
// std::cout << " Engine " <<site << " state " <<eng<<std::endl;
}
#endif
static RngEngine Reseed(RngEngine &eng)
{
std::vector<uint32_t> newseed;
std::uniform_int_distribution<uint32_t> uid;
return Reseed(eng,newseed,uid);
}
static RngEngine Reseed(RngEngine &eng,std::vector<uint32_t> & newseed,
std::uniform_int_distribution<uint32_t> &uid)
{
const int reseeds=4;
static RngEngine Reseed(RngEngine &eng)
{
std::vector<uint32_t> newseed;
std::uniform_int_distribution<uint32_t> uid;
return Reseed(eng,newseed,uid);
}
static RngEngine Reseed(RngEngine &eng,std::vector<uint32_t> & newseed,
std::uniform_int_distribution<uint32_t> &uid)
{
const int reseeds=4;
newseed.resize(reseeds);
for(int i=0;i<reseeds;i++){
newseed[i] = uid(eng);
}
std::seed_seq sseq(newseed.begin(),newseed.end());
return RngEngine(sseq);
}
newseed.resize(reseeds);
for(int i=0;i<reseeds;i++){
newseed[i] = uid(eng);
}
std::seed_seq sseq(newseed.begin(),newseed.end());
return RngEngine(sseq);
}
void GetState(std::vector<RngStateType> & saved,RngEngine &eng) {
saved.resize(RngStateCount);
std::stringstream ss;
ss<<eng;
ss.seekg(0,ss.beg);
for(int i=0;i<RngStateCount;i++){
ss>>saved[i];
}
void GetState(std::vector<RngStateType> & saved,RngEngine &eng) {
saved.resize(RngStateCount);
std::stringstream ss;
ss<<eng;
ss.seekg(0,ss.beg);
for(int i=0;i<RngStateCount;i++){
ss>>saved[i];
}
void GetState(std::vector<RngStateType> & saved,int gen) {
GetState(saved,_generators[gen]);
}
void GetState(std::vector<RngStateType> & saved,int gen) {
GetState(saved,_generators[gen]);
}
void SetState(std::vector<RngStateType> & saved,RngEngine &eng){
assert(saved.size()==RngStateCount);
std::stringstream ss;
for(int i=0;i<RngStateCount;i++){
ss<< saved[i]<<" ";
}
void SetState(std::vector<RngStateType> & saved,RngEngine &eng){
assert(saved.size()==RngStateCount);
std::stringstream ss;
for(int i=0;i<RngStateCount;i++){
ss<< saved[i]<<" ";
}
ss.seekg(0,ss.beg);
ss>>eng;
}
void SetState(std::vector<RngStateType> & saved,int gen){
SetState(saved,_generators[gen]);
}
void SetEngine(RngEngine &Eng, int gen){
_generators[gen]=Eng;
}
void GetEngine(RngEngine &Eng, int gen){
Eng=_generators[gen];
}
template<class source> void Seed(source &src, int gen)
{
_generators[gen] = RngEngine(src);
}
};
ss.seekg(0,ss.beg);
ss>>eng;
}
void SetState(std::vector<RngStateType> & saved,int gen){
SetState(saved,_generators[gen]);
}
void SetEngine(RngEngine &Eng, int gen){
_generators[gen]=Eng;
}
void GetEngine(RngEngine &Eng, int gen){
Eng=_generators[gen];
}
template<class source> void Seed(source &src, int gen)
{
_generators[gen] = RngEngine(src);
}
};
class GridSerialRNG : public GridRNGbase {
public:
class GridSerialRNG : public GridRNGbase {
public:
GridSerialRNG() : GridRNGbase() {
_generators.resize(1);
_uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
_gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
_bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
_uid.resize(1,std::uniform_int_distribution<uint32_t>() );
}
GridSerialRNG() : GridRNGbase() {
_generators.resize(1);
_uniform.resize(1,std::uniform_real_distribution<RealD>{0,1});
_gaussian.resize(1,std::normal_distribution<RealD>(0.0,1.0) );
_bernoulli.resize(1,std::discrete_distribution<int32_t>{1,1});
_uid.resize(1,std::uniform_int_distribution<uint32_t>() );
}
template <class sobj,class distribution> inline void fill(sobj &l,std::vector<distribution> &dist){
template <class sobj,class distribution> inline void fill(sobj &l,std::vector<distribution> &dist){
typedef typename sobj::scalar_type scalar_type;
typedef typename sobj::scalar_type scalar_type;
int words = sizeof(sobj)/sizeof(scalar_type);
int words = sizeof(sobj)/sizeof(scalar_type);
scalar_type *buf = (scalar_type *) & l;
scalar_type *buf = (scalar_type *) & l;
dist[0].reset();
for(int idx=0;idx<words;idx++){
fillScalar(buf[idx],dist[0],_generators[0]);
}
dist[0].reset();
for(int idx=0;idx<words;idx++){
fillScalar(buf[idx],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
};
}
template <class distribution> inline void fill(ComplexF &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
template <class distribution> inline void fill(ComplexF &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(ComplexD &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(RealF &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(RealD &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
// vector fill
template <class distribution> inline void fill(vComplexF &l,std::vector<distribution> &dist){
RealF *pointer=(RealF *)&l;
dist[0].reset();
for(int i=0;i<2*vComplexF::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
template <class distribution> inline void fill(ComplexD &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vComplexD &l,std::vector<distribution> &dist){
RealD *pointer=(RealD *)&l;
dist[0].reset();
for(int i=0;i<2*vComplexD::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
template <class distribution> inline void fill(RealF &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vRealF &l,std::vector<distribution> &dist){
RealF *pointer=(RealF *)&l;
dist[0].reset();
for(int i=0;i<vRealF::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
template <class distribution> inline void fill(RealD &l,std::vector<distribution> &dist){
dist[0].reset();
fillScalar(l,dist[0],_generators[0]);
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
// vector fill
template <class distribution> inline void fill(vComplexF &l,std::vector<distribution> &dist){
RealF *pointer=(RealF *)&l;
dist[0].reset();
for(int i=0;i<2*vComplexF::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vComplexD &l,std::vector<distribution> &dist){
RealD *pointer=(RealD *)&l;
dist[0].reset();
for(int i=0;i<2*vComplexD::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vRealF &l,std::vector<distribution> &dist){
RealF *pointer=(RealF *)&l;
dist[0].reset();
for(int i=0;i<vRealF::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vRealD &l,std::vector<distribution> &dist){
RealD *pointer=(RealD *)&l;
dist[0].reset();
for(int i=0;i<vRealD::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
template <class distribution> inline void fill(vRealD &l,std::vector<distribution> &dist){
RealD *pointer=(RealD *)&l;
dist[0].reset();
for(int i=0;i<vRealD::Nsimd();i++){
fillScalar(pointer[i],dist[0],_generators[0]);
}
CartesianCommunicator::BroadcastWorld(0,(void *)&l,sizeof(l));
}
void SeedFixedIntegers(const std::vector<int> &seeds){
CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
std::seed_seq src(seeds.begin(),seeds.end());
Seed(src,0);
}
void SeedFixedIntegers(const std::vector<int> &seeds){
CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
std::seed_seq src(seeds.begin(),seeds.end());
Seed(src,0);
}
void SeedUniqueString(const std::string &s){
std::vector<int> seeds;
@ -330,65 +337,67 @@ namespace Grid {
std::cout << GridLogMessage << "Seed SHA256: " << sha.str() << std::endl;
SeedFixedIntegers(seeds);
}
};
};
class GridParallelRNG : public GridRNGbase {
class GridParallelRNG : public GridRNGbase {
private:
double _time_counter;
GridBase *_grid;
unsigned int _vol;
double _time_counter;
public:
GridBase *Grid(void) const { return _grid; }
int generator_idx(int os,int is) {
return is*_grid->oSites()+os;
}
public:
GridBase *_grid;
unsigned int _vol;
GridParallelRNG(GridBase *grid) : GridRNGbase() {
_grid = grid;
_vol =_grid->iSites()*_grid->oSites();
int generator_idx(int os,int is) {
return is*_grid->oSites()+os;
}
_generators.resize(_vol);
_uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
_gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
_bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
_uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
}
GridParallelRNG(GridBase *grid) : GridRNGbase() {
_grid = grid;
_vol =_grid->iSites()*_grid->oSites();
template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist){
_generators.resize(_vol);
_uniform.resize(_vol,std::uniform_real_distribution<RealD>{0,1});
_gaussian.resize(_vol,std::normal_distribution<RealD>(0.0,1.0) );
_bernoulli.resize(_vol,std::discrete_distribution<int32_t>{1,1});
_uid.resize(_vol,std::uniform_int_distribution<uint32_t>() );
}
typedef typename vobj::scalar_object scalar_object;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
template <class vobj,class distribution> inline void fill(Lattice<vobj> &l,std::vector<distribution> &dist){
double inner_time_counter = usecond();
typedef typename vobj::scalar_object scalar_object;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int multiplicity = RNGfillable_general(_grid, l.Grid()); // l has finer or same grid
int Nsimd = _grid->Nsimd(); // guaranteed to be the same for l.Grid() too
int osites = _grid->oSites(); // guaranteed to be <= l.Grid()->oSites() by a factor multiplicity
int words = sizeof(scalar_object) / sizeof(scalar_type);
double inner_time_counter = usecond();
auto l_v = l.View();
thread_for( ss, osites, {
ExtractBuffer<scalar_object> buf(Nsimd);
for (int m = 0; m < multiplicity; m++) { // Draw from same generator multiplicity times
int multiplicity = RNGfillable_general(_grid, l._grid); // l has finer or same grid
int Nsimd = _grid->Nsimd(); // guaranteed to be the same for l._grid too
int osites = _grid->oSites(); // guaranteed to be <= l._grid->oSites() by a factor multiplicity
int words = sizeof(scalar_object) / sizeof(scalar_type);
int sm = multiplicity * ss + m; // Maps the generator site to the fine site
parallel_for(int ss=0;ss<osites;ss++){
std::vector<scalar_object> buf(Nsimd);
for (int m = 0; m < multiplicity; m++) { // Draw from same generator multiplicity times
int sm = multiplicity * ss + m; // Maps the generator site to the fine site
for (int si = 0; si < Nsimd; si++) {
for (int si = 0; si < Nsimd; si++) {
int gdx = generator_idx(ss, si); // index of generator state
scalar_type *pointer = (scalar_type *)&buf[si];
dist[gdx].reset();
for (int idx = 0; idx < words; idx++)
fillScalar(pointer[idx], dist[gdx], _generators[gdx]);
}
// merge into SIMD lanes, FIXME suboptimal implementation
merge(l._odata[sm], buf);
}
int gdx = generator_idx(ss, si); // index of generator state
scalar_type *pointer = (scalar_type *)&buf[si];
dist[gdx].reset();
for (int idx = 0; idx < words; idx++)
fillScalar(pointer[idx], dist[gdx], _generators[gdx]);
}
// merge into SIMD lanes, FIXME suboptimal implementation
merge(l_v[sm], buf);
}
});
// });
_time_counter += usecond()- inner_time_counter;
};
_time_counter += usecond()- inner_time_counter;
}
void SeedUniqueString(const std::string &s){
std::vector<int> seeds;
@ -398,119 +407,119 @@ namespace Grid {
std::cout << GridLogMessage << "Seed SHA256: " << GridChecksum::sha256_string(seeds) << std::endl;
SeedFixedIntegers(seeds);
}
void SeedFixedIntegers(const std::vector<int> &seeds){
void SeedFixedIntegers(const std::vector<int> &seeds){
// Everyone generates the same seed_seq based on input seeds
CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
// Everyone generates the same seed_seq based on input seeds
CartesianCommunicator::BroadcastWorld(0,(void *)&seeds[0],sizeof(int)*seeds.size());
std::seed_seq source(seeds.begin(),seeds.end());
std::seed_seq source(seeds.begin(),seeds.end());
RngEngine master_engine(source);
RngEngine master_engine(source);
#ifdef RNG_FAST_DISCARD
////////////////////////////////////////////////
// Skip ahead through a single stream.
// Applicable to SITMO and other has based/crypto RNGs
// Should be applicable to Mersenne Twister, but the C++11
// MT implementation does not implement fast discard even though
// in principle this is possible
////////////////////////////////////////////////
// Everybody loops over global volume.
parallel_for(int gidx=0;gidx<_grid->_gsites;gidx++){
////////////////////////////////////////////////
// Skip ahead through a single stream.
// Applicable to SITMO and other has based/crypto RNGs
// Should be applicable to Mersenne Twister, but the C++11
// MT implementation does not implement fast discard even though
// in principle this is possible
////////////////////////////////////////////////
// Everybody loops over global volume.
thread_for( gidx, _grid->_gsites, {
// Where is it?
int rank,o_idx,i_idx;
std::vector<int> gcoor;
int rank;
int o_idx;
int i_idx;
Coordinate gcoor;
_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
// If this is one of mine we take it
if( rank == _grid->ThisRank() ){
int l_idx=generator_idx(o_idx,i_idx);
_generators[l_idx] = master_engine;
Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
}
}
});
#else
////////////////////////////////////////////////////////////////
// Machine and thread decomposition dependent seeding is efficient
// and maximally parallel; but NOT reproducible from machine to machine.
// Not ideal, but fastest way to reseed all nodes.
////////////////////////////////////////////////////////////////
{
// Obtain one Reseed per processor
int Nproc = _grid->ProcessorCount();
std::vector<RngEngine> seeders(Nproc);
int me= _grid->ThisRank();
for(int p=0;p<Nproc;p++){
seeders[p] = Reseed(master_engine);
}
master_engine = seeders[me];
////////////////////////////////////////////////////////////////
// Machine and thread decomposition dependent seeding is efficient
// and maximally parallel; but NOT reproducible from machine to machine.
// Not ideal, but fastest way to reseed all nodes.
////////////////////////////////////////////////////////////////
{
// Obtain one Reseed per processor
int Nproc = _grid->ProcessorCount();
std::vector<RngEngine> seeders(Nproc);
int me= _grid->ThisRank();
for(int p=0;p<Nproc;p++){
seeders[p] = Reseed(master_engine);
}
master_engine = seeders[me];
}
{
// Obtain one reseeded generator per thread
int Nthread = GridThread::GetThreads();
std::vector<RngEngine> seeders(Nthread);
for(int t=0;t<Nthread;t++){
seeders[t] = Reseed(master_engine);
}
{
// Obtain one reseeded generator per thread
int Nthread = GridThread::GetThreads();
std::vector<RngEngine> seeders(Nthread);
for(int t=0;t<Nthread;t++){
seeders[t] = Reseed(master_engine);
}
parallel_for(int t=0;t<Nthread;t++) {
// set up one per local site in threaded fashion
std::vector<uint32_t> newseeds;
std::uniform_int_distribution<uint32_t> uid;
for(int l=0;l<_grid->lSites();l++) {
if ( (l%Nthread)==t ) {
_generators[l] = Reseed(seeders[t],newseeds,uid);
}
thread_for( t, Nthread, {
// set up one per local site in threaded fashion
std::vector<uint32_t> newseeds;
std::uniform_int_distribution<uint32_t> uid;
for(int l=0;l<_grid->lSites();l++) {
if ( (l%Nthread)==t ) {
_generators[l] = Reseed(seeders[t],newseeds,uid);
}
}
}
});
}
#endif
}
void Report(){
std::cout << GridLogMessage << "Time spent in the fill() routine by GridParallelRNG: "<< _time_counter/1e3 << " ms" << std::endl;
}
////////////////////////////////////////////////////////////////////////
// Support for rigorous test of RNG's
// Return uniform random uint32_t from requested site generator
////////////////////////////////////////////////////////////////////////
uint32_t GlobalU01(int gsite){
uint32_t the_number;
// who
int rank,o_idx,i_idx;
Coordinate gcoor;
_grid->GlobalIndexToGlobalCoor(gsite,gcoor);
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
// draw
int l_idx=generator_idx(o_idx,i_idx);
if( rank == _grid->ThisRank() ){
the_number = _uid[l_idx](_generators[l_idx]);
}
void Report(){
std::cout << GridLogMessage << "Time spent in the fill() routine by GridParallelRNG: "<< _time_counter/1e3 << " ms" << std::endl;
}
////////////////////////////////////////////////////////////////////////
// Support for rigorous test of RNG's
// Return uniform random uint32_t from requested site generator
////////////////////////////////////////////////////////////////////////
uint32_t GlobalU01(int gsite){
uint32_t the_number;
// who
std::vector<int> gcoor;
int rank,o_idx,i_idx;
_grid->GlobalIndexToGlobalCoor(gsite,gcoor);
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
// draw
int l_idx=generator_idx(o_idx,i_idx);
if( rank == _grid->ThisRank() ){
the_number = _uid[l_idx](_generators[l_idx]);
}
// share & return
_grid->Broadcast(rank,(void *)&the_number,sizeof(the_number));
return the_number;
}
// share & return
_grid->Broadcast(rank,(void *)&the_number,sizeof(the_number));
return the_number;
}
};
};
template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._uniform); }
template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._gaussian); }
template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
template <class vobj> inline void random(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._uniform); }
template <class vobj> inline void gaussian(GridParallelRNG &rng,Lattice<vobj> &l) { rng.fill(l,rng._gaussian); }
template <class vobj> inline void bernoulli(GridParallelRNG &rng,Lattice<vobj> &l){ rng.fill(l,rng._bernoulli);}
template <class sobj> inline void random(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._uniform ); }
template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
template <class sobj> inline void random(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._uniform ); }
template <class sobj> inline void gaussian(GridSerialRNG &rng,sobj &l) { rng.fill(l,rng._gaussian ); }
template <class sobj> inline void bernoulli(GridSerialRNG &rng,sobj &l){ rng.fill(l,rng._bernoulli); }
}
NAMESPACE_END(Grid);
#endif

