mirror of
https://github.com/paboyle/Grid.git
synced 2025-06-14 05:07:05 +01:00
Compare commits
2 Commits
feature/di
...
feature/ca
| Author | SHA1 | Date | |
|---|---|---|---|
| 59282f25ec | |||
| b0bd173899 |
@ -44,22 +44,14 @@ directory
|
||||
#ifdef __NVCC__
|
||||
//disables nvcc specific warning in json.hpp
|
||||
#pragma clang diagnostic ignored "-Wdeprecated-register"
|
||||
|
||||
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
|
||||
//disables nvcc specific warning in json.hpp
|
||||
#pragma nv_diag_suppress unsigned_compare_with_zero
|
||||
#pragma nv_diag_suppress cast_to_qualified_type
|
||||
//disables nvcc specific warning in many files
|
||||
#pragma nv_diag_suppress esa_on_defaulted_function_ignored
|
||||
#pragma nv_diag_suppress extra_semicolon
|
||||
#else
|
||||
//disables nvcc specific warning in json.hpp
|
||||
#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
|
||||
#endif
|
||||
|
||||
//Eigen only
|
||||
#endif
|
||||
|
||||
// Disable vectorisation in Eigen on the Power8/9 and PowerPC
|
||||
|
||||
@ -36,7 +36,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
#include <Grid/GridCore.h>
|
||||
#include <Grid/qcd/QCD.h>
|
||||
#include <Grid/qcd/spin/Spin.h>
|
||||
#include <Grid/qcd/gparity/Gparity.h>
|
||||
#include <Grid/qcd/utils/Utils.h>
|
||||
#include <Grid/qcd/representations/Representations.h>
|
||||
NAMESPACE_CHECK(GridQCDCore);
|
||||
|
||||
@ -16,7 +16,6 @@
|
||||
#include <functional>
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
#include <strings.h>
|
||||
#include <stdio.h>
|
||||
#include <signal.h>
|
||||
#include <ctime>
|
||||
|
||||
@ -14,11 +14,7 @@
|
||||
/* NVCC save and restore compile environment*/
|
||||
#ifdef __NVCC__
|
||||
#pragma push
|
||||
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
|
||||
#pragma nv_diag_suppress code_is_unreachable
|
||||
#else
|
||||
#pragma diag_suppress code_is_unreachable
|
||||
#endif
|
||||
#pragma push_macro("__CUDA_ARCH__")
|
||||
#pragma push_macro("__NVCC__")
|
||||
#pragma push_macro("__CUDACC__")
|
||||
|
||||
@ -54,7 +54,6 @@ NAMESPACE_CHECK(BiCGSTAB);
|
||||
#include <Grid/algorithms/iterative/SchurRedBlack.h>
|
||||
#include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
|
||||
#include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
|
||||
#include <Grid/algorithms/iterative/ConjugateGradientMultiShiftMixedPrec.h>
|
||||
#include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
|
||||
#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
|
||||
#include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
|
||||
|
||||
@ -120,9 +120,6 @@ public:
|
||||
SolverTimer.Start();
|
||||
int k;
|
||||
for (k = 1; k <= MaxIterations; k++) {
|
||||
|
||||
GridStopWatch IterationTimer;
|
||||
IterationTimer.Start();
|
||||
c = cp;
|
||||
|
||||
MatrixTimer.Start();
|
||||
@ -155,14 +152,8 @@ public:
|
||||
LinearCombTimer.Stop();
|
||||
LinalgTimer.Stop();
|
||||
|
||||
IterationTimer.Stop();
|
||||
if ( (k % 500) == 0 ) {
|
||||
std::cout << GridLogMessage << "ConjugateGradient: Iteration " << k
|
||||
std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
|
||||
<< " residual " << sqrt(cp/ssq) << " target " << Tolerance << std::endl;
|
||||
} else {
|
||||
std::cout << GridLogIterative << "ConjugateGradient: Iteration " << k
|
||||
<< " residual " << sqrt(cp/ssq) << " target " << Tolerance << " took " << IterationTimer.Elapsed() << std::endl;
|
||||
}
|
||||
|
||||
// Stopping condition
|
||||
if (cp <= rsq) {
|
||||
@ -179,13 +170,13 @@ public:
|
||||
<< "\tTrue residual " << true_residual
|
||||
<< "\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;
|
||||
std::cout << GridLogMessage << "\tInner " << InnerTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tAxpyNorm " << AxpyNormTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogIterative << "Time breakdown "<<std::endl;
|
||||
std::cout << GridLogIterative << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogIterative << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogIterative << "\tLinalg " << LinalgTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogIterative << "\tInner " << InnerTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogIterative << "\tAxpyNorm " << AxpyNormTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogIterative << "\tLinearComb " << LinearCombTimer.Elapsed() <<std::endl;
|
||||
|
||||
if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
|
||||
|
||||
|
||||
@ -49,7 +49,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
Integer TotalInnerIterations; //Number of inner CG iterations
|
||||
Integer TotalOuterIterations; //Number of restarts
|
||||
Integer TotalFinalStepIterations; //Number of CG iterations in final patch-up step
|
||||
RealD TrueResidual;
|
||||
|
||||
//Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
|
||||
LinearFunction<FieldF> *guesser;
|
||||
@ -69,7 +68,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
}
|
||||
|
||||
void operator() (const FieldD &src_d_in, FieldD &sol_d){
|
||||
std::cout << GridLogMessage << "MixedPrecisionConjugateGradient: Starting mixed precision CG with outer tolerance " << Tolerance << " and inner tolerance " << InnerTolerance << std::endl;
|
||||
TotalInnerIterations = 0;
|
||||
|
||||
GridStopWatch TotalTimer;
|
||||
@ -99,7 +97,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
FieldF sol_f(SinglePrecGrid);
|
||||
sol_f.Checkerboard() = cb;
|
||||
|
||||
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting initial inner CG with tolerance " << inner_tol << std::endl;
|
||||
ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
|
||||
CG_f.ErrorOnNoConverge = false;
|
||||
|
||||
@ -133,7 +130,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
(*guesser)(src_f, sol_f);
|
||||
|
||||
//Inner CG
|
||||
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " << outer_iter << " starting inner CG with tolerance " << inner_tol << std::endl;
|
||||
CG_f.Tolerance = inner_tol;
|
||||
InnerCGtimer.Start();
|
||||
CG_f(Linop_f, src_f, sol_f);
|
||||
@ -154,7 +150,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
|
||||
CG_d(Linop_d, src_d_in, sol_d);
|
||||
TotalFinalStepIterations = CG_d.IterationsToComplete;
|
||||
TrueResidual = CG_d.TrueResidual;
|
||||
|
||||
TotalTimer.Stop();
|
||||
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Inner CG iterations " << TotalInnerIterations << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations << std::endl;
|
||||
|
||||
@ -44,7 +44,7 @@ public:
|
||||
|
||||
using OperatorFunction<Field>::operator();
|
||||
|
||||
// RealD Tolerance;
|
||||
RealD Tolerance;
|
||||
Integer MaxIterations;
|
||||
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
|
||||
std::vector<int> IterationsToCompleteShift; // Iterations for this shift
|
||||
@ -52,7 +52,7 @@ public:
|
||||
MultiShiftFunction shifts;
|
||||
std::vector<RealD> TrueResidualShift;
|
||||
|
||||
ConjugateGradientMultiShift(Integer maxit, const MultiShiftFunction &_shifts) :
|
||||
ConjugateGradientMultiShift(Integer maxit,MultiShiftFunction &_shifts) :
|
||||
MaxIterations(maxit),
|
||||
shifts(_shifts)
|
||||
{
|
||||
@ -182,9 +182,6 @@ public:
|
||||
for(int s=0;s<nshift;s++) {
|
||||
axpby(psi[s],0.,-bs[s]*alpha[s],src,src);
|
||||
}
|
||||
|
||||
std::cout << GridLogIterative << "ConjugateGradientMultiShift: initial rn (|src|^2) =" << rn << " qq (|MdagM src|^2) =" << qq << " d ( dot(src, [MdagM + m_0]src) ) =" << d << " c=" << c << std::endl;
|
||||
|
||||
|
||||
///////////////////////////////////////
|
||||
// Timers
|
||||
@ -324,8 +321,8 @@ public:
|
||||
|
||||
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
|
||||
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tAXPY " << AXPYTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tAXPY " << AXPYTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tMarix " << MatrixTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tShift " << ShiftTimer.Elapsed() <<std::endl;
|
||||
|
||||
IterationsToComplete = k;
|
||||
|
||||
@ -1,409 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/algorithms/iterative/ConjugateGradientMultiShift.h
|
||||
|
||||
Copyright (C) 2015
|
||||
|
||||
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Christopher Kelly <ckelly@bnl.gov>
|
||||
|
||||
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_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
|
||||
#define GRID_CONJUGATE_GRADIENT_MULTI_SHIFT_MIXEDPREC_H
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
//CK 2020: A variant of the multi-shift conjugate gradient with the matrix multiplication in single precision.
|
||||
//The residual is stored in single precision, but the search directions and solution are stored in double precision.
|
||||
//Every update_freq iterations the residual is corrected in double precision.
|
||||
|
||||
//For safety the a final regular CG is applied to clean up if necessary
|
||||
|
||||
//Linop to add shift to input linop, used in cleanup CG
|
||||
namespace ConjugateGradientMultiShiftMixedPrecSupport{
|
||||
template<typename Field>
|
||||
class ShiftedLinop: public LinearOperatorBase<Field>{
|
||||
public:
|
||||
LinearOperatorBase<Field> &linop_base;
|
||||
RealD shift;
|
||||
|
||||
ShiftedLinop(LinearOperatorBase<Field> &_linop_base, RealD _shift): linop_base(_linop_base), shift(_shift){}
|
||||
|
||||
void OpDiag (const Field &in, Field &out){ assert(0); }
|
||||
void OpDir (const Field &in, Field &out,int dir,int disp){ assert(0); }
|
||||
void OpDirAll (const Field &in, std::vector<Field> &out){ assert(0); }
|
||||
|
||||
void Op (const Field &in, Field &out){ assert(0); }
|
||||
void AdjOp (const Field &in, Field &out){ assert(0); }
|
||||
|
||||
void HermOp(const Field &in, Field &out){
|
||||
linop_base.HermOp(in, out);
|
||||
axpy(out, shift, in, out);
|
||||
}
|
||||
|
||||
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
|
||||
HermOp(in,out);
|
||||
ComplexD dot = innerProduct(in,out);
|
||||
n1=real(dot);
|
||||
n2=norm2(out);
|
||||
}
|
||||
};
|
||||
};
|
||||
|
||||
|
||||
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 ConjugateGradientMultiShiftMixedPrec : public OperatorMultiFunction<FieldD>,
|
||||
public OperatorFunction<FieldD>
|
||||
{
|
||||
public:
|
||||
|
||||
using OperatorFunction<FieldD>::operator();
|
||||
|
||||
RealD Tolerance;
|
||||
Integer MaxIterations;
|
||||
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
|
||||
std::vector<int> IterationsToCompleteShift; // Iterations for this shift
|
||||
int verbose;
|
||||
MultiShiftFunction shifts;
|
||||
std::vector<RealD> TrueResidualShift;
|
||||
|
||||
int ReliableUpdateFreq; //number of iterations between reliable updates
|
||||
|
||||
GridBase* SinglePrecGrid; //Grid for single-precision fields
|
||||
LinearOperatorBase<FieldF> &Linop_f; //single precision
|
||||
|
||||
ConjugateGradientMultiShiftMixedPrec(Integer maxit, const MultiShiftFunction &_shifts,
|
||||
GridBase* _SinglePrecGrid, LinearOperatorBase<FieldF> &_Linop_f,
|
||||
int _ReliableUpdateFreq
|
||||
) :
|
||||
MaxIterations(maxit), shifts(_shifts), SinglePrecGrid(_SinglePrecGrid), Linop_f(_Linop_f), ReliableUpdateFreq(_ReliableUpdateFreq)
|
||||
{
|
||||
verbose=1;
|
||||
IterationsToCompleteShift.resize(_shifts.order);
|
||||
TrueResidualShift.resize(_shifts.order);
|
||||
}
|
||||
|
||||
void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, FieldD &psi)
|
||||
{
|
||||
GridBase *grid = src.Grid();
|
||||
int nshift = shifts.order;
|
||||
std::vector<FieldD> results(nshift,grid);
|
||||
(*this)(Linop,src,results,psi);
|
||||
}
|
||||
void operator() (LinearOperatorBase<FieldD> &Linop, const FieldD &src, std::vector<FieldD> &results, FieldD &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;
|
||||
}
|
||||
|
||||
void operator() (LinearOperatorBase<FieldD> &Linop_d, const FieldD &src_d, std::vector<FieldD> &psi_d)
|
||||
{
|
||||
GridBase *DoublePrecGrid = src_d.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);
|
||||
|
||||
//Double precision search directions
|
||||
FieldD p_d(DoublePrecGrid);
|
||||
std::vector<FieldD> ps_d(nshift, DoublePrecGrid);// Search directions (double precision)
|
||||
|
||||
FieldD tmp_d(DoublePrecGrid);
|
||||
FieldD r_d(DoublePrecGrid);
|
||||
FieldD mmp_d(DoublePrecGrid);
|
||||
|
||||
assert(psi_d.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];
|
||||
|
||||
const int primary =0;
|
||||
|
||||
//Primary shift fields CG iteration
|
||||
RealD a,b,c,d;
|
||||
RealD cp,bp,qq; //prev
|
||||
|
||||
// Matrix mult fields
|
||||
FieldF r_f(SinglePrecGrid);
|
||||
FieldF p_f(SinglePrecGrid);
|
||||
FieldF tmp_f(SinglePrecGrid);
|
||||
FieldF mmp_f(SinglePrecGrid);
|
||||
FieldF src_f(SinglePrecGrid);
|
||||
precisionChange(src_f, src_d);
|
||||
|
||||
// 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_d);
|
||||
|
||||
// Handle trivial case of zero src.
|
||||
if( cp == 0. ){
|
||||
for(int s=0;s<nshift;s++){
|
||||
psi_d[s] = Zero();
|
||||
IterationsToCompleteShift[s] = 1;
|
||||
TrueResidualShift[s] = 0.;
|
||||
}
|
||||
return;
|
||||
}
|
||||
|
||||
for(int s=0;s<nshift;s++){
|
||||
rsq[s] = cp * mresidual[s] * mresidual[s];
|
||||
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift "<< s <<" target resid "<<rsq[s]<<std::endl;
|
||||
ps_d[s] = src_d;
|
||||
}
|
||||
// r and p for primary
|
||||
r_f=src_f; //residual maintained in single
|
||||
p_f=src_f;
|
||||
p_d = src_d; //primary copy --- make this a reference to ps_d to save axpys
|
||||
|
||||
//MdagM+m[0]
|
||||
Linop_f.HermOpAndNorm(p_f,mmp_f,d,qq); // mmp = MdagM p d=real(dot(p, mmp)), qq=norm2(mmp)
|
||||
axpy(mmp_f,mass[0],p_f,mmp_f);
|
||||
RealD rn = norm2(p_f);
|
||||
d += rn*mass[0];
|
||||
|
||||
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];
|
||||
}
|
||||
|
||||
// r += b[0] A.p[0]
|
||||
// c= norm(r)
|
||||
c=axpy_norm(r_f,b,mmp_f,r_f);
|
||||
|
||||
for(int s=0;s<nshift;s++) {
|
||||
axpby(psi_d[s],0.,-bs[s]*alpha[s],src_d,src_d);
|
||||
}
|
||||
|
||||
///////////////////////////////////////
|
||||
// Timers
|
||||
///////////////////////////////////////
|
||||
GridStopWatch AXPYTimer, ShiftTimer, QRTimer, MatrixTimer, SolverTimer, PrecChangeTimer, CleanupTimer;
|
||||
|
||||
SolverTimer.Start();
|
||||
|
||||
// Iteration loop
|
||||
int k;
|
||||
|
||||
for (k=1;k<=MaxIterations;k++){
|
||||
a = c /cp;
|
||||
|
||||
//Update double precision search direction by residual
|
||||
PrecChangeTimer.Start();
|
||||
precisionChange(r_d, r_f);
|
||||
PrecChangeTimer.Stop();
|
||||
|
||||
AXPYTimer.Start();
|
||||
axpy(p_d,a,p_d,r_d);
|
||||
|
||||
for(int s=0;s<nshift;s++){
|
||||
if ( ! converged[s] ) {
|
||||
if (s==0){
|
||||
axpy(ps_d[s],a,ps_d[s],r_d);
|
||||
} else{
|
||||
RealD as =a *z[s][iz]*bs[s] /(z[s][1-iz]*b);
|
||||
axpby(ps_d[s],z[s][iz],as,r_d,ps_d[s]);
|
||||
}
|
||||
}
|
||||
}
|
||||
AXPYTimer.Stop();
|
||||
|
||||
PrecChangeTimer.Start();
|
||||
precisionChange(p_f, p_d); //get back single prec search direction for linop
|
||||
PrecChangeTimer.Stop();
|
||||
|
||||
cp=c;
|
||||
MatrixTimer.Start();
|
||||
Linop_f.HermOp(p_f,mmp_f);
|
||||
d=real(innerProduct(p_f,mmp_f));
|
||||
MatrixTimer.Stop();
|
||||
|
||||
AXPYTimer.Start();
|
||||
axpy(mmp_f,mass[0],p_f,mmp_f);
|
||||
AXPYTimer.Stop();
|
||||
RealD rn = norm2(p_f);
|
||||
d += rn*mass[0];
|
||||
|
||||
bp=b;
|
||||
b=-cp/d;
|
||||
|
||||
// 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
|
||||
}
|
||||
}
|
||||
ShiftTimer.Stop();
|
||||
|
||||
//Update double precision solutions
|
||||
AXPYTimer.Start();
|
||||
for(int s=0;s<nshift;s++){
|
||||
int ss = s;
|
||||
if( (!converged[s]) ) {
|
||||
axpy(psi_d[ss],-bs[s]*alpha[s],ps_d[s],psi_d[ss]);
|
||||
}
|
||||
}
|
||||
|
||||
//Perform reliable update if necessary; otherwise update residual from single-prec mmp
|
||||
RealD c_f = axpy_norm(r_f,b,mmp_f,r_f);
|
||||
AXPYTimer.Stop();
|
||||
|
||||
c = c_f;
|
||||
|
||||
if(k % ReliableUpdateFreq == 0){
|
||||
//Replace r with true residual
|
||||
MatrixTimer.Start();
|
||||
Linop_d.HermOp(psi_d[0],mmp_d);
|
||||
MatrixTimer.Stop();
|
||||
|
||||
AXPYTimer.Start();
|
||||
axpy(mmp_d,mass[0],psi_d[0],mmp_d);
|
||||
|
||||
RealD c_d = axpy_norm(r_d, -1.0, mmp_d, src_d);
|
||||
AXPYTimer.Stop();
|
||||
|
||||
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<< ", replaced |r|^2 = "<<c_f <<" with |r|^2 = "<<c_d<<std::endl;
|
||||
|
||||
PrecChangeTimer.Start();
|
||||
precisionChange(r_f, r_d);
|
||||
PrecChangeTimer.Stop();
|
||||
c = c_d;
|
||||
}
|
||||
|
||||
// Convergence checks
|
||||
int all_converged = 1;
|
||||
for(int s=0;s<nshift;s++){
|
||||
|
||||
if ( (!converged[s]) ){
|
||||
IterationsToCompleteShift[s] = k;
|
||||
|
||||
RealD css = c * z[s][iz]* z[s][iz];
|
||||
|
||||
if(css<rsq[s]){
|
||||
if ( ! converged[s] )
|
||||
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec k="<<k<<" Shift "<<s<<" has converged"<<std::endl;
|
||||
converged[s]=1;
|
||||
} else {
|
||||
all_converged=0;
|
||||
}
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
if ( all_converged ){
|
||||
|
||||
SolverTimer.Stop();
|
||||
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: All shifts have converged iteration "<<k<<std::endl;
|
||||
std::cout<<GridLogMessage<< "ConjugateGradientMultiShiftMixedPrec: Checking solutions"<<std::endl;
|
||||
|
||||
// Check answers
|
||||
for(int s=0; s < nshift; s++) {
|
||||
Linop_d.HermOpAndNorm(psi_d[s],mmp_d,d,qq);
|
||||
axpy(tmp_d,mass[s],psi_d[s],mmp_d);
|
||||
axpy(r_d,-alpha[s],src_d,tmp_d);
|
||||
RealD rn = norm2(r_d);
|
||||
RealD cn = norm2(src_d);
|
||||
TrueResidualShift[s] = std::sqrt(rn/cn);
|
||||
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: shift["<<s<<"] true residual "<< TrueResidualShift[s] << " target " << mresidual[s] << std::endl;
|
||||
|
||||
//If we have not reached the desired tolerance, do a (mixed precision) CG cleanup
|
||||
if(rn >= rsq[s]){
|
||||
CleanupTimer.Start();
|
||||
std::cout<<GridLogMessage<<"ConjugateGradientMultiShiftMixedPrec: performing cleanup step for shift " << s << std::endl;
|
||||
|
||||
//Setup linear operators for final cleanup
|
||||
ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldD> Linop_shift_d(Linop_d, mass[s]);
|
||||
ConjugateGradientMultiShiftMixedPrecSupport::ShiftedLinop<FieldF> Linop_shift_f(Linop_f, mass[s]);
|
||||
|
||||
MixedPrecisionConjugateGradient<FieldD,FieldF> cg(mresidual[s], MaxIterations, MaxIterations, SinglePrecGrid, Linop_shift_f, Linop_shift_d);
|
||||
cg(src_d, psi_d[s]);
|
||||
|
||||
TrueResidualShift[s] = cg.TrueResidual;
|
||||
CleanupTimer.Stop();
|
||||
}
|
||||
}
|
||||
|
||||
std::cout << GridLogMessage << "ConjugateGradientMultiShiftMixedPrec: Time Breakdown for body"<<std::endl;
|
||||
std::cout << GridLogMessage << "\tSolver " << SolverTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\t\tAXPY " << AXPYTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\t\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\t\tShift " << ShiftTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\t\tPrecision Change " << PrecChangeTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tFinal Cleanup " << CleanupTimer.Elapsed() <<std::endl;
|
||||
std::cout << GridLogMessage << "\tSolver+Cleanup " << SolverTimer.Elapsed() + CleanupTimer.Elapsed() << std::endl;
|
||||
|
||||
IterationsToComplete = k;
|
||||
|
||||
return;
|
||||
}
|
||||
|
||||
|
||||
}
|
||||
// ugly hack
|
||||
std::cout<<GridLogMessage<<"CG multi shift did not converge"<<std::endl;
|
||||
// assert(0);
|
||||
}
|
||||
|
||||
};
|
||||
NAMESPACE_END(Grid);
|
||||
#endif
|
||||
@ -113,43 +113,7 @@ public:
|
||||
blockPromote(guess_coarse,guess,subspace);
|
||||
guess.Checkerboard() = src.Checkerboard();
|
||||
};
|
||||
|
||||
void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
|
||||
int Nevec = (int)evec_coarse.size();
|
||||
int Nsrc = (int)src.size();
|
||||
// make temp variables
|
||||
std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
|
||||
std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());
|
||||
//Preporcessing
|
||||
std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
|
||||
for (int j=0;j<Nsrc;j++)
|
||||
{
|
||||
guess_coarse[j] = Zero();
|
||||
std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
|
||||
blockProject(src_coarse[j],src[j],subspace);
|
||||
}
|
||||
//deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
|
||||
std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
|
||||
for (int i=0;i<Nevec;i++)
|
||||
{
|
||||
std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
|
||||
const CoarseField & tmp = evec_coarse[i];
|
||||
for (int j=0;j<Nsrc;j++)
|
||||
{
|
||||
axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
|
||||
}
|
||||
}
|
||||
//postprocessing
|
||||
std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
|
||||
for (int j=0;j<Nsrc;j++)
|
||||
{
|
||||
std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
|
||||
blockPromote(guess_coarse[j],guess[j],subspace);
|
||||
guess[j].Checkerboard() = src[j].Checkerboard();
|
||||
}
|
||||
};
|
||||
|
||||
};
|
||||
};
|
||||
|
||||
|
||||
|
||||
|
||||
@ -44,7 +44,6 @@ public:
|
||||
int, MinRes); // Must restart
|
||||
};
|
||||
|
||||
//This class is the input parameter class for some testing programs
|
||||
struct LocalCoherenceLanczosParams : Serializable {
|
||||
public:
|
||||
GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
|
||||
@ -146,24 +145,16 @@ public:
|
||||
LinearOperatorBase<FineField> &_Linop;
|
||||
RealD _coarse_relax_tol;
|
||||
std::vector<FineField> &_subspace;
|
||||
|
||||
int _largestEvalIdxForReport; //The convergence of the LCL is based on the evals of the coarse grid operator, not those of the underlying fine grid operator
|
||||
//As a result we do not know what the eval range of the fine operator is until the very end, making tuning the Cheby bounds very difficult
|
||||
//To work around this issue, every restart we separately reconstruct the fine operator eval for the lowest and highest evec and print these
|
||||
//out alongside the evals of the coarse operator. To do so we need to know the index of the largest eval (i.e. Nstop-1)
|
||||
//NOTE: If largestEvalIdxForReport=-1 (default) then this is not performed
|
||||
|
||||
ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField> &Poly,
|
||||
OperatorFunction<FineField> &smoother,
|
||||
LinearOperatorBase<FineField> &Linop,
|
||||
std::vector<FineField> &subspace,
|
||||
RealD coarse_relax_tol=5.0e3,
|
||||
int largestEvalIdxForReport=-1)
|
||||
RealD coarse_relax_tol=5.0e3)
|
||||
: _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
|
||||
_coarse_relax_tol(coarse_relax_tol), _largestEvalIdxForReport(largestEvalIdxForReport)
|
||||
_coarse_relax_tol(coarse_relax_tol)
|
||||
{ };
|
||||
|
||||
//evalMaxApprox: approximation of largest eval of the fine Chebyshev operator (suitably wrapped by block projection)
|
||||
int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
|
||||
{
|
||||
CoarseField v(B);
|
||||
@ -186,26 +177,12 @@ public:
|
||||
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
|
||||
<<std::endl;
|
||||
|
||||
if(_largestEvalIdxForReport != -1 && (j==0 || j==_largestEvalIdxForReport)){
|
||||
std::cout<<GridLogIRL << "Estimating true eval of fine grid operator for eval idx " << j << std::endl;
|
||||
RealD tmp_eval;
|
||||
ReconstructEval(j,eresid,B,tmp_eval,1.0); //don't use evalMaxApprox of coarse operator! (cf below)
|
||||
}
|
||||
|
||||
int conv=0;
|
||||
if( (vv<eresid*eresid) ) conv = 1;
|
||||
return conv;
|
||||
}
|
||||
|
||||
//This function is called at the end of the coarse grid Lanczos. It promotes the coarse eigenvector 'B' to the fine grid,
|
||||
//applies a smoother to the result then computes the computes the *fine grid* eigenvalue (output as 'eval').
|
||||
|
||||
//evalMaxApprox should be the approximation of the largest eval of the fine Hermop. However when this function is called by IRL it actually passes the largest eval of the *Chebyshev* operator (as this is the max approx used for the TestConvergence above)
|
||||
//As the largest eval of the Chebyshev is typically several orders of magnitude larger this makes the convergence test pass even when it should not.
|
||||
//We therefore ignore evalMaxApprox here and use a value of 1.0 (note this value is already used by TestCoarse)
|
||||
int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
|
||||
int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
|
||||
{
|
||||
evalMaxApprox = 1.0; //cf above
|
||||
GridBase *FineGrid = _subspace[0].Grid();
|
||||
int checkerboard = _subspace[0].Checkerboard();
|
||||
FineField fB(FineGrid);fB.Checkerboard() =checkerboard;
|
||||
@ -224,13 +201,13 @@ public:
|
||||
eval = vnum/vden;
|
||||
fv -= eval*fB;
|
||||
RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
|
||||
if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
|
||||
|
||||
|
||||
std::cout.precision(13);
|
||||
std::cout<<GridLogIRL << "[" << std::setw(3)<<j<<"] "
|
||||
<<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
|
||||
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv << " target " << eresid*eresid
|
||||
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
|
||||
<<std::endl;
|
||||
if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
|
||||
if( (vv<eresid*eresid) ) return 1;
|
||||
return 0;
|
||||
}
|
||||
@ -308,10 +285,6 @@ public:
|
||||
evals_coarse.resize(0);
|
||||
};
|
||||
|
||||
//The block inner product is the inner product on the fine grid locally summed over the blocks
|
||||
//to give a Lattice<Scalar> on the coarse grid. This function orthnormalizes the fine-grid subspace
|
||||
//vectors under the block inner product. This step must be performed after computing the fine grid
|
||||
//eigenvectors and before computing the coarse grid eigenvectors.