64
Grid/lattice/Lattice_trace.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,8 +23,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_TRACE_H
#define GRID_LATTICE_TRACE_H
@ -32,36 +32,38 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
// Tracing, transposing, peeking, poking
///////////////////////////////////////////////
namespace Grid {
NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////////////////////////////////////////////////
// Trace
////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
inline auto trace(const Lattice<vobj> &lhs)
-> Lattice<decltype(trace(lhs._odata[0]))>
{
Lattice<decltype(trace(lhs._odata[0]))> ret(lhs._grid);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
ret._odata[ss] = trace(lhs._odata[ss]);
}
return ret;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
// Trace
////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
inline auto trace(const Lattice<vobj> &lhs) -> Lattice<decltype(trace(vobj()))>
{
Lattice<decltype(trace(vobj()))> ret(lhs.Grid());
auto ret_v = ret.View();
auto lhs_v = lhs.View();
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
coalescedWrite(ret_v[ss], trace(lhs_v(ss)));
});
return ret;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
// Trace Index level dependent operation
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Index,class vobj>
inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(lhs._odata[0]))>
{
Lattice<decltype(traceIndex<Index>(lhs._odata[0]))> ret(lhs._grid);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
ret._odata[ss] = traceIndex<Index>(lhs._odata[ss]);
}
return ret;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
// Trace Index level dependent operation
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Index,class vobj>
inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<Index>(vobj()))>
{
Lattice<decltype(traceIndex<Index>(vobj()))> ret(lhs.Grid());
auto ret_v = ret.View();
auto lhs_v = lhs.View();
accelerator_for( ss, lhs_v.size(), vobj::Nsimd(), {
coalescedWrite(ret_v[ss], traceIndex<Index>(lhs_v(ss)));
});
return ret;
};
}
NAMESPACE_END(Grid);
#endif

503
Grid/lattice/Lattice_transfer.h
View File

File diff suppressed because it is too large Load Diff

61
Grid/lattice/Lattice_transpose.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,8 +24,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_TRANSPOSE_H
#define GRID_LATTICE_TRANSPOSE_H
@ -33,31 +33,36 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
// Transpose
///////////////////////////////////////////////
namespace Grid {
NAMESPACE_BEGIN(Grid);
////////////////////////////////////////////////////////////////////////////////////////////////////
// Transpose
////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs._grid);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
ret._odata[ss] = transpose(lhs._odata[ss]);
}
return ret;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
// Transpose
////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
inline Lattice<vobj> transpose(const Lattice<vobj> &lhs){
Lattice<vobj> ret(lhs.Grid());
auto ret_v = ret.View();
auto lhs_v = lhs.View();
accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss], transpose(lhs_v(ss)));
});
return ret;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
// Index level dependent transpose
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Index,class vobj>
inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(lhs._odata[0]))>
{
Lattice<decltype(transposeIndex<Index>(lhs._odata[0]))> ret(lhs._grid);
parallel_for(int ss=0;ss<lhs._grid->oSites();ss++){
ret._odata[ss] = transposeIndex<Index>(lhs._odata[ss]);
}
return ret;
};
}
////////////////////////////////////////////////////////////////////////////////////////////////////
// Index level dependent transpose
////////////////////////////////////////////////////////////////////////////////////////////////////
template<int Index,class vobj>
inline auto TransposeIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(transposeIndex<Index>(vobj()))>
{
Lattice<decltype(transposeIndex<Index>(vobj()))> ret(lhs.Grid());
auto ret_v = ret.View();
auto lhs_v = lhs.View();
accelerator_for(ss,lhs_v.size(),vobj::Nsimd(),{
coalescedWrite(ret_v[ss] , transposeIndex<Index>(lhs_v(ss)));
});
return ret;
};
NAMESPACE_END(Grid);
#endif

98
Grid/lattice/Lattice_unary.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -26,59 +26,55 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LATTICE_UNARY_H
#define GRID_LATTICE_UNARY_H
namespace Grid {
template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs,RealD y){
Lattice<obj> ret(rhs._grid);
ret.checkerboard = rhs.checkerboard;
conformable(ret,rhs);
parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
ret._odata[ss]=pow(rhs._odata[ss],y);
}
return ret;
}
template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs,Integer y){
Lattice<obj> ret(rhs._grid);
ret.checkerboard = rhs.checkerboard;
conformable(ret,rhs);
parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
ret._odata[ss]=mod(rhs._odata[ss],y);
}
return ret;
}
template<class obj> Lattice<obj> div(const Lattice<obj> &rhs,Integer y){
Lattice<obj> ret(rhs._grid);
ret.checkerboard = rhs.checkerboard;
conformable(ret,rhs);
parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
ret._odata[ss]=div(rhs._odata[ss],y);
}
return ret;
}
template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
Lattice<obj> ret(rhs._grid);
ret.checkerboard = rhs.checkerboard;
conformable(ret,rhs);
parallel_for(int ss=0;ss<rhs._grid->oSites();ss++){
ret._odata[ss]=Exponentiate(rhs._odata[ss],alpha, Nexp);
}
return ret;
}
NAMESPACE_BEGIN(Grid);
template<class obj> Lattice<obj> pow(const Lattice<obj> &rhs_i,RealD y){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),1,{
ret[ss]=pow(rhs[ss],y);
});
return ret_i;
}
template<class obj> Lattice<obj> mod(const Lattice<obj> &rhs_i,Integer y){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],mod(rhs(ss),y));
});
return ret_i;
}
template<class obj> Lattice<obj> div(const Lattice<obj> &rhs_i,Integer y){
Lattice<obj> ret_i(rhs_i.Grid());
auto ret = ret_i.View();
auto rhs = rhs_i.View();
ret.Checkerboard() = rhs_i.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],div(rhs(ss),y));
});
return ret_i;
}
template<class obj> Lattice<obj> expMat(const Lattice<obj> &rhs_i, RealD alpha, Integer Nexp = DEFAULT_MAT_EXP){
Lattice<obj> ret_i(rhs_i.Grid());
auto rhs = rhs_i.View();
auto ret = ret_i.View();
ret.Checkerboard() = rhs.Checkerboard();
accelerator_for(ss,rhs.size(),obj::Nsimd(),{
coalescedWrite(ret[ss],Exponentiate(rhs(ss),alpha, Nexp));
});
return ret_i;
}
NAMESPACE_END(Grid);
#endif

27
Grid/log/Log.cc
View File

@ -28,27 +28,27 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
/* END LEGAL */
#include <Grid/GridCore.h>
#include <Grid/util/CompilerCompatible.h>
#include <cxxabi.h>
#include <memory>
namespace Grid {
NAMESPACE_BEGIN(Grid);
std::string demangle(const char* name) {
std::string demangle(const char* name) {
int status = -4; // some arbitrary value to eliminate the compiler warning
int status = -4; // some arbitrary value to eliminate the compiler warning
// enable c++11 by passing the flag -std=c++11 to g++
std::unique_ptr<char, void(*)(void*)> res {
abi::__cxa_demangle(name, NULL, NULL, &status),
std::free
};
// enable c++11 by passing the flag -std=c++11 to g++
std::unique_ptr<char, void(*)(void*)> res {
abi::__cxa_demangle(name, NULL, NULL, &status),
std::free
};
return (status==0) ? res.get() : name ;
}
return (status==0) ? res.get() : name ;
}
GridStopWatch Logger::GlobalStopWatch;
int Logger::timestamp;
@ -109,8 +109,9 @@ void Grid_quiesce_nodes(void) {
}
void Grid_unquiesce_nodes(void) {
#ifdef GRID_COMMS_MPI
#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT)
std::cout.clear();
#endif
}
}
NAMESPACE_END(Grid);

89
Grid/log/Log.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -25,8 +25,8 @@
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 */
*************************************************************************************/
/* END LEGAL */
#include <map>
@ -37,13 +37,12 @@
#include <execinfo.h>
#endif
namespace Grid {
NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////////////////////////////////////////////
// Dress the output; use std::chrono for time stamping via the StopWatch class
//////////////////////////////////////////////////////////////////////////////////////////////////
class Colours{
protected:
bool is_active;
@ -57,15 +56,15 @@ public:
void Active(bool activate){
is_active=activate;
if (is_active){
colour["BLACK"] ="\033[30m";
colour["RED"] ="\033[31m";
colour["GREEN"] ="\033[32m";
colour["YELLOW"] ="\033[33m";
colour["BLUE"] ="\033[34m";
colour["PURPLE"] ="\033[35m";
colour["CYAN"] ="\033[36m";
colour["WHITE"] ="\033[37m";
colour["NORMAL"] ="\033[0;39m";
colour["BLACK"] ="\033[30m";
colour["RED"] ="\033[31m";
colour["GREEN"] ="\033[32m";
colour["YELLOW"] ="\033[33m";
colour["BLUE"] ="\033[34m";
colour["PURPLE"] ="\033[35m";
colour["CYAN"] ="\033[36m";
colour["WHITE"] ="\033[37m";
colour["NORMAL"] ="\033[0;39m";
} else {
colour["BLACK"] ="";
colour["RED"] ="";
@ -102,14 +101,14 @@ public:
std::string colour() {return Painter.colour[COLOUR];}
Logger(std::string topNm, int on, std::string nm, Colours& col_class, std::string col) : active(on),
name(nm),
topName(topNm),
Painter(col_class),
timing_mode(0),
COLOUR(col)
{
StopWatch = & GlobalStopWatch;
};
name(nm),
topName(topNm),
Painter(col_class),
timing_mode(0),
COLOUR(col)
{
StopWatch = & GlobalStopWatch;
};
void Active(int on) {active = on;};
int isActive(void) {return active;};
@ -164,7 +163,7 @@ public:
class GridLogger: public Logger {
public:
GridLogger(int on, std::string nm, Colours&col_class, std::string col_key = "NORMAL"):
Logger("Grid", on, nm, col_class, col_key){};
Logger("Grid", on, nm, col_class, col_key){};
};
void GridLogConfigure(std::vector<std::string> &logstreams);
@ -181,39 +180,39 @@ extern GridLogger GridLogIterative ;
extern GridLogger GridLogIntegrator ;
extern Colours GridLogColours;
std::string demangle(const char* name) ;
std::string demangle(const char* name) ;
#define _NBACKTRACE (256)
extern void * Grid_backtrace_buffer[_NBACKTRACE];
#define BACKTRACEFILE() {\
char string[20]; \
std::sprintf(string,"backtrace.%d",CartesianCommunicator::RankWorld()); \
std::FILE * fp = std::fopen(string,"w"); \
BACKTRACEFP(fp)\
std::fclose(fp); \
}
#define BACKTRACEFILE() { \
char string[20]; \
std::sprintf(string,"backtrace.%d",CartesianCommunicator::RankWorld()); \
std::FILE * fp = std::fopen(string,"w"); \
BACKTRACEFP(fp) \
std::fclose(fp); \
}
#ifdef HAVE_EXECINFO_H
#define BACKTRACEFP(fp) { \
int symbols = backtrace (Grid_backtrace_buffer,_NBACKTRACE);\
char **strings = backtrace_symbols(Grid_backtrace_buffer,symbols);\
for (int i = 0; i < symbols; i++){\
std::fprintf (fp,"BackTrace Strings: %d %s\n",i, demangle(strings[i]).c_str()); std::fflush(fp); \
}\
}
#define BACKTRACEFP(fp) { \
int symbols = backtrace (Grid_backtrace_buffer,_NBACKTRACE); \
char **strings = backtrace_symbols(Grid_backtrace_buffer,symbols); \
for (int i = 0; i < symbols; i++){ \
std::fprintf (fp,"BackTrace Strings: %d %s\n",i, demangle(strings[i]).c_str()); std::fflush(fp); \
} \
}
#else
#define BACKTRACEFP(fp) { \
std::fprintf (fp,"BT %d %lx\n",0, __builtin_return_address(0)); std::fflush(fp); \
std::fprintf (fp,"BT %d %lx\n",1, __builtin_return_address(1)); std::fflush(fp); \
std::fprintf (fp,"BT %d %lx\n",2, __builtin_return_address(2)); std::fflush(fp); \
std::fprintf (fp,"BT %d %lx\n",3, __builtin_return_address(3)); std::fflush(fp); \
}
#define BACKTRACEFP(fp) { \
std::fprintf (fp,"BT %d %lx\n",0, __builtin_return_address(0)); std::fflush(fp); \
std::fprintf (fp,"BT %d %lx\n",1, __builtin_return_address(1)); std::fflush(fp); \
std::fprintf (fp,"BT %d %lx\n",2, __builtin_return_address(2)); std::fflush(fp); \
std::fprintf (fp,"BT %d %lx\n",3, __builtin_return_address(3)); std::fflush(fp); \
}
#endif
#define BACKTRACE() BACKTRACEFP(stdout)
NAMESPACE_END(Grid);
}
#endif

107
Grid/parallelIO/BinaryIO.h
View File

@ -26,8 +26,7 @@
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_BINARY_IO_H
#define GRID_BINARY_IO_H
#pragma once
#if defined(GRID_COMMS_MPI) || defined(GRID_COMMS_MPI3) || defined(GRID_COMMS_MPIT)
#define USE_MPI_IO
@ -42,8 +41,7 @@
#include <arpa/inet.h>
#include <algorithm>
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////
// Byte reversal garbage
@ -91,7 +89,7 @@ class BinaryIO {
{
typedef typename vobj::scalar_object sobj;
GridBase *grid = lat._grid;
GridBase *grid = lat.Grid();
uint64_t lsites = grid->lSites();
std::vector<sobj> scalardata(lsites);
@ -111,21 +109,20 @@ class BinaryIO {
lsites = 1;
}
PARALLEL_REGION
thread_region
{
uint32_t nersc_csum_thr = 0;
PARALLEL_FOR_LOOP_INTERN
for (uint64_t local_site = 0; local_site < lsites; local_site++)
thread_for_in_region( local_site, lsites,
{
uint32_t *site_buf = (uint32_t *)&fbuf[local_site];
for (uint64_t j = 0; j < size32; j++)
{
nersc_csum_thr = nersc_csum_thr + site_buf[j];
}
}
});
PARALLEL_CRITICAL
thread_critical
{
nersc_csum += nersc_csum_thr;
}
@ -134,28 +131,25 @@ PARALLEL_CRITICAL
template<class fobj> static inline void ScidacChecksum(GridBase *grid,std::vector<fobj> &fbuf,uint32_t &scidac_csuma,uint32_t &scidac_csumb)
{
const uint64_t size32 = sizeof(fobj)/sizeof(uint32_t);
int nd = grid->_ndimension;
uint64_t lsites =grid->lSites();
if (fbuf.size()==1) {
lsites=1;
}
std::vector<int> local_vol =grid->LocalDimensions();
std::vector<int> local_start =grid->LocalStarts();
std::vector<int> global_vol =grid->FullDimensions();
Coordinate local_vol =grid->LocalDimensions();
Coordinate local_start =grid->LocalStarts();
Coordinate global_vol =grid->FullDimensions();
PARALLEL_REGION
thread_region
{
std::vector<int> coor(nd);
Coordinate coor(nd);
uint32_t scidac_csuma_thr=0;
uint32_t scidac_csumb_thr=0;
uint32_t site_crc=0;
PARALLEL_FOR_LOOP_INTERN
for(uint64_t local_site=0;local_site<lsites;local_site++){
thread_for_in_region( local_site, lsites,
{
uint32_t * site_buf = (uint32_t *)&fbuf[local_site];
@ -182,9 +176,9 @@ PARALLEL_FOR_LOOP_INTERN
// std::cout << "Site "<<local_site << std::hex<<site_buf[0] <<site_buf[1]<<std::dec <<std::endl;
scidac_csuma_thr ^= site_crc<<gsite29 | site_crc>>(32-gsite29);
scidac_csumb_thr ^= site_crc<<gsite31 | site_crc>>(32-gsite31);
}
});
PARALLEL_CRITICAL
thread_critical
{
scidac_csuma^= scidac_csuma_thr;
scidac_csumb^= scidac_csumb_thr;
@ -202,9 +196,9 @@ PARALLEL_CRITICAL
{
uint32_t * f = (uint32_t *)file_object;
uint64_t count = bytes/sizeof(uint32_t);
parallel_for(uint64_t i=0;i<count;i++){
thread_for( i, count, {
f[i] = ntohl(f[i]);
}
});
}
// LE must Swap and switch to host
static inline void le32toh_v(void *file_object,uint64_t bytes)
@ -212,13 +206,13 @@ PARALLEL_CRITICAL
uint32_t *fp = (uint32_t *)file_object;
uint64_t count = bytes/sizeof(uint32_t);
parallel_for(uint64_t i=0;i<count;i++){
thread_for(i,count,{
uint32_t f;
f = fp[i];
// got network order and the network to host
f = ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ;
fp[i] = ntohl(f);
}
});
}
// BE is same as network
@ -226,9 +220,9 @@ PARALLEL_CRITICAL
{
uint64_t * f = (uint64_t *)file_object;
uint64_t count = bytes/sizeof(uint64_t);
parallel_for(uint64_t i=0;i<count;i++){
thread_for( i, count, {
f[i] = Grid_ntohll(f[i]);
}
});
}
// LE must swap and switch;
@ -236,7 +230,7 @@ PARALLEL_CRITICAL
{
uint64_t *fp = (uint64_t *)file_object;
uint64_t count = bytes/sizeof(uint64_t);
parallel_for(uint64_t i=0;i<count;i++){
thread_for( i, count, {
uint64_t f,g;
f = fp[i];
// got network order and the network to host
@ -245,7 +239,7 @@ PARALLEL_CRITICAL
f = f >> 32;
g|= ((f&0xFF)<<24) | ((f&0xFF00)<<8) | ((f&0xFF0000)>>8) | ((f&0xFF000000UL)>>24) ;
fp[i] = Grid_ntohll(g);
}
});
}
/////////////////////////////////////////////////////////////////////////////
// Real action:
@ -281,13 +275,13 @@ PARALLEL_CRITICAL
int nrank = grid->ProcessorCount();
int myrank = grid->ThisRank();
std::vector<int> psizes = grid->ProcessorGrid();
std::vector<int> pcoor = grid->ThisProcessorCoor();
std::vector<int> gLattice= grid->GlobalDimensions();
std::vector<int> lLattice= grid->LocalDimensions();
Coordinate psizes = grid->ProcessorGrid();
Coordinate pcoor = grid->ThisProcessorCoor();
Coordinate gLattice= grid->GlobalDimensions();
Coordinate lLattice= grid->LocalDimensions();
std::vector<int> lStart(ndim);
std::vector<int> gStart(ndim);
Coordinate lStart(ndim);
Coordinate gStart(ndim);
// Flatten the file
uint64_t lsites = grid->lSites();
@ -546,7 +540,7 @@ PARALLEL_CRITICAL
typedef typename vobj::scalar_object sobj;
typedef typename vobj::Realified::scalar_type word; word w=0;
GridBase *grid = Umu._grid;
GridBase *grid = Umu.Grid();
uint64_t lsites = grid->lSites();
std::vector<sobj> scalardata(lsites);
@ -558,7 +552,7 @@ PARALLEL_CRITICAL
GridStopWatch timer;
timer.Start();
parallel_for(uint64_t x=0;x<lsites;x++) munge(iodata[x], scalardata[x]);
thread_for(x,lsites, { munge(iodata[x], scalardata[x]); });
vectorizeFromLexOrdArray(scalardata,Umu);
grid->Barrier();
@ -582,7 +576,7 @@ PARALLEL_CRITICAL
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::Realified::scalar_type word; word w=0;
GridBase *grid = Umu._grid;
GridBase *grid = Umu.Grid();
uint64_t lsites = grid->lSites(), offsetCopy = offset;
int attemptsLeft = std::max(0, BinaryIO::latticeWriteMaxRetry);
bool checkWrite = (BinaryIO::latticeWriteMaxRetry >= 0);
@ -596,7 +590,7 @@ PARALLEL_CRITICAL
GridStopWatch timer; timer.Start();
unvectorizeToLexOrdArray(scalardata,Umu);
parallel_for(uint64_t x=0;x<lsites;x++) munge(scalardata[x],iodata[x]);
thread_for(x, lsites, { munge(scalardata[x],iodata[x]); });
grid->Barrier();
timer.Stop();
@ -619,7 +613,7 @@ PARALLEL_CRITICAL
{
std::cout << GridLogMessage << "writeLatticeObject: read test checksum failure, re-writing (" << attemptsLeft << " attempt(s) remaining)" << std::endl;
offset = offsetCopy;
parallel_for(uint64_t x=0;x<lsites;x++) munge(scalardata[x],iodata[x]);
thread_for(x,lsites, { munge(scalardata[x],iodata[x]); });
}
else
{
@ -637,8 +631,8 @@ PARALLEL_CRITICAL
/////////////////////////////////////////////////////////////////////////////
// Read a RNG; use IOobject and lexico map to an array of state
//////////////////////////////////////////////////////////////////////////////////////
static inline void readRNG(GridSerialRNG &serial,
GridParallelRNG &parallel,
static inline void readRNG(GridSerialRNG &serial_rng,
GridParallelRNG &parallel_rng,
std::string file,
uint64_t offset,
uint32_t &nersc_csum,
@ -652,7 +646,7 @@ PARALLEL_CRITICAL
std::string format = "IEEE32BIG";
GridBase *grid = parallel._grid;
GridBase *grid = parallel_rng.Grid();
uint64_t gsites = grid->gSites();
uint64_t lsites = grid->lSites();
@ -669,11 +663,11 @@ PARALLEL_CRITICAL
nersc_csum,scidac_csuma,scidac_csumb);
timer.Start();
parallel_for(uint64_t lidx=0;lidx<lsites;lidx++){
thread_for(lidx,lsites,{
std::vector<RngStateType> tmp(RngStateCount);
std::copy(iodata[lidx].begin(),iodata[lidx].end(),tmp.begin());
parallel.SetState(tmp,lidx);
}
parallel_rng.SetState(tmp,lidx);
});
timer.Stop();
iodata.resize(1);
@ -683,7 +677,7 @@ PARALLEL_CRITICAL
{
std::vector<RngStateType> tmp(RngStateCount);
std::copy(iodata[0].begin(),iodata[0].end(),tmp.begin());
serial.SetState(tmp,0);
serial_rng.SetState(tmp,0);
}
nersc_csum = nersc_csum + nersc_csum_tmp;
@ -699,8 +693,8 @@ PARALLEL_CRITICAL
/////////////////////////////////////////////////////////////////////////////
// Write a RNG; lexico map to an array of state and use IOobject
//////////////////////////////////////////////////////////////////////////////////////
static inline void writeRNG(GridSerialRNG &serial,
GridParallelRNG &parallel,
static inline void writeRNG(GridSerialRNG &serial_rng,
GridParallelRNG &parallel_rng,
std::string file,
uint64_t offset,
uint32_t &nersc_csum,
@ -712,7 +706,7 @@ PARALLEL_CRITICAL
const int RngStateCount = GridSerialRNG::RngStateCount;
typedef std::array<RngStateType,RngStateCount> RNGstate;
GridBase *grid = parallel._grid;
GridBase *grid = parallel_rng.Grid();
uint64_t gsites = grid->gSites();
uint64_t lsites = grid->lSites();
@ -727,11 +721,11 @@ PARALLEL_CRITICAL
timer.Start();
std::vector<RNGstate> iodata(lsites);
parallel_for(uint64_t lidx=0;lidx<lsites;lidx++){
thread_for(lidx,lsites,{
std::vector<RngStateType> tmp(RngStateCount);
parallel.GetState(tmp,lidx);
parallel_rng.GetState(tmp,lidx);
std::copy(tmp.begin(),tmp.end(),iodata[lidx].begin());
}
});
timer.Stop();
IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
@ -739,7 +733,7 @@ PARALLEL_CRITICAL
iodata.resize(1);
{
std::vector<RngStateType> tmp(RngStateCount);
serial.GetState(tmp,0);
serial_rng.GetState(tmp,0);
std::copy(tmp.begin(),tmp.end(),iodata[0].begin());
}
IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_MASTER_APPEND,
@ -756,5 +750,4 @@ PARALLEL_CRITICAL
}
};
}
#endif
NAMESPACE_END(Grid);

35
Grid/parallelIO/IldgIO.h
View File

@ -24,8 +24,7 @@ See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_ILDG_IO_H
#define GRID_ILDG_IO_H
#pragma once
#ifdef HAVE_LIME
#include <algorithm>
@ -43,8 +42,7 @@ extern "C" {
#include "lime.h"
}
namespace Grid {
namespace QCD {
NAMESPACE_BEGIN(Grid);
#define GRID_FIELD_NORM "FieldNormMetaData"
#define GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) \
@ -140,7 +138,7 @@ assert(GRID_FIELD_NORM_CALC(FieldNormMetaData_, n2ck) < 1.0e-5);
/////////////////////////////////////
// Scidac Private File structure
/////////////////////////////////////
_scidacFile = scidacFile(field._grid);
_scidacFile = scidacFile(field.Grid());
/////////////////////////////////////
// Scidac Private Record structure
@ -227,10 +225,10 @@ class GridLimeReader : public BinaryIO {
// std::cout << GridLogMessage<< " readLimeLatticeBinaryObject matches ! " <<std::endl;
uint64_t PayloadSize = sizeof(sobj) * field._grid->_gsites;
uint64_t PayloadSize = sizeof(sobj) * field.Grid()->_gsites;
// std::cout << "R sizeof(sobj)= " <<sizeof(sobj)<<std::endl;
// std::cout << "R Gsites " <<field._grid->_gsites<<std::endl;
// std::cout << "R Gsites " <<field.Grid()->_gsites<<std::endl;
// std::cout << "R Payload expected " <<PayloadSize<<std::endl;
// std::cout << "R file size " <<file_bytes <<std::endl;
@ -406,7 +404,7 @@ class GridLimeWriter : public BinaryIO
////////////////////////////////////////////////////
// Write a generic lattice field and csum
// This routine is Collectively called by all nodes
// in communicator used by the field._grid
// in communicator used by the field.Grid()
////////////////////////////////////////////////////
template<class vobj>
void writeLimeLatticeBinaryObject(Lattice<vobj> &field,std::string record_name)
@ -425,8 +423,8 @@ class GridLimeWriter : public BinaryIO
// v) Continue writing scidac record.
////////////////////////////////////////////////////////////////////
GridBase *grid = field._grid;
assert(boss_node == field._grid->IsBoss() );
GridBase *grid = field.Grid();
assert(boss_node == field.Grid()->IsBoss() );
FieldNormMetaData FNMD; FNMD.norm2 = norm2(field);
@ -443,7 +441,7 @@ class GridLimeWriter : public BinaryIO
}
// std::cout << "W sizeof(sobj)" <<sizeof(sobj)<<std::endl;
// std::cout << "W Gsites " <<field._grid->_gsites<<std::endl;
// std::cout << "W Gsites " <<field.Grid()->_gsites<<std::endl;
// std::cout << "W Payload expected " <<PayloadSize<<std::endl;
////////////////////////////////////////////////
@ -515,7 +513,7 @@ class ScidacWriter : public GridLimeWriter {
void writeScidacFieldRecord(Lattice<vobj> &field,userRecord _userRecord,
const unsigned int recordScientificPrec = 0)
{
GridBase * grid = field._grid;
GridBase * grid = field.Grid();
////////////////////////////////////////
// fill the Grid header
@ -557,7 +555,7 @@ class ScidacReader : public GridLimeReader {
void readScidacFieldRecord(Lattice<vobj> &field,userRecord &_userRecord)
{
typedef typename vobj::scalar_object sobj;
GridBase * grid = field._grid;
GridBase * grid = field.Grid();
////////////////////////////////////////
// fill the Grid header
@ -624,7 +622,7 @@ class IldgWriter : public ScidacWriter {
template <class vsimd>
void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,int sequence,std::string LFN,std::string description)
{
GridBase * grid = Umu._grid;
GridBase * grid = Umu.Grid();
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
typedef iLorentzColourMatrix<vsimd> vobj;
typedef typename vobj::scalar_object sobj;
@ -717,9 +715,9 @@ class IldgReader : public GridLimeReader {
typedef LorentzColourMatrixF fobj;
typedef LorentzColourMatrixD dobj;
GridBase *grid = Umu._grid;
GridBase *grid = Umu.Grid();
std::vector<int> dims = Umu._grid->FullDimensions();
Coordinate dims = Umu.Grid()->FullDimensions();
assert(dims.size()==4);
@ -853,6 +851,7 @@ class IldgReader : public GridLimeReader {
// Minimally must find binary segment and checksum
// Since this is an ILDG reader require ILDG format
//////////////////////////////////////////////////////
assert(found_ildgLFN);
assert(found_ildgBinary);
assert(found_ildgFormat);
assert(found_scidacChecksum);
@ -930,9 +929,9 @@ class IldgReader : public GridLimeReader {
}
};
}}
NAMESPACE_END(Grid);
//HAVE_LIME
#endif
#endif

34
Grid/parallelIO/IldgIOtypes.h
View File

@ -23,7 +23,7 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
/* END LEGAL */
#ifndef GRID_ILDGTYPES_IO_H
#define GRID_ILDGTYPES_IO_H
@ -32,7 +32,7 @@ extern "C" { // for linkage
#include "lime.h"
}
namespace Grid {
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////
// Data representation of records that enter ILDG and SciDac formats
@ -51,12 +51,12 @@ namespace Grid {
// Unused SCIDAC records names; could move to support this functionality
#define SCIDAC_SITELIST "scidac-sitelist"
////////////////////////////////////////////////////////////
const int GRID_IO_SINGLEFILE = 0; // hardcode lift from QIO compat
const int GRID_IO_MULTIFILE = 1; // hardcode lift from QIO compat
const int GRID_IO_FIELD = 0; // hardcode lift from QIO compat
const int GRID_IO_GLOBAL = 1; // hardcode lift from QIO compat
////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////////
const int GRID_IO_SINGLEFILE = 0; // hardcode lift from QIO compat
const int GRID_IO_MULTIFILE = 1; // hardcode lift from QIO compat
const int GRID_IO_FIELD = 0; // hardcode lift from QIO compat
const int GRID_IO_GLOBAL = 1; // hardcode lift from QIO compat
////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////
// QIO uses mandatory "private" records fixed format
@ -74,7 +74,7 @@ struct emptyUserRecord : Serializable {
// <?xml version="1.0" encoding="UTF-8"?><scidacFile><version>1.1</version><spacetime>4</spacetime><dims>16 16 16 32 </dims><volfmt>0</volfmt></scidacFile>
////////////////////////
struct scidacFile : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(scidacFile,
double, version,
int, spacetime,
@ -91,7 +91,7 @@ struct scidacFile : Serializable {
return dimensions;
}
void setDimensions(std::vector<int> dimensions) {
void setDimensions(Coordinate dimensions) {
char delimiter = ' ';
std::stringstream stream;
for(int i=0;i<dimensions.size();i++){
@ -124,7 +124,7 @@ struct scidacFile : Serializable {
///////////////////////////////////////////////////////////////////////
struct scidacRecord : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(scidacRecord,
double, version,
std::string, date,
@ -160,7 +160,7 @@ public:
// USQCD info
////////////////////////
struct usqcdInfo : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdInfo,
double, version,
double, plaq,
@ -174,7 +174,7 @@ struct usqcdInfo : Serializable {
// Scidac Checksum
////////////////////////
struct scidacChecksum : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(scidacChecksum,
double, version,
std::string, suma,
@ -201,7 +201,7 @@ struct scidacChecksum : Serializable {
// From http://www.physics.utah.edu/~detar/scidac/qio_2p3.pdf
////////////////////////////////////////////////////////////////////////////////////////
struct usqcdPropFile : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdPropFile,
double, version,
std::string, type,
@ -211,7 +211,7 @@ struct usqcdPropFile : Serializable {
};
};
struct usqcdSourceInfo : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdSourceInfo,
double, version,
std::string, info);
@ -220,7 +220,7 @@ struct usqcdSourceInfo : Serializable {
};
};
struct usqcdPropInfo : Serializable {
public:
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(usqcdPropInfo,
double, version,
int, spin,
@ -232,6 +232,6 @@ struct usqcdPropInfo : Serializable {
};
#endif
}
NAMESPACE_END(Grid);
#endif
#endif