|
||||
void Orthogonalise(void ) {
|
||||
CoarseScalar InnerProd(_CoarseGrid);
|
||||
std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
|
||||
@ -355,8 +328,6 @@ public:
|
||||
}
|
||||
}
|
||||
|
||||
//While this method serves to check the coarse eigenvectors, it also recomputes the eigenvalues from the smoothed reconstructed eigenvectors
|
||||
//hence the smoother can be tuned after running the coarse Lanczos by using a different smoother here
|
||||
void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax)
|
||||
{
|
||||
assert(evals_fine.size() == nbasis);
|
||||
@ -405,31 +376,25 @@ public:
|
||||
evals_fine.resize(nbasis);
|
||||
subspace.resize(nbasis,_FineGrid);
|
||||
}
|
||||
|
||||
|
||||
//cheby_op: Parameters of the fine grid Chebyshev polynomial used for the Lanczos acceleration
|
||||
//cheby_smooth: Parameters of a separate Chebyshev polynomial used after the Lanczos has completed to smooth out high frequency noise in the reconstructed fine grid eigenvectors prior to computing the eigenvalue
|
||||
//relax: Reconstructed eigenvectors (post smoothing) are naturally not as precise as true eigenvectors. This factor acts as a multiplier on the stopping condition when determining whether the results satisfy the user provided stopping condition
|
||||
void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
|
||||
int Nstop, int Nk, int Nm,RealD resid,
|
||||
RealD MaxIt, RealD betastp, int MinRes)
|
||||
{
|
||||
Chebyshev<FineField> Cheby(cheby_op); //Chebyshev of fine operator on fine grid
|
||||
ProjectedHermOp<Fobj,CComplex,nbasis> Op(_FineOp,subspace); //Fine operator on coarse grid with intermediate fine grid conversion
|
||||
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace); //Chebyshev of fine operator on coarse grid with intermediate fine grid conversion
|
||||
Chebyshev<FineField> Cheby(cheby_op);
|
||||
ProjectedHermOp<Fobj,CComplex,nbasis> Op(_FineOp,subspace);
|
||||
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace);
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
Chebyshev<FineField> ChebySmooth(cheby_smooth); //lower order Chebyshev of fine operator on fine grid used to smooth regenerated eigenvectors
|
||||
ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax,Nstop-1);
|
||||
Chebyshev<FineField> ChebySmooth(cheby_smooth);
|
||||
ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
|
||||
|
||||
evals_coarse.resize(Nm);
|
||||
evec_coarse.resize(Nm,_CoarseGrid);
|
||||
|
||||
CoarseField src(_CoarseGrid); src=1.0;
|
||||
|
||||
//Note the "tester" here is also responsible for generating the fine grid eigenvalues which are output into the "evals_coarse" array
|
||||
ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
|
||||
int Nconv=0;
|
||||
IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
|
||||
@ -440,14 +405,6 @@ public:
|
||||
std::cout << i << " Coarse eval = " << evals_coarse[i] << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
//Get the fine eigenvector 'i' by reconstruction
|
||||
void getFineEvecEval(FineField &evec, RealD &eval, const int i) const{
|
||||
blockPromote(evec_coarse[i],evec,subspace);
|
||||
eval = evals_coarse[i];
|
||||
}
|
||||
|
||||
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -29,8 +29,6 @@ template<class Field> class PowerMethod
|
||||
RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
|
||||
RealD vden = norm2(src_n);
|
||||
RealD na = vnum/vden;
|
||||
|
||||
std::cout << GridLogIterative << "PowerMethod: Current approximation of largest eigenvalue " << na << std::endl;
|
||||
|
||||
if ( (fabs(evalMaxApprox/na - 1.0) < 0.001) || (i==_MAX_ITER_EST_-1) ) {
|
||||
evalMaxApprox = na;
|
||||
|
||||
@ -113,6 +113,11 @@ private:
|
||||
static uint64_t DeviceToHostBytes;
|
||||
static uint64_t HostToDeviceXfer;
|
||||
static uint64_t DeviceToHostXfer;
|
||||
|
||||
static uint64_t DeviceAccesses;
|
||||
static uint64_t HostAccesses;
|
||||
static uint64_t DeviceAccessBytes;
|
||||
static uint64_t HostAccessBytes;
|
||||
|
||||
private:
|
||||
#ifndef GRID_UVM
|
||||
@ -152,6 +157,7 @@ private:
|
||||
|
||||
// static void LRUupdate(AcceleratorViewEntry &AccCache);
|
||||
static void LRUinsert(AcceleratorViewEntry &AccCache);
|
||||
static void LRUinsertback(AcceleratorViewEntry &AccCache);
|
||||
static void LRUremove(AcceleratorViewEntry &AccCache);
|
||||
|
||||
// manage entries in the table
|
||||
|
||||
@ -23,6 +23,11 @@ uint64_t MemoryManager::HostToDeviceBytes;
|
||||
uint64_t MemoryManager::DeviceToHostBytes;
|
||||
uint64_t MemoryManager::HostToDeviceXfer;
|
||||
uint64_t MemoryManager::DeviceToHostXfer;
|
||||
uint64_t MemoryManager::DeviceAccesses;
|
||||
uint64_t MemoryManager::HostAccesses;
|
||||
uint64_t MemoryManager::DeviceAccessBytes;
|
||||
uint64_t MemoryManager::HostAccessBytes;
|
||||
|
||||
|
||||
////////////////////////////////////
|
||||
// Priority ordering for unlocked entries
|
||||
@ -86,6 +91,14 @@ void MemoryManager::LRUinsert(AcceleratorViewEntry &AccCache)
|
||||
AccCache.LRU_valid = 1;
|
||||
DeviceLRUBytes+=AccCache.bytes;
|
||||
}
|
||||
void MemoryManager::LRUinsertback(AcceleratorViewEntry &AccCache)
|
||||
{
|
||||
assert(AccCache.LRU_valid==0);
|
||||
LRU.push_back(AccCache.CpuPtr);
|
||||
AccCache.LRU_entry = --LRU.end();
|
||||
AccCache.LRU_valid = 1;
|
||||
DeviceLRUBytes+=AccCache.bytes;
|
||||
}
|
||||
void MemoryManager::LRUremove(AcceleratorViewEntry &AccCache)
|
||||
{
|
||||
assert(AccCache.LRU_valid==1);
|
||||
@ -129,6 +142,7 @@ void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
|
||||
dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);
|
||||
assert(AccCache.accLock==0);
|
||||
assert(AccCache.cpuLock==0);
|
||||
|
||||
if(AccCache.state==AccDirty) {
|
||||
Flush(AccCache);
|
||||
}
|
||||
@ -231,6 +245,9 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
|
||||
EntryCreate(CpuPtr,bytes,mode,hint);
|
||||
}
|
||||
|
||||
DeviceAccesses++;
|
||||
DeviceAccessBytes+=bytes;
|
||||
|
||||
auto AccCacheIterator = EntryLookup(CpuPtr);
|
||||
auto & AccCache = AccCacheIterator->second;
|
||||
if (!AccCache.AccPtr) {
|
||||
@ -349,6 +366,10 @@ void MemoryManager::CpuViewClose(uint64_t CpuPtr)
|
||||
assert(AccCache.accLock==0);
|
||||
|
||||
AccCache.cpuLock--;
|
||||
|
||||
if(AccCache.cpuLock==0) {
|
||||
LRUinsertback(AccCache);
|
||||
}
|
||||
}
|
||||
/*
|
||||
* Action State StateNext Flush Clone
|
||||
@ -371,6 +392,9 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
|
||||
EntryCreate(CpuPtr,bytes,mode,transient);
|
||||
}
|
||||
|
||||
HostAccesses++;
|
||||
HostAccessBytes+=bytes;
|
||||
|
||||
auto AccCacheIterator = EntryLookup(CpuPtr);
|
||||
auto & AccCache = AccCacheIterator->second;
|
||||
|
||||
@ -416,6 +440,12 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
|
||||
|
||||
AccCache.transient= transient? EvictNext : 0;
|
||||
|
||||
// If view is opened on host remove from LRU
|
||||
// Host close says evict next from device
|
||||
if(AccCache.LRU_valid==1){
|
||||
LRUremove(AccCache);
|
||||
}
|
||||
|
||||
return AccCache.CpuPtr;
|
||||
}
|
||||
void MemoryManager::NotifyDeletion(void *_ptr)
|
||||
|
||||
@ -12,6 +12,10 @@ uint64_t MemoryManager::HostToDeviceBytes;
|
||||
uint64_t MemoryManager::DeviceToHostBytes;
|
||||
uint64_t MemoryManager::HostToDeviceXfer;
|
||||
uint64_t MemoryManager::DeviceToHostXfer;
|
||||
uint64_t MemoryManager::DeviceAccesses;
|
||||
uint64_t MemoryManager::HostAccesses;
|
||||
uint64_t MemoryManager::DeviceAccessBytes;
|
||||
uint64_t MemoryManager::HostAccessBytes;
|
||||
|
||||
void MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
|
||||
void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
|
||||
|
||||
@ -53,11 +53,10 @@ public:
|
||||
// Communicator should know nothing of the physics grid, only processor grid.
|
||||
////////////////////////////////////////////
|
||||
int _Nprocessors; // How many in all
|
||||
int _processor; // linear processor rank
|
||||
unsigned long _ndimension;
|
||||
Coordinate _shm_processors; // Which dimensions get relayed out over processors lanes.
|
||||
Coordinate _processors; // Which dimensions get relayed out over processors lanes.
|
||||
int _processor; // linear processor rank
|
||||
Coordinate _processor_coor; // linear processor coordinate
|
||||
unsigned long _ndimension;
|
||||
static Grid_MPI_Comm communicator_world;
|
||||
Grid_MPI_Comm communicator;
|
||||
std::vector<Grid_MPI_Comm> communicator_halo;
|
||||
@ -98,9 +97,8 @@ public:
|
||||
int BossRank(void) ;
|
||||
int ThisRank(void) ;
|
||||
const Coordinate & ThisProcessorCoor(void) ;
|
||||
const Coordinate & ShmGrid(void) { return _shm_processors; } ;
|
||||
const Coordinate & ProcessorGrid(void) ;
|
||||
int ProcessorCount(void) ;
|
||||
int ProcessorCount(void) ;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
// very VERY rarely (Log, serial RNG) we need world without a grid
|
||||
@ -144,16 +142,16 @@ public:
|
||||
int bytes);
|
||||
|
||||
double StencilSendToRecvFrom(void *xmit,
|
||||
int xmit_to_rank,int do_xmit,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,int do_recv,
|
||||
int recv_from_rank,
|
||||
int bytes,int dir);
|
||||
|
||||
double StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
|
||||
void *xmit,
|
||||
int xmit_to_rank,int do_xmit,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,int do_recv,
|
||||
int recv_from_rank,
|
||||
int bytes,int dir);
|
||||
|
||||
|
||||
|
||||
@ -106,7 +106,7 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
|
||||
// Remap using the shared memory optimising routine
|
||||
// The remap creates a comm which must be freed
|
||||
////////////////////////////////////////////////////
|
||||
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm,_shm_processors);
|
||||
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm);
|
||||
InitFromMPICommunicator(processors,optimal_comm);
|
||||
SetCommunicator(optimal_comm);
|
||||
///////////////////////////////////////////////////
|
||||
@ -124,13 +124,12 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
|
||||
int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
|
||||
Coordinate parent_processor_coor(_ndimension,0);
|
||||
Coordinate parent_processors (_ndimension,1);
|
||||
Coordinate shm_processors (_ndimension,1);
|
||||
|
||||
// Can make 5d grid from 4d etc...
|
||||
int pad = _ndimension-parent_ndimension;
|
||||
for(int d=0;d<parent_ndimension;d++){
|
||||
parent_processor_coor[pad+d]=parent._processor_coor[d];
|
||||
parent_processors [pad+d]=parent._processors[d];
|
||||
shm_processors [pad+d]=parent._shm_processors[d];
|
||||
}
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@ -155,7 +154,6 @@ CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const
|
||||
ccoor[d] = parent_processor_coor[d] % processors[d];
|
||||
scoor[d] = parent_processor_coor[d] / processors[d];
|
||||
ssize[d] = parent_processors[d] / processors[d];
|
||||
if ( processors[d] < shm_processors[d] ) shm_processors[d] = processors[d]; // subnode splitting.
|
||||
}
|
||||
|
||||
// rank within subcomm ; srank is rank of subcomm within blocks of subcomms
|
||||
@ -337,22 +335,22 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
|
||||
}
|
||||
// Basic Halo comms primitive
|
||||
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
|
||||
int dest, int dox,
|
||||
int dest,
|
||||
void *recv,
|
||||
int from, int dor,
|
||||
int from,
|
||||
int bytes,int dir)
|
||||
{
|
||||
std::vector<CommsRequest_t> list;
|
||||
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,dox,recv,from,dor,bytes,dir);
|
||||
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
|
||||
StencilSendToRecvFromComplete(list,dir);
|
||||
return offbytes;
|
||||
}
|
||||
|
||||
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
|
||||
void *xmit,
|
||||
int dest,int dox,
|
||||
int dest,
|
||||
void *recv,
|
||||
int from,int dor,
|
||||
int from,
|
||||
int bytes,int dir)
|
||||
{
|
||||
int ncomm =communicator_halo.size();
|
||||
@ -372,33 +370,30 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
|
||||
double off_node_bytes=0.0;
|
||||
int tag;
|
||||
|
||||
if ( dor ) {
|
||||
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
|
||||
tag= dir+from*32;
|
||||
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
|
||||
assert(ierr==0);
|
||||
list.push_back(rrq);
|
||||
off_node_bytes+=bytes;
|
||||
}
|
||||
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
|
||||
tag= dir+from*32;
|
||||
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
|
||||
assert(ierr==0);
|
||||
list.push_back(rrq);
|
||||
off_node_bytes+=bytes;
|
||||
}
|
||||
|
||||
if (dox) {
|
||||
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
|
||||
tag= dir+_processor*32;
|
||||
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
|
||||
assert(ierr==0);
|
||||
list.push_back(xrq);
|
||||
off_node_bytes+=bytes;
|
||||
} else {
|
||||
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
|
||||
assert(shm!=NULL);
|
||||
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
|
||||
}
|
||||
|
||||
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
|
||||
tag= dir+_processor*32;
|
||||
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
|
||||
assert(ierr==0);
|
||||
list.push_back(xrq);
|
||||
off_node_bytes+=bytes;
|
||||
} else {
|
||||
// TODO : make a OMP loop on CPU, call threaded bcopy
|
||||
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
|
||||
assert(shm!=NULL);
|
||||
// std::cout <<"acceleratorCopyDeviceToDeviceAsynch"<< std::endl;
|
||||
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
|
||||
}
|
||||
|
||||
|
||||
if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
|
||||
this->StencilSendToRecvFromComplete(list,dir);
|
||||
list.resize(0);
|
||||
}
|
||||
|
||||
return off_node_bytes;
|
||||
|
||||
@ -45,14 +45,12 @@ void CartesianCommunicator::Init(int *argc, char *** arv)
|
||||
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors,const CartesianCommunicator &parent,int &srank)
|
||||
: CartesianCommunicator(processors)
|
||||
{
|
||||
_shm_processors = Coordinate(processors.size(),1);
|
||||
srank=0;
|
||||
SetCommunicator(communicator_world);
|
||||
}
|
||||
|
||||
CartesianCommunicator::CartesianCommunicator(const Coordinate &processors)
|
||||
{
|
||||
_shm_processors = Coordinate(processors.size(),1);
|
||||
_processors = processors;
|
||||
_ndimension = processors.size(); assert(_ndimension>=1);
|
||||
_processor_coor.resize(_ndimension);
|
||||
@ -113,18 +111,18 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
|
||||
}
|
||||
|
||||
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
|
||||
int xmit_to_rank,int dox,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,int dor,
|
||||
int recv_from_rank,
|
||||
int bytes, int dir)
|
||||
{
|
||||
return 2.0*bytes;
|
||||
}
|
||||
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
|
||||
void *xmit,
|
||||
int xmit_to_rank,int dox,
|
||||
int xmit_to_rank,
|
||||
void *recv,
|
||||
int recv_from_rank,int dor,
|
||||
int recv_from_rank,
|
||||
int bytes, int dir)
|
||||
{
|
||||
return 2.0*bytes;
|
||||
|
||||
@ -93,10 +93,9 @@ public:
|
||||
// Create an optimal reordered communicator that makes MPI_Cart_create get it right
|
||||
//////////////////////////////////////////////////////////////////////////////////////
|
||||
static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
|
||||
// Turns MPI_COMM_WORLD into right layout for Cartesian
|
||||
static void OptimalCommunicator (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);
|
||||
static void OptimalCommunicatorHypercube (const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);
|
||||
static void OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &ShmDims);
|
||||
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
|
||||
static void GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims);
|
||||
///////////////////////////////////////////////////
|
||||
// Provide shared memory facilities off comm world
|
||||
|
||||
@ -152,7 +152,7 @@ int Log2Size(int TwoToPower,int MAXLOG2)
|
||||
}
|
||||
return log2size;
|
||||
}
|
||||
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
|
||||
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
|
||||
@ -165,8 +165,8 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
|
||||
gethostname(name,namelen);
|
||||
int nscan = sscanf(name,"r%di%dn%d",&R,&I,&N) ;
|
||||
|
||||
if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm,SHM);
|
||||
else OptimalCommunicatorSharedMemory(processors,optimal_comm,SHM);
|
||||
if(nscan==3 && HPEhypercube ) OptimalCommunicatorHypercube(processors,optimal_comm);
|
||||
else OptimalCommunicatorSharedMemory(processors,optimal_comm);
|
||||
}
|
||||
static inline int divides(int a,int b)
|
||||
{
|
||||
@ -221,7 +221,7 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
|
||||
dim=(dim+1) %ndimension;
|
||||
}
|
||||
}
|
||||
void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
|
||||
void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
|
||||
{
|
||||
////////////////////////////////////////////////////////////////
|
||||
// Assert power of two shm_size.
|
||||
@ -294,8 +294,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
|
||||
Coordinate HyperCoor(ndimension);
|
||||
|
||||
GetShmDims(WorldDims,ShmDims);
|
||||
SHM = ShmDims;
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////
|
||||
// Establish torus of processes and nodes with sub-blockings
|
||||
////////////////////////////////////////////////////////////////
|
||||
@ -342,7 +341,7 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
|
||||
int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
|
||||
assert(ierr==0);
|
||||
}
|
||||
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
|
||||
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
|
||||
{
|
||||
////////////////////////////////////////////////////////////////
|
||||
// Identify subblock of ranks on node spreading across dims
|
||||
@ -354,8 +353,6 @@ void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &proce
|
||||
Coordinate ShmCoor(ndimension); Coordinate NodeCoor(ndimension); Coordinate WorldCoor(ndimension);
|
||||
|
||||
GetShmDims(WorldDims,ShmDims);
|
||||
SHM=ShmDims;
|
||||
|
||||
////////////////////////////////////////////////////////////////
|
||||
// Establish torus of processes and nodes with sub-blockings
|
||||
////////////////////////////////////////////////////////////////
|
||||
|
||||
@ -48,10 +48,9 @@ void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
|
||||
_ShmSetup=1;
|
||||
}
|
||||
|
||||
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm,Coordinate &SHM)
|
||||
void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
|
||||
{
|
||||
optimal_comm = WorldComm;
|
||||
SHM = Coordinate(processors.size(),1);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@ -46,4 +46,3 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
#include <Grid/lattice/Lattice_unary.h>
|
||||
#include <Grid/lattice/Lattice_transfer.h>
|
||||
#include <Grid/lattice/Lattice_basis.h>
|
||||
#include <Grid/lattice/Lattice_crc.h>
|
||||
|
||||
@ -1,55 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/lattice/Lattice_crc.h
|
||||
|
||||
Copyright (C) 2021
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#pragma once
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
|
||||
{
|
||||
auto ff = localNorm2(f);
|
||||
if ( mu==-1 ) mu = f.Grid()->Nd()-1;
|
||||
typedef typename vobj::tensor_reduced normtype;
|
||||
typedef typename normtype::scalar_object scalar;
|
||||
std::vector<scalar> sff;
|
||||
sliceSum(ff,sff,mu);
|
||||
for(int t=0;t<sff.size();t++){
|
||||
std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
template<class vobj> uint32_t crc(Lattice<vobj> & buf)
|
||||
{
|
||||
autoView( buf_v , buf, CpuRead);
|
||||
return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
|
||||
}
|
||||
|
||||
#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
|
||||
@ -125,12 +125,6 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
|
||||
//////////////////////////////////////////////////////////
|
||||
// Peek a scalar object from the SIMD array
|
||||
//////////////////////////////////////////////////////////
|
||||
template<class vobj>
|
||||
typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){
|
||||
typename vobj::scalar_object s;
|
||||
peekSite(s,l,site);
|
||||
return s;
|
||||
}
|
||||
template<class vobj,class sobj>
|
||||
void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
|
||||
|
||||
|
||||
@ -142,15 +142,6 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
|
||||
return sumD_cpu(arg,osites);
|
||||
#endif
|
||||
}
|
||||
template<class vobj>
|
||||
inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
|
||||
{
|
||||
#if defined(GRID_CUDA)||defined(GRID_HIP)
|
||||
return sumD_gpu_large(arg,osites);
|
||||
#else
|
||||
return sumD_cpu(arg,osites);
|
||||
#endif
|
||||
}
|
||||
|
||||
template<class vobj>
|
||||
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
|
||||
@ -168,22 +159,6 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
|
||||
return ssum;
|
||||
}
|
||||
|
||||
template<class vobj>
|
||||
inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
|
||||
{
|
||||
#if defined(GRID_CUDA)||defined(GRID_HIP)
|
||||
autoView( arg_v, arg, AcceleratorRead);
|
||||
Integer osites = arg.Grid()->oSites();
|
||||
auto ssum= sum_gpu_large(&arg_v[0],osites);
|
||||
#else
|
||||
autoView(arg_v, arg, CpuRead);
|
||||
Integer osites = arg.Grid()->oSites();
|
||||
auto ssum= sum_cpu(&arg_v[0],osites);
|
||||
#endif
|
||||
arg.Grid()->GlobalSum(ssum);
|
||||
return ssum;
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Deterministic Reduction operations
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@ -232,7 +207,6 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
|
||||
const uint64_t sites = grid->oSites();
|
||||
|
||||
// Might make all code paths go this way.
|
||||
#if 0
|
||||
typedef decltype(innerProductD(vobj(),vobj())) inner_t;
|
||||
Vector<inner_t> inner_tmp(sites);
|
||||
auto inner_tmp_v = &inner_tmp[0];
|
||||
@ -242,31 +216,15 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
|
||||
autoView( right_v,right, AcceleratorRead);
|
||||
|
||||
// GPU - SIMT lane compliance...
|
||||
accelerator_for( ss, sites, nsimd,{
|
||||
auto x_l = left_v(ss);
|
||||
auto y_l = right_v(ss);
|
||||
coalescedWrite(inner_tmp_v[ss],innerProductD(x_l,y_l));
|
||||
accelerator_for( ss, sites, 1,{
|
||||
auto x_l = left_v[ss];
|
||||
auto y_l = right_v[ss];
|
||||
inner_tmp_v[ss]=innerProductD(x_l,y_l);
|
||||
});
|
||||
}
|
||||
#else
|
||||
typedef decltype(innerProduct(vobj(),vobj())) inner_t;
|
||||
Vector<inner_t> inner_tmp(sites);
|
||||
auto inner_tmp_v = &inner_tmp[0];
|
||||
|
||||
{
|
||||
autoView( left_v , left, AcceleratorRead);
|
||||
autoView( right_v,right, AcceleratorRead);
|
||||
|
||||
// GPU - SIMT lane compliance...
|
||||
accelerator_for( ss, sites, nsimd,{
|
||||
auto x_l = left_v(ss);
|
||||
auto y_l = right_v(ss);
|
||||
coalescedWrite(inner_tmp_v[ss],innerProduct(x_l,y_l));
|
||||
});
|
||||
}
|
||||
#endif
|
||||
// This is in single precision and fails some tests
|
||||
auto anrm = sumD(inner_tmp_v,sites);
|
||||
auto anrm = sum(inner_tmp_v,sites);
|
||||
nrm = anrm;
|
||||
return nrm;
|
||||
}
|
||||
@ -300,7 +258,7 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
|
||||
conformable(x,y);
|
||||
|
||||
typedef typename vobj::scalar_type scalar_type;
|
||||
// typedef typename vobj::vector_typeD vector_type;
|
||||
typedef typename vobj::vector_typeD vector_type;
|
||||
RealD nrm;
|
||||
|
||||
GridBase *grid = x.Grid();
|
||||
@ -312,29 +270,17 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
|
||||
autoView( x_v, x, AcceleratorRead);
|
||||
autoView( y_v, y, AcceleratorRead);
|
||||
autoView( z_v, z, AcceleratorWrite);
|
||||
#if 0
|
||||
|
||||
typedef decltype(innerProductD(x_v[0],y_v[0])) inner_t;
|
||||
Vector<inner_t> inner_tmp(sites);
|
||||
auto inner_tmp_v = &inner_tmp[0];
|
||||
|
||||
accelerator_for( ss, sites, nsimd,{
|
||||
auto tmp = a*x_v(ss)+b*y_v(ss);
|
||||
coalescedWrite(inner_tmp_v[ss],innerProductD(tmp,tmp));
|
||||
coalescedWrite(z_v[ss],tmp);
|
||||
accelerator_for( ss, sites, 1,{
|
||||
auto tmp = a*x_v[ss]+b*y_v[ss];
|
||||
inner_tmp_v[ss]=innerProductD(tmp,tmp);
|
||||
z_v[ss]=tmp;
|
||||
});
|
||||
nrm = real(TensorRemove(sum(inner_tmp_v,sites)));
|
||||
#else
|
||||
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(inner_tmp_v,sites)));
|
||||
#endif
|
||||
grid->GlobalSum(nrm);
|
||||
return nrm;
|
||||
}
|
||||
|
||||
@ -23,7 +23,7 @@ unsigned int nextPow2(Iterator x) {
|
||||
}
|
||||
|
||||
template <class Iterator>
|
||||
int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
|
||||
void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
|
||||
|
||||
int device;
|
||||
#ifdef GRID_CUDA
|
||||
@ -37,13 +37,13 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
|
||||
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 << "\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);
|
||||
@ -53,12 +53,12 @@ int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &
|
||||
threads = warpSize;
|
||||
if ( threads*sizeofsobj > sharedMemPerBlock ) {
|
||||
std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
|
||||
return 0;
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
|
||||
// keep all the streaming multiprocessors busy
|
||||
blocks = nextPow2(multiProcessorCount);
|
||||
return 1;
|
||||
|
||||
}
|
||||
|
||||
template <class sobj, class Iterator>
|
||||
@ -198,7 +198,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
|
||||
// Possibly promote to double and sum
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
template <class vobj>
|
||||
inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
|
||||
inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
|
||||
{
|
||||
typedef typename vobj::scalar_objectD sobj;
|
||||
typedef decltype(lat) Iterator;
|
||||
@ -207,9 +207,7 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
|
||||
Integer size = osites*nsimd;
|
||||
|
||||
Integer numThreads, numBlocks;
|
||||
int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
|
||||
assert(ok);
|
||||
|
||||
getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
|
||||
Integer smemSize = numThreads * sizeof(sobj);
|
||||
|
||||
Vector<sobj> buffer(numBlocks);
|
||||
@ -220,54 +218,6 @@ inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osi
|
||||
auto result = buffer_v[0];
|
||||
return result;
|
||||
}
|
||||
|
||||
template <class vobj>
|
||||
inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
|
||||
{
|
||||
typedef typename vobj::vector_type vector;
|
||||
typedef typename vobj::scalar_typeD scalarD;
|
||||
typedef typename vobj::scalar_objectD sobj;
|
||||
sobj ret;
|
||||
scalarD *ret_p = (scalarD *)&ret;
|
||||
|
||||
const int words = sizeof(vobj)/sizeof(vector);
|
||||
|
||||
Vector<vector> buffer(osites);
|
||||
vector *dat = (vector *)lat;
|
||||
vector *buf = &buffer[0];
|
||||
iScalar<vector> *tbuf =(iScalar<vector> *) &buffer[0];
|
||||
for(int w=0;w<words;w++) {
|
||||
|
||||
accelerator_for(ss,osites,1,{
|
||||
buf[ss] = dat[ss*words+w];
|
||||
});
|
||||
|
||||
ret_p[w] = sumD_gpu_small(tbuf,osites);
|
||||
}
|
||||
return ret;
|
||||
}
|
||||
|
||||
template <class vobj>
|
||||
inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
|
||||
{
|
||||
typedef typename vobj::vector_type vector;
|
||||
typedef typename vobj::scalar_typeD scalarD;
|
||||
typedef typename vobj::scalar_objectD sobj;
|
||||
sobj ret;
|
||||
|
||||
Integer nsimd= vobj::Nsimd();
|
||||
Integer size = osites*nsimd;
|
||||
Integer numThreads, numBlocks;
|
||||
int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
|
||||
|
||||
if ( ok ) {
|
||||
ret = sumD_gpu_small(lat,osites);
|
||||
} else {
|
||||
ret = sumD_gpu_large(lat,osites);
|
||||
}
|
||||
return ret;
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Return as same precision as input performing reduction in double precision though
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
@ -280,13 +230,6 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
|
||||
return result;
|
||||
}
|
||||
|
||||
template <class vobj>
|
||||
inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
|
||||
{
|
||||
typedef typename vobj::scalar_object sobj;
|
||||
sobj result;
|
||||
result = sumD_gpu_large(lat,osites);
|
||||
return result;
|
||||
}
|
||||
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -424,32 +424,9 @@ public:
|
||||
// MT implementation does not implement fast discard even though
|
||||
// in principle this is possible
|
||||
////////////////////////////////////////////////
|
||||
#if 1
|
||||
thread_for( lidx, _grid->lSites(), {
|
||||
|
||||
int gidx;
|
||||
int o_idx;
|
||||
int i_idx;
|
||||
int rank;
|
||||
Coordinate pcoor;
|
||||
Coordinate lcoor;
|
||||
Coordinate gcoor;
|
||||
_grid->LocalIndexToLocalCoor(lidx,lcoor);
|
||||
pcoor=_grid->ThisProcessorCoor();
|
||||
_grid->ProcessorCoorLocalCoorToGlobalCoor(pcoor,lcoor,gcoor);
|
||||
_grid->GlobalCoorToGlobalIndex(gcoor,gidx);
|
||||
|
||||
_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
|
||||
assert(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
|
||||
// Everybody loops over global volume.
|
||||
thread_for( gidx, _grid->_gsites, {
|
||||
|
||||
// Where is it?
|
||||
int rank;
|
||||
int o_idx;
|
||||
@ -466,7 +443,6 @@ public:
|
||||
Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
|
||||
}
|
||||
});
|
||||
#endif
|
||||
#else
|
||||
////////////////////////////////////////////////////////////////
|
||||
// Machine and thread decomposition dependent seeding is efficient
|
||||
|
||||
@ -855,7 +855,7 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
|
||||
|
||||
|
||||
template<class vobj>
|
||||
void Replicate(const Lattice<vobj> &coarse,Lattice<vobj> & fine)
|
||||
void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
|
||||
{
|
||||
typedef typename vobj::scalar_object sobj;
|
||||
|
||||
|
||||
@ -65,34 +65,29 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
|
||||
GridLogger GridLogError (1, "Error" , GridLogColours, "RED");
|
||||
GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
|
||||
GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
|
||||
GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
|
||||
GridLogger GridLogDebug (1, "Debug", GridLogColours, "PURPLE");
|
||||
GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
|
||||
GridLogger GridLogIterative (1, "Iterative", GridLogColours, "BLUE");
|
||||
GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
|
||||
GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
|
||||
|
||||
void GridLogConfigure(std::vector<std::string> &logstreams) {
|
||||
GridLogError.Active(1);
|
||||
GridLogError.Active(0);
|
||||
GridLogWarning.Active(0);
|
||||
GridLogMessage.Active(1); // at least the messages should be always on
|
||||
GridLogMemory.Active(0); // at least the messages should be always on
|
||||
GridLogIterative.Active(0);
|
||||
GridLogDebug.Active(0);
|
||||
GridLogPerformance.Active(0);
|
||||
GridLogIntegrator.Active(1);
|
||||
GridLogColours.Active(0);
|
||||
GridLogHMC.Active(1);
|
||||
|
||||
for (int i = 0; i < logstreams.size(); i++) {
|
||||
if (logstreams[i] == std::string("Memory")) GridLogMemory.Active(1);
|
||||
if (logstreams[i] == std::string("Error")) GridLogError.Active(1);
|
||||
if (logstreams[i] == std::string("Warning")) GridLogWarning.Active(1);
|
||||
if (logstreams[i] == std::string("NoMessage")) GridLogMessage.Active(0);
|
||||
if (logstreams[i] == std::string("Iterative")) GridLogIterative.Active(1);
|
||||
if (logstreams[i] == std::string("Debug")) GridLogDebug.Active(1);
|
||||
if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
|
||||
if (logstreams[i] == std::string("NoIntegrator")) GridLogIntegrator.Active(0);
|
||||
if (logstreams[i] == std::string("NoHMC")) GridLogHMC.Active(0);
|
||||
if (logstreams[i] == std::string("Integrator")) GridLogIntegrator.Active(1);
|
||||
if (logstreams[i] == std::string("Colours")) GridLogColours.Active(1);
|
||||
}
|
||||
}
|
||||
|
||||
@ -182,8 +182,6 @@ extern GridLogger GridLogDebug ;
|
||||
extern GridLogger GridLogPerformance;
|
||||
extern GridLogger GridLogIterative ;
|
||||
extern GridLogger GridLogIntegrator ;
|
||||
extern GridLogger GridLogHMC;
|
||||
extern GridLogger GridLogMemory;
|
||||
extern Colours GridLogColours;
|
||||
|
||||
std::string demangle(const char* name) ;
|
||||
|
||||
@ -31,7 +31,6 @@ directory
|
||||
#include <fstream>
|
||||
#include <iomanip>
|
||||
#include <iostream>
|
||||
#include <string>
|
||||
#include <map>
|
||||
|
||||
#include <pwd.h>
|
||||
@ -655,8 +654,7 @@ class IldgWriter : public ScidacWriter {
|
||||
// Fill ILDG header data struct
|
||||
//////////////////////////////////////////////////////
|
||||
ildgFormat ildgfmt ;
|
||||
const std::string stNC = std::to_string( Nc ) ;
|
||||
ildgfmt.field = std::string("su"+stNC+"gauge");
|
||||
ildgfmt.field = std::string("su3gauge");
|
||||
|
||||
if ( format == std::string("IEEE32BIG") ) {
|
||||
ildgfmt.precision = 32;
|
||||
@ -873,8 +871,7 @@ class IldgReader : public GridLimeReader {
|
||||
} else {
|
||||
|
||||
assert(found_ildgFormat);
|
||||
const std::string stNC = std::to_string( Nc ) ;
|
||||
assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") );
|
||||
assert ( ildgFormat_.field == std::string("su3gauge") );
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////
|
||||
// Populate our Grid metadata as best we can
|
||||
@ -882,7 +879,7 @@ class IldgReader : public GridLimeReader {
|
||||
|
||||
std::ostringstream vers; vers << ildgFormat_.version;
|
||||
FieldMetaData_.hdr_version = vers.str();
|
||||
FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC);
|
||||
FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
|
||||
|
||||
FieldMetaData_.nd=4;
|
||||
FieldMetaData_.dimension.resize(4);
|
||||
|
||||
@ -6,8 +6,8 @@
|
||||
|
||||
Copyright (C) 2015
|
||||
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
|
||||
|
||||
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
|
||||
@ -182,8 +182,8 @@ class GaugeStatistics
|
||||
public:
|
||||
void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
|
||||
{
|
||||
header.link_trace = WilsonLoops<Impl>::linkTrace(data);
|
||||
header.plaquette = WilsonLoops<Impl>::avgPlaquette(data);
|
||||
header.link_trace=WilsonLoops<Impl>::linkTrace(data);
|
||||
header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
|
||||
}
|
||||
};
|
||||
typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
|
||||
@ -203,24 +203,20 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
|
||||
//////////////////////////////////////////////////////////////////////
|
||||
inline void reconstruct3(LorentzColourMatrix & cm)
|
||||
{
|
||||
assert( Nc < 4 && Nc > 1 ) ;
|
||||
const int x=0;
|
||||
const int y=1;
|
||||
const int z=2;
|
||||
for(int mu=0;mu<Nd;mu++){
|
||||
#if Nc == 2
|
||||
cm(mu)()(1,0) = -adj(cm(mu)()(0,y)) ;
|
||||
cm(mu)()(1,1) = adj(cm(mu)()(0,x)) ;
|
||||
#else
|
||||
const int x=0 , y=1 , z=2 ; // a little disinenuous labelling
|
||||
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
|
||||
#endif
|
||||
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>, Nc-1>, Nd >;
|
||||
template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
|
||||
|
||||
typedef iLorentzColour2x3<Complex> LorentzColour2x3;
|
||||
typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
|
||||
@ -282,6 +278,7 @@ struct GaugeSimpleMunger{
|
||||
|
||||
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++) {
|
||||
@ -320,8 +317,8 @@ 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<Nc-1;i++){
|
||||
for(int j=0;j<Nc;j++){
|
||||
for(int i=0;i<2;i++){
|
||||
for(int j=0;j<3;j++){
|
||||
out(mu)()(i,j) = in(mu)(i)(j);
|
||||
}}
|
||||
}
|
||||
@ -333,8 +330,8 @@ 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<Nc-1;i++){
|
||||
for(int j=0;j<Nc;j++){
|
||||
for(int i=0;i<2;i++){
|
||||
for(int j=0;j<3;j++){
|
||||
out(mu)(i)(j) = in(mu)()(i,j);
|
||||
}}
|
||||
}
|
||||
|
||||
@ -9,7 +9,6 @@
|
||||
Author: Matt Spraggs <matthew.spraggs@gmail.com>
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
|
||||
|
||||
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
|
||||
@ -31,8 +30,6 @@
|
||||
#ifndef GRID_NERSC_IO_H
|
||||
#define GRID_NERSC_IO_H
|
||||
|
||||
#include <string>
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
using namespace Grid;
|
||||
@ -42,10 +39,8 @@ using namespace Grid;
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
class NerscIO : public BinaryIO {
|
||||
public:
|
||||
typedef Lattice<vLorentzColourMatrixD> GaugeField;
|
||||
|
||||
// Enable/disable exiting if the plaquette in the header does not match the value computed (default true)
|
||||
static bool & exitOnReadPlaquetteMismatch(){ static bool v=true; return v; }
|
||||
typedef Lattice<vLorentzColourMatrixD> GaugeField;
|
||||
|
||||
static inline void truncate(std::string file){
|
||||
std::ofstream fout(file,std::ios::out);
|
||||
@ -150,17 +145,15 @@ public:
|
||||
|
||||
std::string format(header.floating_point);
|
||||
|
||||
const int ieee32big = (format == std::string("IEEE32BIG"));
|
||||
const int ieee32 = (format == std::string("IEEE32"));
|
||||
const int ieee64big = (format == std::string("IEEE64BIG"));
|
||||
const int ieee64 = (format == std::string("IEEE64") || \
|
||||
format == std::string("IEEE64LITTLE"));
|
||||
int ieee32big = (format == std::string("IEEE32BIG"));
|
||||
int ieee32 = (format == std::string("IEEE32"));
|
||||
int ieee64big = (format == std::string("IEEE64BIG"));
|
||||
int ieee64 = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
|
||||
|
||||
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
|
||||
// depending on datatype, set up munger;
|
||||
// munger is a function of <floating point, Real, data_type>
|
||||
const std::string stNC = std::to_string( Nc ) ;
|
||||
if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) {
|
||||
if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
|
||||
if ( ieee32 || ieee32big ) {
|
||||
BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F>
|
||||
(Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
|
||||
@ -171,7 +164,7 @@ public:
|
||||
(Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
|
||||
nersc_csum,scidac_csuma,scidac_csumb);
|
||||
}
|
||||
} else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) {
|
||||
} else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
|
||||
if ( ieee32 || ieee32big ) {
|
||||
BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
|
||||
(Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
|
||||
@ -205,7 +198,7 @@ public:
|
||||
std::cerr << " nersc_csum " <<std::hex<< nersc_csum << " " << header.checksum<< std::dec<< std::endl;
|
||||
exit(0);
|
||||
}
|
||||
if(exitOnReadPlaquetteMismatch()) assert(fabs(clone.plaquette -header.plaquette ) < 1.0e-5 );
|
||||
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 );
|
||||
|
||||
@ -216,29 +209,27 @@ public:
|
||||
template<class GaugeStats=PeriodicGaugeStatistics>
|
||||
static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
|
||||
std::string file,
|
||||
std::string ens_label = std::string("DWF"),
|
||||
std::string ens_id = std::string("UKQCD"),
|
||||
unsigned int sequence_number = 1)
|
||||
std::string ens_label = std::string("DWF"))
|
||||
{
|
||||
writeConfiguration(Umu,file,0,1,ens_label,ens_id,sequence_number);
|
||||
writeConfiguration(Umu,file,0,1,ens_label);
|
||||
}
|
||||
template<class GaugeStats=PeriodicGaugeStatistics>
|
||||
static inline void writeConfiguration(Lattice<vLorentzColourMatrixD > &Umu,
|
||||
std::string file,
|
||||
int two_row,
|
||||
int bits32,
|
||||
std::string ens_label = std::string("DWF"),
|
||||
std::string ens_id = std::string("UKQCD"),
|
||||
unsigned int sequence_number = 1)
|
||||
std::string ens_label = std::string("DWF"))
|
||||
{
|
||||
typedef vLorentzColourMatrixD vobj;
|
||||
typedef typename vobj::scalar_object sobj;
|
||||
|
||||
FieldMetaData header;
|
||||
header.sequence_number = sequence_number;
|
||||
header.ensemble_id = ens_id;
|
||||
///////////////////////////////////////////
|
||||
// Following should become arguments
|
||||
///////////////////////////////////////////
|
||||
header.sequence_number = 1;
|
||||
header.ensemble_id = std::string("UKQCD");
|
||||
header.ensemble_label = ens_label;
|
||||
header.hdr_version = "1.0" ;
|
||||
|
||||
typedef LorentzColourMatrixD fobj3D;
|
||||
typedef LorentzColour2x3D fobj2D;
|
||||
@ -252,14 +243,10 @@ public:
|
||||
|
||||
uint64_t offset;
|
||||
|
||||
// Sod it -- always write NcxNc double
|
||||
header.floating_point = std::string("IEEE64BIG");
|
||||
const std::string stNC = std::to_string( Nc ) ;
|
||||
if( two_row ) {
|
||||
header.data_type = std::string("4D_SU" + stNC + "_GAUGE" );
|
||||
} else {
|
||||
header.data_type = std::string("4D_SU" + stNC + "_GAUGE_" + stNC + "x" + stNC );
|
||||
}
|
||||
// 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);
|
||||
@ -267,15 +254,8 @@ public:
|
||||
grid->Broadcast(0,(void *)&offset,sizeof(offset));
|
||||
|
||||
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
|
||||
if( two_row ) {
|
||||
Gauge3x2unmunger<fobj2D,sobj> munge;
|
||||
BinaryIO::writeLatticeObject<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point,
|
||||
nersc_csum,scidac_csuma,scidac_csumb);
|
||||
} else {
|
||||
GaugeSimpleUnmunger<fobj3D,sobj> munge;
|
||||
BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
|
||||
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);
|
||||
@ -307,7 +287,8 @@ public:
|
||||
header.plaquette=0.0;
|
||||
MachineCharacteristics(header);
|
||||
|
||||
uint64_t offset;
|
||||
uint64_t offset;
|
||||
|
||||
#ifdef RNG_RANLUX
|
||||
header.floating_point = std::string("UINT64");
|
||||
header.data_type = std::string("RANLUX48");
|
||||
@ -347,7 +328,7 @@ public:
|
||||
|
||||
GridBase *grid = parallel.Grid();
|
||||
|
||||
uint64_t offset = readHeader(file,grid,header);
|
||||
uint64_t offset = readHeader(file,grid,header);
|
||||
|
||||
FieldMetaData clone(header);
|
||||
|
||||
|
||||
@ -72,9 +72,17 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
|
||||
inline uint64_t cyclecount(void){
|
||||
return 0;
|
||||
}
|
||||
#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
|
||||
#define __SSC_STOP __SSC_MARK(0x110)
|
||||
#define __SSC_START __SSC_MARK(0x111)
|
||||
|
||||
|
||||
#else
|
||||
|
||||
#define __SSC_MARK(mark)
|
||||
#define __SSC_STOP
|
||||
#define __SSC_START
|
||||
|
||||
/*
|
||||
* cycle counters arch dependent
|
||||
*/
|
||||
|
||||
@ -39,9 +39,9 @@ NAMESPACE_BEGIN(Grid)
|
||||
// C++11 time facilities better?
|
||||
inline double usecond(void) {
|
||||
struct timeval tv;
|
||||
tv.tv_sec = 0;
|
||||
tv.tv_usec = 0;
|
||||
#ifdef TIMERS_ON
|
||||
gettimeofday(&tv,NULL);
|
||||
#endif
|
||||
return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
|
||||
}
|
||||
|
||||
|
||||
@ -16,12 +16,8 @@
|
||||
|
||||
#ifdef __NVCC__
|
||||
#pragma push
|
||||
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
|
||||
#pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
|
||||
#else
|
||||
#pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#include "pugixml.h"
|
||||
|
||||
|
||||
@ -63,7 +63,6 @@ static constexpr int Ngp=2; // gparity index range
|
||||
#define ColourIndex (2)
|
||||
#define SpinIndex (1)
|
||||
#define LorentzIndex (0)
|
||||
#define GparityFlavourIndex (0)
|
||||
|
||||
// Also should make these a named enum type
|
||||
static constexpr int DaggerNo=0;
|
||||
@ -88,8 +87,6 @@ template<typename T> struct isCoarsened {
|
||||
template <typename T> using IfCoarsened = Invoke<std::enable_if< isCoarsened<T>::value,int> > ;
|
||||
template <typename T> using IfNotCoarsened = Invoke<std::enable_if<!isCoarsened<T>::value,int> > ;
|
||||
|
||||
const int GparityFlavourTensorIndex = 3; //TensorLevel counts from the bottom!
|
||||
|
||||
// ChrisK very keen to add extra space for Gparity doubling.
|
||||
//
|
||||
// Also add domain wall index, in a way where Wilson operator
|
||||
@ -113,10 +110,8 @@ template<typename vtype> using iHalfSpinColourVector = iScalar<iVector<iVec
|
||||
template<typename vtype> using iSpinColourSpinColourMatrix = iScalar<iMatrix<iMatrix<iMatrix<iMatrix<vtype, Nc>, Ns>, Nc>, Ns> >;
|
||||
|
||||
|
||||
template<typename vtype> using iGparityFlavourVector = iVector<iScalar<iScalar<vtype> >, 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 >;
|
||||
template<typename vtype> using iGparityFlavourMatrix = iMatrix<iScalar<iScalar<vtype> >, Ngp>;
|
||||
|
||||
// Spin matrix
|
||||
typedef iSpinMatrix<Complex > SpinMatrix;
|
||||
@ -181,16 +176,6 @@ typedef iDoubleStoredColourMatrix<vComplex > vDoubleStoredColourMatrix;
|
||||
typedef iDoubleStoredColourMatrix<vComplexF> vDoubleStoredColourMatrixF;
|
||||
typedef iDoubleStoredColourMatrix<vComplexD> vDoubleStoredColourMatrixD;
|
||||
|
||||
//G-parity flavour matrix
|
||||
typedef iGparityFlavourMatrix<Complex> GparityFlavourMatrix;
|
||||
typedef iGparityFlavourMatrix<ComplexF> GparityFlavourMatrixF;
|
||||
typedef iGparityFlavourMatrix<ComplexD> GparityFlavourMatrixD;
|
||||
|
||||
typedef iGparityFlavourMatrix<vComplex> vGparityFlavourMatrix;
|
||||
typedef iGparityFlavourMatrix<vComplexF> vGparityFlavourMatrixF;
|
||||
typedef iGparityFlavourMatrix<vComplexD> vGparityFlavourMatrixD;
|
||||
|
||||
|
||||
// Spin vector
|
||||
typedef iSpinVector<Complex > SpinVector;
|
||||
typedef iSpinVector<ComplexF> SpinVectorF;
|
||||
@ -235,16 +220,6 @@ typedef iHalfSpinColourVector<ComplexD> HalfSpinColourVectorD;
|
||||
typedef iHalfSpinColourVector<vComplex > vHalfSpinColourVector;
|
||||
typedef iHalfSpinColourVector<vComplexF> vHalfSpinColourVectorF;
|
||||
typedef iHalfSpinColourVector<vComplexD> vHalfSpinColourVectorD;
|
||||
|
||||
//G-parity flavour vector
|
||||
typedef iGparityFlavourVector<Complex > GparityFlavourVector;
|
||||
typedef iGparityFlavourVector<ComplexF> GparityFlavourVectorF;
|
||||
typedef iGparityFlavourVector<ComplexD> GparityFlavourVectorD;
|
||||
|
||||
typedef iGparityFlavourVector<vComplex > vGparityFlavourVector;
|
||||
typedef iGparityFlavourVector<vComplexF> vGparityFlavourVectorF;
|
||||
typedef iGparityFlavourVector<vComplexD> vGparityFlavourVectorD;
|
||||
|
||||
|
||||
// singlets
|
||||
typedef iSinglet<Complex > TComplex; // FIXME This is painful. Tensor singlet complex type.
|
||||
|
||||
@ -40,29 +40,6 @@ class Action
|
||||
|
||||
public:
|
||||
bool is_smeared = false;
|
||||
RealD deriv_norm_sum;
|
||||
RealD deriv_max_sum;
|
||||
int deriv_num;
|
||||
RealD deriv_us;
|
||||
RealD S_us;
|
||||
RealD refresh_us;
|
||||
void reset_timer(void) {
|
||||
deriv_us = S_us = refresh_us = 0.0;
|
||||
deriv_num=0;
|
||||
deriv_norm_sum = deriv_max_sum=0.0;
|
||||
}
|
||||
void deriv_log(RealD nrm, RealD max) { deriv_max_sum+=max; deriv_norm_sum+=nrm; deriv_num++;}
|
||||
RealD deriv_max_average(void) { return deriv_max_sum/deriv_num; };
|
||||
RealD deriv_norm_average(void) { return deriv_norm_sum/deriv_num; };
|
||||
RealD deriv_timer(void) { return deriv_us; };
|
||||
RealD S_timer(void) { return deriv_us; };
|
||||
RealD refresh_timer(void) { return deriv_us; };
|
||||
void deriv_timer_start(void) { deriv_us-=usecond(); }
|
||||
void deriv_timer_stop(void) { deriv_us+=usecond(); }
|
||||
void refresh_timer_start(void) { refresh_us-=usecond(); }
|
||||
void refresh_timer_stop(void) { refresh_us+=usecond(); }
|
||||
void S_timer_start(void) { S_us-=usecond(); }
|
||||
void S_timer_stop(void) { S_us+=usecond(); }
|
||||
// Heatbath?
|
||||
virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) = 0; // refresh pseudofermions
|
||||
virtual RealD S(const GaugeField& U) = 0; // evaluate the action
|
||||
|
||||
@ -37,10 +37,6 @@ NAMESPACE_CHECK(ActionSet);
|
||||
#include <Grid/qcd/action/ActionParams.h>
|
||||
NAMESPACE_CHECK(ActionParams);
|
||||
|
||||
#include <Grid/qcd/action/filters/MomentumFilter.h>
|
||||
#include <Grid/qcd/action/filters/DirichletFilter.h>
|
||||
#include <Grid/qcd/action/filters/DDHMCFilter.h>
|
||||
|
||||
////////////////////////////////////////////
|
||||
// Gauge Actions
|
||||
////////////////////////////////////////////
|
||||
|
||||
@ -37,33 +37,24 @@ NAMESPACE_BEGIN(Grid);
|
||||
// These can move into a params header and be given MacroMagic serialisation
|
||||
struct GparityWilsonImplParams {
|
||||
Coordinate twists;
|
||||
//mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
|
||||
Coordinate dirichlet; // Blocksize of dirichlet BCs
|
||||
GparityWilsonImplParams() : twists(Nd, 0), dirichlet(Nd, 0) {};
|
||||
GparityWilsonImplParams() : twists(Nd, 0) {};
|
||||
};
|
||||
|
||||
struct WilsonImplParams {
|
||||
bool overlapCommsCompute;
|
||||
Coordinate dirichlet; // Blocksize of dirichlet BCs
|
||||
AcceleratorVector<Real,Nd> twist_n_2pi_L;
|
||||
AcceleratorVector<Complex,Nd> boundary_phases;
|
||||
WilsonImplParams() {
|
||||
dirichlet.resize(Nd,0);
|
||||
boundary_phases.resize(Nd, 1.0);
|
||||
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);
|
||||
dirichlet.resize(Nd,0);
|
||||
}
|
||||
};
|
||||
|
||||
struct StaggeredImplParams {
|
||||
Coordinate dirichlet; // Blocksize of dirichlet BCs
|
||||
StaggeredImplParams()
|
||||
{
|
||||
dirichlet.resize(Nd,0);
|
||||
};
|
||||
StaggeredImplParams() {};
|
||||
};
|
||||
|
||||
struct OneFlavourRationalParams : Serializable {
|
||||
@ -72,11 +63,9 @@ struct StaggeredImplParams {
|
||||
RealD, hi,
|
||||
int, MaxIter,
|
||||
RealD, tolerance,
|
||||
RealD, mdtolerance,
|
||||
int, degree,
|
||||
int, precision,
|
||||
int, BoundsCheckFreq,
|
||||
RealD, BoundsCheckTol);
|
||||
int, BoundsCheckFreq);
|
||||
|
||||
// MaxIter and tolerance, vectors??
|
||||
|
||||
@ -87,62 +76,16 @@ struct StaggeredImplParams {
|
||||
RealD tol = 1.0e-8,
|
||||
int _degree = 10,
|
||||
int _precision = 64,
|
||||
int _BoundsCheckFreq=20,
|
||||
RealD mdtol = 1.0e-6,
|
||||
double _BoundsCheckTol=1e-6)
|
||||
int _BoundsCheckFreq=20)
|
||||
: lo(_lo),
|
||||
hi(_hi),
|
||||
MaxIter(_maxit),
|
||||
tolerance(tol),
|
||||
mdtolerance(mdtol),
|
||||
degree(_degree),
|
||||
precision(_precision),
|
||||
BoundsCheckFreq(_BoundsCheckFreq),
|
||||
BoundsCheckTol(_BoundsCheckTol){};
|
||||
BoundsCheckFreq(_BoundsCheckFreq){};
|
||||
};
|
||||
|
||||
/*Action parameters for the generalized rational action
|
||||
The approximation is for (M^dag M)^{1/inv_pow}
|
||||
where inv_pow is the denominator of the fractional power.
|
||||
Default inv_pow=2 for square root, making this equivalent to
|
||||
the OneFlavourRational action
|
||||
*/
|
||||
struct RationalActionParams : Serializable {
|
||||
GRID_SERIALIZABLE_CLASS_MEMBERS(RationalActionParams,
|
||||
int, inv_pow,
|
||||
RealD, lo, //low eigenvalue bound of rational approx
|
||||
RealD, hi, //high eigenvalue bound of rational approx
|
||||
int, MaxIter, //maximum iterations in msCG
|
||||
RealD, action_tolerance, //msCG tolerance in action evaluation
|
||||
int, action_degree, //rational approx tolerance in action evaluation
|
||||
RealD, md_tolerance, //msCG tolerance in MD integration
|
||||
int, md_degree, //rational approx tolerance in MD integration
|
||||
int, precision, //precision of floating point arithmetic
|
||||
int, BoundsCheckFreq); //frequency the approximation is tested (with Metropolis degree/tolerance); 0 disables the check
|
||||
// constructor
|
||||
RationalActionParams(int _inv_pow = 2,
|
||||
RealD _lo = 0.0,
|
||||
RealD _hi = 1.0,
|
||||
int _maxit = 1000,
|
||||
RealD _action_tolerance = 1.0e-8,
|
||||
int _action_degree = 10,
|
||||
RealD _md_tolerance = 1.0e-8,
|
||||
int _md_degree = 10,
|
||||
int _precision = 64,
|
||||
int _BoundsCheckFreq=20)
|
||||
: inv_pow(_inv_pow),
|
||||
lo(_lo),
|
||||
hi(_hi),
|
||||
MaxIter(_maxit),
|
||||
action_tolerance(_action_tolerance),
|
||||
action_degree(_action_degree),
|
||||
md_tolerance(_md_tolerance),
|
||||
md_degree(_md_degree),
|
||||
precision(_precision),
|
||||
BoundsCheckFreq(_BoundsCheckFreq){};
|
||||
};
|
||||
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
#endif
|
||||
|
||||
@ -68,17 +68,9 @@ public:
|
||||
///////////////////////////////////////////////////////////////
|
||||
// Support for MADWF tricks
|
||||
///////////////////////////////////////////////////////////////
|
||||
RealD Mass(void) { return (mass_plus + mass_minus) / 2.0; };
|
||||
RealD MassPlus(void) { return mass_plus; };
|
||||
RealD MassMinus(void) { return mass_minus; };
|
||||
|
||||
RealD Mass(void) { return mass; };
|
||||
void SetMass(RealD _mass) {
|
||||
mass_plus=mass_minus=_mass;
|
||||
SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c); // Reset coeffs
|
||||
} ;
|
||||
void SetMass(RealD _mass_plus, RealD _mass_minus) {
|
||||
mass_plus=_mass_plus;
|
||||
mass_minus=_mass_minus;
|
||||
mass=_mass;
|
||||
SetCoefficientsInternal(_zolo_hi,_gamma,_b,_c); // Reset coeffs
|
||||
} ;
|
||||
void P(const FermionField &psi, FermionField &chi);
|
||||
@ -116,7 +108,7 @@ public:
|
||||
void MeooeDag5D (const FermionField &in, FermionField &out);
|
||||
|
||||
// protected:
|
||||
RealD mass_plus, mass_minus;
|
||||
RealD mass;
|
||||
|
||||
// Save arguments to SetCoefficientsInternal
|
||||
Vector<Coeff_t> _gamma;
|
||||
|
||||
@ -1,435 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
|
||||
|
||||
Copyright (C) 2017 - 2022
|
||||
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
|
||||
Author: Mattia Bruno <mattia.bruno@cern.ch>
|
||||
|
||||
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.h>
|
||||
#include <Grid/qcd/spin/Dirac.h>
|
||||
#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
|
||||
|
||||
////////////////////////////////////////////
|
||||
// Standard Clover
|
||||
// (4+m0) + csw * clover_term
|
||||
// Exp Clover
|
||||
// (4+m0) * exp(csw/(4+m0) clover_term)
|
||||
// = (4+m0) + csw * clover_term + ...
|
||||
////////////////////////////////////////////
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
|
||||
//////////////////////////////////
|
||||
// Generic Standard Clover
|
||||
//////////////////////////////////
|
||||
|
||||
template<class Impl>
|
||||
class CloverHelpers: public WilsonCloverHelpers<Impl> {
|
||||
public:
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
INHERIT_CLOVER_TYPES(Impl);
|
||||
|
||||
typedef WilsonCloverHelpers<Impl> Helpers;
|
||||
|
||||
static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) {
|
||||
GridBase *grid = CloverTerm.Grid();
|
||||
CloverTerm += diag_mass;
|
||||
|
||||
int lvol = grid->lSites();
|
||||
int DimRep = Impl::Dimension;
|
||||
{
|
||||
autoView(CTv,CloverTerm,CpuRead);
|
||||
autoView(CTIv,CloverTermInv,CpuWrite);
|
||||
thread_for(site, lvol, {
|
||||
Coordinate lcoor;
|
||||
grid->LocalIndexToLocalCoor(site, lcoor);
|
||||
Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
|
||||
Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
|
||||
typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
|
||||
peekLocalSite(Qx, CTv, lcoor);
|
||||
|
||||
for (int j = 0; j < Ns; j++)
|
||||
for (int k = 0; k < Ns; k++)
|
||||
for (int a = 0; a < DimRep; a++)
|
||||
for (int b = 0; b < DimRep; b++){
|
||||
auto zz = Qx()(j, k)(a, b);
|
||||
EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
|
||||
}
|
||||
|
||||
EigenInvCloverOp = EigenCloverOp.inverse();
|
||||
for (int j = 0; j < Ns; j++)
|
||||
for (int k = 0; k < Ns; k++)
|
||||
for (int a = 0; a < DimRep; a++)
|
||||
for (int b = 0; b < DimRep; b++)
|
||||
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
|
||||
pokeLocalSite(Qxinv, CTIv, lcoor);
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
|
||||
return Helpers::Cmunu(U, lambda, mu, nu);
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
|
||||
//////////////////////////////////
|
||||
// Generic Exp Clover
|
||||
//////////////////////////////////
|
||||
|
||||
template<class Impl>
|
||||
class ExpCloverHelpers: public WilsonCloverHelpers<Impl> {
|
||||
public:
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
INHERIT_CLOVER_TYPES(Impl);
|
||||
|
||||
template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
|
||||
typedef WilsonCloverHelpers<Impl> Helpers;
|
||||
|
||||
// Can this be avoided?