371
Grid/parallelIO/MetaData.h
View File

@ -36,23 +36,24 @@
#include <sys/utsname.h>
#include <pwd.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////
// Precision mapping
///////////////////////////////////////////////////////
template<class vobj> static std::string getFormatString (void)
{
std::string format;
typedef typename getPrecision<vobj>::real_scalar_type stype;
if ( sizeof(stype) == sizeof(float) ) {
format = std::string("IEEE32BIG");
}
if ( sizeof(stype) == sizeof(double) ) {
format = std::string("IEEE64BIG");
}
return format;
///////////////////////////////////////////////////////
// Precision mapping
///////////////////////////////////////////////////////
template<class vobj> static std::string getFormatString (void)
{
std::string format;
typedef typename getPrecision<vobj>::real_scalar_type stype;
if ( sizeof(stype) == sizeof(float) ) {
format = std::string("IEEE32BIG");
}
if ( sizeof(stype) == sizeof(double) ) {
format = std::string("IEEE64BIG");
}
return format;
};
////////////////////////////////////////////////////////////////////////////////
// header specification/interpretation
////////////////////////////////////////////////////////////////////////////////
@ -93,146 +94,145 @@ namespace Grid {
link_trace(0.), plaquette(0.), checksum(0),
scidac_checksuma(0), scidac_checksumb(0), sequence_number(0)
{}
};
};
namespace QCD {
// PB disable using namespace - this is a header and forces namesapce visibility for all
// including files
//using namespace Grid;
using namespace Grid;
//////////////////////////////////////////////////////////////////////
// Bit and Physical Checksumming and QA of data
//////////////////////////////////////////////////////////////////////
inline void GridMetaData(GridBase *grid,FieldMetaData &header)
{
int nd = grid->_ndimension;
header.nd = nd;
header.dimension.resize(nd);
header.boundary.resize(nd);
header.data_start = 0;
for(int d=0;d<nd;d++) {
header.dimension[d] = grid->_fdimensions[d];
}
for(int d=0;d<nd;d++) {
header.boundary[d] = std::string("PERIODIC");
}
}
inline void MachineCharacteristics(FieldMetaData &header)
{
// Who
struct passwd *pw = getpwuid (getuid());
if (pw) header.creator = std::string(pw->pw_name);
//////////////////////////////////////////////////////////////////////
// Bit and Physical Checksumming and QA of data
//////////////////////////////////////////////////////////////////////
inline void GridMetaData(GridBase *grid,FieldMetaData &header)
{
int nd = grid->_ndimension;
header.nd = nd;
header.dimension.resize(nd);
header.boundary.resize(nd);
header.data_start = 0;
for(int d=0;d<nd;d++) {
header.dimension[d] = grid->_fdimensions[d];
}
for(int d=0;d<nd;d++) {
header.boundary[d] = std::string("PERIODIC");
}
}
// When
std::time_t t = std::time(nullptr);
std::tm tm_ = *std::localtime(&t);
std::ostringstream oss;
// oss << std::put_time(&tm_, "%c %Z");
header.creation_date = oss.str();
header.archive_date = header.creation_date;
inline void MachineCharacteristics(FieldMetaData &header)
{
// Who
struct passwd *pw = getpwuid (getuid());
if (pw) header.creator = std::string(pw->pw_name);
// When
std::time_t t = std::time(nullptr);
std::tm tm_ = *std::localtime(&t);
std::ostringstream oss;
// oss << std::put_time(&tm_, "%c %Z");
header.creation_date = oss.str();
header.archive_date = header.creation_date;
// What
struct utsname name; uname(&name);
header.creator_hardware = std::string(name.nodename)+"-";
header.creator_hardware+= std::string(name.machine)+"-";
header.creator_hardware+= std::string(name.sysname)+"-";
header.creator_hardware+= std::string(name.release);
}
// What
struct utsname name; uname(&name);
header.creator_hardware = std::string(name.nodename)+"-";
header.creator_hardware+= std::string(name.machine)+"-";
header.creator_hardware+= std::string(name.sysname)+"-";
header.creator_hardware+= std::string(name.release);
}
#define dump_meta_data(field, s) \
s << "BEGIN_HEADER" << std::endl; \
s << "HDR_VERSION = " << field.hdr_version << std::endl; \
s << "DATATYPE = " << field.data_type << std::endl; \
s << "STORAGE_FORMAT = " << field.storage_format << std::endl; \
for(int i=0;i<4;i++){ \
s << "DIMENSION_" << i+1 << " = " << field.dimension[i] << std::endl ; \
} \
s << "LINK_TRACE = " << std::setprecision(10) << field.link_trace << std::endl; \
s << "PLAQUETTE = " << std::setprecision(10) << field.plaquette << std::endl; \
for(int i=0;i<4;i++){ \
s << "BOUNDARY_"<<i+1<<" = " << field.boundary[i] << std::endl; \
} \
s << "BEGIN_HEADER" << std::endl; \
s << "HDR_VERSION = " << field.hdr_version << std::endl; \
s << "DATATYPE = " << field.data_type << std::endl; \
s << "STORAGE_FORMAT = " << field.storage_format << std::endl; \
for(int i=0;i<4;i++){ \
s << "DIMENSION_" << i+1 << " = " << field.dimension[i] << std::endl ; \
} \
s << "LINK_TRACE = " << std::setprecision(10) << field.link_trace << std::endl; \
s << "PLAQUETTE = " << std::setprecision(10) << field.plaquette << std::endl; \
for(int i=0;i<4;i++){ \
s << "BOUNDARY_"<<i+1<<" = " << field.boundary[i] << std::endl; \
} \
\
s << "CHECKSUM = "<< std::hex << std::setw(10) << field.checksum << std::dec<<std::endl; \
s << "SCIDAC_CHECKSUMA = "<< std::hex << std::setw(10) << field.scidac_checksuma << std::dec<<std::endl; \
s << "SCIDAC_CHECKSUMB = "<< std::hex << std::setw(10) << field.scidac_checksumb << std::dec<<std::endl; \
s << "ENSEMBLE_ID = " << field.ensemble_id << std::endl; \
s << "ENSEMBLE_LABEL = " << field.ensemble_label << std::endl; \
s << "SEQUENCE_NUMBER = " << field.sequence_number << std::endl; \
s << "CREATOR = " << field.creator << std::endl; \
s << "CREATOR_HARDWARE = "<< field.creator_hardware << std::endl; \
s << "CREATION_DATE = " << field.creation_date << std::endl; \
s << "ARCHIVE_DATE = " << field.archive_date << std::endl; \
s << "FLOATING_POINT = " << field.floating_point << std::endl; \
s << "END_HEADER" << std::endl;
s << "CHECKSUM = "<< std::hex << std::setw(10) << field.checksum << std::dec<<std::endl; \
s << "SCIDAC_CHECKSUMA = "<< std::hex << std::setw(10) << field.scidac_checksuma << std::dec<<std::endl; \
s << "SCIDAC_CHECKSUMB = "<< std::hex << std::setw(10) << field.scidac_checksumb << std::dec<<std::endl; \
s << "ENSEMBLE_ID = " << field.ensemble_id << std::endl; \
s << "ENSEMBLE_LABEL = " << field.ensemble_label << std::endl; \
s << "SEQUENCE_NUMBER = " << field.sequence_number << std::endl; \
s << "CREATOR = " << field.creator << std::endl; \
s << "CREATOR_HARDWARE = "<< field.creator_hardware << std::endl; \
s << "CREATION_DATE = " << field.creation_date << std::endl; \
s << "ARCHIVE_DATE = " << field.archive_date << std::endl; \
s << "FLOATING_POINT = " << field.floating_point << std::endl; \
s << "END_HEADER" << std::endl;
template<class vobj> inline void PrepareMetaData(Lattice<vobj> & field, FieldMetaData &header)
{
GridBase *grid = field._grid;
GridBase *grid = field.Grid();
std::string format = getFormatString<vobj>();
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
MachineCharacteristics(header);
}
inline void GaugeStatistics(Lattice<vLorentzColourMatrixF> & data,FieldMetaData &header)
{
// How to convert data precision etc...
header.link_trace=Grid::QCD::WilsonLoops<PeriodicGimplF>::linkTrace(data);
header.plaquette =Grid::QCD::WilsonLoops<PeriodicGimplF>::avgPlaquette(data);
}
inline void GaugeStatistics(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
{
// How to convert data precision etc...
header.link_trace=Grid::QCD::WilsonLoops<PeriodicGimplD>::linkTrace(data);
header.plaquette =Grid::QCD::WilsonLoops<PeriodicGimplD>::avgPlaquette(data);
}
template<> inline void PrepareMetaData<vLorentzColourMatrixF>(Lattice<vLorentzColourMatrixF> & field, FieldMetaData &header)
{
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
MachineCharacteristics(header);
}
inline void GaugeStatistics(Lattice<vLorentzColourMatrixF> & data,FieldMetaData &header)
{
// How to convert data precision etc...
header.link_trace=WilsonLoops<PeriodicGimplF>::linkTrace(data);
header.plaquette =WilsonLoops<PeriodicGimplF>::avgPlaquette(data);
}
inline void GaugeStatistics(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
{
// How to convert data precision etc...
header.link_trace=WilsonLoops<PeriodicGimplD>::linkTrace(data);
header.plaquette =WilsonLoops<PeriodicGimplD>::avgPlaquette(data);
}
template<> inline void PrepareMetaData<vLorentzColourMatrixF>(Lattice<vLorentzColourMatrixF> & field, FieldMetaData &header)
{
GridBase *grid = field._grid;
std::string format = getFormatString<vLorentzColourMatrixF>();
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
GaugeStatistics(field,header);
MachineCharacteristics(header);
}
template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header)
{
GridBase *grid = field._grid;
std::string format = getFormatString<vLorentzColourMatrixD>();
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
GaugeStatistics(field,header);
MachineCharacteristics(header);
}
GridBase *grid = field.Grid();
std::string format = getFormatString<vLorentzColourMatrixF>();
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
GaugeStatistics(field,header);
MachineCharacteristics(header);
}
template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzColourMatrixD> & field, FieldMetaData &header)
{
GridBase *grid = field.Grid();
std::string format = getFormatString<vLorentzColourMatrixD>();
header.floating_point = format;
header.checksum = 0x0; // Nersc checksum unused in ILDG, Scidac
GridMetaData(grid,header);
GaugeStatistics(field,header);
MachineCharacteristics(header);
}
//////////////////////////////////////////////////////////////////////
// Utilities ; these are QCD aware
//////////////////////////////////////////////////////////////////////
inline void reconstruct3(LorentzColourMatrix & cm)
{
const int x=0;
const int y=1;
const int z=2;
for(int mu=0;mu<Nd;mu++){
cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
}
}
//////////////////////////////////////////////////////////////////////
// Utilities ; these are QCD aware
//////////////////////////////////////////////////////////////////////
inline void reconstruct3(LorentzColourMatrix & cm)
{
const int x=0;
const int y=1;
const int z=2;
for(int mu=0;mu<Nd;mu++){
cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
}
}
////////////////////////////////////////////////////////////////////////////////
// Some data types for intermediate storage
////////////////////////////////////////////////////////////////////////////////
template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
////////////////////////////////////////////////////////////////////////////////
// Some data types for intermediate storage
////////////////////////////////////////////////////////////////////////////////
template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
typedef iLorentzColour2x3<Complex> LorentzColour2x3;
typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
typedef iLorentzColour2x3<ComplexD> LorentzColour2x3D;
typedef iLorentzColour2x3<Complex> LorentzColour2x3;
typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
typedef iLorentzColour2x3<ComplexD> LorentzColour2x3D;
/////////////////////////////////////////////////////////////////////////////////
// Simple classes for precision conversion
@ -276,56 +276,55 @@ struct BinarySimpleMunger {
};
template<class fobj,class sobj>
struct GaugeSimpleMunger{
void operator()(fobj &in, sobj &out) {
for (int mu = 0; mu < Nd; 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 GaugeSimpleMunger{
void operator()(fobj &in, sobj &out) {
for (int mu = 0; mu < Nd; 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 GaugeSimpleUnmunger {
template <class fobj, class sobj>
struct GaugeSimpleUnmunger {
void operator()(sobj &in, fobj &out) {
for (int mu = 0; mu < Nd; mu++) {
for (int i = 0; i < Nc; i++) {
for (int j = 0; j < Nc; j++) {
out(mu)()(i, j) = in(mu)()(i, j);
}}
}
};
};
void operator()(sobj &in, fobj &out) {
for (int mu = 0; mu < Nd; 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){
for(int mu=0;mu<Nd;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
out(mu)()(i,j) = in(mu)(i)(j);
}}
}
reconstruct3(out);
}
};
template<class fobj,class sobj>
struct Gauge3x2unmunger{
void operator() (sobj &in,fobj &out){
for(int mu=0;mu<Nd;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
out(mu)(i)(j) = in(mu)()(i,j);
}}
}
}
};
template<class fobj,class sobj>
struct Gauge3x2munger{
void operator() (fobj &in,sobj &out){
for(int mu=0;mu<Nd;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
out(mu)()(i,j) = in(mu)(i)(j);
}}
}
reconstruct3(out);
}
};
template<class fobj,class sobj>
struct Gauge3x2unmunger{
void operator() (sobj &in,fobj &out){
for(int mu=0;mu<Nd;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
out(mu)(i)(j) = in(mu)()(i,j);
}}
}
}
};
NAMESPACE_END(Grid);
}

480
Grid/parallelIO/NerscIO.h
View File

@ -30,334 +30,330 @@
#ifndef GRID_NERSC_IO_H
#define GRID_NERSC_IO_H
namespace Grid {
namespace QCD {
NAMESPACE_BEGIN(Grid);
using namespace Grid;
using namespace Grid;
////////////////////////////////////////////////////////////////////////////////
// Write and read from fstream; comput header offset for payload
////////////////////////////////////////////////////////////////////////////////
class NerscIO : public BinaryIO {
public:
////////////////////////////////////////////////////////////////////////////////
// Write and read from fstream; comput header offset for payload
////////////////////////////////////////////////////////////////////////////////
class NerscIO : public BinaryIO {
public:
static inline void truncate(std::string file){
std::ofstream fout(file,std::ios::out);
}
static inline void truncate(std::string file){
std::ofstream fout(file,std::ios::out);
}
static inline unsigned int writeHeader(FieldMetaData &field,std::string file)
{
std::ofstream fout(file,std::ios::out|std::ios::in);
fout.seekp(0,std::ios::beg);
dump_meta_data(field, fout);
field.data_start = fout.tellp();
return field.data_start;
}
static inline unsigned int writeHeader(FieldMetaData &field,std::string file)
{
std::ofstream fout(file,std::ios::out|std::ios::in);
fout.seekp(0,std::ios::beg);
dump_meta_data(field, fout);
field.data_start = fout.tellp();
return field.data_start;
}
// for the header-reader
static inline int readHeader(std::string file,GridBase *grid, FieldMetaData &field)
{
uint64_t offset=0;
std::map<std::string,std::string> header;
std::string line;
// for the header-reader
static inline int readHeader(std::string file,GridBase *grid, FieldMetaData &field)
{
std::map<std::string,std::string> header;
std::string line;
//////////////////////////////////////////////////
// read the header
//////////////////////////////////////////////////
std::ifstream fin(file);
//////////////////////////////////////////////////
// read the header
//////////////////////////////////////////////////
std::ifstream fin(file);
getline(fin,line); // read one line and insist is
getline(fin,line); // read one line and insist is
removeWhitespace(line);
std::cout << GridLogMessage << "* " << line << std::endl;
removeWhitespace(line);
std::cout << GridLogMessage << "* " << line << std::endl;
assert(line==std::string("BEGIN_HEADER"));
assert(line==std::string("BEGIN_HEADER"));
do {
do {
getline(fin,line); // read one line
std::cout << GridLogMessage << "* "<<line<< std::endl;
int eq = line.find("=");
if(eq >0) {
std::string key=line.substr(0,eq);
std::string val=line.substr(eq+1);
removeWhitespace(key);
removeWhitespace(val);
std::string key=line.substr(0,eq);
std::string val=line.substr(eq+1);
removeWhitespace(key);
removeWhitespace(val);
header[key] = val;
}
header[key] = val;
}
} while( line.find("END_HEADER") == std::string::npos );
field.data_start = fin.tellg();
field.data_start = fin.tellg();
//////////////////////////////////////////////////
// chomp the values
//////////////////////////////////////////////////
field.hdr_version = header["HDR_VERSION"];
field.data_type = header["DATATYPE"];
field.storage_format = header["STORAGE_FORMAT"];
//////////////////////////////////////////////////
// chomp the values
//////////////////////////////////////////////////
field.hdr_version = header["HDR_VERSION"];
field.data_type = header["DATATYPE"];
field.storage_format = header["STORAGE_FORMAT"];
field.dimension[0] = std::stol(header["DIMENSION_1"]);
field.dimension[1] = std::stol(header["DIMENSION_2"]);
field.dimension[2] = std::stol(header["DIMENSION_3"]);
field.dimension[3] = std::stol(header["DIMENSION_4"]);
field.dimension[0] = std::stol(header["DIMENSION_1"]);
field.dimension[1] = std::stol(header["DIMENSION_2"]);
field.dimension[2] = std::stol(header["DIMENSION_3"]);
field.dimension[3] = std::stol(header["DIMENSION_4"]);
assert(grid->_ndimension == 4);
for(int d=0;d<4;d++){
assert(grid->_ndimension == 4);
for(int d=0;d<4;d++){
assert(grid->_fdimensions[d]==field.dimension[d]);
}
field.link_trace = std::stod(header["LINK_TRACE"]);
field.plaquette = std::stod(header["PLAQUETTE"]);
field.link_trace = std::stod(header["LINK_TRACE"]);
field.plaquette = std::stod(header["PLAQUETTE"]);
field.boundary[0] = header["BOUNDARY_1"];
field.boundary[1] = header["BOUNDARY_2"];
field.boundary[2] = header["BOUNDARY_3"];
field.boundary[3] = header["BOUNDARY_4"];
field.boundary[0] = header["BOUNDARY_1"];
field.boundary[1] = header["BOUNDARY_2"];
field.boundary[2] = header["BOUNDARY_3"];
field.boundary[3] = header["BOUNDARY_4"];
field.checksum = std::stoul(header["CHECKSUM"],0,16);
field.ensemble_id = header["ENSEMBLE_ID"];
field.ensemble_label = header["ENSEMBLE_LABEL"];
field.sequence_number = std::stol(header["SEQUENCE_NUMBER"]);
field.creator = header["CREATOR"];
field.creator_hardware = header["CREATOR_HARDWARE"];
field.creation_date = header["CREATION_DATE"];
field.archive_date = header["ARCHIVE_DATE"];
field.floating_point = header["FLOATING_POINT"];
field.checksum = std::stoul(header["CHECKSUM"],0,16);
field.ensemble_id = header["ENSEMBLE_ID"];
field.ensemble_label = header["ENSEMBLE_LABEL"];
field.sequence_number = std::stol(header["SEQUENCE_NUMBER"]);
field.creator = header["CREATOR"];
field.creator_hardware = header["CREATOR_HARDWARE"];
field.creation_date = header["CREATION_DATE"];
field.archive_date = header["ARCHIVE_DATE"];
field.floating_point = header["FLOATING_POINT"];
return field.data_start;
return field.data_start;
}
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Now the meat: the object readers
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vsimd>
static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
FieldMetaData& header,
std::string file)
{
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
GridBase *grid = Umu.Grid();
uint64_t offset = readHeader(file,Umu.Grid(),header);
FieldMetaData clone(header);
std::string format(header.floating_point);
int ieee32big = (format == std::string("IEEE32BIG"));
int ieee32 = (format == std::string("IEEE32"));
int ieee64big = (format == std::string("IEEE64BIG"));
int ieee64 = (format == std::string("IEEE64"));
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
// depending on datatype, set up munger;
// munger is a function of <floating point, Real, data_type>
if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
if ( ieee32 || ieee32big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F>
(Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
if ( ieee64 || ieee64big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D>
(Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
} else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
if ( ieee32 || ieee32big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
(Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
if ( ieee64 || ieee64big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
(Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
} else {
assert(0);
}
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Now the meat: the object readers
/////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
GaugeStatistics(Umu,clone);
template<class vsimd>
static inline void readConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
FieldMetaData& header,
std::string file)
{
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
<<" header "<<std::hex<<header.checksum<<std::dec <<std::endl;
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" plaquette "<<clone.plaquette
<<" header "<<header.plaquette<<std::endl;
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" link_trace "<<clone.link_trace
<<" header "<<header.link_trace<<std::endl;
GridBase *grid = Umu._grid;
uint64_t offset = readHeader(file,Umu._grid,header);
FieldMetaData clone(header);
std::string format(header.floating_point);
int ieee32big = (format == std::string("IEEE32BIG"));
int ieee32 = (format == std::string("IEEE32"));
int ieee64big = (format == std::string("IEEE64BIG"));
int ieee64 = (format == std::string("IEEE64"));
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
// depending on datatype, set up munger;
// munger is a function of <floating point, Real, data_type>
if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
if ( ieee32 || ieee32big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3F>
(Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
if ( ieee64 || ieee64big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>, LorentzColour2x3D>
(Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
} else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
if ( ieee32 || ieee32big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixF>
(Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
if ( ieee64 || ieee64big ) {
BinaryIO::readLatticeObject<iLorentzColourMatrix<vsimd>,LorentzColourMatrixD>
(Umu,file,GaugeSimpleMunger<LorentzColourMatrixD,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
} else {
assert(0);
}
GaugeStatistics(Umu,clone);
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" checksum "<<std::hex<<nersc_csum<< std::dec
<<" header "<<std::hex<<header.checksum<<std::dec <<std::endl;
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" plaquette "<<clone.plaquette
<<" header "<<header.plaquette<<std::endl;
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<<" link_trace "<<clone.link_trace
<<" header "<<header.link_trace<<std::endl;
if ( fabs(clone.plaquette -header.plaquette ) >= 1.0e-5 ) {
std::cout << " Plaquette mismatch "<<std::endl;
std::cout << Umu[0]<<std::endl;
std::cout << Umu[1]<<std::endl;
}
if ( nersc_csum != header.checksum ) {
std::cerr << " checksum mismatch " << std::endl;
std::cerr << " plaqs " << clone.plaquette << " " << header.plaquette << std::endl;
std::cerr << " trace " << clone.link_trace<< " " << header.link_trace<< std::endl;
std::cerr << " nersc_csum " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
exit(0);
}
assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
assert(nersc_csum == header.checksum );
if ( fabs(clone.plaquette -header.plaquette ) >= 1.0e-5 ) {
std::cout << " Plaquette mismatch "<<std::endl;
}
if ( nersc_csum != header.checksum ) {
std::cerr << " checksum mismatch " << std::endl;
std::cerr << " plaqs " << clone.plaquette << " " << header.plaquette << std::endl;
std::cerr << " trace " << clone.link_trace<< " " << header.link_trace<< std::endl;
std::cerr << " nersc_csum " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
exit(0);
}
assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
assert(fabs(clone.link_trace-header.link_trace) < 1.0e-6 );
assert(nersc_csum == header.checksum );
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
}
std::cout<<GridLogMessage <<"NERSC Configuration "<<file<< " and plaquette, link trace, and checksum agree"<<std::endl;
}
template<class vsimd>
static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
std::string file,
int two_row,
int bits32)
{
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
template<class vsimd>
static inline void writeConfiguration(Lattice<iLorentzColourMatrix<vsimd> > &Umu,
std::string file,
int two_row,
int bits32)
{
typedef Lattice<iLorentzColourMatrix<vsimd> > GaugeField;
typedef iLorentzColourMatrix<vsimd> vobj;
typedef typename vobj::scalar_object sobj;
typedef iLorentzColourMatrix<vsimd> vobj;
typedef typename vobj::scalar_object sobj;
FieldMetaData header;
///////////////////////////////////////////
// Following should become arguments
///////////////////////////////////////////
header.sequence_number = 1;
header.ensemble_id = "UKQCD";
header.ensemble_label = "DWF";
FieldMetaData header;
///////////////////////////////////////////
// Following should become arguments
///////////////////////////////////////////
header.sequence_number = 1;
header.ensemble_id = "UKQCD";
header.ensemble_label = "DWF";
typedef LorentzColourMatrixD fobj3D;
typedef LorentzColour2x3D fobj2D;
typedef LorentzColourMatrixD fobj3D;
typedef LorentzColour2x3D fobj2D;
GridBase *grid = Umu._grid;
GridBase *grid = Umu.Grid();
GridMetaData(grid,header);
assert(header.nd==4);
GaugeStatistics(Umu,header);
MachineCharacteristics(header);
GridMetaData(grid,header);
assert(header.nd==4);
GaugeStatistics(Umu,header);
MachineCharacteristics(header);
uint64_t offset;
// Sod it -- always write 3x3 double
header.floating_point = std::string("IEEE64BIG");
header.data_type = std::string("4D_SU3_GAUGE_3x3");
GaugeSimpleUnmunger<fobj3D,sobj> munge;
// Sod it -- always write 3x3 double
header.floating_point = std::string("IEEE64BIG");
header.data_type = std::string("4D_SU3_GAUGE_3x3");
GaugeSimpleUnmunger<fobj3D,sobj> munge;
if ( grid->IsBoss() ) {
truncate(file);
offset = writeHeader(header,file);
offset = writeHeader(header,file);
}
grid->Broadcast(0,(void *)&offset,sizeof(offset));
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
nersc_csum,scidac_csuma,scidac_csumb);
header.checksum = nersc_csum;
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
nersc_csum,scidac_csuma,scidac_csumb);
header.checksum = nersc_csum;
if ( grid->IsBoss() ) {
writeHeader(header,file);
writeHeader(header,file);
}
std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "
<<std::hex<<header.checksum
<<std::dec<<" plaq "<< header.plaquette <<std::endl;
std::cout<<GridLogMessage <<"Written NERSC Configuration on "<< file << " checksum "
<<std::hex<<header.checksum
<<std::dec<<" plaq "<< header.plaquette <<std::endl;
}
///////////////////////////////
// RNG state
///////////////////////////////
static inline void writeRNGState(GridSerialRNG &serial,GridParallelRNG &parallel,std::string file)
{
typedef typename GridParallelRNG::RngStateType RngStateType;
}
///////////////////////////////
// RNG state
///////////////////////////////
static inline void writeRNGState(GridSerialRNG &serial,GridParallelRNG &parallel,std::string file)
{
typedef typename GridParallelRNG::RngStateType RngStateType;
// Following should become arguments
FieldMetaData header;
header.sequence_number = 1;
header.ensemble_id = "UKQCD";
header.ensemble_label = "DWF";
// Following should become arguments
FieldMetaData header;
header.sequence_number = 1;
header.ensemble_id = "UKQCD";
header.ensemble_label = "DWF";
GridBase *grid = parallel._grid;
GridBase *grid = parallel.Grid();
GridMetaData(grid,header);
assert(header.nd==4);
header.link_trace=0.0;
header.plaquette=0.0;
MachineCharacteristics(header);
GridMetaData(grid,header);
assert(header.nd==4);
header.link_trace=0.0;
header.plaquette=0.0;
MachineCharacteristics(header);
uint64_t offset;
#ifdef RNG_RANLUX
header.floating_point = std::string("UINT64");
header.data_type = std::string("RANLUX48");
header.floating_point = std::string("UINT64");
header.data_type = std::string("RANLUX48");
#endif
#ifdef RNG_MT19937
header.floating_point = std::string("UINT32");
header.data_type = std::string("MT19937");
header.floating_point = std::string("UINT32");
header.data_type = std::string("MT19937");
#endif
#ifdef RNG_SITMO
header.floating_point = std::string("UINT64");
header.data_type = std::string("SITMO");
header.floating_point = std::string("UINT64");
header.data_type = std::string("SITMO");
#endif
if ( grid->IsBoss() ) {
truncate(file);
offset = writeHeader(header,file);
truncate(file);
offset = writeHeader(header,file);
}
grid->Broadcast(0,(void *)&offset,sizeof(offset));
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::writeRNG(serial,parallel,file,offset,nersc_csum,scidac_csuma,scidac_csumb);
header.checksum = nersc_csum;
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::writeRNG(serial,parallel,file,offset,nersc_csum,scidac_csuma,scidac_csumb);
header.checksum = nersc_csum;
if ( grid->IsBoss() ) {
offset = writeHeader(header,file);
offset = writeHeader(header,file);
}
std::cout<<GridLogMessage
<<"Written NERSC RNG STATE "<<file<< " checksum "
<<std::hex<<header.checksum
<<std::dec<<std::endl;
std::cout<<GridLogMessage
<<"Written NERSC RNG STATE "<<file<< " checksum "
<<std::hex<<header.checksum
<<std::dec<<std::endl;
}
}
static inline void readRNGState(GridSerialRNG &serial,GridParallelRNG & parallel,FieldMetaData& header,std::string file)
{
typedef typename GridParallelRNG::RngStateType RngStateType;
static inline void readRNGState(GridSerialRNG &serial,GridParallelRNG & parallel,FieldMetaData& header,std::string file)
{
typedef typename GridParallelRNG::RngStateType RngStateType;
GridBase *grid = parallel._grid;
GridBase *grid = parallel.Grid();
uint64_t offset = readHeader(file,grid,header);
FieldMetaData clone(header);
FieldMetaData clone(header);
std::string format(header.floating_point);
std::string data_type(header.data_type);
std::string format(header.floating_point);
std::string data_type(header.data_type);
#ifdef RNG_RANLUX
assert(format == std::string("UINT64"));
assert(data_type == std::string("RANLUX48"));
assert(format == std::string("UINT64"));
assert(data_type == std::string("RANLUX48"));
#endif
#ifdef RNG_MT19937
assert(format == std::string("UINT32"));
assert(data_type == std::string("MT19937"));
assert(format == std::string("UINT32"));
assert(data_type == std::string("MT19937"));
#endif
#ifdef RNG_SITMO
assert(format == std::string("UINT64"));
assert(data_type == std::string("SITMO"));
assert(format == std::string("UINT64"));
assert(data_type == std::string("SITMO"));
#endif
// depending on datatype, set up munger;
// munger is a function of <floating point, Real, data_type>
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::readRNG(serial,parallel,file,offset,nersc_csum,scidac_csuma,scidac_csumb);
// depending on datatype, set up munger;
// munger is a function of <floating point, Real, data_type>
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::readRNG(serial,parallel,file,offset,nersc_csum,scidac_csuma,scidac_csumb);
if ( nersc_csum != header.checksum ) {
std::cerr << "checksum mismatch "<<std::hex<< nersc_csum <<" "<<header.checksum<<std::dec<<std::endl;
exit(0);
}
assert(nersc_csum == header.checksum );
if ( nersc_csum != header.checksum ) {
std::cerr << "checksum mismatch "<<std::hex<< nersc_csum <<" "<<header.checksum<<std::dec<<std::endl;
exit(0);
}
assert(nersc_csum == header.checksum );
std::cout<<GridLogMessage <<"Read NERSC RNG file "<<file<< " format "<< data_type <<std::endl;
}
std::cout<<GridLogMessage <<"Read NERSC RNG file "<<file<< " format "<< data_type <<std::endl;
}
};
};
NAMESPACE_END(QCD);
}}
#endif

35
Grid/perfmon/PerfCount.cc
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -23,13 +23,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#include <Grid/GridCore.h>
#include <Grid/perfmon/PerfCount.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
#define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
#define RawConfig(A,B) (A<<8|B)
@ -39,16 +39,16 @@ const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::Performan
{ PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES , "CACHE_MISSES......." , CACHE_REFERENCES},
{ PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES , "CPUCYCLES.........." , INSTRUCTIONS},
{ PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS , "INSTRUCTIONS......." , CPUCYCLES },
// 4
// 4
#ifdef KNL
{ PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", CPUCYCLES },
{ PERF_TYPE_RAW, RawConfig(0x01,0x04), "L1_MISS_LOADS......", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x02,0x04), "L2_HIT_LOADS.......", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x04,0x04), "L2_MISS_LOADS......", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x10,0x04), "UTLB_MISS_LOADS....", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x08,0x04), "DTLB_MISS_LOADS....", L1D_READ_ACCESS },
// 11
{ PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", CPUCYCLES },
{ PERF_TYPE_RAW, RawConfig(0x01,0x04), "L1_MISS_LOADS......", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x40,0x04), "ALL_LOADS..........", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x02,0x04), "L2_HIT_LOADS.......", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x04,0x04), "L2_MISS_LOADS......", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x10,0x04), "UTLB_MISS_LOADS....", L1D_READ_ACCESS },
{ PERF_TYPE_RAW, RawConfig(0x08,0x04), "DTLB_MISS_LOADS....", L1D_READ_ACCESS },
// 11
#else
{ PERF_TYPE_HW_CACHE, CacheControl(L1D,READ,ACCESS) , "L1D_READ_ACCESS....",INSTRUCTIONS},
{ PERF_TYPE_HW_CACHE, CacheControl(L1D,READ,MISS) , "L1D_READ_MISS......",L1D_READ_ACCESS},
@ -57,19 +57,20 @@ const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::Performan
{ PERF_TYPE_HW_CACHE, CacheControl(L1D,PREFETCH,MISS) , "L1D_PREFETCH_MISS..",L1D_READ_ACCESS},
{ PERF_TYPE_HW_CACHE, CacheControl(L1D,PREFETCH,ACCESS) , "L1D_PREFETCH_ACCESS",L1D_READ_ACCESS},
{ PERF_TYPE_HW_CACHE, CacheControl(L1D,PREFETCH,ACCESS) , "L1D_PREFETCH_ACCESS",L1D_READ_ACCESS},
// 11
// 11
#endif
{ PERF_TYPE_HW_CACHE, CacheControl(LL,READ,MISS) , "LL_READ_MISS.......",L1D_READ_ACCESS},
{ PERF_TYPE_HW_CACHE, CacheControl(LL,READ,ACCESS) , "LL_READ_ACCESS.....",L1D_READ_ACCESS},
{ PERF_TYPE_HW_CACHE, CacheControl(LL,WRITE,MISS) , "LL_WRITE_MISS......",L1D_READ_ACCESS},
{ PERF_TYPE_HW_CACHE, CacheControl(LL,WRITE,ACCESS) , "LL_WRITE_ACCESS....",L1D_READ_ACCESS},
//15
//15
{ PERF_TYPE_HW_CACHE, CacheControl(LL,PREFETCH,MISS) , "LL_PREFETCH_MISS...",L1D_READ_ACCESS},
{ PERF_TYPE_HW_CACHE, CacheControl(LL,PREFETCH,ACCESS) , "LL_PREFETCH_ACCESS.",L1D_READ_ACCESS},
{ PERF_TYPE_HW_CACHE, CacheControl(L1I,READ,MISS) , "L1I_READ_MISS......",INSTRUCTIONS},
{ PERF_TYPE_HW_CACHE, CacheControl(L1I,READ,ACCESS) , "L1I_READ_ACCESS....",INSTRUCTIONS}
//19
//19
// { PERF_TYPE_HARDWARE, PERF_COUNT_HW_STALLED_CYCLES_FRONTEND, "STALL_CYCLES" },
#endif
};
}
NAMESPACE_END(Grid);

28
Grid/perfmon/PerfCount.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -25,8 +25,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_PERFCOUNT_H
#define GRID_PERFCOUNT_H
@ -47,7 +47,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <x86intrin.h>
#endif
namespace Grid {
NAMESPACE_BEGIN(Grid);
#ifdef __linux__
static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
@ -84,11 +84,14 @@ inline uint64_t cyclecount(void){
#ifdef __bgq__
inline uint64_t cyclecount(void){
uint64_t tmp;
asm volatile ("mfspr %0,0x10C" : "=&r" (tmp) );
return tmp;
uint64_t tmp;
asm volatile ("mfspr %0,0x10C" : "=&r" (tmp) );
return tmp;
}
#elif defined __x86_64__
#ifdef GRID_NVCC
accelerator_inline uint64_t __rdtsc(void) { return 0; }
#endif
inline uint64_t cyclecount(void){
return __rdtsc();
// unsigned int dummy;
@ -97,7 +100,7 @@ inline uint64_t cyclecount(void){
#else
inline uint64_t cyclecount(void){
return 0;
return 0;
}
#endif
@ -212,7 +215,7 @@ public:
::ioctl(cyclefd, PERF_EVENT_IOC_DISABLE, 0);
ign=::read(fd, &count, sizeof(long long));
ign+=::read(cyclefd, &cycles, sizeof(long long));
assert(ign=2*sizeof(long long));
assert(ign==2*sizeof(long long));
}
elapsed = cyclecount() - begin;
#else
@ -225,8 +228,8 @@ public:
int N = PerformanceCounterConfigs[PCT].normalisation;
const char * sn = PerformanceCounterConfigs[N].name ;
const char * sc = PerformanceCounterConfigs[PCT].name;
std::printf("tsc = %llu %s = %llu %s = %20llu\n (%s/%s) rate = %lf\n", elapsed,sn ,cycles,
sc, count, sc,sn, (double)count/(double)cycles);
std::printf("tsc = %llu %s = %llu %s = %20llu\n (%s/%s) rate = %lf\n", elapsed,sn ,cycles,
sc, count, sc,sn, (double)count/(double)cycles);
#else
std::printf("%llu cycles \n", elapsed );
#endif
@ -241,5 +244,6 @@ public:
};
}
NAMESPACE_END(Grid);
#endif

67
Grid/perfmon/Stat.cc
View File

@ -2,7 +2,7 @@
#include <Grid/perfmon/PerfCount.h>
#include <Grid/perfmon/Stat.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
bool PmuStat::pmu_initialized=false;
@ -175,39 +175,39 @@ void PmuStat::KNLevsetup(const char *ename, int &fd, int event, int umask)
}
void PmuStat::KNLsetup(void){
void PmuStat::KNLsetup(void){
int ret;
char fname[1024];
int ret;
char fname[1024];
// MC RPQ inserts and WPQ inserts (reads & writes)
for (int mc = 0; mc < NMC; ++mc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_imc_%d",mc);
// RPQ Inserts
KNLevsetup(fname, gbl.mc_rd[mc], 0x1, 0x1);
// WPQ Inserts
KNLevsetup(fname, gbl.mc_wr[mc], 0x2, 0x1);
}
// EDC RPQ inserts and WPQ inserts
for (int edc=0; edc < NEDC; ++edc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_eclk_%d",edc);
// RPQ inserts
KNLevsetup(fname, gbl.edc_rd[edc], 0x1, 0x1);
// WPQ inserts
KNLevsetup(fname, gbl.edc_wr[edc], 0x2, 0x1);
}
// EDC HitE, HitM, MissE, MissM
for (int edc=0; edc < NEDC; ++edc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_uclk_%d", edc);
KNLevsetup(fname, gbl.edc_hite[edc], 0x2, 0x1);
KNLevsetup(fname, gbl.edc_hitm[edc], 0x2, 0x2);
KNLevsetup(fname, gbl.edc_misse[edc], 0x2, 0x4);
KNLevsetup(fname, gbl.edc_missm[edc], 0x2, 0x8);
}
}
// MC RPQ inserts and WPQ inserts (reads & writes)
for (int mc = 0; mc < NMC; ++mc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_imc_%d",mc);
// RPQ Inserts
KNLevsetup(fname, gbl.mc_rd[mc], 0x1, 0x1);
// WPQ Inserts
KNLevsetup(fname, gbl.mc_wr[mc], 0x2, 0x1);
}
// EDC RPQ inserts and WPQ inserts
for (int edc=0; edc < NEDC; ++edc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_eclk_%d",edc);
// RPQ inserts
KNLevsetup(fname, gbl.edc_rd[edc], 0x1, 0x1);
// WPQ inserts
KNLevsetup(fname, gbl.edc_wr[edc], 0x2, 0x1);
}
// EDC HitE, HitM, MissE, MissM
for (int edc=0; edc < NEDC; ++edc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_uclk_%d", edc);
KNLevsetup(fname, gbl.edc_hite[edc], 0x2, 0x1);
KNLevsetup(fname, gbl.edc_hitm[edc], 0x2, 0x2);
KNLevsetup(fname, gbl.edc_misse[edc], 0x2, 0x4);
KNLevsetup(fname, gbl.edc_missm[edc], 0x2, 0x8);
}
}
uint64_t PmuStat::KNLreadctr(int fd)
{
@ -242,4 +242,5 @@ void PmuStat::KNLreadctrs(ctrs &c)
}
#endif
}
NAMESPACE_END(Grid);