|
||||
static void IdentityTimesC(const CloverField& in, RealD c) {
|
||||
int DimRep = Impl::Dimension;
|
||||
|
||||
autoView(in_v, in, AcceleratorWrite);
|
||||
|
||||
accelerator_for(ss, in.Grid()->oSites(), 1, {
|
||||
for (int sa=0; sa<Ns; sa++)
|
||||
for (int ca=0; ca<DimRep; ca++)
|
||||
in_v[ss]()(sa,sa)(ca,ca) = c;
|
||||
});
|
||||
}
|
||||
|
||||
static int getNMAX(RealD prec, RealD R) {
|
||||
/* compute stop condition for exponential */
|
||||
int NMAX=1;
|
||||
RealD cond=R*R/2.;
|
||||
|
||||
while (cond*std::exp(R)>prec) {
|
||||
NMAX++;
|
||||
cond*=R/(double)(NMAX+1);
|
||||
}
|
||||
return NMAX;
|
||||
}
|
||||
|
||||
static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
|
||||
static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
|
||||
|
||||
static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
|
||||
GridBase* grid = Clover.Grid();
|
||||
CloverField ExpClover(grid);
|
||||
|
||||
int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
|
||||
|
||||
Clover *= (1.0/diag_mass);
|
||||
|
||||
// Taylor expansion, slow but generic
|
||||
// Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
|
||||
// qN = cN
|
||||
// qn = cn + qn+1 X
|
||||
std::vector<RealD> cn(NMAX+1);
|
||||
cn[0] = 1.0;
|
||||
for (int i=1; i<=NMAX; i++)
|
||||
cn[i] = cn[i-1] / RealD(i);
|
||||
|
||||
ExpClover = Zero();
|
||||
IdentityTimesC(ExpClover, cn[NMAX]);
|
||||
for (int i=NMAX-1; i>=0; i--)
|
||||
ExpClover = ExpClover * Clover + cn[i];
|
||||
|
||||
// prepare inverse
|
||||
CloverInv = (-1.0)*Clover;
|
||||
|
||||
Clover = ExpClover * diag_mass;
|
||||
|
||||
ExpClover = Zero();
|
||||
IdentityTimesC(ExpClover, cn[NMAX]);
|
||||
for (int i=NMAX-1; i>=0; i--)
|
||||
ExpClover = ExpClover * CloverInv + cn[i];
|
||||
|
||||
CloverInv = ExpClover * (1.0/diag_mass);
|
||||
|
||||
}
|
||||
|
||||
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
|
||||
assert(0);
|
||||
return lambda;
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
|
||||
//////////////////////////////////
|
||||
// Compact Standard Clover
|
||||
//////////////////////////////////
|
||||
|
||||
|
||||
template<class Impl>
|
||||
class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>,
|
||||
public WilsonCloverHelpers<Impl> {
|
||||
public:
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
INHERIT_CLOVER_TYPES(Impl);
|
||||
INHERIT_COMPACT_CLOVER_TYPES(Impl);
|
||||
|
||||
typedef WilsonCloverHelpers<Impl> Helpers;
|
||||
typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
|
||||
|
||||
static void MassTerm(CloverField& Clover, RealD diag_mass) {
|
||||
Clover += diag_mass;
|
||||
}
|
||||
|
||||
static void Exponentiate_Clover(CloverDiagonalField& Diagonal,
|
||||
CloverTriangleField& Triangle,
|
||||
RealD csw_t, RealD diag_mass) {
|
||||
|
||||
// Do nothing
|
||||
}
|
||||
|
||||
// TODO: implement Cmunu for better performances with compact layout, but don't do it
|
||||
// here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers
|
||||
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
|
||||
return Helpers::Cmunu(U, lambda, mu, nu);
|
||||
}
|
||||
};
|
||||
|
||||
//////////////////////////////////
|
||||
// Compact Exp Clover
|
||||
//////////////////////////////////
|
||||
|
||||
template<class Impl>
|
||||
class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> {
|
||||
public:
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
INHERIT_CLOVER_TYPES(Impl);
|
||||
INHERIT_COMPACT_CLOVER_TYPES(Impl);
|
||||
|
||||
template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
|
||||
typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
|
||||
|
||||
static void MassTerm(CloverField& Clover, RealD diag_mass) {
|
||||
// do nothing!
|
||||
// mass term is multiplied to exp(Clover) below
|
||||
}
|
||||
|
||||
static int getNMAX(RealD prec, RealD R) {
|
||||
/* compute stop condition for exponential */
|
||||
int NMAX=1;
|
||||
RealD cond=R*R/2.;
|
||||
|
||||
while (cond*std::exp(R)>prec) {
|
||||
NMAX++;
|
||||
cond*=R/(double)(NMAX+1);
|
||||
}
|
||||
return NMAX;
|
||||
}
|
||||
|
||||
static int getNMAX(Lattice<iImplCloverDiagonal<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
|
||||
static int getNMAX(Lattice<iImplCloverDiagonal<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
|
||||
|
||||
static void ExponentiateHermitean6by6(const iMatrix<ComplexD,6> &arg, const RealD& alpha, const std::vector<RealD>& cN, const int Niter, iMatrix<ComplexD,6>& dest){
|
||||
|
||||
typedef iMatrix<ComplexD,6> mat;
|
||||
|
||||
RealD qn[6];
|
||||
RealD qnold[6];
|
||||
RealD p[5];
|
||||
RealD trA2, trA3, trA4;
|
||||
|
||||
mat A2, A3, A4, A5;
|
||||
A2 = alpha * alpha * arg * arg;
|
||||
A3 = alpha * arg * A2;
|
||||
A4 = A2 * A2;
|
||||
A5 = A2 * A3;
|
||||
|
||||
trA2 = toReal( trace(A2) );
|
||||
trA3 = toReal( trace(A3) );
|
||||
trA4 = toReal( trace(A4));
|
||||
|
||||
p[0] = toReal( trace(A3 * A3)) / 6.0 - 0.125 * trA4 * trA2 - trA3 * trA3 / 18.0 + trA2 * trA2 * trA2/ 48.0;
|
||||
p[1] = toReal( trace(A5)) / 5.0 - trA3 * trA2 / 6.0;
|
||||
p[2] = toReal( trace(A4)) / 4.0 - 0.125 * trA2 * trA2;
|
||||
p[3] = trA3 / 3.0;
|
||||
p[4] = 0.5 * trA2;
|
||||
|
||||
qnold[0] = cN[Niter];
|
||||
qnold[1] = 0.0;
|
||||
qnold[2] = 0.0;
|
||||
qnold[3] = 0.0;
|
||||
qnold[4] = 0.0;
|
||||
qnold[5] = 0.0;
|
||||
|
||||
for(int i = Niter-1; i >= 0; i--)
|
||||
{
|
||||
qn[0] = p[0] * qnold[5] + cN[i];
|
||||
qn[1] = p[1] * qnold[5] + qnold[0];
|
||||
qn[2] = p[2] * qnold[5] + qnold[1];
|
||||
qn[3] = p[3] * qnold[5] + qnold[2];
|
||||
qn[4] = p[4] * qnold[5] + qnold[3];
|
||||
qn[5] = qnold[4];
|
||||
|
||||
qnold[0] = qn[0];
|
||||
qnold[1] = qn[1];
|
||||
qnold[2] = qn[2];
|
||||
qnold[3] = qn[3];
|
||||
qnold[4] = qn[4];
|
||||
qnold[5] = qn[5];
|
||||
}
|
||||
|
||||
mat unit(1.0);
|
||||
|
||||
dest = (qn[0] * unit + qn[1] * alpha * arg + qn[2] * A2 + qn[3] * A3 + qn[4] * A4 + qn[5] * A5);
|
||||
|
||||
}
|
||||
|
||||
static void Exponentiate_Clover(CloverDiagonalField& Diagonal, CloverTriangleField& Triangle, RealD csw_t, RealD diag_mass) {
|
||||
|
||||
GridBase* grid = Diagonal.Grid();
|
||||
int NMAX = getNMAX(Diagonal, 3.*csw_t/diag_mass);
|
||||
|
||||
//
|
||||
// Implementation completely in Daniel's layout
|
||||
//
|
||||
|
||||
// Taylor expansion with Cayley-Hamilton recursion
|
||||
// underlying Horner scheme as above
|
||||
std::vector<RealD> cn(NMAX+1);
|
||||
cn[0] = 1.0;
|
||||
for (int i=1; i<=NMAX; i++){
|
||||
cn[i] = cn[i-1] / RealD(i);
|
||||
}
|
||||
|
||||
// Taken over from Daniel's implementation
|
||||
conformable(Diagonal, Triangle);
|
||||
|
||||
long lsites = grid->lSites();
|
||||
{
|
||||
typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
|
||||
typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
|
||||
typedef iMatrix<ComplexD,6> mat;
|
||||
|
||||
autoView(diagonal_v, Diagonal, CpuRead);
|
||||
autoView(triangle_v, Triangle, CpuRead);
|
||||
autoView(diagonalExp_v, Diagonal, CpuWrite);
|
||||
autoView(triangleExp_v, Triangle, CpuWrite);
|
||||
|
||||
thread_for(site, lsites, { // NOTE: Not on GPU because of (peek/poke)LocalSite
|
||||
|
||||
mat srcCloverOpUL(0.0); // upper left block
|
||||
mat srcCloverOpLR(0.0); // lower right block
|
||||
mat ExpCloverOp;
|
||||
|
||||
scalar_object_diagonal diagonal_tmp = Zero();
|
||||
scalar_object_diagonal diagonal_exp_tmp = Zero();
|
||||
scalar_object_triangle triangle_tmp = Zero();
|
||||
scalar_object_triangle triangle_exp_tmp = Zero();
|
||||
|
||||
Coordinate lcoor;
|
||||
grid->LocalIndexToLocalCoor(site, lcoor);
|
||||
|
||||
peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
|
||||
peekLocalSite(triangle_tmp, triangle_v, lcoor);
|
||||
|
||||
int block;
|
||||
block = 0;
|
||||
for(int i = 0; i < 6; i++){
|
||||
for(int j = 0; j < 6; j++){
|
||||
if (i == j){
|
||||
srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
|
||||
}
|
||||
else{
|
||||
srcCloverOpUL(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
|
||||
}
|
||||
}
|
||||
}
|
||||
block = 1;
|
||||
for(int i = 0; i < 6; i++){
|
||||
for(int j = 0; j < 6; j++){
|
||||
if (i == j){
|
||||
srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
|
||||
}
|
||||
else{
|
||||
srcCloverOpLR(i,j) = static_cast<ComplexD>(TensorRemove(CompactHelpers::triangle_elem(triangle_tmp, block, i, j)));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// exp(Clover)
|
||||
|
||||
ExponentiateHermitean6by6(srcCloverOpUL,1.0/diag_mass,cn,NMAX,ExpCloverOp);
|
||||
|
||||
block = 0;
|
||||
for(int i = 0; i < 6; i++){
|
||||
for(int j = 0; j < 6; j++){
|
||||
if (i == j){
|
||||
diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
|
||||
}
|
||||
else if(i < j){
|
||||
triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
ExponentiateHermitean6by6(srcCloverOpLR,1.0/diag_mass,cn,NMAX,ExpCloverOp);
|
||||
|
||||
block = 1;
|
||||
for(int i = 0; i < 6; i++){
|
||||
for(int j = 0; j < 6; j++){
|
||||
if (i == j){
|
||||
diagonal_exp_tmp()(block)(i) = ExpCloverOp(i,j);
|
||||
}
|
||||
else if(i < j){
|
||||
triangle_exp_tmp()(block)(CompactHelpers::triangle_index(i, j)) = ExpCloverOp(i,j);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
pokeLocalSite(diagonal_exp_tmp, diagonalExp_v, lcoor);
|
||||
pokeLocalSite(triangle_exp_tmp, triangleExp_v, lcoor);
|
||||
});
|
||||
}
|
||||
|
||||
Diagonal *= diag_mass;
|
||||
Triangle *= diag_mass;
|
||||
}
|
||||
|
||||
|
||||
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
|
||||
assert(0);
|
||||
return lambda;
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1,241 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermion.h
|
||||
|
||||
Copyright (C) 2020 - 2022
|
||||
|
||||
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
|
||||
Author: Nils Meyer <nils.meyer@ur.de>
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
|
||||
#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
|
||||
#include <Grid/qcd/action/fermion/CloverHelpers.h>
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
// see Grid/qcd/action/fermion/WilsonCloverFermion.h for description
|
||||
//
|
||||
// Modifications done here:
|
||||
//
|
||||
// Original: clover term = 12x12 matrix per site
|
||||
//
|
||||
// But: Only two diagonal 6x6 hermitian blocks are non-zero (also true for original, verified by running)
|
||||
// Sufficient to store/transfer only the real parts of the diagonal and one triangular part
|
||||
// 2 * (6 + 15 * 2) = 72 real or 36 complex words to be stored/transfered
|
||||
//
|
||||
// Here: Above but diagonal as complex numbers, i.e., need to store/transfer
|
||||
// 2 * (6 * 2 + 15 * 2) = 84 real or 42 complex words
|
||||
//
|
||||
// Words per site and improvement compared to original (combined with the input and output spinors):
|
||||
//
|
||||
// - Original: 2*12 + 12*12 = 168 words -> 1.00 x less
|
||||
// - Minimal: 2*12 + 36 = 60 words -> 2.80 x less
|
||||
// - Here: 2*12 + 42 = 66 words -> 2.55 x less
|
||||
//
|
||||
// These improvements directly translate to wall-clock time
|
||||
//
|
||||
// Data layout:
|
||||
//
|
||||
// - diagonal and triangle part as separate lattice fields,
|
||||
// this was faster than as 1 combined field on all tested machines
|
||||
// - diagonal: as expected
|
||||
// - triangle: store upper right triangle in row major order
|
||||
// - graphical:
|
||||
// 0 1 2 3 4
|
||||
// 5 6 7 8
|
||||
// 9 10 11 = upper right triangle indices
|
||||
// 12 13
|
||||
// 14
|
||||
// 0
|
||||
// 1
|
||||
// 2
|
||||
// 3 = diagonal indices
|
||||
// 4
|
||||
// 5
|
||||
// 0
|
||||
// 1 5
|
||||
// 2 6 9 = lower left triangle indices
|
||||
// 3 7 10 12
|
||||
// 4 8 11 13 14
|
||||
//
|
||||
// Impact on total memory consumption:
|
||||
// - Original: (2 * 1 + 8 * 1/2) 12x12 matrices = 6 12x12 matrices = 864 complex words per site
|
||||
// - Here: (2 * 1 + 4 * 1/2) diagonal parts = 4 diagonal parts = 24 complex words per site
|
||||
// + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts = 60 complex words per site
|
||||
// = 84 complex words per site
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
|
||||
public WilsonCloverHelpers<Impl>,
|
||||
public CompactWilsonCloverHelpers<Impl> {
|
||||
/////////////////////////////////////////////
|
||||
// Sizes
|
||||
/////////////////////////////////////////////
|
||||
|
||||
public:
|
||||
|
||||
INHERIT_COMPACT_CLOVER_SIZES(Impl);
|
||||
|
||||
/////////////////////////////////////////////
|
||||
// Type definitions
|
||||
/////////////////////////////////////////////
|
||||
|
||||
public:
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
INHERIT_CLOVER_TYPES(Impl);
|
||||
INHERIT_COMPACT_CLOVER_TYPES(Impl);
|
||||
|
||||
typedef WilsonFermion<Impl> WilsonBase;
|
||||
typedef WilsonCloverHelpers<Impl> Helpers;
|
||||
typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
|
||||
|
||||
/////////////////////////////////////////////
|
||||
// Constructors
|
||||
/////////////////////////////////////////////
|
||||
|
||||
public:
|
||||
|
||||
CompactWilsonCloverFermion(GaugeField& _Umu,
|
||||
GridCartesian& Fgrid,
|
||||
GridRedBlackCartesian& Hgrid,
|
||||
const RealD _mass,
|
||||
const RealD _csw_r = 0.0,
|
||||
const RealD _csw_t = 0.0,
|
||||
const RealD _cF = 1.0,
|
||||
const WilsonAnisotropyCoefficients& clover_anisotropy = WilsonAnisotropyCoefficients(),
|
||||
const ImplParams& impl_p = ImplParams());
|
||||
|
||||
/////////////////////////////////////////////
|
||||
// Member functions (implementing interface)
|
||||
/////////////////////////////////////////////
|
||||
|
||||
public:
|
||||
|
||||
virtual void Instantiatable() {};
|
||||
int ConstEE() override { return 0; };
|
||||
int isTrivialEE() override { return 0; };
|
||||
|
||||
void Dhop(const FermionField& in, FermionField& out, int dag) override;
|
||||
|
||||
void DhopOE(const FermionField& in, FermionField& out, int dag) override;
|
||||
|
||||
void DhopEO(const FermionField& in, FermionField& out, int dag) override;
|
||||
|
||||
void DhopDir(const FermionField& in, FermionField& out, int dir, int disp) override;
|
||||
|
||||
void DhopDirAll(const FermionField& in, std::vector<FermionField>& out) /* override */;
|
||||
|
||||
void M(const FermionField& in, FermionField& out) override;
|
||||
|
||||
void Mdag(const FermionField& in, FermionField& out) override;
|
||||
|
||||
void Meooe(const FermionField& in, FermionField& out) override;
|
||||
|
||||
void MeooeDag(const FermionField& in, FermionField& out) override;
|
||||
|
||||
void Mooee(const FermionField& in, FermionField& out) override;
|
||||
|
||||
void MooeeDag(const FermionField& in, FermionField& out) override;
|
||||
|
||||
void MooeeInv(const FermionField& in, FermionField& out) override;
|
||||
|
||||
void MooeeInvDag(const FermionField& in, FermionField& out) override;
|
||||
|
||||
void Mdir(const FermionField& in, FermionField& out, int dir, int disp) override;
|
||||
|
||||
void MdirAll(const FermionField& in, std::vector<FermionField>& out) override;
|
||||
|
||||
void MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) override;
|
||||
|
||||
void MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
|
||||
|
||||
void MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) override;
|
||||
|
||||
/////////////////////////////////////////////
|
||||
// Member functions (internals)
|
||||
/////////////////////////////////////////////
|
||||
|
||||
void MooeeInternal(const FermionField& in,
|
||||
FermionField& out,
|
||||
const CloverDiagonalField& diagonal,
|
||||
const CloverTriangleField& triangle);
|
||||
|
||||
/////////////////////////////////////////////
|
||||
// Helpers
|
||||
/////////////////////////////////////////////
|
||||
|
||||
void ImportGauge(const GaugeField& _Umu) override;
|
||||
|
||||
/////////////////////////////////////////////
|
||||
// Helpers
|
||||
/////////////////////////////////////////////
|
||||
|
||||
private:
|
||||
|
||||
template<class Field>
|
||||
const MaskField* getCorrectMaskField(const Field &in) const {
|
||||
if(in.Grid()->_isCheckerBoarded) {
|
||||
if(in.Checkerboard() == Odd) {
|
||||
return &this->BoundaryMaskOdd;
|
||||
} else {
|
||||
return &this->BoundaryMaskEven;
|
||||
}
|
||||
} else {
|
||||
return &this->BoundaryMask;
|
||||
}
|
||||
}
|
||||
|
||||
template<class Field>
|
||||
void ApplyBoundaryMask(Field& f) {
|
||||
const MaskField* m = getCorrectMaskField(f); assert(m != nullptr);
|
||||
assert(m != nullptr);
|
||||
CompactHelpers::ApplyBoundaryMask(f, *m);
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////
|
||||
// Member Data
|
||||
/////////////////////////////////////////////
|
||||
|
||||
public:
|
||||
|
||||
RealD csw_r;
|
||||
RealD csw_t;
|
||||
RealD cF;
|
||||
|
||||
bool open_boundaries;
|
||||
|
||||
CloverDiagonalField Diagonal, DiagonalEven, DiagonalOdd;
|
||||
CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;
|
||||
|
||||
CloverTriangleField Triangle, TriangleEven, TriangleOdd;
|
||||
CloverTriangleField TriangleInv, TriangleInvEven, TriangleInvOdd;
|
||||
|
||||
FermionField Tmp;
|
||||
|
||||
MaskField BoundaryMask, BoundaryMaskEven, BoundaryMaskOdd;
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -53,7 +53,6 @@ NAMESPACE_CHECK(Wilson);
|
||||
#include <Grid/qcd/action/fermion/WilsonTMFermion.h> // 4d wilson like
|
||||
NAMESPACE_CHECK(WilsonTM);
|
||||
#include <Grid/qcd/action/fermion/WilsonCloverFermion.h> // 4d wilson clover fermions
|
||||
#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h> // 4d compact wilson clover fermions
|
||||
NAMESPACE_CHECK(WilsonClover);
|
||||
#include <Grid/qcd/action/fermion/WilsonFermion5D.h> // 5d base used by all 5d overlap types
|
||||
NAMESPACE_CHECK(Wilson5D);
|
||||
@ -138,52 +137,21 @@ typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
|
||||
typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
|
||||
|
||||
// Clover fermions
|
||||
template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
|
||||
template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
|
||||
typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR;
|
||||
typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF;
|
||||
typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD;
|
||||
|
||||
typedef WilsonClover<WilsonImplR> WilsonCloverFermionR;
|
||||
typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
|
||||
typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
|
||||
typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR;
|
||||
typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF;
|
||||
typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD;
|
||||
|
||||
typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
|
||||
typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
|
||||
typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
|
||||
typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
|
||||
typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
|
||||
typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
|
||||
|
||||
typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
|
||||
typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
|
||||
typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
|
||||
|
||||
typedef WilsonClover<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
|
||||
typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
|
||||
typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
|
||||
|
||||
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
|
||||
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
|
||||
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
|
||||
|
||||
// Compact Clover fermions
|
||||
template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
|
||||
template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
|
||||
|
||||
typedef CompactWilsonClover<WilsonImplR> CompactWilsonCloverFermionR;
|
||||
typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
|
||||
typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
|
||||
|
||||
typedef CompactWilsonExpClover<WilsonImplR> CompactWilsonExpCloverFermionR;
|
||||
typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
|
||||
typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
|
||||
|
||||
typedef CompactWilsonClover<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
|
||||
typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
|
||||
typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
|
||||
|
||||
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
|
||||
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
|
||||
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
|
||||
|
||||
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
|
||||
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
|
||||
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
|
||||
typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
|
||||
typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
|
||||
typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
|
||||
|
||||
// Domain Wall fermions
|
||||
typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
|
||||
|
||||
@ -49,8 +49,6 @@ public:
|
||||
|
||||
virtual FermionField &tmp(void) = 0;
|
||||
|
||||
virtual void DirichletBlock(const Coordinate & _Block) { assert(0); };
|
||||
|
||||
GridBase * Grid(void) { return FermionGrid(); }; // this is all the linalg routines need to know
|
||||
GridBase * RedBlackGrid(void) { return FermionRedBlackGrid(); };
|
||||
|
||||
|
||||
@ -30,18 +30,6 @@ directory
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
/*
|
||||
Policy implementation for G-parity boundary conditions
|
||||
|
||||
Rather than treating the gauge field as a flavored field, the Grid implementation of G-parity treats the gauge field as a regular
|
||||
field with complex conjugate boundary conditions. In order to ensure the second flavor interacts with the conjugate links and the first
|
||||
with the regular links we overload the functionality of doubleStore, whose purpose is to store the gauge field and the barrel-shifted gauge field
|
||||
to avoid communicating links when applying the Dirac operator, such that the double-stored field contains also a flavor index which maps to
|
||||
either the link or the conjugate link. This flavored field is then used by multLink to apply the correct link to a spinor.
|
||||
|
||||
Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction.
|
||||
mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
|
||||
*/
|
||||
template <class S, class Representation = FundamentalRepresentation, class Options=CoeffReal>
|
||||
class GparityWilsonImpl : public ConjugateGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > {
|
||||
public:
|
||||
@ -125,7 +113,7 @@ public:
|
||||
|| ((distance== 1)&&(icoor[direction]==1))
|
||||
|| ((distance==-1)&&(icoor[direction]==0));
|
||||
|
||||
permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu] && mmu < Nd-1; //only if we are going around the world in a spatial direction
|
||||
permute_lane = permute_lane && SE->_around_the_world && St.parameters.twists[mmu]; //only if we are going around the world
|
||||
|
||||
//Apply the links
|
||||
int f_upper = permute_lane ? 1 : 0;
|
||||
@ -151,10 +139,10 @@ public:
|
||||
assert((distance == 1) || (distance == -1)); // nearest neighbour stencil hard code
|
||||
assert((sl == 1) || (sl == 2));
|
||||
|
||||
//If this site is an global boundary site, perform the G-parity flavor twist
|
||||
if ( mmu < Nd-1 && SE->_around_the_world && St.parameters.twists[mmu] ) {
|
||||
if ( SE->_around_the_world && St.parameters.twists[mmu] ) {
|
||||
|
||||
if ( sl == 2 ) {
|
||||
//Only do the twist for lanes on the edge of the physical node
|
||||
|
||||
ExtractBuffer<sobj> vals(Nsimd);
|
||||
|
||||
extract(chi,vals);
|
||||
@ -209,19 +197,6 @@ public:
|
||||
reg = memory;
|
||||
}
|
||||
|
||||
|
||||
//Poke 'poke_f0' onto flavor 0 and 'poke_f1' onto flavor 1 in direction mu of the doubled gauge field Uds
|
||||
inline void pokeGparityDoubledGaugeField(DoubledGaugeField &Uds, const GaugeLinkField &poke_f0, const GaugeLinkField &poke_f1, const int mu){
|
||||
autoView(poke_f0_v, poke_f0, CpuRead);
|
||||
autoView(poke_f1_v, poke_f1, CpuRead);
|
||||
autoView(Uds_v, Uds, CpuWrite);
|
||||
thread_foreach(ss,poke_f0_v,{
|
||||
Uds_v[ss](0)(mu) = poke_f0_v[ss]();
|
||||
Uds_v[ss](1)(mu) = poke_f1_v[ss]();
|
||||
});
|
||||
}
|
||||
|
||||
|
||||
inline void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
|
||||
{
|
||||
conformable(Uds.Grid(),GaugeGrid);
|
||||
@ -232,19 +207,14 @@ public:
|
||||
GaugeLinkField Uconj(GaugeGrid);
|
||||
|
||||
Lattice<iScalar<vInteger> > coor(GaugeGrid);
|
||||
|
||||
//Here the first Nd-1 directions are treated as "spatial", and a twist value of 1 indicates G-parity BCs in that direction.
|
||||
//mu=Nd-1 is assumed to be the time direction and a twist value of 1 indicates antiperiodic BCs
|
||||
for(int mu=0;mu<Nd-1;mu++){
|
||||
|
||||
if( Params.twists[mu] ){
|
||||
LatticeCoordinate(coor,mu);
|
||||
}
|
||||
|
||||
for(int mu=0;mu<Nd;mu++){
|
||||
|
||||
LatticeCoordinate(coor,mu);
|
||||
|
||||
U = PeekIndex<LorentzIndex>(Umu,mu);
|
||||
Uconj = conjugate(U);
|
||||
|
||||
// Implement the isospin rotation sign on the boundary between f=1 and f=0
|
||||
// This phase could come from a simple bc 1,1,-1,1 ..
|
||||
int neglink = GaugeGrid->GlobalDimensions()[mu]-1;
|
||||
if ( Params.twists[mu] ) {
|
||||
@ -259,7 +229,7 @@ public:
|
||||
thread_foreach(ss,U_v,{
|
||||
Uds_v[ss](0)(mu) = U_v[ss]();
|
||||
Uds_v[ss](1)(mu) = Uconj_v[ss]();
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
U = adj(Cshift(U ,mu,-1)); // correct except for spanning the boundary
|
||||
@ -290,38 +260,6 @@ public:
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
{ //periodic / antiperiodic temporal BCs
|
||||
int mu = Nd-1;
|
||||
int L = GaugeGrid->GlobalDimensions()[mu];
|
||||
int Lmu = L - 1;
|
||||
|
||||
LatticeCoordinate(coor, mu);
|
||||
|
||||
U = PeekIndex<LorentzIndex>(Umu, mu); //Get t-directed links
|
||||
|
||||
GaugeLinkField *Upoke = &U;
|
||||
|
||||
if(Params.twists[mu]){ //antiperiodic
|
||||
Utmp = where(coor == Lmu, -U, U);
|
||||
Upoke = &Utmp;
|
||||
}
|
||||
|
||||
Uconj = conjugate(*Upoke); //second flavor interacts with conjugate links
|
||||
pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu);
|
||||
|
||||
//Get the barrel-shifted field
|
||||
Utmp = adj(Cshift(U, mu, -1)); //is a forward shift!