111
Grid/perfmon/Stat.h
View File

@ -5,7 +5,7 @@
#define _KNIGHTS_LANDING_ROOTONLY
#endif
namespace Grid {
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////////////////////////
// Extra KNL counters from MCDRAM
@ -15,14 +15,14 @@ namespace Grid {
#define NEDC 8
struct ctrs
{
uint64_t mcrd[NMC];
uint64_t mcwr[NMC];
uint64_t edcrd[NEDC];
uint64_t edcwr[NEDC];
uint64_t edchite[NEDC];
uint64_t edchitm[NEDC];
uint64_t edcmisse[NEDC];
uint64_t edcmissm[NEDC];
uint64_t mcrd[NMC];
uint64_t mcwr[NMC];
uint64_t edcrd[NEDC];
uint64_t edcwr[NEDC];
uint64_t edchite[NEDC];
uint64_t edchitm[NEDC];
uint64_t edcmisse[NEDC];
uint64_t edcmissm[NEDC];
};
// Peter/Azusa:
// Our modification of a code provided by Larry Meadows from Intel
@ -44,61 +44,62 @@ struct knl_gbl_
class PmuStat
{
uint64_t counters[8][256];
uint64_t counters[8][256];
#ifdef _KNIGHTS_LANDING_
static struct knl_gbl_ gbl;
static struct knl_gbl_ gbl;
#endif
const char *name;
const char *name;
uint64_t reads; // memory reads
uint64_t writes; // memory writes
uint64_t mrstart; // memory read counter at start of parallel region
uint64_t mrend; // memory read counter at end of parallel region
uint64_t mwstart; // memory write counter at start of parallel region
uint64_t mwend; // memory write counter at end of parallel region
uint64_t reads; // memory reads
uint64_t writes; // memory writes
uint64_t mrstart; // memory read counter at start of parallel region
uint64_t mrend; // memory read counter at end of parallel region
uint64_t mwstart; // memory write counter at start of parallel region
uint64_t mwend; // memory write counter at end of parallel region
// cumulative counters
uint64_t count; // number of invocations
uint64_t tregion; // total time in parallel region (from thread 0)
uint64_t tcycles; // total cycles inside parallel region
uint64_t inst, ref, cyc; // fixed counters
uint64_t pmc0, pmc1;// pmu
// add memory counters here
// temp variables
uint64_t tstart; // tsc at start of parallel region
uint64_t tend; // tsc at end of parallel region
// map for ctrs values
// 0 pmc0 start
// 1 pmc0 end
// 2 pmc1 start
// 3 pmc1 end
// 4 tsc start
// 5 tsc end
static bool pmu_initialized;
// cumulative counters
uint64_t count; // number of invocations
uint64_t tregion; // total time in parallel region (from thread 0)
uint64_t tcycles; // total cycles inside parallel region
uint64_t inst, ref, cyc; // fixed counters
uint64_t pmc0, pmc1;// pmu
// add memory counters here
// temp variables
uint64_t tstart; // tsc at start of parallel region
uint64_t tend; // tsc at end of parallel region
// map for ctrs values
// 0 pmc0 start
// 1 pmc0 end
// 2 pmc1 start
// 3 pmc1 end
// 4 tsc start
// 5 tsc end
static bool pmu_initialized;
public:
static bool is_init(void){ return pmu_initialized;}
static void pmu_init(void);
static void pmu_fini(void);
static void pmu_start(void);
static void pmu_stop(void);
void accum(int nthreads);
static void xmemctrs(uint64_t *mr, uint64_t *mw);
void start(void);
void enter(int t);
void exit(int t);
void print(void);
void init(const char *regname);
void clear(void);
static bool is_init(void){ return pmu_initialized;}
static void pmu_init(void);
static void pmu_fini(void);
static void pmu_start(void);
static void pmu_stop(void);
void accum(int nthreads);
static void xmemctrs(uint64_t *mr, uint64_t *mw);
void start(void);
void enter(int t);
void exit(int t);
void print(void);
void init(const char *regname);
void clear(void);
#ifdef _KNIGHTS_LANDING_
static void KNLsetup(void);
static uint64_t KNLreadctr(int fd);
static void KNLreadctrs(ctrs &c);
static void KNLevsetup(const char *ename, int &fd, int event, int umask);
static void KNLsetup(void);
static uint64_t KNLreadctr(int fd);
static void KNLreadctrs(ctrs &c);
static void KNLevsetup(const char *ename, int &fd, int event, int umask);
#endif
};
};
NAMESPACE_END(Grid);
}
#endif

17
Grid/perfmon/Timer.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,8 +24,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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 */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_TIME_H
#define GRID_TIME_H
@ -33,11 +33,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <ctime>
#include <chrono>
namespace Grid {
// Dress the output; use std::chrono
NAMESPACE_BEGIN(Grid)
// Dress the output; use std::chrono
// C++11 time facilities better?
inline double usecond(void) {
struct timeval tv;
@ -80,7 +78,7 @@ inline std::ostream& operator<< (std::ostream & stream, const GridUsecs & now)
stream.fill(fill);
return stream;
}
class GridStopWatch {
private:
@ -125,5 +123,6 @@ public:
}
};
}
NAMESPACE_END(Grid)
#endif

13
Grid/pugixml/pugixml.cc
View File

@ -14,7 +14,12 @@
#ifndef SOURCE_PUGIXML_CPP
#define SOURCE_PUGIXML_CPP
#include <Grid/pugixml/pugixml.h>
#ifdef __NVCC__
#pragma push
#pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
#endif
#include "pugixml.h"
#include <stdlib.h>
#include <stdio.h>
@ -202,7 +207,7 @@ PUGI__NS_BEGIN
// Without a template<> we'll get multiple definitions of the same static
template <typename T> allocation_function xml_memory_management_function_storage<T>::allocate = default_allocate;
template <typename T> deallocation_function xml_memory_management_function_storage<T>::deallocate = default_deallocate;
template struct xml_memory_management_function_storage<int>;
typedef xml_memory_management_function_storage<int> xml_memory;
PUGI__NS_END
@ -12768,6 +12773,10 @@ namespace pugi
#undef PUGI__THROW_ERROR
#undef PUGI__CHECK_ERROR
#ifdef GRID_NVCC
#pragma pop
#endif
#endif
/**