|
||||
Upoke = &Utmp;
|
||||
|
||||
if(Params.twists[mu]){
|
||||
U = where(coor == 0, -Utmp, Utmp); //boundary phase
|
||||
Upoke = &U;
|
||||
}
|
||||
|
||||
Uconj = conjugate(*Upoke);
|
||||
pokeGparityDoubledGaugeField(Uds, *Upoke, Uconj, mu + 4);
|
||||
}
|
||||
}
|
||||
|
||||
inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A, int mu) {
|
||||
@ -360,48 +298,28 @@ public:
|
||||
inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
|
||||
assert(0);
|
||||
}
|
||||
|
||||
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField Ã, int mu) {
|
||||
int Ls=Btilde.Grid()->_fdimensions[0];
|
||||
|
||||
{
|
||||
GridBase *GaugeGrid = mat.Grid();
|
||||
Lattice<iScalar<vInteger> > coor(GaugeGrid);
|
||||
|
||||
if( Params.twists[mu] ){
|
||||
LatticeCoordinate(coor,mu);
|
||||
}
|
||||
|
||||
autoView( mat_v , mat, AcceleratorWrite);
|
||||
autoView( Btilde_v , Btilde, AcceleratorRead);
|
||||
autoView( Atilde_v , Atilde, AcceleratorRead);
|
||||
accelerator_for(sss,mat.Grid()->oSites(), FermionField::vector_type::Nsimd(),{
|
||||
int sU=sss;
|
||||
typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
|
||||
ColorMatrixType sum;
|
||||
zeroit(sum);
|
||||
for(int s=0;s<Ls;s++){
|
||||
int sF = s+Ls*sU;
|
||||
for(int spn=0;spn<Ns;spn++){ //sum over spin
|
||||
//Flavor 0
|
||||
auto bb = coalescedRead(Btilde_v[sF](0)(spn) ); //color vector
|
||||
auto aa = coalescedRead(Atilde_v[sF](0)(spn) );
|
||||
sum = sum + outerProduct(bb,aa);
|
||||
|
||||
//Flavor 1
|
||||
bb = coalescedRead(Btilde_v[sF](1)(spn) );
|
||||
aa = coalescedRead(Atilde_v[sF](1)(spn) );
|
||||
sum = sum + conjugate(outerProduct(bb,aa));
|
||||
}
|
||||
}
|
||||
coalescedWrite(mat_v[sU](mu)(), sum);
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField Ã, int mu) {
|
||||
|
||||
int Ls = Btilde.Grid()->_fdimensions[0];
|
||||
|
||||
GaugeLinkField tmp(mat.Grid());
|
||||
tmp = Zero();
|
||||
{
|
||||
autoView( tmp_v , tmp, CpuWrite);
|
||||
autoView( Atilde_v , Atilde, CpuRead);
|
||||
autoView( Btilde_v , Btilde, CpuRead);
|
||||
thread_for(ss,tmp.Grid()->oSites(),{
|
||||
for (int s = 0; s < Ls; s++) {
|
||||
int sF = s + Ls * ss;
|
||||
auto ttmp = traceIndex<SpinIndex>(outerProduct(Btilde_v[sF], Atilde_v[sF]));
|
||||
tmp_v[ss]() = tmp_v[ss]() + ttmp(0, 0) + conjugate(ttmp(1, 1));
|
||||
}
|
||||
});
|
||||
}
|
||||
PokeIndex<LorentzIndex>(mat, tmp, mu);
|
||||
return;
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
|
||||
@ -4,11 +4,10 @@
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.h
|
||||
|
||||
Copyright (C) 2017 - 2022
|
||||
Copyright (C) 2017
|
||||
|
||||
Author: Guido Cossu <guido.cossu@ed.ac.uk>
|
||||
Author: David Preti <>
|
||||
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
|
||||
|
||||
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
|
||||
@ -30,9 +29,7 @@
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
|
||||
#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
|
||||
#include <Grid/qcd/action/fermion/CloverHelpers.h>
|
||||
#include <Grid/Grid.h>
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
@ -52,16 +49,19 @@ NAMESPACE_BEGIN(Grid);
|
||||
// csw_r = csw_t to recover the isotropic version
|
||||
//////////////////////////////////////////////////////////////////
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
class WilsonCloverFermion : public WilsonFermion<Impl>,
|
||||
public WilsonCloverHelpers<Impl>
|
||||
template <class Impl>
|
||||
class WilsonCloverFermion : public WilsonFermion<Impl>
|
||||
{
|
||||
public:
|
||||
// Types definitions
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
INHERIT_CLOVER_TYPES(Impl);
|
||||
template <typename vtype>
|
||||
using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
|
||||
typedef iImplClover<Simd> SiteCloverType;
|
||||
typedef Lattice<SiteCloverType> CloverFieldType;
|
||||
|
||||
typedef WilsonFermion<Impl> WilsonBase;
|
||||
typedef WilsonCloverHelpers<Impl> Helpers;
|
||||
public:
|
||||
typedef WilsonFermion<Impl> WilsonBase;
|
||||
|
||||
virtual int ConstEE(void) { return 0; };
|
||||
virtual void Instantiatable(void){};
|
||||
@ -72,7 +72,42 @@ public:
|
||||
const RealD _csw_r = 0.0,
|
||||
const RealD _csw_t = 0.0,
|
||||
const WilsonAnisotropyCoefficients &clover_anisotropy = WilsonAnisotropyCoefficients(),
|
||||
const ImplParams &impl_p = ImplParams());
|
||||
const ImplParams &impl_p = ImplParams()) : WilsonFermion<Impl>(_Umu,
|
||||
Fgrid,
|
||||
Hgrid,
|
||||
_mass, impl_p, clover_anisotropy),
|
||||
CloverTerm(&Fgrid),
|
||||
CloverTermInv(&Fgrid),
|
||||
CloverTermEven(&Hgrid),
|
||||
CloverTermOdd(&Hgrid),
|
||||
CloverTermInvEven(&Hgrid),
|
||||
CloverTermInvOdd(&Hgrid),
|
||||
CloverTermDagEven(&Hgrid),
|
||||
CloverTermDagOdd(&Hgrid),
|
||||
CloverTermInvDagEven(&Hgrid),
|
||||
CloverTermInvDagOdd(&Hgrid)
|
||||
{
|
||||
assert(Nd == 4); // require 4 dimensions
|
||||
|
||||
if (clover_anisotropy.isAnisotropic)
|
||||
{
|
||||
csw_r = _csw_r * 0.5 / clover_anisotropy.xi_0;
|
||||
diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
|
||||
}
|
||||
else
|
||||
{
|
||||
csw_r = _csw_r * 0.5;
|
||||
diag_mass = 4.0 + _mass;
|
||||
}
|
||||
csw_t = _csw_t * 0.5;
|
||||
|
||||
if (csw_r == 0)
|
||||
std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
|
||||
if (csw_t == 0)
|
||||
std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
|
||||
|
||||
ImportGauge(_Umu);
|
||||
}
|
||||
|
||||
virtual void M(const FermionField &in, FermionField &out);
|
||||
virtual void Mdag(const FermionField &in, FermionField &out);
|
||||
@ -89,21 +124,250 @@ public:
|
||||
void ImportGauge(const GaugeField &_Umu);
|
||||
|
||||
// Derivative parts unpreconditioned pseudofermions
|
||||
void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag);
|
||||
void MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
|
||||
{
|
||||
conformable(X.Grid(), Y.Grid());
|
||||
conformable(X.Grid(), force.Grid());
|
||||
GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
|
||||
GaugeField clover_force(force.Grid());
|
||||
PropagatorField Lambda(force.Grid());
|
||||
|
||||
public:
|
||||
// Guido: Here we are hitting some performance issues:
|
||||
// need to extract the components of the DoubledGaugeField
|
||||
// for each call
|
||||
// Possible solution
|
||||
// Create a vector object to store them? (cons: wasting space)
|
||||
std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
|
||||
|
||||
Impl::extractLinkField(U, this->Umu);
|
||||
|
||||
force = Zero();
|
||||
// Derivative of the Wilson hopping term
|
||||
this->DhopDeriv(force, X, Y, dag);
|
||||
|
||||
///////////////////////////////////////////////////////////
|
||||
// Clover term derivative
|
||||
///////////////////////////////////////////////////////////
|
||||
Impl::outerProductImpl(Lambda, X, Y);
|
||||
//std::cout << "Lambda:" << Lambda << std::endl;
|
||||
|
||||
Gamma::Algebra sigma[] = {
|
||||
Gamma::Algebra::SigmaXY,
|
||||
Gamma::Algebra::SigmaXZ,
|
||||
Gamma::Algebra::SigmaXT,
|
||||
Gamma::Algebra::MinusSigmaXY,
|
||||
Gamma::Algebra::SigmaYZ,
|
||||
Gamma::Algebra::SigmaYT,
|
||||
Gamma::Algebra::MinusSigmaXZ,
|
||||
Gamma::Algebra::MinusSigmaYZ,
|
||||
Gamma::Algebra::SigmaZT,
|
||||
Gamma::Algebra::MinusSigmaXT,
|
||||
Gamma::Algebra::MinusSigmaYT,
|
||||
Gamma::Algebra::MinusSigmaZT};
|
||||
|
||||
/*
|
||||
sigma_{\mu \nu}=
|
||||
| 0 sigma[0] sigma[1] sigma[2] |
|
||||
| sigma[3] 0 sigma[4] sigma[5] |
|
||||
| sigma[6] sigma[7] 0 sigma[8] |
|
||||
| sigma[9] sigma[10] sigma[11] 0 |
|
||||
*/
|
||||
|
||||
int count = 0;
|
||||
clover_force = Zero();
|
||||
for (int mu = 0; mu < 4; mu++)
|
||||
{
|
||||
force_mu = Zero();
|
||||
for (int nu = 0; nu < 4; nu++)
|
||||
{
|
||||
if (mu == nu)
|
||||
continue;
|
||||
|
||||
RealD factor;
|
||||
if (nu == 4 || mu == 4)
|
||||
{
|
||||
factor = 2.0 * csw_t;
|
||||
}
|
||||
else
|
||||
{
|
||||
factor = 2.0 * csw_r;
|
||||
}
|
||||
PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
|
||||
Impl::TraceSpinImpl(lambda, Slambda); // traceSpin ok
|
||||
force_mu -= factor*Cmunu(U, lambda, mu, nu); // checked
|
||||
count++;
|
||||
}
|
||||
|
||||
pokeLorentz(clover_force, U[mu] * force_mu, mu);
|
||||
}
|
||||
//clover_force *= csw;
|
||||
force += clover_force;
|
||||
}
|
||||
|
||||
// Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
|
||||
GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
|
||||
{
|
||||
conformable(lambda.Grid(), U[0].Grid());
|
||||
GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
|
||||
// insertion in upper staple
|
||||
// please check redundancy of shift operations
|
||||
|
||||
// C1+
|
||||
tmp = lambda * U[nu];
|
||||
out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
|
||||
|
||||
// C2+
|
||||
tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
|
||||
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
|
||||
|
||||
// C3+
|
||||
tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
|
||||
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
|
||||
|
||||
// C4+
|
||||
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
|
||||
|
||||
// insertion in lower staple
|
||||
// C1-
|
||||
out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
|
||||
|
||||
// C2-
|
||||
tmp = adj(lambda) * U[nu];
|
||||
out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
|
||||
|
||||
// C3-
|
||||
tmp = lambda * U[nu];
|
||||
out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
|
||||
|
||||
// C4-
|
||||
out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
|
||||
|
||||
return out;
|
||||
}
|
||||
|
||||
protected:
|
||||
// here fixing the 4 dimensions, make it more general?
|
||||
|
||||
RealD csw_r; // Clover coefficient - spatial
|
||||
RealD csw_t; // Clover coefficient - temporal
|
||||
RealD diag_mass; // Mass term
|
||||
CloverField CloverTerm, CloverTermInv; // Clover term
|
||||
CloverField CloverTermEven, CloverTermOdd; // Clover term EO
|
||||
CloverField CloverTermInvEven, CloverTermInvOdd; // Clover term Inv EO
|
||||
CloverField CloverTermDagEven, CloverTermDagOdd; // Clover term Dag EO
|
||||
CloverField CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
|
||||
};
|
||||
CloverFieldType CloverTerm, CloverTermInv; // Clover term
|
||||
CloverFieldType CloverTermEven, CloverTermOdd; // Clover term EO
|
||||
CloverFieldType CloverTermInvEven, CloverTermInvOdd; // Clover term Inv EO
|
||||
CloverFieldType CloverTermDagEven, CloverTermDagOdd; // Clover term Dag EO
|
||||
CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
|
||||
|
||||
public:
|
||||
// eventually these can be compressed into 6x6 blocks instead of the 12x12
|
||||
// using the DeGrand-Rossi basis for the gamma matrices
|
||||
CloverFieldType fillCloverYZ(const GaugeLinkField &F)
|
||||
{
|
||||
CloverFieldType T(F.Grid());
|
||||
T = Zero();
|
||||
autoView(T_v,T,AcceleratorWrite);
|
||||
autoView(F_v,F,AcceleratorRead);
|
||||
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
|
||||
{
|
||||
T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
|
||||
T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
|
||||
T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
|
||||
T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
CloverFieldType fillCloverXZ(const GaugeLinkField &F)
|
||||
{
|
||||
CloverFieldType T(F.Grid());
|
||||
T = Zero();
|
||||
|
||||
autoView(T_v, T,AcceleratorWrite);
|
||||
autoView(F_v, F,AcceleratorRead);
|
||||
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
|
||||
{
|
||||
T_v[i]()(0, 1) = -F_v[i]()();
|
||||
T_v[i]()(1, 0) = F_v[i]()();
|
||||
T_v[i]()(2, 3) = -F_v[i]()();
|
||||
T_v[i]()(3, 2) = F_v[i]()();
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
CloverFieldType fillCloverXY(const GaugeLinkField &F)
|
||||
{
|
||||
CloverFieldType T(F.Grid());
|
||||
T = Zero();
|
||||
|
||||
autoView(T_v,T,AcceleratorWrite);
|
||||
autoView(F_v,F,AcceleratorRead);
|
||||
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
|
||||
{
|
||||
T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
|
||||
T_v[i]()(1, 1) = timesI(F_v[i]()());
|
||||
T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
|
||||
T_v[i]()(3, 3) = timesI(F_v[i]()());
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
CloverFieldType fillCloverXT(const GaugeLinkField &F)
|
||||
{
|
||||
CloverFieldType T(F.Grid());
|
||||
T = Zero();
|
||||
|
||||
autoView( T_v , T, AcceleratorWrite);
|
||||
autoView( F_v , F, AcceleratorRead);
|
||||
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
|
||||
{
|
||||
T_v[i]()(0, 1) = timesI(F_v[i]()());
|
||||
T_v[i]()(1, 0) = timesI(F_v[i]()());
|
||||
T_v[i]()(2, 3) = timesMinusI(F_v[i]()());
|
||||
T_v[i]()(3, 2) = timesMinusI(F_v[i]()());
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
CloverFieldType fillCloverYT(const GaugeLinkField &F)
|
||||
{
|
||||
CloverFieldType T(F.Grid());
|
||||
T = Zero();
|
||||
|
||||
autoView( T_v ,T,AcceleratorWrite);
|
||||
autoView( F_v ,F,AcceleratorRead);
|
||||
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
|
||||
{
|
||||
T_v[i]()(0, 1) = -(F_v[i]()());
|
||||
T_v[i]()(1, 0) = (F_v[i]()());
|
||||
T_v[i]()(2, 3) = (F_v[i]()());
|
||||
T_v[i]()(3, 2) = -(F_v[i]()());
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
CloverFieldType fillCloverZT(const GaugeLinkField &F)
|
||||
{
|
||||
CloverFieldType T(F.Grid());
|
||||
|
||||
T = Zero();
|
||||
|
||||
autoView( T_v , T,AcceleratorWrite);
|
||||
autoView( F_v , F,AcceleratorRead);
|
||||
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
|
||||
{
|
||||
T_v[i]()(0, 0) = timesI(F_v[i]()());
|
||||
T_v[i]()(1, 1) = timesMinusI(F_v[i]()());
|
||||
T_v[i]()(2, 2) = timesMinusI(F_v[i]()());
|
||||
T_v[i]()(3, 3) = timesI(F_v[i]()());
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
};
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
|
||||
|
||||
@ -1,763 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonCloverHelpers.h
|
||||
|
||||
Copyright (C) 2021 - 2022
|
||||
|
||||
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
|
||||
|
||||
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
|
||||
|
||||
// Helper routines that implement common clover functionality
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
template<class Impl> class WilsonCloverHelpers {
|
||||
public:
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
INHERIT_CLOVER_TYPES(Impl);
|
||||
|
||||
// Computing C_{\mu \nu}(x) as in Eq.(B.39) in Zbigniew Sroczynski's PhD thesis
|
||||
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu)
|
||||
{
|
||||
conformable(lambda.Grid(), U[0].Grid());
|
||||
GaugeLinkField out(lambda.Grid()), tmp(lambda.Grid());
|
||||
// insertion in upper staple
|
||||
// please check redundancy of shift operations
|
||||
|
||||
// C1+
|
||||
tmp = lambda * U[nu];
|
||||
out = Impl::ShiftStaple(Impl::CovShiftForward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
|
||||
|
||||
// C2+
|
||||
tmp = U[mu] * Impl::ShiftStaple(adj(lambda), mu);
|
||||
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(tmp, mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu);
|
||||
|
||||
// C3+
|
||||
tmp = U[nu] * Impl::ShiftStaple(adj(lambda), nu);
|
||||
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(tmp, nu))), mu);
|
||||
|
||||
// C4+
|
||||
out += Impl::ShiftStaple(Impl::CovShiftForward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, Impl::CovShiftIdentityBackward(U[nu], nu))), mu) * lambda;
|
||||
|
||||
// insertion in lower staple
|
||||
// C1-
|
||||
out -= Impl::ShiftStaple(lambda, mu) * Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
|
||||
|
||||
// C2-
|
||||
tmp = adj(lambda) * U[nu];
|
||||
out -= Impl::ShiftStaple(Impl::CovShiftBackward(tmp, nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu);
|
||||
|
||||
// C3-
|
||||
tmp = lambda * U[nu];
|
||||
out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, tmp)), mu);
|
||||
|
||||
// C4-
|
||||
out -= Impl::ShiftStaple(Impl::CovShiftBackward(U[nu], nu, Impl::CovShiftBackward(U[mu], mu, U[nu])), mu) * lambda;
|
||||
|
||||
return out;
|
||||
}
|
||||
|
||||
static CloverField fillCloverYZ(const GaugeLinkField &F)
|
||||
{
|
||||
CloverField T(F.Grid());
|
||||
T = Zero();
|
||||
autoView(T_v,T,AcceleratorWrite);
|
||||
autoView(F_v,F,AcceleratorRead);
|
||||
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
|
||||
{
|
||||
coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
static CloverField fillCloverXZ(const GaugeLinkField &F)
|
||||
{
|
||||
CloverField T(F.Grid());
|
||||
T = Zero();
|
||||
|
||||
autoView(T_v, T,AcceleratorWrite);
|
||||
autoView(F_v, F,AcceleratorRead);
|
||||
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
|
||||
{
|
||||
coalescedWrite(T_v[i]()(0, 1), coalescedRead(-F_v[i]()()));
|
||||
coalescedWrite(T_v[i]()(1, 0), coalescedRead(F_v[i]()()));
|
||||
coalescedWrite(T_v[i]()(2, 3), coalescedRead(-F_v[i]()()));
|
||||
coalescedWrite(T_v[i]()(3, 2), coalescedRead(F_v[i]()()));
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
static CloverField fillCloverXY(const GaugeLinkField &F)
|
||||
{
|
||||
CloverField T(F.Grid());
|
||||
T = Zero();
|
||||
|
||||
autoView(T_v,T,AcceleratorWrite);
|
||||
autoView(F_v,F,AcceleratorRead);
|
||||
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
|
||||
{
|
||||
coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
static CloverField fillCloverXT(const GaugeLinkField &F)
|
||||
{
|
||||
CloverField T(F.Grid());
|
||||
T = Zero();
|
||||
|
||||
autoView( T_v , T, AcceleratorWrite);
|
||||
autoView( F_v , F, AcceleratorRead);
|
||||
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
|
||||
{
|
||||
coalescedWrite(T_v[i]()(0, 1), coalescedRead(timesI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(1, 0), coalescedRead(timesI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(2, 3), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(3, 2), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
static CloverField fillCloverYT(const GaugeLinkField &F)
|
||||
{
|
||||
CloverField T(F.Grid());
|
||||
T = Zero();
|
||||
|
||||
autoView( T_v ,T,AcceleratorWrite);
|
||||
autoView( F_v ,F,AcceleratorRead);
|
||||
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
|
||||
{
|
||||
coalescedWrite(T_v[i]()(0, 1), coalescedRead(-(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(1, 0), coalescedRead((F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(2, 3), coalescedRead((F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(3, 2), coalescedRead(-(F_v[i]()())));
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
static CloverField fillCloverZT(const GaugeLinkField &F)
|
||||
{
|
||||
CloverField T(F.Grid());
|
||||
|
||||
T = Zero();
|
||||
|
||||
autoView( T_v , T,AcceleratorWrite);
|
||||
autoView( F_v , F,AcceleratorRead);
|
||||
accelerator_for(i, T.Grid()->oSites(),CloverField::vector_type::Nsimd(),
|
||||
{
|
||||
coalescedWrite(T_v[i]()(0, 0), coalescedRead(timesI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(1, 1), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(2, 2), coalescedRead(timesMinusI(F_v[i]()())));
|
||||
coalescedWrite(T_v[i]()(3, 3), coalescedRead(timesI(F_v[i]()())));
|
||||
});
|
||||
|
||||
return T;
|
||||
}
|
||||
|
||||
template<class _Spinor>
|
||||
static accelerator_inline void multClover(_Spinor& phi, const SiteClover& C, const _Spinor& chi) {
|
||||
auto CC = coalescedRead(C);
|
||||
mult(&phi, &CC, &chi);
|
||||
}
|
||||
|
||||
template<class _SpinorField>
|
||||
inline void multCloverField(_SpinorField& out, const CloverField& C, const _SpinorField& phi) {
|
||||
const int Nsimd = SiteSpinor::Nsimd();
|
||||
autoView(out_v, out, AcceleratorWrite);
|
||||
autoView(phi_v, phi, AcceleratorRead);
|
||||
autoView(C_v, C, AcceleratorRead);
|
||||
typedef decltype(coalescedRead(out_v[0])) calcSpinor;
|
||||
accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
|
||||
calcSpinor tmp;
|
||||
multClover(tmp,C_v[sss],phi_v(sss));
|
||||
coalescedWrite(out_v[sss],tmp);
|
||||
});
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////
|
||||
|
||||
template<class Impl> class CompactWilsonCloverHelpers {
|
||||
public:
|
||||
|
||||
INHERIT_COMPACT_CLOVER_SIZES(Impl);
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
INHERIT_CLOVER_TYPES(Impl);
|
||||
INHERIT_COMPACT_CLOVER_TYPES(Impl);
|
||||
|
||||
#if 0
|
||||
static accelerator_inline typename SiteCloverTriangle::vector_type triangle_elem(const SiteCloverTriangle& triangle, int block, int i, int j) {
|
||||
assert(i != j);
|
||||
if(i < j) {
|
||||
return triangle()(block)(triangle_index(i, j));
|
||||
} else { // i > j
|
||||
return conjugate(triangle()(block)(triangle_index(i, j)));
|
||||
}
|
||||
}
|
||||
#else
|
||||
template<typename vobj>
|
||||
static accelerator_inline vobj triangle_elem(const iImplCloverTriangle<vobj>& triangle, int block, int i, int j) {
|
||||
assert(i != j);
|
||||
if(i < j) {
|
||||
return triangle()(block)(triangle_index(i, j));
|
||||
} else { // i > j
|
||||
return conjugate(triangle()(block)(triangle_index(i, j)));
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
static accelerator_inline int triangle_index(int i, int j) {
|
||||
if(i == j)
|
||||
return 0;
|
||||
else if(i < j)
|
||||
return Nred * (Nred - 1) / 2 - (Nred - i) * (Nred - i - 1) / 2 + j - i - 1;
|
||||
else // i > j
|
||||
return Nred * (Nred - 1) / 2 - (Nred - j) * (Nred - j - 1) / 2 + i - j - 1;
|
||||
}
|
||||
|
||||
static void MooeeKernel_gpu(int Nsite,
|
||||
int Ls,
|
||||
const FermionField& in,
|
||||
FermionField& out,
|
||||
const CloverDiagonalField& diagonal,
|
||||
const CloverTriangleField& triangle) {
|
||||
autoView(diagonal_v, diagonal, AcceleratorRead);
|
||||
autoView(triangle_v, triangle, AcceleratorRead);
|
||||
autoView(in_v, in, AcceleratorRead);
|
||||
autoView(out_v, out, AcceleratorWrite);
|
||||
|
||||
typedef decltype(coalescedRead(out_v[0])) CalcSpinor;
|
||||
|
||||
const uint64_t NN = Nsite * Ls;
|
||||
|
||||
accelerator_for(ss, NN, Simd::Nsimd(), {
|
||||
int sF = ss;
|
||||
int sU = ss/Ls;
|
||||
CalcSpinor res;
|
||||
CalcSpinor in_t = in_v(sF);
|
||||
auto diagonal_t = diagonal_v(sU);
|
||||
auto triangle_t = triangle_v(sU);
|
||||
for(int block=0; block<Nhs; block++) {
|
||||
int s_start = block*Nhs;
|
||||
for(int i=0; i<Nred; i++) {
|
||||
int si = s_start + i/Nc, ci = i%Nc;
|
||||
res()(si)(ci) = diagonal_t()(block)(i) * in_t()(si)(ci);
|
||||
for(int j=0; j<Nred; j++) {
|
||||
if (j == i) continue;
|
||||
int sj = s_start + j/Nc, cj = j%Nc;
|
||||
res()(si)(ci) = res()(si)(ci) + triangle_elem(triangle_t, block, i, j) * in_t()(sj)(cj);
|
||||
};
|
||||
};
|
||||
};
|
||||
coalescedWrite(out_v[sF], res);
|
||||
});
|
||||
}
|
||||
|
||||
static void MooeeKernel_cpu(int Nsite,
|
||||
int Ls,
|
||||
const FermionField& in,
|
||||
FermionField& out,
|
||||
const CloverDiagonalField& diagonal,
|
||||
const CloverTriangleField& triangle) {
|
||||
autoView(diagonal_v, diagonal, CpuRead);
|
||||
autoView(triangle_v, triangle, CpuRead);
|
||||
autoView(in_v, in, CpuRead);
|
||||
autoView(out_v, out, CpuWrite);
|
||||
|
||||
typedef SiteSpinor CalcSpinor;
|
||||
|
||||
#if defined(A64FX) || defined(A64FXFIXEDSIZE)
|
||||
#define PREFETCH_CLOVER(BASE) { \
|
||||
uint64_t base; \
|
||||
int pf_dist_L1 = 1; \
|
||||
int pf_dist_L2 = -5; /* -> penalty -> disable */ \
|
||||
\
|
||||
if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) { \
|
||||
base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 0), SV_PLDL1STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 256), SV_PLDL1STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 512), SV_PLDL1STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 768), SV_PLDL1STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL1STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL1STRM); \
|
||||
} \
|
||||
\
|
||||
if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) { \
|
||||
base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 0), SV_PLDL2STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 256), SV_PLDL2STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 512), SV_PLDL2STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 768), SV_PLDL2STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 1024), SV_PLDL2STRM); \
|
||||
svprfd(svptrue_b64(), (int64_t*)(base + 1280), SV_PLDL2STRM); \
|
||||
} \
|
||||
}
|
||||
// TODO: Implement/generalize this for other architectures
|
||||
// I played around a bit on KNL (see below) but didn't bring anything
|
||||
// #elif defined(AVX512)
|
||||
// #define PREFETCH_CLOVER(BASE) { \
|
||||
// uint64_t base; \
|
||||
// int pf_dist_L1 = 1; \
|
||||
// int pf_dist_L2 = +4; \
|
||||
// \
|
||||
// if ((pf_dist_L1 >= 0) && (sU + pf_dist_L1 < Nsite)) { \
|
||||
// base = (uint64_t)&diag_t()(pf_dist_L1+BASE)(0); \
|
||||
// _mm_prefetch((const char*)(base + 0), _MM_HINT_T0); \
|
||||
// _mm_prefetch((const char*)(base + 64), _MM_HINT_T0); \
|
||||
// _mm_prefetch((const char*)(base + 128), _MM_HINT_T0); \
|
||||
// _mm_prefetch((const char*)(base + 192), _MM_HINT_T0); \
|
||||
// _mm_prefetch((const char*)(base + 256), _MM_HINT_T0); \
|
||||
// _mm_prefetch((const char*)(base + 320), _MM_HINT_T0); \
|
||||
// } \
|
||||
// \
|
||||
// if ((pf_dist_L2 >= 0) && (sU + pf_dist_L2 < Nsite)) { \
|
||||
// base = (uint64_t)&diag_t()(pf_dist_L2+BASE)(0); \
|
||||
// _mm_prefetch((const char*)(base + 0), _MM_HINT_T1); \
|
||||
// _mm_prefetch((const char*)(base + 64), _MM_HINT_T1); \
|
||||
// _mm_prefetch((const char*)(base + 128), _MM_HINT_T1); \
|
||||
// _mm_prefetch((const char*)(base + 192), _MM_HINT_T1); \
|
||||
// _mm_prefetch((const char*)(base + 256), _MM_HINT_T1); \
|
||||
// _mm_prefetch((const char*)(base + 320), _MM_HINT_T1); \
|
||||
// } \
|
||||
// }
|
||||
#else
|
||||
#define PREFETCH_CLOVER(BASE)
|
||||
#endif
|
||||
|
||||
const uint64_t NN = Nsite * Ls;
|
||||
|
||||
thread_for(ss, NN, {
|
||||
int sF = ss;
|
||||
int sU = ss/Ls;
|
||||
CalcSpinor res;
|
||||
CalcSpinor in_t = in_v[sF];
|
||||
auto diag_t = diagonal_v[sU]; // "diag" instead of "diagonal" here to make code below easier to read
|
||||
auto triangle_t = triangle_v[sU];
|
||||
|
||||
// upper half
|
||||
PREFETCH_CLOVER(0);
|
||||
|
||||
auto in_cc_0_0 = conjugate(in_t()(0)(0)); // Nils: reduces number
|
||||
auto in_cc_0_1 = conjugate(in_t()(0)(1)); // of conjugates from
|
||||
auto in_cc_0_2 = conjugate(in_t()(0)(2)); // 30 to 20
|
||||
auto in_cc_1_0 = conjugate(in_t()(1)(0));
|
||||
auto in_cc_1_1 = conjugate(in_t()(1)(1));
|
||||
|
||||
res()(0)(0) = diag_t()(0)( 0) * in_t()(0)(0)
|
||||
+ triangle_t()(0)( 0) * in_t()(0)(1)
|
||||
+ triangle_t()(0)( 1) * in_t()(0)(2)
|
||||
+ triangle_t()(0)( 2) * in_t()(1)(0)
|
||||
+ triangle_t()(0)( 3) * in_t()(1)(1)
|
||||
+ triangle_t()(0)( 4) * in_t()(1)(2);
|
||||
|
||||
res()(0)(1) = triangle_t()(0)( 0) * in_cc_0_0;
|
||||
res()(0)(1) = diag_t()(0)( 1) * in_t()(0)(1)
|
||||
+ triangle_t()(0)( 5) * in_t()(0)(2)
|
||||
+ triangle_t()(0)( 6) * in_t()(1)(0)
|
||||
+ triangle_t()(0)( 7) * in_t()(1)(1)
|
||||
+ triangle_t()(0)( 8) * in_t()(1)(2)
|
||||
+ conjugate( res()(0)( 1));
|
||||
|
||||
res()(0)(2) = triangle_t()(0)( 1) * in_cc_0_0
|
||||
+ triangle_t()(0)( 5) * in_cc_0_1;
|
||||
res()(0)(2) = diag_t()(0)( 2) * in_t()(0)(2)
|
||||
+ triangle_t()(0)( 9) * in_t()(1)(0)
|
||||
+ triangle_t()(0)(10) * in_t()(1)(1)
|
||||
+ triangle_t()(0)(11) * in_t()(1)(2)
|
||||
+ conjugate( res()(0)( 2));
|
||||
|
||||
res()(1)(0) = triangle_t()(0)( 2) * in_cc_0_0
|
||||
+ triangle_t()(0)( 6) * in_cc_0_1
|
||||
+ triangle_t()(0)( 9) * in_cc_0_2;
|
||||
res()(1)(0) = diag_t()(0)( 3) * in_t()(1)(0)
|
||||
+ triangle_t()(0)(12) * in_t()(1)(1)
|
||||
+ triangle_t()(0)(13) * in_t()(1)(2)
|
||||
+ conjugate( res()(1)( 0));
|
||||
|
||||
res()(1)(1) = triangle_t()(0)( 3) * in_cc_0_0
|
||||
+ triangle_t()(0)( 7) * in_cc_0_1
|
||||
+ triangle_t()(0)(10) * in_cc_0_2
|
||||
+ triangle_t()(0)(12) * in_cc_1_0;
|
||||
res()(1)(1) = diag_t()(0)( 4) * in_t()(1)(1)
|
||||
+ triangle_t()(0)(14) * in_t()(1)(2)
|
||||
+ conjugate( res()(1)( 1));
|
||||
|
||||
res()(1)(2) = triangle_t()(0)( 4) * in_cc_0_0
|
||||
+ triangle_t()(0)( 8) * in_cc_0_1
|
||||
+ triangle_t()(0)(11) * in_cc_0_2
|
||||
+ triangle_t()(0)(13) * in_cc_1_0
|
||||
+ triangle_t()(0)(14) * in_cc_1_1;
|
||||
res()(1)(2) = diag_t()(0)( 5) * in_t()(1)(2)
|
||||
+ conjugate( res()(1)( 2));
|
||||
|
||||
vstream(out_v[sF]()(0)(0), res()(0)(0));
|
||||
vstream(out_v[sF]()(0)(1), res()(0)(1));
|
||||
vstream(out_v[sF]()(0)(2), res()(0)(2));
|
||||
vstream(out_v[sF]()(1)(0), res()(1)(0));
|
||||
vstream(out_v[sF]()(1)(1), res()(1)(1));
|
||||
vstream(out_v[sF]()(1)(2), res()(1)(2));
|
||||
|
||||
// lower half
|
||||
PREFETCH_CLOVER(1);
|
||||
|
||||
auto in_cc_2_0 = conjugate(in_t()(2)(0));
|
||||
auto in_cc_2_1 = conjugate(in_t()(2)(1));
|
||||
auto in_cc_2_2 = conjugate(in_t()(2)(2));
|
||||
auto in_cc_3_0 = conjugate(in_t()(3)(0));
|
||||
auto in_cc_3_1 = conjugate(in_t()(3)(1));
|
||||
|
||||
res()(2)(0) = diag_t()(1)( 0) * in_t()(2)(0)
|
||||
+ triangle_t()(1)( 0) * in_t()(2)(1)
|
||||
+ triangle_t()(1)( 1) * in_t()(2)(2)
|
||||
+ triangle_t()(1)( 2) * in_t()(3)(0)
|
||||
+ triangle_t()(1)( 3) * in_t()(3)(1)
|
||||
+ triangle_t()(1)( 4) * in_t()(3)(2);
|
||||
|
||||
res()(2)(1) = triangle_t()(1)( 0) * in_cc_2_0;
|
||||
res()(2)(1) = diag_t()(1)( 1) * in_t()(2)(1)
|
||||
+ triangle_t()(1)( 5) * in_t()(2)(2)
|
||||
+ triangle_t()(1)( 6) * in_t()(3)(0)
|
||||
+ triangle_t()(1)( 7) * in_t()(3)(1)
|
||||
+ triangle_t()(1)( 8) * in_t()(3)(2)
|
||||
+ conjugate( res()(2)( 1));
|
||||
|
||||
res()(2)(2) = triangle_t()(1)( 1) * in_cc_2_0
|
||||
+ triangle_t()(1)( 5) * in_cc_2_1;
|
||||
res()(2)(2) = diag_t()(1)( 2) * in_t()(2)(2)
|
||||
+ triangle_t()(1)( 9) * in_t()(3)(0)
|
||||
+ triangle_t()(1)(10) * in_t()(3)(1)
|
||||
+ triangle_t()(1)(11) * in_t()(3)(2)
|
||||
+ conjugate( res()(2)( 2));
|
||||
|
||||
res()(3)(0) = triangle_t()(1)( 2) * in_cc_2_0
|
||||
+ triangle_t()(1)( 6) * in_cc_2_1
|
||||
+ triangle_t()(1)( 9) * in_cc_2_2;
|
||||
res()(3)(0) = diag_t()(1)( 3) * in_t()(3)(0)
|
||||
+ triangle_t()(1)(12) * in_t()(3)(1)
|
||||
+ triangle_t()(1)(13) * in_t()(3)(2)
|
||||
+ conjugate( res()(3)( 0));
|
||||
|
||||
res()(3)(1) = triangle_t()(1)( 3) * in_cc_2_0
|
||||
+ triangle_t()(1)( 7) * in_cc_2_1
|
||||
+ triangle_t()(1)(10) * in_cc_2_2
|
||||
+ triangle_t()(1)(12) * in_cc_3_0;
|
||||
res()(3)(1) = diag_t()(1)( 4) * in_t()(3)(1)
|
||||
+ triangle_t()(1)(14) * in_t()(3)(2)
|
||||
+ conjugate( res()(3)( 1));
|
||||
|
||||
res()(3)(2) = triangle_t()(1)( 4) * in_cc_2_0
|
||||
+ triangle_t()(1)( 8) * in_cc_2_1
|
||||
+ triangle_t()(1)(11) * in_cc_2_2
|
||||
+ triangle_t()(1)(13) * in_cc_3_0
|
||||
+ triangle_t()(1)(14) * in_cc_3_1;
|
||||
res()(3)(2) = diag_t()(1)( 5) * in_t()(3)(2)
|
||||
+ conjugate( res()(3)( 2));
|
||||
|
||||
vstream(out_v[sF]()(2)(0), res()(2)(0));
|
||||
vstream(out_v[sF]()(2)(1), res()(2)(1));
|
||||
vstream(out_v[sF]()(2)(2), res()(2)(2));
|
||||
vstream(out_v[sF]()(3)(0), res()(3)(0));
|
||||
vstream(out_v[sF]()(3)(1), res()(3)(1));
|
||||
vstream(out_v[sF]()(3)(2), res()(3)(2));
|
||||
});
|
||||
}
|
||||
|
||||
static void MooeeKernel(int Nsite,
|
||||
int Ls,
|
||||
const FermionField& in,
|
||||
FermionField& out,
|
||||
const CloverDiagonalField& diagonal,
|
||||
const CloverTriangleField& triangle) {
|
||||
#if defined(GRID_CUDA) || defined(GRID_HIP)
|
||||
MooeeKernel_gpu(Nsite, Ls, in, out, diagonal, triangle);
|
||||
#else
|
||||
MooeeKernel_cpu(Nsite, Ls, in, out, diagonal, triangle);
|
||||
#endif
|
||||
}
|
||||
|
||||
static void Invert(const CloverDiagonalField& diagonal,
|
||||
const CloverTriangleField& triangle,
|
||||
CloverDiagonalField& diagonalInv,
|
||||
CloverTriangleField& triangleInv) {
|
||||
conformable(diagonal, diagonalInv);
|
||||
conformable(triangle, triangleInv);
|
||||
conformable(diagonal, triangle);
|
||||
|
||||
diagonalInv.Checkerboard() = diagonal.Checkerboard();
|
||||
triangleInv.Checkerboard() = triangle.Checkerboard();
|
||||
|
||||
GridBase* grid = diagonal.Grid();
|
||||
|
||||
long lsites = grid->lSites();
|
||||
|
||||
typedef typename SiteCloverDiagonal::scalar_object scalar_object_diagonal;
|
||||
typedef typename SiteCloverTriangle::scalar_object scalar_object_triangle;
|
||||
|
||||
autoView(diagonal_v, diagonal, CpuRead);
|
||||
autoView(triangle_v, triangle, CpuRead);
|
||||
autoView(diagonalInv_v, diagonalInv, CpuWrite);
|
||||
autoView(triangleInv_v, triangleInv, CpuWrite);
|
||||
|
||||
thread_for(site, lsites, { // NOTE: Not on GPU because of Eigen & (peek/poke)LocalSite
|
||||
Eigen::MatrixXcd clover_inv_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
|
||||
Eigen::MatrixXcd clover_eigen = Eigen::MatrixXcd::Zero(Ns*Nc, Ns*Nc);
|
||||
|
||||
scalar_object_diagonal diagonal_tmp = Zero();
|
||||
scalar_object_diagonal diagonal_inv_tmp = Zero();
|
||||
scalar_object_triangle triangle_tmp = Zero();
|
||||
scalar_object_triangle triangle_inv_tmp = Zero();
|
||||
|
||||
Coordinate lcoor;
|
||||
grid->LocalIndexToLocalCoor(site, lcoor);
|
||||
|
||||
peekLocalSite(diagonal_tmp, diagonal_v, lcoor);
|
||||
peekLocalSite(triangle_tmp, triangle_v, lcoor);
|
||||
|
||||
// TODO: can we save time here by inverting the two 6x6 hermitian matrices separately?