794
Grid/qcd/QCD.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -27,114 +27,112 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_QCD_BASE_H
#define GRID_QCD_BASE_H
namespace Grid{
namespace QCD {
*************************************************************************************/
/* END LEGAL */
#pragma once
static const int Xdir = 0;
static const int Ydir = 1;
static const int Zdir = 2;
static const int Tdir = 3;
NAMESPACE_BEGIN(Grid);
static const int Xp = 0;
static const int Yp = 1;
static const int Zp = 2;
static const int Tp = 3;
static const int Xm = 4;
static const int Ym = 5;
static const int Zm = 6;
static const int Tm = 7;
static constexpr int Xdir = 0;
static constexpr int Ydir = 1;
static constexpr int Zdir = 2;
static constexpr int Tdir = 3;
static const int Nc=3;
static const int Ns=4;
static const int Nd=4;
static const int Nhs=2; // half spinor
static const int Nds=8; // double stored gauge field
static const int Ngp=2; // gparity index range
static constexpr int Xp = 0;
static constexpr int Yp = 1;
static constexpr int Zp = 2;
static constexpr int Tp = 3;
static constexpr int Xm = 4;
static constexpr int Ym = 5;
static constexpr int Zm = 6;
static constexpr int Tm = 7;
//////////////////////////////////////////////////////////////////////////////
// QCD iMatrix types
// Index conventions: Lorentz x Spin x Colour
// note: static const int or constexpr will work for type deductions
// with the intel compiler (up to version 17)
//////////////////////////////////////////////////////////////////////////////
#define ColourIndex 2
#define SpinIndex 1
#define LorentzIndex 0
static constexpr int Nc=3;
static constexpr int Ns=4;
static constexpr int Nd=4;
static constexpr int Nhs=2; // half spinor
static constexpr int Nds=8; // double stored gauge field
static constexpr int Ngp=2; // gparity index range
// Also should make these a named enum type
static const int DaggerNo=0;
static const int DaggerYes=1;
static const int InverseNo=0;
static const int InverseYes=1;
//////////////////////////////////////////////////////////////////////////////
// QCD iMatrix types
// Index conventions: Lorentz x Spin x Colour
// note: static constexpr int or constexpr will work for type deductions
// with the intel compiler (up to version 17)
//////////////////////////////////////////////////////////////////////////////
#define ColourIndex (2)
#define SpinIndex (1)
#define LorentzIndex (0)
// Useful traits is this a spin index
//typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
// Also should make these a named enum type
static constexpr int DaggerNo=0;
static constexpr int DaggerYes=1;
static constexpr int InverseNo=0;
static constexpr int InverseYes=1;
const int SpinorIndex = 2;
template<typename T> struct isSpinor {
static const bool value = (SpinorIndex==T::TensorLevel);
};
template <typename T> using IfSpinor = Invoke<std::enable_if< isSpinor<T>::value,int> > ;
template <typename T> using IfNotSpinor = Invoke<std::enable_if<!isSpinor<T>::value,int> > ;
// Useful traits is this a spin index
//typename std::enable_if<matchGridTensorIndex<iVector<vtype,Ns>,SpinorIndex>::value,iVector<vtype,Ns> >::type *SFINAE;
// ChrisK very keen to add extra space for Gparity doubling.
//
// Also add domain wall index, in a way where Wilson operator
// naturally distributes across the 5th dimensions.
//
// That probably makes for GridRedBlack4dCartesian grid.
const int SpinorIndex = 2;
template<typename T> struct isSpinor {
static constexpr bool value = (SpinorIndex==T::TensorLevel);
};
template <typename T> using IfSpinor = Invoke<std::enable_if< isSpinor<T>::value,int> > ;
template <typename T> using IfNotSpinor = Invoke<std::enable_if<!isSpinor<T>::value,int> > ;
// s,sp,c,spc,lc
// ChrisK very keen to add extra space for Gparity doubling.
//
// Also add domain wall index, in a way where Wilson operator
// naturally distributes across the 5th dimensions.
//
// That probably makes for GridRedBlack4dCartesian grid.
template<typename vtype> using iSinglet = iScalar<iScalar<iScalar<vtype> > >;
template<typename vtype> using iSpinMatrix = iScalar<iMatrix<iScalar<vtype>, Ns> >;
template<typename vtype> using iColourMatrix = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
template<typename vtype> using iSpinColourMatrix = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >;
template<typename vtype> using iLorentzColourMatrix = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ;
template<typename vtype> using iDoubleStoredColourMatrix = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ;
template<typename vtype> using iSpinVector = iScalar<iVector<iScalar<vtype>, Ns> >;
template<typename vtype> using iColourVector = iScalar<iScalar<iVector<vtype, Nc> > >;
template<typename vtype> using iSpinColourVector = iScalar<iVector<iVector<vtype, Nc>, Ns> >;
template<typename vtype> using iHalfSpinVector = iScalar<iVector<iScalar<vtype>, Nhs> >;
template<typename vtype> using iHalfSpinColourVector = iScalar<iVector<iVector<vtype, Nc>, Nhs> >;
// s,sp,c,spc,lc
template<typename vtype> using iSinglet = iScalar<iScalar<iScalar<vtype> > >;
template<typename vtype> using iSpinMatrix = iScalar<iMatrix<iScalar<vtype>, Ns> >;
template<typename vtype> using iColourMatrix = iScalar<iScalar<iMatrix<vtype, Nc> > > ;
template<typename vtype> using iSpinColourMatrix = iScalar<iMatrix<iMatrix<vtype, Nc>, Ns> >;
template<typename vtype> using iLorentzColourMatrix = iVector<iScalar<iMatrix<vtype, Nc> >, Nd > ;
template<typename vtype> using iDoubleStoredColourMatrix = iVector<iScalar<iMatrix<vtype, Nc> >, Nds > ;
template<typename vtype> using iSpinVector = iScalar<iVector<iScalar<vtype>, Ns> >;
template<typename vtype> using iColourVector = iScalar<iScalar<iVector<vtype, Nc> > >;
template<typename vtype> using iSpinColourVector = iScalar<iVector<iVector<vtype, Nc>, Ns> >;
template<typename vtype> using iHalfSpinVector = iScalar<iVector<iScalar<vtype>, Nhs> >;
template<typename vtype> using iHalfSpinColourVector = iScalar<iVector<iVector<vtype, Nc>, Nhs> >;
template<typename vtype> using iSpinColourSpinColourMatrix = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >;
template<typename vtype> using iGparitySpinColourVector = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >;
template<typename vtype> using iGparityHalfSpinColourVector = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >;
template<typename vtype> using iGparitySpinColourVector = iVector<iVector<iVector<vtype, Nc>, Ns>, Ngp >;
template<typename vtype> using iGparityHalfSpinColourVector = iVector<iVector<iVector<vtype, Nc>, Nhs>, Ngp >;
// Spin matrix
typedef iSpinMatrix<Complex > SpinMatrix;
typedef iSpinMatrix<ComplexF > SpinMatrixF;
typedef iSpinMatrix<ComplexD > SpinMatrixD;
// Spin matrix
typedef iSpinMatrix<Complex > SpinMatrix;
typedef iSpinMatrix<ComplexF > SpinMatrixF;
typedef iSpinMatrix<ComplexD > SpinMatrixD;
typedef iSpinMatrix<vComplex > vSpinMatrix;
typedef iSpinMatrix<vComplexF> vSpinMatrixF;
typedef iSpinMatrix<vComplexD> vSpinMatrixD;
typedef iSpinMatrix<vComplex > vSpinMatrix;
typedef iSpinMatrix<vComplexF> vSpinMatrixF;
typedef iSpinMatrix<vComplexD> vSpinMatrixD;
// Colour Matrix
typedef iColourMatrix<Complex > ColourMatrix;
typedef iColourMatrix<ComplexF > ColourMatrixF;
typedef iColourMatrix<ComplexD > ColourMatrixD;
// Colour Matrix
typedef iColourMatrix<Complex > ColourMatrix;
typedef iColourMatrix<ComplexF > ColourMatrixF;
typedef iColourMatrix<ComplexD > ColourMatrixD;
typedef iColourMatrix<vComplex > vColourMatrix;
typedef iColourMatrix<vComplexF> vColourMatrixF;
typedef iColourMatrix<vComplexD> vColourMatrixD;
typedef iColourMatrix<vComplex > vColourMatrix;
typedef iColourMatrix<vComplexF> vColourMatrixF;
typedef iColourMatrix<vComplexD> vColourMatrixD;
// SpinColour matrix
typedef iSpinColourMatrix<Complex > SpinColourMatrix;
typedef iSpinColourMatrix<ComplexF > SpinColourMatrixF;
typedef iSpinColourMatrix<ComplexD > SpinColourMatrixD;
typedef iSpinColourMatrix<vComplex > vSpinColourMatrix;
typedef iSpinColourMatrix<vComplexF> vSpinColourMatrixF;
typedef iSpinColourMatrix<vComplexD> vSpinColourMatrixD;
// SpinColour matrix
typedef iSpinColourMatrix<Complex > SpinColourMatrix;
typedef iSpinColourMatrix<ComplexF > SpinColourMatrixF;
typedef iSpinColourMatrix<ComplexD > SpinColourMatrixD;
typedef iSpinColourMatrix<vComplex > vSpinColourMatrix;
typedef iSpinColourMatrix<vComplexF> vSpinColourMatrixF;
typedef iSpinColourMatrix<vComplexD> vSpinColourMatrixD;
// SpinColourSpinColour matrix
typedef iSpinColourSpinColourMatrix<Complex > SpinColourSpinColourMatrix;
typedef iSpinColourSpinColourMatrix<ComplexF > SpinColourSpinColourMatrixF;
@ -153,383 +151,379 @@ namespace QCD {
typedef iSpinColourSpinColourMatrix<vComplexF> vSpinColourSpinColourMatrixF;
typedef iSpinColourSpinColourMatrix<vComplexD> vSpinColourSpinColourMatrixD;
// LorentzColour
typedef iLorentzColourMatrix<Complex > LorentzColourMatrix;
typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF;
typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD;
// LorentzColour
typedef iLorentzColourMatrix<Complex > LorentzColourMatrix;
typedef iLorentzColourMatrix<ComplexF > LorentzColourMatrixF;
typedef iLorentzColourMatrix<ComplexD > LorentzColourMatrixD;
typedef iLorentzColourMatrix<vComplex > vLorentzColourMatrix;
typedef iLorentzColourMatrix<vComplexF> vLorentzColourMatrixF;
typedef iLorentzColourMatrix<vComplexD> vLorentzColourMatrixD;
typedef iLorentzColourMatrix<vComplex > vLorentzColourMatrix;
typedef iLorentzColourMatrix<vComplexF> vLorentzColourMatrixF;
typedef iLorentzColourMatrix<vComplexD> vLorentzColourMatrixD;
// DoubleStored gauge field
typedef iDoubleStoredColourMatrix<Complex > DoubleStoredColourMatrix;
typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF;
typedef iDoubleStoredColourMatrix<ComplexD > DoubleStoredColourMatrixD;
// DoubleStored gauge field
typedef iDoubleStoredColourMatrix<Complex > DoubleStoredColourMatrix;
typedef iDoubleStoredColourMatrix<ComplexF > DoubleStoredColourMatrixF;
typedef iDoubleStoredColourMatrix<ComplexD > DoubleStoredColourMatrixD;
typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
// Spin vector
typedef iSpinVector<Complex > SpinVector;
typedef iSpinVector<ComplexF> SpinVectorF;
typedef iSpinVector<ComplexD> SpinVectorD;
// Spin vector
typedef iSpinVector<Complex > SpinVector;
typedef iSpinVector<ComplexF> SpinVectorF;
typedef iSpinVector<ComplexD> SpinVectorD;
typedef iSpinVector<vComplex > vSpinVector;
typedef iSpinVector<vComplexF> vSpinVectorF;
typedef iSpinVector<vComplexD> vSpinVectorD;
typedef iSpinVector<vComplex > vSpinVector;
typedef iSpinVector<vComplexF> vSpinVectorF;
typedef iSpinVector<vComplexD> vSpinVectorD;
// Colour vector
typedef iColourVector<Complex > ColourVector;
typedef iColourVector<ComplexF> ColourVectorF;
typedef iColourVector<ComplexD> ColourVectorD;
// Colour vector
typedef iColourVector<Complex > ColourVector;
typedef iColourVector<ComplexF> ColourVectorF;
typedef iColourVector<ComplexD> ColourVectorD;
typedef iColourVector<vComplex > vColourVector;
typedef iColourVector<vComplexF> vColourVectorF;
typedef iColourVector<vComplexD> vColourVectorD;
typedef iColourVector<vComplex > vColourVector;
typedef iColourVector<vComplexF> vColourVectorF;
typedef iColourVector<vComplexD> vColourVectorD;
// SpinColourVector
typedef iSpinColourVector<Complex > SpinColourVector;
typedef iSpinColourVector<ComplexF> SpinColourVectorF;
typedef iSpinColourVector<ComplexD> SpinColourVectorD;
// SpinColourVector
typedef iSpinColourVector<Complex > SpinColourVector;
typedef iSpinColourVector<ComplexF> SpinColourVectorF;
typedef iSpinColourVector<ComplexD> SpinColourVectorD;
typedef iSpinColourVector<vComplex > vSpinColourVector;
typedef iSpinColourVector<vComplexF> vSpinColourVectorF;
typedef iSpinColourVector<vComplexD> vSpinColourVectorD;
typedef iSpinColourVector<vComplex > vSpinColourVector;
typedef iSpinColourVector<vComplexF> vSpinColourVectorF;
typedef iSpinColourVector<vComplexD> vSpinColourVectorD;
// HalfSpin vector
typedef iHalfSpinVector<Complex > HalfSpinVector;
typedef iHalfSpinVector<ComplexF> HalfSpinVectorF;
typedef iHalfSpinVector<ComplexD> HalfSpinVectorD;
// HalfSpin vector
typedef iHalfSpinVector<Complex > HalfSpinVector;
typedef iHalfSpinVector<ComplexF> HalfSpinVectorF;
typedef iHalfSpinVector<ComplexD> HalfSpinVectorD;
typedef iHalfSpinVector<vComplex > vHalfSpinVector;
typedef iHalfSpinVector<vComplexF> vHalfSpinVectorF;
typedef iHalfSpinVector<vComplexD> vHalfSpinVectorD;
typedef iHalfSpinVector<vComplex > vHalfSpinVector;
typedef iHalfSpinVector<vComplexF> vHalfSpinVectorF;
typedef iHalfSpinVector<vComplexD> vHalfSpinVectorD;
// HalfSpinColour vector
typedef iHalfSpinColourVector<Complex > HalfSpinColourVector;
typedef iHalfSpinColourVector<ComplexF> HalfSpinColourVectorF;
typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
// HalfSpinColour vector
typedef iHalfSpinColourVector<Complex > HalfSpinColourVector;
typedef iHalfSpinColourVector<ComplexF> HalfSpinColourVectorF;
typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
// singlets
typedef iSinglet<Complex > TComplex; // FIXME This is painful. Tensor singlet complex type.
typedef iSinglet<ComplexF> TComplexF; // FIXME This is painful. Tensor singlet complex type.
typedef iSinglet<ComplexD> TComplexD; // FIXME This is painful. Tensor singlet complex type.
// singlets
typedef iSinglet<Complex > TComplex; // FIXME This is painful. Tensor singlet complex type.
typedef iSinglet<ComplexF> TComplexF; // FIXME This is painful. Tensor singlet complex type.
typedef iSinglet<ComplexD> TComplexD; // FIXME This is painful. Tensor singlet complex type.
typedef iSinglet<vComplex > vTComplex ; // what if we don't know the tensor structure
typedef iSinglet<vComplexF> vTComplexF; // what if we don't know the tensor structure
typedef iSinglet<vComplexD> vTComplexD; // what if we don't know the tensor structure
typedef iSinglet<vComplex > vTComplex ; // what if we don't know the tensor structure
typedef iSinglet<vComplexF> vTComplexF; // what if we don't know the tensor structure
typedef iSinglet<vComplexD> vTComplexD; // what if we don't know the tensor structure
typedef iSinglet<Real > TReal; // Shouldn't need these; can I make it work without?
typedef iSinglet<RealF> TRealF; // Shouldn't need these; can I make it work without?
typedef iSinglet<RealD> TRealD; // Shouldn't need these; can I make it work without?
typedef iSinglet<Real > TReal; // Shouldn't need these; can I make it work without?
typedef iSinglet<RealF> TRealF; // Shouldn't need these; can I make it work without?
typedef iSinglet<RealD> TRealD; // Shouldn't need these; can I make it work without?
typedef iSinglet<vReal > vTReal;
typedef iSinglet<vRealF> vTRealF;
typedef iSinglet<vRealD> vTRealD;
typedef iSinglet<vReal > vTReal;
typedef iSinglet<vRealF> vTRealF;
typedef iSinglet<vRealD> vTRealD;
typedef iSinglet<vInteger> vTInteger;
typedef iSinglet<Integer > TInteger;
typedef iSinglet<vInteger> vTInteger;
typedef iSinglet<Integer > TInteger;
// Lattices of these
typedef Lattice<vColourMatrix> LatticeColourMatrix;
typedef Lattice<vColourMatrixF> LatticeColourMatrixF;
typedef Lattice<vColourMatrixD> LatticeColourMatrixD;
// Lattices of these
typedef Lattice<vColourMatrix> LatticeColourMatrix;
typedef Lattice<vColourMatrixF> LatticeColourMatrixF;
typedef Lattice<vColourMatrixD> LatticeColourMatrixD;
typedef Lattice<vSpinMatrix> LatticeSpinMatrix;
typedef Lattice<vSpinMatrixF> LatticeSpinMatrixF;
typedef Lattice<vSpinMatrixD> LatticeSpinMatrixD;
typedef Lattice<vSpinMatrix> LatticeSpinMatrix;
typedef Lattice<vSpinMatrixF> LatticeSpinMatrixF;
typedef Lattice<vSpinMatrixD> LatticeSpinMatrixD;
typedef Lattice<vSpinColourMatrix> LatticeSpinColourMatrix;
typedef Lattice<vSpinColourMatrixF> LatticeSpinColourMatrixF;
typedef Lattice<vSpinColourMatrixD> LatticeSpinColourMatrixD;
typedef Lattice<vSpinColourMatrix> LatticeSpinColourMatrix;
typedef Lattice<vSpinColourMatrixF> LatticeSpinColourMatrixF;
typedef Lattice<vSpinColourMatrixD> LatticeSpinColourMatrixD;
typedef Lattice<vSpinColourSpinColourMatrix> LatticeSpinColourSpinColourMatrix;
typedef Lattice<vSpinColourSpinColourMatrixF> LatticeSpinColourSpinColourMatrixF;
typedef Lattice<vSpinColourSpinColourMatrixD> LatticeSpinColourSpinColourMatrixD;
typedef Lattice<vSpinColourSpinColourMatrix> LatticeSpinColourSpinColourMatrix;
typedef Lattice<vSpinColourSpinColourMatrixF> LatticeSpinColourSpinColourMatrixF;
typedef Lattice<vSpinColourSpinColourMatrixD> LatticeSpinColourSpinColourMatrixD;
typedef Lattice<vLorentzColourMatrix> LatticeLorentzColourMatrix;
typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF;
typedef Lattice<vLorentzColourMatrixD> LatticeLorentzColourMatrixD;
typedef Lattice<vLorentzColourMatrix> LatticeLorentzColourMatrix;
typedef Lattice<vLorentzColourMatrixF> LatticeLorentzColourMatrixF;
typedef Lattice<vLorentzColourMatrixD> LatticeLorentzColourMatrixD;
// DoubleStored gauge field
typedef Lattice<vDoubleStoredColourMatrix> LatticeDoubleStoredColourMatrix;
typedef Lattice<vDoubleStoredColourMatrixF> LatticeDoubleStoredColourMatrixF;
typedef Lattice<vDoubleStoredColourMatrixD> LatticeDoubleStoredColourMatrixD;
// DoubleStored gauge field
typedef Lattice<vDoubleStoredColourMatrix> LatticeDoubleStoredColourMatrix;
typedef Lattice<vDoubleStoredColourMatrixF> LatticeDoubleStoredColourMatrixF;
typedef Lattice<vDoubleStoredColourMatrixD> LatticeDoubleStoredColourMatrixD;
typedef Lattice<vSpinVector> LatticeSpinVector;
typedef Lattice<vSpinVectorF> LatticeSpinVectorF;
typedef Lattice<vSpinVectorD> LatticeSpinVectorD;
typedef Lattice<vSpinVector> LatticeSpinVector;
typedef Lattice<vSpinVectorF> LatticeSpinVectorF;
typedef Lattice<vSpinVectorD> LatticeSpinVectorD;
typedef Lattice<vColourVector> LatticeColourVector;
typedef Lattice<vColourVectorF> LatticeColourVectorF;
typedef Lattice<vColourVectorD> LatticeColourVectorD;
typedef Lattice<vColourVector> LatticeColourVector;
typedef Lattice<vColourVectorF> LatticeColourVectorF;
typedef Lattice<vColourVectorD> LatticeColourVectorD;
typedef Lattice<vSpinColourVector> LatticeSpinColourVector;
typedef Lattice<vSpinColourVectorF> LatticeSpinColourVectorF;
typedef Lattice<vSpinColourVectorD> LatticeSpinColourVectorD;
typedef Lattice<vSpinColourVector> LatticeSpinColourVector;
typedef Lattice<vSpinColourVectorF> LatticeSpinColourVectorF;
typedef Lattice<vSpinColourVectorD> LatticeSpinColourVectorD;
typedef Lattice<vHalfSpinVector> LatticeHalfSpinVector;
typedef Lattice<vHalfSpinVectorF> LatticeHalfSpinVectorF;
typedef Lattice<vHalfSpinVectorD> LatticeHalfSpinVectorD;
typedef Lattice<vHalfSpinVector> LatticeHalfSpinVector;
typedef Lattice<vHalfSpinVectorF> LatticeHalfSpinVectorF;
typedef Lattice<vHalfSpinVectorD> LatticeHalfSpinVectorD;
typedef Lattice<vHalfSpinColourVector> LatticeHalfSpinColourVector;
typedef Lattice<vHalfSpinColourVectorF> LatticeHalfSpinColourVectorF;
typedef Lattice<vHalfSpinColourVectorD> LatticeHalfSpinColourVectorD;
typedef Lattice<vHalfSpinColourVector> LatticeHalfSpinColourVector;
typedef Lattice<vHalfSpinColourVectorF> LatticeHalfSpinColourVectorF;
typedef Lattice<vHalfSpinColourVectorD> LatticeHalfSpinColourVectorD;
typedef Lattice<vTReal> LatticeReal;
typedef Lattice<vTRealF> LatticeRealF;
typedef Lattice<vTRealD> LatticeRealD;
typedef Lattice<vTReal> LatticeReal;
typedef Lattice<vTRealF> LatticeRealF;
typedef Lattice<vTRealD> LatticeRealD;
typedef Lattice<vTComplex> LatticeComplex;
typedef Lattice<vTComplexF> LatticeComplexF;
typedef Lattice<vTComplexD> LatticeComplexD;
typedef Lattice<vTComplex> LatticeComplex;
typedef Lattice<vTComplexF> LatticeComplexF;
typedef Lattice<vTComplexD> LatticeComplexD;
typedef Lattice<vTInteger> LatticeInteger; // Predicates for "where"
typedef Lattice<vTInteger> LatticeInteger; // Predicates for "where"
///////////////////////////////////////////
// Physical names for things
///////////////////////////////////////////
typedef LatticeHalfSpinColourVector LatticeHalfFermion;
typedef LatticeHalfSpinColourVectorF LatticeHalfFermionF;
typedef LatticeHalfSpinColourVectorF LatticeHalfFermionD;
///////////////////////////////////////////
// Physical names for things
///////////////////////////////////////////
typedef LatticeHalfSpinColourVector LatticeHalfFermion;
typedef LatticeHalfSpinColourVectorF LatticeHalfFermionF;
typedef LatticeHalfSpinColourVectorF LatticeHalfFermionD;
typedef LatticeSpinColourVector LatticeFermion;
typedef LatticeSpinColourVectorF LatticeFermionF;
typedef LatticeSpinColourVectorD LatticeFermionD;
typedef LatticeSpinColourVector LatticeFermion;
typedef LatticeSpinColourVectorF LatticeFermionF;
typedef LatticeSpinColourVectorD LatticeFermionD;
typedef LatticeSpinColourMatrix LatticePropagator;
typedef LatticeSpinColourMatrixF LatticePropagatorF;
typedef LatticeSpinColourMatrixD LatticePropagatorD;
typedef LatticeSpinColourMatrix LatticePropagator;
typedef LatticeSpinColourMatrixF LatticePropagatorF;
typedef LatticeSpinColourMatrixD LatticePropagatorD;
typedef LatticeLorentzColourMatrix LatticeGaugeField;
typedef LatticeLorentzColourMatrixF LatticeGaugeFieldF;
typedef LatticeLorentzColourMatrixD LatticeGaugeFieldD;
typedef LatticeLorentzColourMatrix LatticeGaugeField;
typedef LatticeLorentzColourMatrixF LatticeGaugeFieldF;
typedef LatticeLorentzColourMatrixD LatticeGaugeFieldD;
typedef LatticeDoubleStoredColourMatrix LatticeDoubledGaugeField;
typedef LatticeDoubleStoredColourMatrixF LatticeDoubledGaugeFieldF;
typedef LatticeDoubleStoredColourMatrixD LatticeDoubledGaugeFieldD;
typedef LatticeDoubleStoredColourMatrix LatticeDoubledGaugeField;
typedef LatticeDoubleStoredColourMatrixF LatticeDoubledGaugeFieldF;
typedef LatticeDoubleStoredColourMatrixD LatticeDoubledGaugeFieldD;
template<class GF> using LorentzScalar = Lattice<iScalar<typename GF::vector_object::element> >;
template<class GF> using LorentzScalar = Lattice<iScalar<typename GF::vector_object::element> >;
// Uhgg... typing this hurt ;)
// (my keyboard got burning hot when I typed this, must be the anti-Fermion)
typedef Lattice<vColourVector> LatticeStaggeredFermion;
typedef Lattice<vColourVectorF> LatticeStaggeredFermionF;
typedef Lattice<vColourVectorD> LatticeStaggeredFermionD;
// Uhgg... typing this hurt ;)
// (my keyboard got burning hot when I typed this, must be the anti-Fermion)
typedef Lattice<vColourVector> LatticeStaggeredFermion;
typedef Lattice<vColourVectorF> LatticeStaggeredFermionF;
typedef Lattice<vColourVectorD> LatticeStaggeredFermionD;
typedef Lattice<vColourMatrix> LatticeStaggeredPropagator;
typedef Lattice<vColourMatrixF> LatticeStaggeredPropagatorF;
typedef Lattice<vColourMatrixD> LatticeStaggeredPropagatorD;
typedef Lattice<vColourMatrix> LatticeStaggeredPropagator;
typedef Lattice<vColourMatrixF> LatticeStaggeredPropagatorF;
typedef Lattice<vColourMatrixD> LatticeStaggeredPropagatorD;
//////////////////////////////////////////////////////////////////////////////
// Peek and Poke named after physics attributes
//////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////
// Peek and Poke named after physics attributes
//////////////////////////////////////////////////////////////////////////////
//spin
template<class vobj> auto peekSpin(const vobj &rhs,int i) -> decltype(PeekIndex<SpinIndex>(rhs,0))
{
return PeekIndex<SpinIndex>(rhs,i);
}
template<class vobj> auto peekSpin(const vobj &rhs,int i,int j) -> decltype(PeekIndex<SpinIndex>(rhs,0,0))
{
return PeekIndex<SpinIndex>(rhs,i,j);
}
template<class vobj> auto peekSpin(const Lattice<vobj> &rhs,int i) -> decltype(PeekIndex<SpinIndex>(rhs,0))
{
return PeekIndex<SpinIndex>(rhs,i);
}
template<class vobj> auto peekSpin(const Lattice<vobj> &rhs,int i,int j) -> decltype(PeekIndex<SpinIndex>(rhs,0,0))
{
return PeekIndex<SpinIndex>(rhs,i,j);
}
//colour
template<class vobj> auto peekColour(const vobj &rhs,int i) -> decltype(PeekIndex<ColourIndex>(rhs,0))
{
return PeekIndex<ColourIndex>(rhs,i);
}
template<class vobj> auto peekColour(const vobj &rhs,int i,int j) -> decltype(PeekIndex<ColourIndex>(rhs,0,0))
{
return PeekIndex<ColourIndex>(rhs,i,j);
}
template<class vobj> auto peekColour(const Lattice<vobj> &rhs,int i) -> decltype(PeekIndex<ColourIndex>(rhs,0))
{
return PeekIndex<ColourIndex>(rhs,i);
}
template<class vobj> auto peekColour(const Lattice<vobj> &rhs,int i,int j) -> decltype(PeekIndex<ColourIndex>(rhs,0,0))
{
return PeekIndex<ColourIndex>(rhs,i,j);
}
//lorentz
template<class vobj> auto peekLorentz(const vobj &rhs,int i) -> decltype(PeekIndex<LorentzIndex>(rhs,0))
{
return PeekIndex<LorentzIndex>(rhs,i);
}
template<class vobj> auto peekLorentz(const Lattice<vobj> &rhs,int i) -> decltype(PeekIndex<LorentzIndex>(rhs,0))
{
return PeekIndex<LorentzIndex>(rhs,i);
}
//spin
template<class vobj> auto peekSpin(const vobj &rhs,int i) -> decltype(PeekIndex<SpinIndex>(rhs,0))
{
return PeekIndex<SpinIndex>(rhs,i);
}
template<class vobj> auto peekSpin(const vobj &rhs,int i,int j) -> decltype(PeekIndex<SpinIndex>(rhs,0,0))
{
return PeekIndex<SpinIndex>(rhs,i,j);
}
template<class vobj> auto peekSpin(const Lattice<vobj> &rhs,int i) -> decltype(PeekIndex<SpinIndex>(rhs,0))
{
return PeekIndex<SpinIndex>(rhs,i);
}
template<class vobj> auto peekSpin(const Lattice<vobj> &rhs,int i,int j) -> decltype(PeekIndex<SpinIndex>(rhs,0,0))
{
return PeekIndex<SpinIndex>(rhs,i,j);
}
//colour
template<class vobj> auto peekColour(const vobj &rhs,int i) -> decltype(PeekIndex<ColourIndex>(rhs,0))
{
return PeekIndex<ColourIndex>(rhs,i);
}
template<class vobj> auto peekColour(const vobj &rhs,int i,int j) -> decltype(PeekIndex<ColourIndex>(rhs,0,0))
{
return PeekIndex<ColourIndex>(rhs,i,j);
}
template<class vobj> auto peekColour(const Lattice<vobj> &rhs,int i) -> decltype(PeekIndex<ColourIndex>(rhs,0))
{
return PeekIndex<ColourIndex>(rhs,i);
}
template<class vobj> auto peekColour(const Lattice<vobj> &rhs,int i,int j) -> decltype(PeekIndex<ColourIndex>(rhs,0,0))
{
return PeekIndex<ColourIndex>(rhs,i,j);
}
//lorentz
template<class vobj> auto peekLorentz(const vobj &rhs,int i) -> decltype(PeekIndex<LorentzIndex>(rhs,0))
{
return PeekIndex<LorentzIndex>(rhs,i);
}
template<class vobj> auto peekLorentz(const Lattice<vobj> &rhs,int i) -> decltype(PeekIndex<LorentzIndex>(rhs,0))
{
return PeekIndex<LorentzIndex>(rhs,i);
}
//////////////////////////////////////////////
// Poke lattice
//////////////////////////////////////////////
template<class vobj>
void pokeColour(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<ColourIndex>(lhs._odata[0],0))> & rhs,
//////////////////////////////////////////////
// Poke lattice
//////////////////////////////////////////////
template<class vobj>
void pokeColour(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<ColourIndex>(vobj(),0))> & rhs,
int i)
{
PokeIndex<ColourIndex>(lhs,rhs,i);
}
template<class vobj>
void pokeColour(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<ColourIndex>(vobj(),0,0))> & rhs,
int i,int j)
{
PokeIndex<ColourIndex>(lhs,rhs,i,j);
}
template<class vobj>
void pokeSpin(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<SpinIndex>(vobj(),0))> & rhs,
int i)
{
PokeIndex<ColourIndex>(lhs,rhs,i);
}
template<class vobj>
void pokeColour(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<ColourIndex>(lhs._odata[0],0,0))> & rhs,
{
PokeIndex<SpinIndex>(lhs,rhs,i);
}
template<class vobj>
void pokeSpin(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<SpinIndex>(vobj(),0,0))> & rhs,
int i,int j)
{
PokeIndex<ColourIndex>(lhs,rhs,i,j);
}
template<class vobj>
void pokeSpin(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<SpinIndex>(lhs._odata[0],0))> & rhs,
int i)
{
PokeIndex<SpinIndex>(lhs,rhs,i);
}
template<class vobj>
void pokeSpin(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<SpinIndex>(lhs._odata[0],0,0))> & rhs,
int i,int j)
{
PokeIndex<SpinIndex>(lhs,rhs,i,j);
}
template<class vobj>
void pokeLorentz(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<LorentzIndex>(lhs._odata[0],0))> & rhs,
int i)
{
PokeIndex<LorentzIndex>(lhs,rhs,i);
}
{
PokeIndex<SpinIndex>(lhs,rhs,i,j);
}
template<class vobj>
void pokeLorentz(Lattice<vobj> &lhs,
const Lattice<decltype(peekIndex<LorentzIndex>(vobj(),0))> & rhs,
int i)
{
PokeIndex<LorentzIndex>(lhs,rhs,i);
}
//////////////////////////////////////////////
// Poke scalars
//////////////////////////////////////////////
template<class vobj> void pokeSpin(vobj &lhs,const decltype(peekIndex<SpinIndex>(lhs,0)) & rhs,int i)
{
pokeIndex<SpinIndex>(lhs,rhs,i);
}
template<class vobj> void pokeSpin(vobj &lhs,const decltype(peekIndex<SpinIndex>(lhs,0,0)) & rhs,int i,int j)
{
pokeIndex<SpinIndex>(lhs,rhs,i,j);
}
//////////////////////////////////////////////
// Poke scalars
//////////////////////////////////////////////
template<class vobj> void pokeSpin(vobj &lhs,const decltype(peekIndex<SpinIndex>(lhs,0)) & rhs,int i)
{
pokeIndex<SpinIndex>(lhs,rhs,i);
}
template<class vobj> void pokeSpin(vobj &lhs,const decltype(peekIndex<SpinIndex>(lhs,0,0)) & rhs,int i,int j)
{
pokeIndex<SpinIndex>(lhs,rhs,i,j);
}
template<class vobj> void pokeColour(vobj &lhs,const decltype(peekIndex<ColourIndex>(lhs,0)) & rhs,int i)
{
pokeIndex<ColourIndex>(lhs,rhs,i);
}
template<class vobj> void pokeColour(vobj &lhs,const decltype(peekIndex<ColourIndex>(lhs,0,0)) & rhs,int i,int j)
{
pokeIndex<ColourIndex>(lhs,rhs,i,j);
}
template<class vobj> void pokeColour(vobj &lhs,const decltype(peekIndex<ColourIndex>(lhs,0)) & rhs,int i)
{
pokeIndex<ColourIndex>(lhs,rhs,i);
}
template<class vobj> void pokeColour(vobj &lhs,const decltype(peekIndex<ColourIndex>(lhs,0,0)) & rhs,int i,int j)
{
pokeIndex<ColourIndex>(lhs,rhs,i,j);
}
template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<LorentzIndex>(lhs,0)) & rhs,int i)
{
pokeIndex<LorentzIndex>(lhs,rhs,i);
}
template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<LorentzIndex>(lhs,0)) & rhs,int i)
{
pokeIndex<LorentzIndex>(lhs,rhs,i);
}
//////////////////////////////////////////////
// Fermion <-> propagator assignements
//////////////////////////////////////////////
//////////////////////////////////////////////
// Fermion <-> propagator assignements
//////////////////////////////////////////////
//template <class Prop, class Ferm>
template <class Fimpl>
void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
{
for(int j = 0; j < Ns; ++j)
{
for(int j = 0; j < Ns; ++j)
{
auto pjs = peekSpin(p, j, s);
auto fj = peekSpin(f, j);
auto pjs = peekSpin(p, j, s);
auto fj = peekSpin(f, j);
for(int i = 0; i < Fimpl::Dimension; ++i)
{
pokeColour(pjs, peekColour(fj, i), i, c);
}
pokeSpin(p, pjs, j, s);
}
{
pokeColour(pjs, peekColour(fj, i), i, c);
}
pokeSpin(p, pjs, j, s);
}
}
//template <class Prop, class Ferm>
template <class Fimpl>
void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
{
for(int j = 0; j < Ns; ++j)
{
for(int j = 0; j < Ns; ++j)
{
auto pjs = peekSpin(p, j, s);
auto fj = peekSpin(f, j);
auto pjs = peekSpin(p, j, s);
auto fj = peekSpin(f, j);
for(int i = 0; i < Fimpl::Dimension; ++i)
{
pokeColour(fj, peekColour(pjs, i, c), i);
}
pokeSpin(f, fj, j);
}
{
pokeColour(fj, peekColour(pjs, i, c), i);
}
pokeSpin(f, fj, j);
}
}
//////////////////////////////////////////////
// transpose array and scalar
//////////////////////////////////////////////
template<int Index,class vobj> inline Lattice<vobj> transposeSpin(const Lattice<vobj> &lhs){
return transposeIndex<SpinIndex>(lhs);
}
template<int Index,class vobj> inline Lattice<vobj> transposeColour(const Lattice<vobj> &lhs){
return transposeIndex<ColourIndex>(lhs);
}
template<int Index,class vobj> inline vobj transposeSpin(const vobj &lhs){
return transposeIndex<SpinIndex>(lhs);
}
template<int Index,class vobj> inline vobj transposeColour(const vobj &lhs){
return transposeIndex<ColourIndex>(lhs);
}
//////////////////////////////////////////////
// transpose array and scalar
//////////////////////////////////////////////
template<int Index,class vobj> inline Lattice<vobj> transposeSpin(const Lattice<vobj> &lhs){
return transposeIndex<SpinIndex>(lhs);
}
template<int Index,class vobj> inline Lattice<vobj> transposeColour(const Lattice<vobj> &lhs){
return transposeIndex<ColourIndex>(lhs);
}
template<int Index,class vobj> inline vobj transposeSpin(const vobj &lhs){
return transposeIndex<SpinIndex>(lhs);
}
template<int Index,class vobj> inline vobj transposeColour(const vobj &lhs){
return transposeIndex<ColourIndex>(lhs);
}
//////////////////////////////////////////
// Trace lattice and non-lattice
//////////////////////////////////////////
template<int Index,class vobj>
inline auto traceSpin(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(lhs._odata[0]))>
{
return traceIndex<SpinIndex>(lhs);
}
template<int Index,class vobj>
inline auto traceColour(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<ColourIndex>(lhs._odata[0]))>
{
return traceIndex<ColourIndex>(lhs);
}
template<int Index,class vobj>
inline auto traceSpin(const vobj &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(lhs))>
{
return traceIndex<SpinIndex>(lhs);
}
template<int Index,class vobj>
inline auto traceColour(const vobj &lhs) -> Lattice<decltype(traceIndex<ColourIndex>(lhs))>
{
return traceIndex<ColourIndex>(lhs);
}
//////////////////////////////////////////
// Trace lattice and non-lattice
//////////////////////////////////////////
template<int Index,class vobj>
inline auto traceSpin(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(vobj()))>
{
return traceIndex<SpinIndex>(lhs);
}
template<int Index,class vobj>
inline auto traceColour(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<ColourIndex>(vobj()))>
{
return traceIndex<ColourIndex>(lhs);
}
template<int Index,class vobj>
inline auto traceSpin(const vobj &lhs) -> Lattice<decltype(traceIndex<SpinIndex>(lhs))>
{
return traceIndex<SpinIndex>(lhs);
}
template<int Index,class vobj>
inline auto traceColour(const vobj &lhs) -> Lattice<decltype(traceIndex<ColourIndex>(lhs))>
{
return traceIndex<ColourIndex>(lhs);
}
//////////////////////////////////////////
// Current types
//////////////////////////////////////////
GRID_SERIALIZABLE_ENUM(Current, undef,
Vector, 0,
Axial, 1,
Tadpole, 2);
//////////////////////////////////////////
// Current types
//////////////////////////////////////////
GRID_SERIALIZABLE_ENUM(Current, undef,
Vector, 0,
Axial, 1,
Tadpole, 2);
} //namespace QCD
} // Grid
NAMESPACE_END(Grid);
#endif