|
||||
for (long s_row=0;s_row<Ns;s_row++) {
|
||||
for (long s_col=0;s_col<Ns;s_col++) {
|
||||
if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
|
||||
int block = s_row / Nhs;
|
||||
int s_row_block = s_row % Nhs;
|
||||
int s_col_block = s_col % Nhs;
|
||||
for (long c_row=0;c_row<Nc;c_row++) {
|
||||
for (long c_col=0;c_col<Nc;c_col++) {
|
||||
int i = s_row_block * Nc + c_row;
|
||||
int j = s_col_block * Nc + c_col;
|
||||
if(i == j)
|
||||
clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(diagonal_tmp()(block)(i)));
|
||||
else
|
||||
clover_eigen(s_row*Nc+c_row, s_col*Nc+c_col) = static_cast<ComplexD>(TensorRemove(triangle_elem(triangle_tmp, block, i, j)));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
clover_inv_eigen = clover_eigen.inverse();
|
||||
|
||||
for (long s_row=0;s_row<Ns;s_row++) {
|
||||
for (long s_col=0;s_col<Ns;s_col++) {
|
||||
if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
|
||||
int block = s_row / Nhs;
|
||||
int s_row_block = s_row % Nhs;
|
||||
int s_col_block = s_col % Nhs;
|
||||
for (long c_row=0;c_row<Nc;c_row++) {
|
||||
for (long c_col=0;c_col<Nc;c_col++) {
|
||||
int i = s_row_block * Nc + c_row;
|
||||
int j = s_col_block * Nc + c_col;
|
||||
if(i == j)
|
||||
diagonal_inv_tmp()(block)(i) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
|
||||
else if(i < j)
|
||||
triangle_inv_tmp()(block)(triangle_index(i, j)) = clover_inv_eigen(s_row*Nc+c_row, s_col*Nc+c_col);
|
||||
else
|
||||
continue;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
pokeLocalSite(diagonal_inv_tmp, diagonalInv_v, lcoor);
|
||||
pokeLocalSite(triangle_inv_tmp, triangleInv_v, lcoor);
|
||||
});
|
||||
}
|
||||
|
||||
static void ConvertLayout(const CloverField& full,
|
||||
CloverDiagonalField& diagonal,
|
||||
CloverTriangleField& triangle) {
|
||||
conformable(full, diagonal);
|
||||
conformable(full, triangle);
|
||||
|
||||
diagonal.Checkerboard() = full.Checkerboard();
|
||||
triangle.Checkerboard() = full.Checkerboard();
|
||||
|
||||
autoView(full_v, full, AcceleratorRead);
|
||||
autoView(diagonal_v, diagonal, AcceleratorWrite);
|
||||
autoView(triangle_v, triangle, AcceleratorWrite);
|
||||
|
||||
// NOTE: this function cannot be 'private' since nvcc forbids this for kernels
|
||||
accelerator_for(ss, full.Grid()->oSites(), 1, {
|
||||
for(int s_row = 0; s_row < Ns; s_row++) {
|
||||
for(int s_col = 0; s_col < Ns; s_col++) {
|
||||
if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
|
||||
int block = s_row / Nhs;
|
||||
int s_row_block = s_row % Nhs;
|
||||
int s_col_block = s_col % Nhs;
|
||||
for(int c_row = 0; c_row < Nc; c_row++) {
|
||||
for(int c_col = 0; c_col < Nc; c_col++) {
|
||||
int i = s_row_block * Nc + c_row;
|
||||
int j = s_col_block * Nc + c_col;
|
||||
if(i == j)
|
||||
diagonal_v[ss]()(block)(i) = full_v[ss]()(s_row, s_col)(c_row, c_col);
|
||||
else if(i < j)
|
||||
triangle_v[ss]()(block)(triangle_index(i, j)) = full_v[ss]()(s_row, s_col)(c_row, c_col);
|
||||
else
|
||||
continue;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
});
|
||||
}
|
||||
|
||||
|
||||
static void ConvertLayout(const CloverDiagonalField& diagonal,
|
||||
const CloverTriangleField& triangle,
|
||||
CloverField& full) {
|
||||
conformable(full, diagonal);
|
||||
conformable(full, triangle);
|
||||
|
||||
full.Checkerboard() = diagonal.Checkerboard();
|
||||
|
||||
full = Zero();
|
||||
|
||||
autoView(diagonal_v, diagonal, AcceleratorRead);
|
||||
autoView(triangle_v, triangle, AcceleratorRead);
|
||||
autoView(full_v, full, AcceleratorWrite);
|
||||
|
||||
// NOTE: this function cannot be 'private' since nvcc forbids this for kernels
|
||||
accelerator_for(ss, full.Grid()->oSites(), 1, {
|
||||
for(int s_row = 0; s_row < Ns; s_row++) {
|
||||
for(int s_col = 0; s_col < Ns; s_col++) {
|
||||
if(abs(s_row - s_col) > 1 || s_row + s_col == 3) continue;
|
||||
int block = s_row / Nhs;
|
||||
int s_row_block = s_row % Nhs;
|
||||
int s_col_block = s_col % Nhs;
|
||||
for(int c_row = 0; c_row < Nc; c_row++) {
|
||||
for(int c_col = 0; c_col < Nc; c_col++) {
|
||||
int i = s_row_block * Nc + c_row;
|
||||
int j = s_col_block * Nc + c_col;
|
||||
if(i == j)
|
||||
full_v[ss]()(s_row, s_col)(c_row, c_col) = diagonal_v[ss]()(block)(i);
|
||||
else
|
||||
full_v[ss]()(s_row, s_col)(c_row, c_col) = triangle_elem(triangle_v[ss], block, i, j);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
});
|
||||
}
|
||||
|
||||
static void ModifyBoundaries(CloverDiagonalField& diagonal, CloverTriangleField& triangle, RealD csw_t, RealD cF, RealD diag_mass) {
|
||||
// Checks/grid
|
||||
double t0 = usecond();
|
||||
conformable(diagonal, triangle);
|
||||
GridBase* grid = diagonal.Grid();
|
||||
|
||||
// Determine the boundary coordinates/sites
|
||||
double t1 = usecond();
|
||||
int t_dir = Nd - 1;
|
||||
Lattice<iScalar<vInteger>> t_coor(grid);
|
||||
LatticeCoordinate(t_coor, t_dir);
|
||||
int T = grid->GlobalDimensions()[t_dir];
|
||||
|
||||
// Set off-diagonal parts at boundary to zero -- OK
|
||||
double t2 = usecond();
|
||||
CloverTriangleField zeroTriangle(grid);
|
||||
zeroTriangle.Checkerboard() = triangle.Checkerboard();
|
||||
zeroTriangle = Zero();
|
||||
triangle = where(t_coor == 0, zeroTriangle, triangle);
|
||||
triangle = where(t_coor == T-1, zeroTriangle, triangle);
|
||||
|
||||
// Set diagonal to unity (scaled correctly) -- OK
|
||||
double t3 = usecond();
|
||||
CloverDiagonalField tmp(grid);
|
||||
tmp.Checkerboard() = diagonal.Checkerboard();
|
||||
tmp = -1.0 * csw_t + diag_mass;
|
||||
diagonal = where(t_coor == 0, tmp, diagonal);
|
||||
diagonal = where(t_coor == T-1, tmp, diagonal);
|
||||
|
||||
// Correct values next to boundary
|
||||
double t4 = usecond();
|
||||
if(cF != 1.0) {
|
||||
tmp = cF - 1.0;
|
||||
tmp += diagonal;
|
||||
diagonal = where(t_coor == 1, tmp, diagonal);
|
||||
diagonal = where(t_coor == T-2, tmp, diagonal);
|
||||
}
|
||||
|
||||
// Report timings
|
||||
double t5 = usecond();
|
||||
#if 0
|
||||
std::cout << GridLogMessage << "CompactWilsonCloverHelpers::ModifyBoundaries timings:"
|
||||
<< " checks = " << (t1 - t0) / 1e6
|
||||
<< ", coordinate = " << (t2 - t1) / 1e6
|
||||
<< ", off-diag zero = " << (t3 - t2) / 1e6
|
||||
<< ", diagonal unity = " << (t4 - t3) / 1e6
|
||||
<< ", near-boundary = " << (t5 - t4) / 1e6
|
||||
<< ", total = " << (t5 - t0) / 1e6
|
||||
<< std::endl;
|
||||
#endif
|
||||
}
|
||||
|
||||
template<class Field, class Mask>
|
||||
static strong_inline void ApplyBoundaryMask(Field& f, const Mask& m) {
|
||||
conformable(f, m);
|
||||
auto grid = f.Grid();
|
||||
const uint32_t Nsite = grid->oSites();
|
||||
const uint32_t Nsimd = grid->Nsimd();
|
||||
autoView(f_v, f, AcceleratorWrite);
|
||||
autoView(m_v, m, AcceleratorRead);
|
||||
// NOTE: this function cannot be 'private' since nvcc forbids this for kernels
|
||||
accelerator_for(ss, Nsite, Nsimd, {
|
||||
coalescedWrite(f_v[ss], m_v(ss) * f_v(ss));
|
||||
});
|
||||
}
|
||||
|
||||
template<class MaskField>
|
||||
static void SetupMasks(MaskField& full, MaskField& even, MaskField& odd) {
|
||||
assert(even.Grid()->_isCheckerBoarded && even.Checkerboard() == Even);
|
||||
assert(odd.Grid()->_isCheckerBoarded && odd.Checkerboard() == Odd);
|
||||
assert(!full.Grid()->_isCheckerBoarded);
|
||||
|
||||
GridBase* grid = full.Grid();
|
||||
int t_dir = Nd-1;
|
||||
Lattice<iScalar<vInteger>> t_coor(grid);
|
||||
LatticeCoordinate(t_coor, t_dir);
|
||||
int T = grid->GlobalDimensions()[t_dir];
|
||||
|
||||
MaskField zeroMask(grid); zeroMask = Zero();
|
||||
full = 1.0;
|
||||
full = where(t_coor == 0, zeroMask, full);
|
||||
full = where(t_coor == T-1, zeroMask, full);
|
||||
|
||||
pickCheckerboard(Even, even, full);
|
||||
pickCheckerboard(Odd, odd, full);
|
||||
}
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1,90 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonCloverTypes.h
|
||||
|
||||
Copyright (C) 2021 - 2022
|
||||
|
||||
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
|
||||
#pragma once
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
template<class Impl>
|
||||
class WilsonCloverTypes {
|
||||
public:
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
|
||||
template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
|
||||
|
||||
typedef iImplClover<Simd> SiteClover;
|
||||
|
||||
typedef Lattice<SiteClover> CloverField;
|
||||
};
|
||||
|
||||
template<class Impl>
|
||||
class CompactWilsonCloverTypes {
|
||||
public:
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
|
||||
static constexpr int Nred = Nc * Nhs; // 6
|
||||
static constexpr int Nblock = Nhs; // 2
|
||||
static constexpr int Ndiagonal = Nred; // 6
|
||||
static constexpr int Ntriangle = (Nred - 1) * Nc; // 15
|
||||
|
||||
template<typename vtype> using iImplCloverDiagonal = iScalar<iVector<iVector<vtype, Ndiagonal>, Nblock>>;
|
||||
template<typename vtype> using iImplCloverTriangle = iScalar<iVector<iVector<vtype, Ntriangle>, Nblock>>;
|
||||
|
||||
typedef iImplCloverDiagonal<Simd> SiteCloverDiagonal;
|
||||
typedef iImplCloverTriangle<Simd> SiteCloverTriangle;
|
||||
typedef iSinglet<Simd> SiteMask;
|
||||
|
||||
typedef Lattice<SiteCloverDiagonal> CloverDiagonalField;
|
||||
typedef Lattice<SiteCloverTriangle> CloverTriangleField;
|
||||
typedef Lattice<SiteMask> MaskField;
|
||||
};
|
||||
|
||||
#define INHERIT_CLOVER_TYPES(Impl) \
|
||||
typedef typename WilsonCloverTypes<Impl>::SiteClover SiteClover; \
|
||||
typedef typename WilsonCloverTypes<Impl>::CloverField CloverField;
|
||||
|
||||
#define INHERIT_COMPACT_CLOVER_TYPES(Impl) \
|
||||
typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverDiagonal SiteCloverDiagonal; \
|
||||
typedef typename CompactWilsonCloverTypes<Impl>::SiteCloverTriangle SiteCloverTriangle; \
|
||||
typedef typename CompactWilsonCloverTypes<Impl>::SiteMask SiteMask; \
|
||||
typedef typename CompactWilsonCloverTypes<Impl>::CloverDiagonalField CloverDiagonalField; \
|
||||
typedef typename CompactWilsonCloverTypes<Impl>::CloverTriangleField CloverTriangleField; \
|
||||
typedef typename CompactWilsonCloverTypes<Impl>::MaskField MaskField; \
|
||||
/* ugly duplication but needed inside functionality classes */ \
|
||||
template<typename vtype> using iImplCloverDiagonal = \
|
||||
iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ndiagonal>, CompactWilsonCloverTypes<Impl>::Nblock>>; \
|
||||
template<typename vtype> using iImplCloverTriangle = \
|
||||
iScalar<iVector<iVector<vtype, CompactWilsonCloverTypes<Impl>::Ntriangle>, CompactWilsonCloverTypes<Impl>::Nblock>>;
|
||||
|
||||
#define INHERIT_COMPACT_CLOVER_SIZES(Impl) \
|
||||
static constexpr int Nred = CompactWilsonCloverTypes<Impl>::Nred; \
|
||||
static constexpr int Nblock = CompactWilsonCloverTypes<Impl>::Nblock; \
|
||||
static constexpr int Ndiagonal = CompactWilsonCloverTypes<Impl>::Ndiagonal; \
|
||||
static constexpr int Ntriangle = CompactWilsonCloverTypes<Impl>::Ntriangle;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -297,7 +297,7 @@ public:
|
||||
void ZeroCountersi(void) { }
|
||||
void Reporti(int calls) { }
|
||||
|
||||
// Vector<int> surface_list;
|
||||
std::vector<int> surface_list;
|
||||
|
||||
WilsonStencil(GridBase *grid,
|
||||
int npoints,
|
||||
@ -307,11 +307,10 @@ public:
|
||||
: CartesianStencil<vobj,cobj,Parameters> (grid,npoints,checkerboard,directions,distances,p)
|
||||
{
|
||||
ZeroCountersi();
|
||||
// surface_list.resize(0);
|
||||
surface_list.resize(0);
|
||||
this->same_node.resize(npoints);
|
||||
};
|
||||
|
||||
/*
|
||||
void BuildSurfaceList(int Ls,int vol4){
|
||||
|
||||
// find same node for SHM
|
||||
@ -332,8 +331,7 @@ public:
|
||||
}
|
||||
}
|
||||
}
|
||||
*/
|
||||
|
||||
|
||||
template < class compressor>
|
||||
void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress)
|
||||
{
|
||||
|
||||
@ -75,10 +75,6 @@ public:
|
||||
FermionField _tmp;
|
||||
FermionField &tmp(void) { return _tmp; }
|
||||
|
||||
int Dirichlet;
|
||||
Coordinate Block;
|
||||
|
||||
/********** Deprecate timers **********/
|
||||
void Report(void);
|
||||
void ZeroCounters(void);
|
||||
double DhopCalls;
|
||||
@ -177,10 +173,7 @@ public:
|
||||
GridCartesian &FourDimGrid,
|
||||
GridRedBlackCartesian &FourDimRedBlackGrid,
|
||||
double _M5,const ImplParams &p= ImplParams());
|
||||
|
||||
virtual void DirichletBlock(const Coordinate & block)
|
||||
{
|
||||
}
|
||||
|
||||
// Constructors
|
||||
/*
|
||||
WilsonFermion5D(int simd,
|
||||
|
||||
@ -47,7 +47,7 @@ CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
|
||||
FiveDimRedBlackGrid,
|
||||
FourDimGrid,
|
||||
FourDimRedBlackGrid,_M5,p),
|
||||
mass_plus(_mass), mass_minus(_mass)
|
||||
mass(_mass)
|
||||
{
|
||||
}
|
||||
|
||||
@ -209,8 +209,8 @@ void CayleyFermion5D<Impl>::M5D (const FermionField &psi, FermionField &chi)
|
||||
{
|
||||
int Ls=this->Ls;
|
||||
Vector<Coeff_t> diag (Ls,1.0);
|
||||
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass_minus;
|
||||
Vector<Coeff_t> lower(Ls,-1.0); lower[0] =mass_plus;
|
||||
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
|
||||
Vector<Coeff_t> lower(Ls,-1.0); lower[0] =mass;
|
||||
M5D(psi,chi,chi,lower,diag,upper);
|
||||
}
|
||||
template<class Impl>
|
||||
@ -220,8 +220,8 @@ void CayleyFermion5D<Impl>::Meooe5D (const FermionField &psi, FermionField &D
|
||||
Vector<Coeff_t> diag = bs;
|
||||
Vector<Coeff_t> upper= cs;
|
||||
Vector<Coeff_t> lower= cs;
|
||||
upper[Ls-1]=-mass_minus*upper[Ls-1];
|
||||
lower[0] =-mass_plus*lower[0];
|
||||
upper[Ls-1]=-mass*upper[Ls-1];
|
||||
lower[0] =-mass*lower[0];
|
||||
M5D(psi,psi,Din,lower,diag,upper);
|
||||
}
|
||||
// FIXME Redunant with the above routine; check this and eliminate
|
||||
@ -235,8 +235,8 @@ template<class Impl> void CayleyFermion5D<Impl>::Meo5D (const FermionField &
|
||||
upper[i]=-ceo[i];
|
||||
lower[i]=-ceo[i];
|
||||
}
|
||||
upper[Ls-1]=-mass_minus*upper[Ls-1];
|
||||
lower[0] =-mass_plus*lower[0];
|
||||
upper[Ls-1]=-mass*upper[Ls-1];
|
||||
lower[0] =-mass*lower[0];
|
||||
M5D(psi,psi,chi,lower,diag,upper);
|
||||
}
|
||||
template<class Impl>
|
||||
@ -250,8 +250,8 @@ void CayleyFermion5D<Impl>::Mooee (const FermionField &psi, FermionField &
|
||||
upper[i]=-cee[i];
|
||||
lower[i]=-cee[i];
|
||||
}
|
||||
upper[Ls-1]=-mass_minus*upper[Ls-1];
|
||||
lower[0] =-mass_plus*lower[0];
|
||||
upper[Ls-1]=-mass*upper[Ls-1];
|
||||
lower[0] =-mass*lower[0];
|
||||
M5D(psi,psi,chi,lower,diag,upper);
|
||||
}
|
||||
template<class Impl>
|
||||
@ -266,9 +266,9 @@ void CayleyFermion5D<Impl>::MooeeDag (const FermionField &psi, FermionField &
|
||||
// Assemble the 5d matrix
|
||||
if ( s==0 ) {
|
||||
upper[s] = -cee[s+1] ;
|
||||
lower[s] = mass_minus*cee[Ls-1];
|
||||
lower[s] = mass*cee[Ls-1];
|
||||
} else if ( s==(Ls-1)) {
|
||||
upper[s] = mass_plus*cee[0];
|
||||
upper[s] = mass*cee[0];
|
||||
lower[s] = -cee[s-1];
|
||||
} else {
|
||||
upper[s]=-cee[s+1];
|
||||
@ -291,8 +291,8 @@ void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
|
||||
Vector<Coeff_t> diag(Ls,1.0);
|
||||
Vector<Coeff_t> upper(Ls,-1.0);
|
||||
Vector<Coeff_t> lower(Ls,-1.0);
|
||||
upper[Ls-1]=-mass_plus*upper[Ls-1];
|
||||
lower[0] =-mass_minus*lower[0];
|
||||
upper[Ls-1]=-mass*upper[Ls-1];
|
||||
lower[0] =-mass*lower[0];
|
||||
M5Ddag(psi,chi,chi,lower,diag,upper);
|
||||
}
|
||||
|
||||
@ -307,9 +307,9 @@ void CayleyFermion5D<Impl>::MeooeDag5D (const FermionField &psi, FermionField
|
||||
for (int s=0;s<Ls;s++){
|
||||
if ( s== 0 ) {
|
||||
upper[s] = cs[s+1];
|
||||
lower[s] =-mass_minus*cs[Ls-1];
|
||||
lower[s] =-mass*cs[Ls-1];
|
||||
} else if ( s==(Ls-1) ) {
|
||||
upper[s] =-mass_plus*cs[0];
|
||||
upper[s] =-mass*cs[0];
|
||||
lower[s] = cs[s-1];
|
||||
} else {
|
||||
upper[s] = cs[s+1];
|
||||
@ -552,7 +552,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
|
||||
|
||||
lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
|
||||
|
||||
leem[i]=mass_minus*cee[Ls-1]/bee[0];
|
||||
leem[i]=mass*cee[Ls-1]/bee[0];
|
||||
for(int j=0;j<i;j++) {
|
||||
assert(bee[j+1]!=Coeff_t(0.0));
|
||||
leem[i]*= aee[j]/bee[j+1];
|
||||
@ -560,7 +560,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
|
||||
|
||||
uee[i] =-aee[i]/bee[i]; // up-diag entry on the ith row
|
||||
|
||||
ueem[i]=mass_plus;
|
||||
ueem[i]=mass;
|
||||
for(int j=1;j<=i;j++) ueem[i]*= cee[j]/bee[j];
|
||||
ueem[i]*= aee[0]/bee[0];
|
||||
|
||||
@ -573,7 +573,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t
|
||||
}
|
||||
|
||||
{
|
||||
Coeff_t delta_d=mass_minus*cee[Ls-1];
|
||||
Coeff_t delta_d=mass*cee[Ls-1];
|
||||
for(int j=0;j<Ls-1;j++) {
|
||||
assert(bee[j] != Coeff_t(0.0));
|
||||
delta_d *= cee[j]/bee[j];
|
||||
@ -642,10 +642,6 @@ void CayleyFermion5D<Impl>::ContractConservedCurrent( PropagatorField &q_in_1,
|
||||
Current curr_type,
|
||||
unsigned int mu)
|
||||
{
|
||||
|
||||
assert(mass_plus == mass_minus);
|
||||
RealD mass = mass_plus;
|
||||
|
||||
#if (!defined(GRID_HIP))
|
||||
Gamma::Algebra Gmu [] = {
|
||||
Gamma::Algebra::GammaX,
|
||||
@ -781,8 +777,6 @@ void CayleyFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
|
||||
assert(mu>=0);
|
||||
assert(mu<Nd);
|
||||
|
||||
assert(mass_plus == mass_minus);
|
||||
RealD mass = mass_plus;
|
||||
|
||||
#if 0
|
||||
int tshift = (mu == Nd-1) ? 1 : 0;
|
||||
|
||||
@ -66,17 +66,18 @@ CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
|
||||
M5Dcalls++;
|
||||
M5Dtime-=usecond();
|
||||
|
||||
uint64_t nloop = grid->oSites();
|
||||
uint64_t nloop = grid->oSites()/Ls;
|
||||
accelerator_for(sss,nloop,Simd::Nsimd(),{
|
||||
uint64_t s = sss%Ls;
|
||||
uint64_t ss= sss-s;
|
||||
uint64_t ss= sss*Ls;
|
||||
typedef decltype(coalescedRead(psi[0])) spinor;
|
||||
spinor tmp1, tmp2;
|
||||
uint64_t idx_u = ss+((s+1)%Ls);
|
||||
uint64_t idx_l = ss+((s+Ls-1)%Ls);
|
||||
spProj5m(tmp1,psi(idx_u));
|
||||
spProj5p(tmp2,psi(idx_l));
|
||||
coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
|
||||
for(int s=0;s<Ls;s++){
|
||||
uint64_t idx_u = ss+((s+1)%Ls);
|
||||
uint64_t idx_l = ss+((s+Ls-1)%Ls);
|
||||
spProj5m(tmp1,psi(idx_u));
|
||||
spProj5p(tmp2,psi(idx_l));
|
||||
coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
|
||||
}
|
||||
});
|
||||
M5Dtime+=usecond();
|
||||
}
|
||||
@ -107,17 +108,18 @@ CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
|
||||
M5Dcalls++;
|
||||
M5Dtime-=usecond();
|
||||
|
||||
uint64_t nloop = grid->oSites();
|
||||
uint64_t nloop = grid->oSites()/Ls;
|
||||
accelerator_for(sss,nloop,Simd::Nsimd(),{
|
||||
uint64_t s = sss%Ls;
|
||||
uint64_t ss= sss-s;
|
||||
uint64_t ss=sss*Ls;
|
||||
typedef decltype(coalescedRead(psi[0])) spinor;
|
||||
spinor tmp1,tmp2;
|
||||
uint64_t idx_u = ss+((s+1)%Ls);
|
||||
uint64_t idx_l = ss+((s+Ls-1)%Ls);
|
||||
spProj5p(tmp1,psi(idx_u));
|
||||
spProj5m(tmp2,psi(idx_l));
|
||||
coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
|
||||
for(int s=0;s<Ls;s++){
|
||||
uint64_t idx_u = ss+((s+1)%Ls);
|
||||
uint64_t idx_l = ss+((s+Ls-1)%Ls);
|
||||
spProj5p(tmp1,psi(idx_u));
|
||||
spProj5m(tmp2,psi(idx_l));
|
||||
coalescedWrite(chi[ss+s],pdiag[s]*phi(ss+s)+pupper[s]*tmp1+plower[s]*tmp2);
|
||||
}
|
||||
});
|
||||
M5Dtime+=usecond();
|
||||
}
|
||||
|
||||
@ -1,373 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/CompactWilsonCloverFermionImplementation.h
|
||||
|
||||
Copyright (C) 2017 - 2022
|
||||
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Guido Cossu <guido.cossu@ed.ac.uk>
|
||||
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
|
||||
#include <Grid/Grid.h>
|
||||
#include <Grid/qcd/spin/Dirac.h>
|
||||
#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
|
||||
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
template<class Impl, class CloverHelpers>
|
||||
CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(GaugeField& _Umu,
|
||||
GridCartesian& Fgrid,
|
||||
GridRedBlackCartesian& Hgrid,
|
||||
const RealD _mass,
|
||||
const RealD _csw_r,
|
||||
const RealD _csw_t,
|
||||
const RealD _cF,
|
||||
const WilsonAnisotropyCoefficients& clover_anisotropy,
|
||||
const ImplParams& impl_p)
|
||||
: WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
|
||||
, csw_r(_csw_r)
|
||||
, csw_t(_csw_t)
|
||||
, cF(_cF)
|
||||
, open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
|
||||
, Diagonal(&Fgrid), Triangle(&Fgrid)
|
||||
, DiagonalEven(&Hgrid), TriangleEven(&Hgrid)
|
||||
, DiagonalOdd(&Hgrid), TriangleOdd(&Hgrid)
|
||||
, DiagonalInv(&Fgrid), TriangleInv(&Fgrid)
|
||||
, DiagonalInvEven(&Hgrid), TriangleInvEven(&Hgrid)
|
||||
, DiagonalInvOdd(&Hgrid), TriangleInvOdd(&Hgrid)
|
||||
, Tmp(&Fgrid)
|
||||
, BoundaryMask(&Fgrid)
|
||||
, BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
|
||||
{
|
||||
assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
|
||||
|
||||
csw_r *= 0.5;
|
||||
csw_t *= 0.5;
|
||||
if (clover_anisotropy.isAnisotropic)
|
||||
csw_r /= clover_anisotropy.xi_0;
|
||||
|
||||
ImportGauge(_Umu);
|
||||
if (open_boundaries) {
|
||||
this->BoundaryMaskEven.Checkerboard() = Even;
|
||||
this->BoundaryMaskOdd.Checkerboard() = Odd;
|
||||
CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
|
||||
}
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
|
||||
WilsonBase::Dhop(in, out, dag);
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
|
||||
WilsonBase::DhopOE(in, out, dag);
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
|
||||
WilsonBase::DhopEO(in, out, dag);
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
|
||||
WilsonBase::DhopDir(in, out, dir, disp);
|
||||
if(this->open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
|
||||
WilsonBase::DhopDirAll(in, out);
|
||||
if(this->open_boundaries) {
|
||||
for(auto& o : out) ApplyBoundaryMask(o);
|
||||
}
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
|
||||
out.Checkerboard() = in.Checkerboard();
|
||||
WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
|
||||
Mooee(in, Tmp);
|
||||
axpy(out, 1.0, out, Tmp);
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& in, FermionField& out) {
|
||||
out.Checkerboard() = in.Checkerboard();
|
||||
WilsonBase::Dhop(in, out, DaggerYes); // call base to save applying bc
|
||||
MooeeDag(in, Tmp);
|
||||
axpy(out, 1.0, out, Tmp);
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
|
||||
WilsonBase::Meooe(in, out);
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
|
||||
WilsonBase::MeooeDag(in, out);
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField& in, FermionField& out) {
|
||||
if(in.Grid()->_isCheckerBoarded) {
|
||||
if(in.Checkerboard() == Odd) {
|
||||
MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
|
||||
} else {
|
||||
MooeeInternal(in, out, DiagonalEven, TriangleEven);
|
||||
}
|
||||
} else {
|
||||
MooeeInternal(in, out, Diagonal, Triangle);
|
||||
}
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField& in, FermionField& out) {
|
||||
Mooee(in, out); // blocks are hermitian