4
Grid/qcd/action/Action.h
View File

@ -37,14 +37,18 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
// Abstract base interface
////////////////////////////////////////////
#include <Grid/qcd/action/ActionCore.h>
NAMESPACE_CHECK(ActionCore);
////////////////////////////////////////////////////////////////////////
// Fermion actions; prevent coupling fermion.cc files to other headers
////////////////////////////////////////////////////////////////////////
#include <Grid/qcd/action/fermion/FermionCore.h>
NAMESPACE_CHECK(FermionCore);
#include <Grid/qcd/action/fermion/Fermion.h>
NAMESPACE_CHECK(Fermion);
////////////////////////////////////////
// Pseudo fermion combinations for HMC
////////////////////////////////////////
#include <Grid/qcd/action/pseudofermion/PseudoFermion.h>
NAMESPACE_CHECK(PseudoFermion);
#endif

10
Grid/qcd/action/ActionBase.h
View File

@ -27,19 +27,18 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
/* END LEGAL */
#ifndef ACTION_BASE_H
#define ACTION_BASE_H
namespace Grid {
namespace QCD {
NAMESPACE_BEGIN(Grid);
template <class GaugeField >
class Action
{
public:
public:
bool is_smeared = false;
// Heatbath?
virtual void refresh(const GaugeField& U, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
@ -50,7 +49,6 @@ class Action
virtual ~Action(){}
};
}
}
NAMESPACE_END(Grid);
#endif // ACTION_BASE_H

8
Grid/qcd/action/ActionCore.h
View File

@ -31,29 +31,37 @@ directory
#define QCD_ACTION_CORE
#include <Grid/qcd/action/ActionBase.h>
NAMESPACE_CHECK(ActionBase);
#include <Grid/qcd/action/ActionSet.h>
NAMESPACE_CHECK(ActionSet);
#include <Grid/qcd/action/ActionParams.h>
NAMESPACE_CHECK(ActionParams);
////////////////////////////////////////////
// Gauge Actions
////////////////////////////////////////////
#include <Grid/qcd/action/gauge/Gauge.h>
NAMESPACE_CHECK(Gauge);
////////////////////////////////////////////
// Fermion prereqs
////////////////////////////////////////////
#include <Grid/qcd/action/fermion/FermionCore.h>
NAMESPACE_CHECK(ActionFermionCore);
////////////////////////////////////////////
// Scalar Actions
////////////////////////////////////////////
#include <Grid/qcd/action/scalar/Scalar.h>
NAMESPACE_CHECK(Scalar);
////////////////////////////////////////////
// Utility functions
////////////////////////////////////////////
#include <Grid/qcd/utils/Metric.h>
NAMESPACE_CHECK(Metric);
#include <Grid/qcd/utils/CovariantLaplacian.h>
NAMESPACE_CHECK(CovariantLaplacian);

55
Grid/qcd/action/ActionParams.h
View File

@ -27,37 +27,35 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
/* END LEGAL */
#ifndef GRID_QCD_ACTION_PARAMS_H
#define GRID_QCD_ACTION_PARAMS_H
namespace Grid {
namespace QCD {
NAMESPACE_BEGIN(Grid);
// These can move into a params header and be given MacroMagic serialisation
struct GparityWilsonImplParams {
bool overlapCommsCompute;
std::vector<int> twists;
GparityWilsonImplParams() : twists(Nd, 0), overlapCommsCompute(false){};
};
// These can move into a params header and be given MacroMagic serialisation
struct GparityWilsonImplParams {
Coordinate twists;
GparityWilsonImplParams() : twists(Nd, 0) {};
};
struct WilsonImplParams {
bool overlapCommsCompute;
std::vector<Real> twist_n_2pi_L;
std::vector<Complex> boundary_phases;
WilsonImplParams() : overlapCommsCompute(false) {
boundary_phases.resize(Nd, 1.0);
struct WilsonImplParams {
bool overlapCommsCompute;
AcceleratorVector<Real,Nd> twist_n_2pi_L;
AcceleratorVector<Complex,Nd> boundary_phases;
WilsonImplParams() {
boundary_phases.resize(Nd, 1.0);
twist_n_2pi_L.resize(Nd, 0.0);
};
WilsonImplParams(const std::vector<Complex> phi) : boundary_phases(phi), overlapCommsCompute(false) {
twist_n_2pi_L.resize(Nd, 0.0);
}
};
WilsonImplParams(const AcceleratorVector<Complex,Nd> phi) : boundary_phases(phi), overlapCommsCompute(false) {
twist_n_2pi_L.resize(Nd, 0.0);
}
};
struct StaggeredImplParams {
StaggeredImplParams() {};
};
struct StaggeredImplParams {
StaggeredImplParams() {};
};
struct OneFlavourRationalParams : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(OneFlavourRationalParams,
@ -69,10 +67,10 @@ namespace QCD {
int, precision,
int, BoundsCheckFreq);
// MaxIter and tolerance, vectors??
// MaxIter and tolerance, vectors??
// constructor
OneFlavourRationalParams( RealD _lo = 0.0,
// constructor
OneFlavourRationalParams( RealD _lo = 0.0,
RealD _hi = 1.0,
int _maxit = 1000,
RealD tol = 1.0e-8,
@ -88,11 +86,6 @@ namespace QCD {
BoundsCheckFreq(_BoundsCheckFreq){};
};
}
}
NAMESPACE_END(Grid);
#endif

16
Grid/qcd/action/ActionSet.h
View File

@ -26,14 +26,11 @@ with this program; if not, write to the Free Software Foundation, Inc.,
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
/* END LEGAL */
#ifndef ACTION_SET_H
#define ACTION_SET_H
namespace Grid {
// Should drop this namespace here
namespace QCD {
NAMESPACE_BEGIN(Grid);
//////////////////////////////////
// Indexing of tuple types
@ -62,7 +59,7 @@ struct Index<T, std::tuple<U, Types...>> {
template <class Field, class Repr = NoHirep >
struct ActionLevel {
public:
public:
unsigned int multiplier;
// Fundamental repr actions separated because of the smearing
@ -77,7 +74,7 @@ struct ActionLevel {
std::vector<ActPtr>& actions;
explicit ActionLevel(unsigned int mul = 1) :
actions(std::get<0>(actions_hirep)), multiplier(mul) {
actions(std::get<0>(actions_hirep)), multiplier(mul) {
// initialize the hirep vectors to zero.
// apply(this->resize, actions_hirep, 0); //need a working resize
assert(mul >= 1);
@ -87,7 +84,7 @@ struct ActionLevel {
void push_back(Action<GenField>* ptr) {
// insert only in the correct vector
std::get< Index < GenField, action_hirep_types>::value >(actions_hirep).push_back(ptr);
};
}
template <class ActPtr>
static void resize(ActPtr ap, unsigned int n) {
@ -110,7 +107,6 @@ struct ActionLevel {
template <class GaugeField, class R>
using ActionSet = std::vector<ActionLevel<GaugeField, R> >;
} // QCD
} // Grid
NAMESPACE_END(Grid);
#endif // ACTION_SET_H

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