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField& in, FermionField& out) {
|
||||
if(in.Grid()->_isCheckerBoarded) {
|
||||
if(in.Checkerboard() == Odd) {
|
||||
MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
|
||||
} else {
|
||||
MooeeInternal(in, out, DiagonalInvEven, TriangleInvEven);
|
||||
}
|
||||
} else {
|
||||
MooeeInternal(in, out, DiagonalInv, TriangleInv);
|
||||
}
|
||||
if(open_boundaries) ApplyBoundaryMask(out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField& in, FermionField& out) {
|
||||
MooeeInv(in, out); // blocks are hermitian
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
|
||||
DhopDir(in, out, dir, disp);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
|
||||
DhopDirAll(in, out);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
|
||||
assert(!open_boundaries); // TODO check for changes required for open bc
|
||||
|
||||
// NOTE: code copied from original clover term
|
||||
conformable(X.Grid(), Y.Grid());
|
||||
conformable(X.Grid(), force.Grid());
|
||||
GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
|
||||
GaugeField clover_force(force.Grid());
|
||||
PropagatorField Lambda(force.Grid());
|
||||
|
||||
// Guido: Here we are hitting some performance issues:
|
||||
// need to extract the components of the DoubledGaugeField
|
||||
// for each call
|
||||
// Possible solution
|
||||
// Create a vector object to store them? (cons: wasting space)
|
||||
std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
|
||||
|
||||
Impl::extractLinkField(U, this->Umu);
|
||||
|
||||
force = Zero();
|
||||
// Derivative of the Wilson hopping term
|
||||
this->DhopDeriv(force, X, Y, dag);
|
||||
|
||||
///////////////////////////////////////////////////////////
|
||||
// Clover term derivative
|
||||
///////////////////////////////////////////////////////////
|
||||
Impl::outerProductImpl(Lambda, X, Y);
|
||||
//std::cout << "Lambda:" << Lambda << std::endl;
|
||||
|
||||
Gamma::Algebra sigma[] = {
|
||||
Gamma::Algebra::SigmaXY,
|
||||
Gamma::Algebra::SigmaXZ,
|
||||
Gamma::Algebra::SigmaXT,
|
||||
Gamma::Algebra::MinusSigmaXY,
|
||||
Gamma::Algebra::SigmaYZ,
|
||||
Gamma::Algebra::SigmaYT,
|
||||
Gamma::Algebra::MinusSigmaXZ,
|
||||
Gamma::Algebra::MinusSigmaYZ,
|
||||
Gamma::Algebra::SigmaZT,
|
||||
Gamma::Algebra::MinusSigmaXT,
|
||||
Gamma::Algebra::MinusSigmaYT,
|
||||
Gamma::Algebra::MinusSigmaZT};
|
||||
|
||||
/*
|
||||
sigma_{\mu \nu}=
|
||||
| 0 sigma[0] sigma[1] sigma[2] |
|
||||
| sigma[3] 0 sigma[4] sigma[5] |
|
||||
| sigma[6] sigma[7] 0 sigma[8] |
|
||||
| sigma[9] sigma[10] sigma[11] 0 |
|
||||
*/
|
||||
|
||||
int count = 0;
|
||||
clover_force = Zero();
|
||||
for (int mu = 0; mu < 4; mu++)
|
||||
{
|
||||
force_mu = Zero();
|
||||
for (int nu = 0; nu < 4; nu++)
|
||||
{
|
||||
if (mu == nu)
|
||||
continue;
|
||||
|
||||
RealD factor;
|
||||
if (nu == 4 || mu == 4)
|
||||
{
|
||||
factor = 2.0 * csw_t;
|
||||
}
|
||||
else
|
||||
{
|
||||
factor = 2.0 * csw_r;
|
||||
}
|
||||
PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
|
||||
Impl::TraceSpinImpl(lambda, Slambda); // traceSpin ok
|
||||
force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu); // checked
|
||||
count++;
|
||||
}
|
||||
|
||||
pokeLorentz(clover_force, U[mu] * force_mu, mu);
|
||||
}
|
||||
//clover_force *= csw;
|
||||
force += clover_force;
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
|
||||
assert(0);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
|
||||
assert(0);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField& in,
|
||||
FermionField& out,
|
||||
const CloverDiagonalField& diagonal,
|
||||
const CloverTriangleField& triangle) {
|
||||
assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
|
||||
out.Checkerboard() = in.Checkerboard();
|
||||
conformable(in, out);
|
||||
conformable(in, diagonal);
|
||||
conformable(in, triangle);
|
||||
|
||||
CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
|
||||
// NOTE: parts copied from original implementation
|
||||
|
||||
// Import gauge into base class
|
||||
double t0 = usecond();
|
||||
WilsonBase::ImportGauge(_Umu); // NOTE: called here and in wilson constructor -> performed twice, but can't avoid that
|
||||
|
||||
// Initialize temporary variables
|
||||
double t1 = usecond();
|
||||
conformable(_Umu.Grid(), this->GaugeGrid());
|
||||
GridBase* grid = _Umu.Grid();
|
||||
typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
|
||||
CloverField TmpOriginal(grid);
|
||||
|
||||
// Compute the field strength terms mu>nu
|
||||
double t2 = usecond();
|
||||
WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
|
||||
WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
|
||||
WilsonLoops<Impl>::FieldStrength(Bz, _Umu, Ydir, Xdir);
|
||||
WilsonLoops<Impl>::FieldStrength(Ex, _Umu, Tdir, Xdir);
|
||||
WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
|
||||
WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
|
||||
|
||||
// Compute the Clover Operator acting on Colour and Spin
|
||||
// multiply here by the clover coefficients for the anisotropy
|
||||
double t3 = usecond();
|
||||
TmpOriginal = Helpers::fillCloverYZ(Bx) * csw_r;
|
||||
TmpOriginal += Helpers::fillCloverXZ(By) * csw_r;
|
||||
TmpOriginal += Helpers::fillCloverXY(Bz) * csw_r;
|
||||
TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
|
||||
TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
|
||||
TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
|
||||
// Handle mass term based on clover policy
|
||||
CloverHelpers::MassTerm(TmpOriginal, this->diag_mass);
|
||||
|
||||
// Convert the data layout of the clover term
|
||||
double t4 = usecond();
|
||||
CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
|
||||
|
||||
// Exponentiate the clover (nothing happens in case of the standard clover)
|
||||
double t5 = usecond();
|
||||
CloverHelpers::Exponentiate_Clover(Diagonal, Triangle, csw_t, this->diag_mass);
|
||||
|
||||
// Possible modify the boundary values
|
||||
double t6 = usecond();
|
||||
if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
|
||||
|
||||
// Invert the Clover term (explicit inversion needed for the improvement in case of open boundary conditions)
|
||||
double t7 = usecond();
|
||||
CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
|
||||
|
||||
// Fill the remaining clover fields
|
||||
double t8 = usecond();
|
||||
pickCheckerboard(Even, DiagonalEven, Diagonal);
|
||||
pickCheckerboard(Even, TriangleEven, Triangle);
|
||||
pickCheckerboard(Odd, DiagonalOdd, Diagonal);
|
||||
pickCheckerboard(Odd, TriangleOdd, Triangle);
|
||||
pickCheckerboard(Even, DiagonalInvEven, DiagonalInv);
|
||||
pickCheckerboard(Even, TriangleInvEven, TriangleInv);
|
||||
pickCheckerboard(Odd, DiagonalInvOdd, DiagonalInv);
|
||||
pickCheckerboard(Odd, TriangleInvOdd, TriangleInv);
|
||||
|
||||
// Report timings
|
||||
double t9 = usecond();
|
||||
|
||||
std::cout << GridLogDebug << "CompactWilsonCloverFermion::ImportGauge timings:" << std::endl;
|
||||
std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "allocations = " << (t2 - t1) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "field strength = " << (t3 - t2) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "fill clover = " << (t4 - t3) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "convert = " << (t5 - t4) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "exponentiation = " << (t6 - t5) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "boundaries = " << (t7 - t6) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "inversions = " << (t8 - t7) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "pick cbs = " << (t9 - t8) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "total = " << (t9 - t0) / 1e6 << std::endl;
|
||||
}
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -2,13 +2,12 @@
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
|
||||
Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.cc
|
||||
|
||||
Copyright (C) 2017 - 2022
|
||||
Copyright (C) 2017
|
||||
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Guido Cossu <guido.cossu@ed.ac.uk>
|
||||
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
|
||||
|
||||
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
|
||||
@ -34,48 +33,9 @@
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
WilsonCloverFermion<Impl, CloverHelpers>::WilsonCloverFermion(GaugeField& _Umu,
|
||||
GridCartesian& Fgrid,
|
||||
GridRedBlackCartesian& Hgrid,
|
||||
const RealD _mass,
|
||||
const RealD _csw_r,
|
||||
const RealD _csw_t,
|
||||
const WilsonAnisotropyCoefficients& clover_anisotropy,
|
||||
const ImplParams& impl_p)
|
||||
: WilsonFermion<Impl>(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
|
||||
, CloverTerm(&Fgrid)
|
||||
, CloverTermInv(&Fgrid)
|
||||
, CloverTermEven(&Hgrid)
|
||||
, CloverTermOdd(&Hgrid)
|
||||
, CloverTermInvEven(&Hgrid)
|
||||
, CloverTermInvOdd(&Hgrid)
|
||||
, CloverTermDagEven(&Hgrid)
|
||||
, CloverTermDagOdd(&Hgrid)
|
||||
, CloverTermInvDagEven(&Hgrid)
|
||||
, CloverTermInvDagOdd(&Hgrid) {
|
||||
assert(Nd == 4); // require 4 dimensions
|
||||
|
||||
if(clover_anisotropy.isAnisotropic) {
|
||||
csw_r = _csw_r * 0.5 / clover_anisotropy.xi_0;
|
||||
diag_mass = _mass + 1.0 + (Nd - 1) * (clover_anisotropy.nu / clover_anisotropy.xi_0);
|
||||
} else {
|
||||
csw_r = _csw_r * 0.5;
|
||||
diag_mass = 4.0 + _mass;
|
||||
}
|
||||
csw_t = _csw_t * 0.5;
|
||||
|
||||
if(csw_r == 0)
|
||||
std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_r = 0" << std::endl;
|
||||
if(csw_t == 0)
|
||||
std::cout << GridLogWarning << "Initializing WilsonCloverFermion with csw_t = 0" << std::endl;
|
||||
|
||||
ImportGauge(_Umu);
|
||||
}
|
||||
|
||||
// *NOT* EO
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, FermionField &out)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
|
||||
{
|
||||
FermionField temp(out.Grid());
|
||||
|
||||
@ -89,8 +49,8 @@ void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, Fermion
|
||||
out += temp;
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, FermionField &out)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
|
||||
{
|
||||
FermionField temp(out.Grid());
|
||||
|
||||
@ -104,16 +64,13 @@ void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, Ferm
|
||||
out += temp;
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Umu)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
|
||||
{
|
||||
double t0 = usecond();
|
||||
WilsonFermion<Impl>::ImportGauge(_Umu);
|
||||
double t1 = usecond();
|
||||
GridBase *grid = _Umu.Grid();
|
||||
typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
|
||||
|
||||
double t2 = usecond();
|
||||
// Compute the field strength terms mu>nu
|
||||
WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
|
||||
WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
|
||||
@ -122,20 +79,52 @@ void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Um
|
||||
WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
|
||||
WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
|
||||
|
||||
double t3 = usecond();
|
||||
// Compute the Clover Operator acting on Colour and Spin
|
||||
// multiply here by the clover coefficients for the anisotropy
|
||||
CloverTerm = Helpers::fillCloverYZ(Bx) * csw_r;
|
||||
CloverTerm += Helpers::fillCloverXZ(By) * csw_r;
|
||||
CloverTerm += Helpers::fillCloverXY(Bz) * csw_r;
|
||||
CloverTerm += Helpers::fillCloverXT(Ex) * csw_t;
|
||||
CloverTerm += Helpers::fillCloverYT(Ey) * csw_t;
|
||||
CloverTerm += Helpers::fillCloverZT(Ez) * csw_t;
|
||||
|
||||
double t4 = usecond();
|
||||
CloverHelpers::Instantiate(CloverTerm, CloverTermInv, csw_t, this->diag_mass);
|
||||
CloverTerm = fillCloverYZ(Bx) * csw_r;
|
||||
CloverTerm += fillCloverXZ(By) * csw_r;
|
||||
CloverTerm += fillCloverXY(Bz) * csw_r;
|
||||
CloverTerm += fillCloverXT(Ex) * csw_t;
|
||||
CloverTerm += fillCloverYT(Ey) * csw_t;
|
||||
CloverTerm += fillCloverZT(Ez) * csw_t;
|
||||
CloverTerm += diag_mass;
|
||||
|
||||
int lvol = _Umu.Grid()->lSites();
|
||||
int DimRep = Impl::Dimension;
|
||||
|
||||
{
|
||||
autoView(CTv,CloverTerm,CpuRead);
|
||||
autoView(CTIv,CloverTermInv,CpuWrite);
|
||||
thread_for(site, lvol, {
|
||||
Coordinate lcoor;
|
||||
grid->LocalIndexToLocalCoor(site, lcoor);
|
||||
Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
|
||||
Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
|
||||
typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
|
||||
peekLocalSite(Qx, CTv, lcoor);
|
||||
//if (csw!=0){
|
||||
for (int j = 0; j < Ns; j++)
|
||||
for (int k = 0; k < Ns; k++)
|
||||
for (int a = 0; a < DimRep; a++)
|
||||
for (int b = 0; b < DimRep; b++){
|
||||
auto zz = Qx()(j, k)(a, b);
|
||||
EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
|
||||
}
|
||||
// if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
|
||||
|
||||
EigenInvCloverOp = EigenCloverOp.inverse();
|
||||
//std::cout << EigenInvCloverOp << std::endl;
|
||||
for (int j = 0; j < Ns; j++)
|
||||
for (int k = 0; k < Ns; k++)
|
||||
for (int a = 0; a < DimRep; a++)
|
||||
for (int b = 0; b < DimRep; b++)
|
||||
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
|
||||
// if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
|
||||
// }
|
||||
pokeLocalSite(Qxinv, CTIv, lcoor);
|
||||
});
|
||||
}
|
||||
|
||||
double t5 = usecond();
|
||||
// Separate the even and odd parts
|
||||
pickCheckerboard(Even, CloverTermEven, CloverTerm);
|
||||
pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
|
||||
@ -148,47 +137,37 @@ void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Um
|
||||
|
||||
pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
|
||||
pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
|
||||
double t6 = usecond();
|
||||
|
||||
std::cout << GridLogDebug << "WilsonCloverFermion::ImportGauge timings:" << std::endl;
|
||||
std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "allocations = " << (t2 - t1) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "field strength = " << (t3 - t2) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "fill clover = " << (t4 - t3) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "instantiation = " << (t5 - t4) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "pick cbs = " << (t6 - t5) / 1e6 << std::endl;
|
||||
std::cout << GridLogDebug << "total = " << (t6 - t0) / 1e6 << std::endl;
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField &in, FermionField &out)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
|
||||
{
|
||||
this->MooeeInternal(in, out, DaggerNo, InverseNo);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField &in, FermionField &out)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
|
||||
{
|
||||
this->MooeeInternal(in, out, DaggerYes, InverseNo);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField &in, FermionField &out)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
|
||||
{
|
||||
this->MooeeInternal(in, out, DaggerNo, InverseYes);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField &in, FermionField &out)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
|
||||
{
|
||||
this->MooeeInternal(in, out, DaggerYes, InverseYes);
|
||||
}
|
||||
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
|
||||
{
|
||||
out.Checkerboard() = in.Checkerboard();
|
||||
CloverField *Clover;
|
||||
CloverFieldType *Clover;
|
||||
assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
|
||||
|
||||
if (dag)
|
||||
@ -203,12 +182,12 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField
|
||||
{
|
||||
Clover = (inv) ? &CloverTermInvDagEven : &CloverTermDagEven;
|
||||
}
|
||||
Helpers::multCloverField(out, *Clover, in);
|
||||
out = *Clover * in;
|
||||
}
|
||||
else
|
||||
{
|
||||
Clover = (inv) ? &CloverTermInv : &CloverTerm;
|
||||
Helpers::multCloverField(out, *Clover, in); // don't bother with adj, hermitian anyway
|
||||
out = adj(*Clover) * in;
|
||||
}
|
||||
}
|
||||
else
|
||||
@ -226,109 +205,29 @@ void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField
|
||||
// std::cout << "Calling clover term Even" << std::endl;
|
||||
Clover = (inv) ? &CloverTermInvEven : &CloverTermEven;
|
||||
}
|
||||
Helpers::multCloverField(out, *Clover, in);
|
||||
out = *Clover * in;
|
||||
// std::cout << GridLogMessage << "*Clover.Checkerboard() " << (*Clover).Checkerboard() << std::endl;
|
||||
}
|
||||
else
|
||||
{
|
||||
Clover = (inv) ? &CloverTermInv : &CloverTerm;
|
||||
Helpers::multCloverField(out, *Clover, in);
|
||||
out = *Clover * in;
|
||||
}
|
||||
}
|
||||
|
||||
} // MooeeInternal
|
||||
|
||||
// Derivative parts unpreconditioned pseudofermions
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
|
||||
{
|
||||
conformable(X.Grid(), Y.Grid());
|
||||
conformable(X.Grid(), force.Grid());
|
||||
GaugeLinkField force_mu(force.Grid()), lambda(force.Grid());
|
||||
GaugeField clover_force(force.Grid());
|
||||
PropagatorField Lambda(force.Grid());
|
||||
|
||||
// Guido: Here we are hitting some performance issues:
|
||||
// need to extract the components of the DoubledGaugeField
|
||||
// for each call
|
||||
// Possible solution
|
||||
// Create a vector object to store them? (cons: wasting space)
|
||||
std::vector<GaugeLinkField> U(Nd, this->Umu.Grid());
|
||||
|
||||
Impl::extractLinkField(U, this->Umu);
|
||||
|
||||
force = Zero();
|
||||
// Derivative of the Wilson hopping term
|
||||
this->DhopDeriv(force, X, Y, dag);
|
||||
|
||||
///////////////////////////////////////////////////////////
|
||||
// Clover term derivative
|
||||
///////////////////////////////////////////////////////////
|
||||
Impl::outerProductImpl(Lambda, X, Y);
|
||||
//std::cout << "Lambda:" << Lambda << std::endl;
|
||||
|
||||
Gamma::Algebra sigma[] = {
|
||||
Gamma::Algebra::SigmaXY,
|
||||
Gamma::Algebra::SigmaXZ,
|
||||
Gamma::Algebra::SigmaXT,
|
||||
Gamma::Algebra::MinusSigmaXY,
|
||||
Gamma::Algebra::SigmaYZ,
|
||||
Gamma::Algebra::SigmaYT,
|
||||
Gamma::Algebra::MinusSigmaXZ,
|
||||
Gamma::Algebra::MinusSigmaYZ,
|
||||
Gamma::Algebra::SigmaZT,
|
||||
Gamma::Algebra::MinusSigmaXT,
|
||||
Gamma::Algebra::MinusSigmaYT,
|
||||
Gamma::Algebra::MinusSigmaZT};
|
||||
|
||||
/*
|
||||
sigma_{\mu \nu}=
|
||||
| 0 sigma[0] sigma[1] sigma[2] |
|
||||
| sigma[3] 0 sigma[4] sigma[5] |
|
||||
| sigma[6] sigma[7] 0 sigma[8] |
|
||||
| sigma[9] sigma[10] sigma[11] 0 |
|
||||
*/
|
||||
|
||||
int count = 0;
|
||||
clover_force = Zero();
|
||||
for (int mu = 0; mu < 4; mu++)
|
||||
{
|
||||
force_mu = Zero();
|
||||
for (int nu = 0; nu < 4; nu++)
|
||||
{
|
||||
if (mu == nu)
|
||||
continue;
|
||||
|
||||
RealD factor;
|
||||
if (nu == 4 || mu == 4)
|
||||
{
|
||||
factor = 2.0 * csw_t;
|
||||
}
|
||||
else
|
||||
{
|
||||
factor = 2.0 * csw_r;
|
||||
}
|
||||
PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
|
||||
Impl::TraceSpinImpl(lambda, Slambda); // traceSpin ok
|
||||
force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu); // checked
|
||||
count++;
|
||||
}
|
||||
|
||||
pokeLorentz(clover_force, U[mu] * force_mu, mu);
|
||||
}
|
||||
//clover_force *= csw;
|
||||
force += clover_force;
|
||||
}
|
||||
|
||||
// Derivative parts
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
|
||||
{
|
||||
assert(0);
|
||||
}
|
||||
|
||||
// Derivative parts
|
||||
template<class Impl, class CloverHelpers>
|
||||
void WilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
|
||||
template <class Impl>
|
||||
void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
|
||||
{
|
||||
assert(0); // not implemented yet
|
||||
}
|
||||
|
||||
@ -60,8 +60,7 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
|
||||
UmuOdd (_FourDimRedBlackGrid),
|
||||
Lebesgue(_FourDimGrid),
|
||||
LebesgueEvenOdd(_FourDimRedBlackGrid),
|
||||
_tmp(&FiveDimRedBlackGrid),
|
||||
Dirichlet(0)
|
||||
_tmp(&FiveDimRedBlackGrid)
|
||||
{
|
||||
// some assertions
|
||||
assert(FiveDimGrid._ndimension==5);
|
||||
@ -92,19 +91,6 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
|
||||
assert(FourDimRedBlackGrid._simd_layout[d] ==FourDimGrid._simd_layout[d]);
|
||||
}
|
||||
|
||||
if ( p.dirichlet.size() == Nd+1) {
|
||||
Coordinate block = p.dirichlet;
|
||||
if ( block[0] || block[1] || block[2] || block[3] || block[4] ){
|
||||
Dirichlet = 1;
|
||||
Block = block;
|
||||
}
|
||||
} else {
|
||||
Coordinate block(Nd+1,0);
|
||||
Block = block;
|
||||
}
|
||||
|
||||
ZeroCounters();
|
||||
|
||||
if (Impl::LsVectorised) {
|
||||
|
||||
int nsimd = Simd::Nsimd();
|
||||
@ -232,14 +218,6 @@ void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
|
||||
{
|
||||
GaugeField HUmu(_Umu.Grid());
|
||||
HUmu = _Umu*(-0.5);
|
||||
if ( Dirichlet ) {
|
||||
std::cout << GridLogMessage << " Dirichlet BCs 5d " <<Block<<std::endl;
|
||||
Coordinate GaugeBlock(Nd);
|
||||
for(int d=0;d<Nd;d++) GaugeBlock[d] = Block[d+1];
|
||||
std::cout << GridLogMessage << " Dirichlet BCs 4d " <<GaugeBlock<<std::endl;
|
||||
DirichletFilter<GaugeField> Filter(GaugeBlock);
|
||||
Filter.applyFilter(HUmu);
|
||||
}
|
||||
Impl::DoubleStore(GaugeGrid(),Umu,HUmu);
|
||||
pickCheckerboard(Even,UmuEven,Umu);
|
||||
pickCheckerboard(Odd ,UmuOdd,Umu);
|
||||
|
||||
@ -4,13 +4,12 @@ Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonFermion.cc
|
||||
|
||||
Copyright (C) 2022
|
||||
Copyright (C) 2015
|
||||
|
||||
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Fabian Joswig <fabian.joswig@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
|
||||
@ -600,47 +599,11 @@ void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
|
||||
Current curr_type,
|
||||
unsigned int mu)
|
||||
{
|
||||
if(curr_type != Current::Vector)
|
||||
{
|
||||
std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
|
||||
exit(1);
|
||||
}
|
||||
|
||||
Gamma g5(Gamma::Algebra::Gamma5);
|
||||
conformable(_grid, q_in_1.Grid());
|
||||
conformable(_grid, q_in_2.Grid());
|
||||
conformable(_grid, q_out.Grid());
|
||||
auto UGrid= this->GaugeGrid();
|
||||
|
||||
PropagatorField tmp_shifted(UGrid);
|
||||
PropagatorField g5Lg5(UGrid);
|
||||
PropagatorField R(UGrid);
|
||||
PropagatorField gmuR(UGrid);
|
||||
|
||||
Gamma::Algebra Gmu [] = {
|
||||
Gamma::Algebra::GammaX,
|
||||
Gamma::Algebra::GammaY,
|
||||
Gamma::Algebra::GammaZ,
|
||||
Gamma::Algebra::GammaT,
|
||||
};
|
||||
Gamma gmu=Gamma(Gmu[mu]);
|
||||
|
||||
g5Lg5=g5*q_in_1*g5;
|
||||
tmp_shifted=Cshift(q_in_2,mu,1);
|
||||
Impl::multLinkField(R,this->Umu,tmp_shifted,mu);
|
||||
gmuR=gmu*R;
|
||||
|
||||
q_out=adj(g5Lg5)*R;
|
||||
q_out-=adj(g5Lg5)*gmuR;
|
||||
|
||||
tmp_shifted=Cshift(q_in_1,mu,1);
|
||||
Impl::multLinkField(g5Lg5,this->Umu,tmp_shifted,mu);
|
||||
g5Lg5=g5*g5Lg5*g5;
|
||||
R=q_in_2;
|
||||
gmuR=gmu*R;
|
||||
|
||||
q_out-=adj(g5Lg5)*R;
|
||||
q_out-=adj(g5Lg5)*gmuR;
|
||||
assert(0);
|
||||
}
|
||||
|
||||
|
||||
@ -654,51 +617,9 @@ void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
|
||||
unsigned int tmax,
|
||||
ComplexField &lattice_cmplx)
|
||||
{
|
||||
if(curr_type != Current::Vector)
|
||||
{
|
||||
std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
|
||||
exit(1);
|
||||
}
|
||||
|
||||
int tshift = (mu == Nd-1) ? 1 : 0;
|
||||
unsigned int LLt = GridDefaultLatt()[Tp];
|
||||
conformable(_grid, q_in.Grid());
|
||||
conformable(_grid, q_out.Grid());
|
||||
auto UGrid= this->GaugeGrid();
|
||||
|
||||
PropagatorField tmp(UGrid);
|
||||
PropagatorField Utmp(UGrid);
|
||||
PropagatorField L(UGrid);
|
||||
PropagatorField zz (UGrid);
|
||||
zz=Zero();
|
||||
LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
|
||||
|
||||
Gamma::Algebra Gmu [] = {
|
||||
Gamma::Algebra::GammaX,
|
||||
Gamma::Algebra::GammaY,
|
||||
Gamma::Algebra::GammaZ,
|
||||
Gamma::Algebra::GammaT,
|
||||
};
|
||||
Gamma gmu=Gamma(Gmu[mu]);
|
||||
|
||||
tmp = Cshift(q_in,mu,1);
|
||||
Impl::multLinkField(Utmp,this->Umu,tmp,mu);
|
||||
tmp = ( Utmp*lattice_cmplx - gmu*Utmp*lattice_cmplx ); // Forward hop
|
||||
tmp = where((lcoor>=tmin),tmp,zz); // Mask the time
|
||||
q_out = where((lcoor<=tmax),tmp,zz); // Position of current complicated
|
||||
|
||||
tmp = q_in *lattice_cmplx;
|
||||
tmp = Cshift(tmp,mu,-1);
|
||||
Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
|
||||
tmp = -( Utmp + gmu*Utmp );
|
||||
// Mask the time
|
||||
if (tmax == LLt - 1 && tshift == 1){ // quick fix to include timeslice 0 if tmax + tshift is over the last timeslice
|
||||
unsigned int t0 = 0;
|
||||
tmp = where(((lcoor==t0) || (lcoor>=tmin+tshift)),tmp,zz);
|
||||
} else {
|
||||
tmp = where((lcoor>=tmin+tshift),tmp,zz);
|
||||
}
|
||||
q_out+= where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
|
||||
assert(0);
|
||||
}
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -440,17 +440,6 @@ void WilsonKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeField &U,S
|
||||
|
||||
#define KERNEL_CALL(A) KERNEL_CALLNB(A); accelerator_barrier();
|
||||
|
||||
#define KERNEL_CALL_EXT(A) \
|
||||
const uint64_t NN = Nsite*Ls; \
|
||||
const uint64_t sz = st.surface_list.size(); \
|
||||
auto ptr = &st.surface_list[0]; \
|
||||
accelerator_forNB( ss, sz, Simd::Nsimd(), { \
|
||||
int sF = ptr[ss]; \
|
||||
int sU = ss/Ls; \
|
||||
WilsonKernels<Impl>::A(st_v,U_v,buf,sF,sU,in_v,out_v); \
|
||||
}); \
|
||||
accelerator_barrier();
|
||||
|
||||
#define ASM_CALL(A) \
|
||||
thread_for( ss, Nsite, { \
|
||||
int sU = ss; \
|
||||
|
||||
@ -1,44 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/ qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master
|
||||
|
||||
Copyright (C) 2017 - 2022
|
||||
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Guido Cossu <guido.cossu@ed.ac.uk>
|
||||
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
|
||||
Author: Mattia Bruno <mattia.bruno@cern.ch>
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
|
||||
#include <Grid/Grid.h>
|
||||
#include <Grid/qcd/spin/Dirac.h>
|
||||
#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/CloverHelpers.h>
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactCloverHelpers<IMPLEMENTATION>>;
|
||||
template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactExpCloverHelpers<IMPLEMENTATION>>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -8,8 +8,7 @@
|
||||
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Guido Cossu <guido.cossu@ed.ac.uk>
|
||||
Author: Mattia Bruno <mattia.bruno@cern.ch>
|
||||
|
||||
|
||||
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
|
||||
@ -32,12 +31,10 @@
|
||||
#include <Grid/qcd/spin/Dirac.h>
|
||||
#include <Grid/qcd/action/fermion/WilsonCloverFermion.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/CloverHelpers.h>
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonCloverFermion<IMPLEMENTATION, CloverHelpers<IMPLEMENTATION>>;
|
||||
template class WilsonCloverFermion<IMPLEMENTATION, ExpCloverHelpers<IMPLEMENTATION>>;
|
||||
template class WilsonCloverFermion<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -1 +0,0 @@
|
||||
../CompactWilsonCloverFermionInstantiation.cc.master
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../CompactWilsonCloverFermionInstantiation.cc.master
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -1 +0,0 @@
|
||||
../WilsonKernelsInstantiation.cc.master
|
||||
@ -0,0 +1,51 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
|
||||
|
||||
Copyright (C) 2015, 2020
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution
|
||||
directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid/qcd/action/fermion/FermionCore.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
|
||||
|
||||
#ifndef AVX512
|
||||
#ifndef QPX
|
||||
#ifndef A64FX
|
||||
#ifndef A64FXFIXEDSIZE
|
||||
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
#include "impl.h"
|
||||
template class WilsonKernels<IMPLEMENTATION>;
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
@ -18,10 +18,6 @@ WILSON_IMPL_LIST=" \
|
||||
GparityWilsonImplF \
|
||||
GparityWilsonImplD "
|
||||
|
||||
COMPACT_WILSON_IMPL_LIST=" \
|
||||
WilsonImplF \
|
||||
WilsonImplD "
|
||||
|
||||
DWF_IMPL_LIST=" \
|
||||
WilsonImplF \
|
||||
WilsonImplD \
|
||||
@ -50,17 +46,7 @@ for impl in $WILSON_IMPL_LIST
|
||||
do
|
||||
for f in $CC_LIST
|
||||
do
|
||||
ln -f -s ../$f.cc.master $impl/$f$impl.cc
|
||||
done
|
||||
done
|
||||
|
||||
CC_LIST="CompactWilsonCloverFermionInstantiation"
|
||||
|
||||
for impl in $COMPACT_WILSON_IMPL_LIST
|
||||
do
|
||||
for f in $CC_LIST
|
||||
do
|
||||
ln -f -s ../$f.cc.master $impl/$f$impl.cc
|
||||
ln -f -s ../$f.cc.master $impl/$f$impl.cc
|
||||
done
|
||||
done
|
||||
|
||||
@ -77,14 +63,14 @@ for impl in $DWF_IMPL_LIST $GDWF_IMPL_LIST
|
||||
do
|
||||
for f in $CC_LIST
|
||||
do
|
||||
ln -f -s ../$f.cc.master $impl/$f$impl.cc
|
||||
ln -f -s ../$f.cc.master $impl/$f$impl.cc
|
||||
done
|
||||
done
|
||||
|
||||
# overwrite the .cc file in Gparity directories
|
||||
for impl in $GDWF_IMPL_LIST
|
||||
do
|
||||
ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc
|
||||
ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc
|
||||
done
|
||||
|
||||
|
||||
@ -98,7 +84,7 @@ for impl in $STAG_IMPL_LIST
|
||||
do
|
||||
for f in $CC_LIST
|
||||
do
|
||||
ln -f -s ../$f.cc.master $impl/$f$impl.cc
|
||||
ln -f -s ../$f.cc.master $impl/$f$impl.cc
|
||||
done
|
||||
done
|
||||
|
||||
|
||||
@ -1,102 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/hmc/integrators/DirichletFilter.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 */
|
||||
//--------------------------------------------------------------------
|
||||
#pragma once
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
////////////////////////////////////////////////////
|
||||
// DDHMC filter with sub-block size B[mu]
|
||||
////////////////////////////////////////////////////
|
||||
|
||||
template<typename GaugeField>
|
||||
struct DDHMCFilter: public MomentumFilterBase<GaugeField>
|
||||
{
|
||||
Coordinate Block;
|
||||
int Width;
|
||||
|
||||
DDHMCFilter(const Coordinate &_Block,int _Width=2): Block(_Block) { Width=_Width; }
|
||||
|
||||
void applyFilter(GaugeField &U) const override
|
||||
{
|
||||
GridBase *grid = U.Grid();
|
||||
Coordinate Global=grid->GlobalDimensions();
|
||||
GaugeField zzz(grid); zzz = Zero();
|
||||
LatticeInteger coor(grid);
|
||||
|
||||
auto zzz_mu = PeekIndex<LorentzIndex>(zzz,0);
|
||||
////////////////////////////////////////////////////
|
||||
// Zero BDY layers
|
||||
////////////////////////////////////////////////////
|
||||
std::cout<<GridLogMessage<<" DDHMC Force Filter Block "<<Block<<" width " <<Width<<std::endl;
|
||||
for(int mu=0;mu<Nd;mu++) {
|
||||
|
||||
Integer B1 = Block[mu];
|
||||
if ( B1 && (B1 <= Global[mu]) ) {
|
||||
LatticeCoordinate(coor,mu);
|
||||
|
||||
////////////////////////////////
|
||||
// OmegaBar - zero all links contained in slice B-1,0 and
|
||||
// mu links connecting to Omega
|
||||
////////////////////////////////
|
||||
if ( Width==1) {
|
||||
U = where(mod(coor,B1)==Integer(B1-1),zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(0) ,zzz,U);
|
||||
auto U_mu = PeekIndex<LorentzIndex>(U,mu);
|
||||
U_mu = where(mod(coor,B1)==Integer(B1-2),zzz_mu,U_mu);
|
||||
PokeIndex<LorentzIndex>(U, U_mu, mu);
|
||||
}
|
||||
if ( Width==2) {
|
||||
U = where(mod(coor,B1)==Integer(B1-2),zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(B1-1),zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(0) ,zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(1) ,zzz,U);
|
||||
auto U_mu = PeekIndex<LorentzIndex>(U,mu);
|
||||
U_mu = where(mod(coor,B1)==Integer(B1-3),zzz_mu,U_mu);
|
||||
PokeIndex<LorentzIndex>(U, U_mu, mu);
|
||||
}
|
||||
if ( Width==3) {
|
||||
U = where(mod(coor,B1)==Integer(B1-3),zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(B1-2),zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(B1-1),zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(0) ,zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(1) ,zzz,U);
|
||||
U = where(mod(coor,B1)==Integer(2) ,zzz,U);
|
||||
auto U_mu = PeekIndex<LorentzIndex>(U,mu);
|
||||
U_mu = where(mod(coor,B1)==Integer(B1-4),zzz_mu,U_mu);
|
||||
PokeIndex<LorentzIndex>(U, U_mu, mu);
|
||||
}
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
}
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -1,71 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/hmc/integrators/DirichletFilter.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 */
|
||||
//--------------------------------------------------------------------
|
||||
#pragma once
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
template<typename MomentaField>
|
||||
struct DirichletFilter: public MomentumFilterBase<MomentaField>
|
||||
{
|
||||
typedef typename MomentaField::vector_type vector_type; //SIMD-vectorized complex type
|
||||
typedef typename MomentaField::scalar_type scalar_type; //scalar complex type
|
||||
|
||||
typedef iScalar<iScalar<iScalar<vector_type> > > ScalarType; //complex phase for each site
|
||||
|
||||
Coordinate Block;
|
||||
|
||||
DirichletFilter(const Coordinate &_Block): Block(_Block){}
|
||||
|
||||
void applyFilter(MomentaField &P) const override
|
||||
{
|
||||
GridBase *grid = P.Grid();
|
||||
typedef decltype(PeekIndex<LorentzIndex>(P, 0)) LatCM;
|
||||
////////////////////////////////////////////////////
|
||||
// Zero strictly links crossing between domains
|
||||
////////////////////////////////////////////////////
|
||||
LatticeInteger coor(grid);
|
||||
LatCM zz(grid); zz = Zero();
|
||||
for(int mu=0;mu<Nd;mu++) {
|
||||
if ( (Block[mu]) && (Block[mu] <= grid->GlobalDimensions()[mu] ) ) {
|
||||
// If costly could provide Grid earlier and precompute masks
|
||||
std::cout << GridLogMessage << " Dirichlet in mu="<<mu<<std::endl;
|
||||
LatticeCoordinate(coor,mu);
|
||||
auto P_mu = PeekIndex<LorentzIndex>(P, mu);
|
||||
P_mu = where(mod(coor,Block[mu])==Integer(Block[mu]-1),zz,P_mu);
|
||||
PokeIndex<LorentzIndex>(P, P_mu, mu);
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -69,11 +69,6 @@ public:
|
||||
return PeriodicBC::ShiftStaple(Link,mu);
|
||||
}
|
||||
|
||||
//Same as Cshift for periodic BCs
|
||||
static inline GaugeLinkField CshiftLink(const GaugeLinkField &Link, int mu, int shift){
|
||||
return PeriodicBC::CshiftLink(Link,mu,shift);
|
||||
}
|
||||
|
||||
static inline bool isPeriodicGaugeField(void) { return true; }
|
||||
};
|
||||
|
||||
@ -115,11 +110,6 @@ public:
|
||||
return PeriodicBC::CovShiftBackward(Link, mu, field);
|
||||
}
|
||||
|
||||
//If mu is a conjugate BC direction
|
||||
//Out(x) = U^dag_\mu(x-mu) | x_\mu != 0
|
||||
// = U^T_\mu(L-1) | x_\mu == 0
|
||||
//else
|
||||
//Out(x) = U^dag_\mu(x-mu mod L)
|
||||
static inline GaugeLinkField
|
||||
CovShiftIdentityBackward(const GaugeLinkField &Link, int mu)
|
||||
{
|
||||
@ -139,13 +129,6 @@ public:
|
||||
return PeriodicBC::CovShiftIdentityForward(Link,mu);
|
||||
}
|
||||
|
||||
|
||||
//If mu is a conjugate BC direction
|
||||
//Out(x) = S_\mu(x+mu) | x_\mu != L-1
|
||||
// = S*_\mu(x+mu) | x_\mu == L-1
|
||||
//else
|
||||
//Out(x) = S_\mu(x+mu mod L)
|
||||
//Note: While this is used for Staples it is also applicable for shifting gauge links or gauge transformation matrices
|
||||
static inline GaugeLinkField ShiftStaple(const GaugeLinkField &Link, int mu)
|
||||
{
|
||||
assert(_conjDirs.size() == Nd);
|
||||
@ -155,27 +138,6 @@ public:
|
||||
return PeriodicBC::ShiftStaple(Link,mu);
|
||||
}
|
||||
|
||||
//Boundary-aware C-shift of gauge links / gauge transformation matrices
|
||||
//For conjugate BC direction
|
||||
//shift = 1
|
||||
//Out(x) = U_\mu(x+\hat\mu) | x_\mu != L-1
|
||||
// = U*_\mu(0) | x_\mu == L-1
|
||||
//shift = -1
|
||||
//Out(x) = U_\mu(x-mu) | x_\mu != 0
|
||||
// = U*_\mu(L-1) | x_\mu == 0
|
||||
//else
|
||||
//shift = 1
|
||||
//Out(x) = U_\mu(x+\hat\mu mod L)
|
||||
//shift = -1
|
||||
//Out(x) = U_\mu(x-\hat\mu mod L)
|
||||
static inline GaugeLinkField CshiftLink(const GaugeLinkField &Link, int mu, int shift){
|
||||
assert(_conjDirs.size() == Nd);
|
||||
if(_conjDirs[mu])
|
||||
return ConjugateBC::CshiftLink(Link,mu,shift);
|
||||
else
|
||||
return PeriodicBC::CshiftLink(Link,mu,shift);
|
||||
}
|
||||
|
||||
static inline void setDirections(std::vector<int> &conjDirs) { _conjDirs=conjDirs; }
|
||||
static inline std::vector<int> getDirections(void) { return _conjDirs; }
|
||||
static inline bool isPeriodicGaugeField(void) { return false; }
|
||||
|
||||
@ -13,31 +13,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
std::cout << GridLogMessage << "Pseudofermion action lamda_max "<<lambda_max<<"( bound "<<hi<<")"<<std::endl;
|
||||
assert( (lambda_max < hi) && " High Bounds Check on operator failed" );
|
||||
}
|
||||
|
||||
template<class Field> void ChebyBoundsCheck(LinearOperatorBase<Field> &HermOp,
|
||||
Field &GaussNoise,
|
||||
RealD lo,RealD hi)
|
||||
{
|
||||
int orderfilter = 1000;
|
||||
Chebyshev<Field> Cheb(lo,hi,orderfilter);
|
||||
|
||||
GridBase *FermionGrid = GaussNoise.Grid();
|
||||
|
||||
Field X(FermionGrid);
|
||||
Field Z(FermionGrid);
|
||||
|
||||
X=GaussNoise;
|
||||
RealD Nx = norm2(X);
|
||||
Cheb(HermOp,X,Z);
|
||||
RealD Nz = norm2(Z);
|
||||
|
||||
std::cout << "************************* "<<std::endl;
|
||||
std::cout << " noise = "<<Nx<<std::endl;
|
||||
std::cout << " Cheb x noise = "<<Nz<<std::endl;
|
||||
std::cout << " Ratio = "<<Nz/Nx<<std::endl;
|
||||
std::cout << "************************* "<<std::endl;
|
||||
assert( ((Nz/Nx)<1.0) && " ChebyBoundsCheck ");
|
||||
}
|
||||
|
||||
template<class Field> void InverseSqrtBoundsCheck(int MaxIter,double tol,
|
||||
LinearOperatorBase<Field> &HermOp,
|
||||
@ -65,65 +40,13 @@ NAMESPACE_BEGIN(Grid);
|
||||
X=X-Y;
|
||||
RealD Nd = norm2(X);
|
||||
std::cout << "************************* "<<std::endl;
|
||||
std::cout << " | noise |^2 = "<<Nx<<std::endl;
|
||||
std::cout << " | (MdagM^-1/2)^2 noise |^2 = "<<Nz<<std::endl;
|
||||
std::cout << " | MdagM (MdagM^-1/2)^2 noise |^2 = "<<Ny<<std::endl;
|
||||
std::cout << " | noise - MdagM (MdagM^-1/2)^2 noise |^2 = "<<Nd<<std::endl;
|
||||
std::cout << " | noise - MdagM (MdagM^-1/2)^2 noise|/|noise| = " << std::sqrt(Nd/Nx) << std::endl;
|
||||
std::cout << " noise = "<<Nx<<std::endl;
|
||||
std::cout << " (MdagM^-1/2)^2 noise = "<<Nz<<std::endl;
|
||||
std::cout << " MdagM (MdagM^-1/2)^2 noise = "<<Ny<<std::endl;
|
||||
std::cout << " noise - MdagM (MdagM^-1/2)^2 noise = "<<Nd<<std::endl;
|
||||
std::cout << "************************* "<<std::endl;
|
||||
assert( (std::sqrt(Nd/Nx)<tol) && " InverseSqrtBoundsCheck ");
|
||||
}
|
||||
|
||||
/* For a HermOp = M^dag M, check the approximation of HermOp^{-1/inv_pow}
|
||||
by computing |X - HermOp * [ Hermop^{-1/inv_pow} ]^{inv_pow} X| < tol
|
||||
for noise X (aka GaussNoise).
|
||||
ApproxNegPow should be the rational approximation for X^{-1/inv_pow}
|
||||
*/
|
||||
template<class Field> void InversePowerBoundsCheck(int inv_pow,
|
||||
int MaxIter,double tol,
|
||||
LinearOperatorBase<Field> &HermOp,
|
||||
Field &GaussNoise,
|
||||
MultiShiftFunction &ApproxNegPow)
|
||||
{
|
||||
GridBase *FermionGrid = GaussNoise.Grid();
|
||||
|
||||
Field X(FermionGrid);
|
||||
Field Y(FermionGrid);
|
||||
Field Z(FermionGrid);
|
||||
|
||||
Field tmp1(FermionGrid), tmp2(FermionGrid);
|
||||
|
||||
X=GaussNoise;
|
||||
RealD Nx = norm2(X);
|
||||
|
||||
ConjugateGradientMultiShift<Field> msCG(MaxIter,ApproxNegPow);
|
||||
|
||||
tmp1 = X;
|
||||
|
||||
Field* in = &tmp1;
|
||||
Field* out = &tmp2;
|
||||
for(int i=0;i<inv_pow;i++){ //apply [ Hermop^{-1/inv_pow} ]^{inv_pow} X = HermOp^{-1} X
|
||||
msCG(HermOp, *in, *out); //backwards conventions!
|
||||
if(i!=inv_pow-1) std::swap(in, out);
|
||||
}
|
||||
Z = *out;
|
||||
|
||||
RealD Nz = norm2(Z);
|
||||
|
||||
HermOp.HermOp(Z,Y);
|
||||
RealD Ny = norm2(Y);
|
||||
|
||||
X=X-Y;
|
||||
RealD Nd = norm2(X);
|
||||
std::cout << "************************* "<<std::endl;
|
||||
std::cout << " | noise |^2 = "<<Nx<<std::endl;
|
||||
std::cout << " | (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2 = "<<Nz<<std::endl;
|
||||
std::cout << " | MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2 = "<<Ny<<std::endl;
|
||||
std::cout << " | noise - MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |^2 = "<<Nd<<std::endl;
|
||||
std::cout << " | noise - MdagM (MdagM^-1/" << inv_pow << ")^" << inv_pow << " noise |/| noise | = "<<std::sqrt(Nd/Nx)<<std::endl;
|
||||
std::cout << "************************* "<<std::endl;
|
||||
assert( (std::sqrt(Nd/Nx)<tol) && " InversePowerBoundsCheck ");
|
||||
}
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
|
||||
@ -1,163 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundaryBoson.h
|
||||
|
||||
Copyright (C) 2021
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#pragma once
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
///////////////////////////////////////
|
||||
// Two flavour ratio
|
||||
///////////////////////////////////////
|
||||
template<class ImplD,class ImplF>
|
||||
class DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion : public Action<typename ImplD::GaugeField> {
|
||||
public:
|
||||
INHERIT_IMPL_TYPES(ImplD);
|
||||
|
||||
private:
|
||||
SchurFactoredFermionOperator<ImplD,ImplF> & NumOp;// the basic operator
|
||||
RealD InnerStoppingCondition;
|
||||
RealD ActionStoppingCondition;
|
||||
RealD DerivativeStoppingCondition;
|
||||
FermionField Phi; // the pseudo fermion field for this trajectory
|
||||
public:
|
||||
DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF> &_NumOp,RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol=1.0e-6)
|
||||
: NumOp(_NumOp),
|
||||
DerivativeStoppingCondition(_DerivativeTol),
|
||||
ActionStoppingCondition(_ActionTol),
|
||||
InnerStoppingCondition(_InnerTol),
|
||||
Phi(_NumOp.FermionGrid()) {};
|
||||
|
||||
virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourBosonPseudoFermion";}
|
||||
|
||||
virtual std::string LogParameters(){
|
||||
std::stringstream sstream;
|
||||
return sstream.str();
|
||||
}
|
||||
|
||||
virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
|
||||
{
|
||||
// P(phi) = e^{- phi^dag P^dag P phi}
|
||||
//
|
||||
// NumOp == P
|
||||
//
|
||||
// Take phi = P^{-1} eta ; eta = P Phi
|
||||
//
|
||||
// P(eta) = e^{- eta^dag eta}
|
||||
//
|
||||
// e^{x^2/2 sig^2} => sig^2 = 0.5.
|
||||
//
|
||||
// So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
|
||||
//
|
||||
RealD scale = std::sqrt(0.5);
|
||||
|
||||
NumOp.tolinner=InnerStoppingCondition;
|
||||
NumOp.tol=ActionStoppingCondition;
|
||||
NumOp.ImportGauge(U);
|
||||
|
||||
FermionField eta(NumOp.FermionGrid());
|
||||
|
||||
gaussian(pRNG,eta); eta=eta*scale;
|
||||
|
||||
NumOp.ProjectBoundaryBar(eta);
|
||||
//DumpSliceNorm("eta",eta);
|
||||
NumOp.RInv(eta,Phi);
|
||||
|
||||
//DumpSliceNorm("Phi",Phi);
|
||||
|
||||
};
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
// S = phi^dag Pdag P phi
|
||||
//////////////////////////////////////////////////////
|
||||
virtual RealD S(const GaugeField &U) {
|
||||
|
||||
NumOp.tolinner=InnerStoppingCondition;
|
||||
NumOp.tol=ActionStoppingCondition;
|
||||
NumOp.ImportGauge(U);
|
||||
|
||||
FermionField Y(NumOp.FermionGrid());
|
||||
|
||||
NumOp.R(Phi,Y);
|
||||
|
||||
RealD action = norm2(Y);
|
||||
|
||||
return action;
|
||||
};
|
||||
|
||||
virtual void deriv(const GaugeField &U,GaugeField & dSdU)
|
||||
{
|
||||
NumOp.tolinner=InnerStoppingCondition;
|
||||
NumOp.tol=DerivativeStoppingCondition;
|
||||
NumOp.ImportGauge(U);
|
||||
|
||||
GridBase *fgrid = NumOp.FermionGrid();
|
||||
GridBase *ugrid = NumOp.GaugeGrid();
|
||||
|
||||
FermionField X(fgrid);
|
||||
FermionField Y(fgrid);
|
||||
FermionField tmp(fgrid);
|
||||
|
||||
GaugeField force(ugrid);
|
||||
|
||||
FermionField DobiDdbPhi(fgrid); // Vector A in my notes
|
||||
FermionField DoiDdDobiDdbPhi(fgrid); // Vector B in my notes
|
||||
FermionField DoidP_Phi(fgrid); // Vector E in my notes
|
||||
FermionField DobidDddDoidP_Phi(fgrid); // Vector F in my notes
|
||||
|
||||
FermionField P_Phi(fgrid);
|
||||
|
||||
// P term
|
||||
NumOp.dBoundaryBar(Phi,tmp);
|
||||
NumOp.dOmegaBarInv(tmp,DobiDdbPhi); // Vector A
|
||||
NumOp.dBoundary(DobiDdbPhi,tmp);
|
||||
NumOp.dOmegaInv(tmp,DoiDdDobiDdbPhi); // Vector B
|
||||
P_Phi = Phi - DoiDdDobiDdbPhi;
|
||||
NumOp.ProjectBoundaryBar(P_Phi);
|
||||
|
||||
// P^dag P term
|
||||
NumOp.dOmegaDagInv(P_Phi,DoidP_Phi); // Vector E
|
||||
NumOp.dBoundaryDag(DoidP_Phi,tmp);
|
||||
NumOp.dOmegaBarDagInv(tmp,DobidDddDoidP_Phi); // Vector F
|
||||
NumOp.dBoundaryBarDag(DobidDddDoidP_Phi,tmp);
|
||||
|
||||
X = DobiDdbPhi;
|
||||
Y = DobidDddDoidP_Phi;
|
||||
NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo); dSdU=force;
|
||||
NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes); dSdU=dSdU+force;
|
||||
|
||||
X = DoiDdDobiDdbPhi;
|
||||
Y = DoidP_Phi;
|
||||
NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo); dSdU=dSdU+force;
|
||||
NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes); dSdU=dSdU+force;
|
||||
|
||||
dSdU *= -1.0;
|
||||
|
||||
};
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -1,158 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundary.h
|
||||
|
||||
Copyright (C) 2021
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#pragma once
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
///////////////////////////////////////
|
||||
// Two flavour ratio
|
||||
///////////////////////////////////////
|
||||
template<class ImplD,class ImplF>
|
||||
class DomainDecomposedBoundaryTwoFlavourPseudoFermion : public Action<typename ImplD::GaugeField> {
|
||||
public:
|
||||
INHERIT_IMPL_TYPES(ImplD);
|
||||
|
||||
private:
|
||||
SchurFactoredFermionOperator<ImplD,ImplF> & DenOp;// the basic operator
|
||||
RealD ActionStoppingCondition;
|
||||
RealD DerivativeStoppingCondition;
|
||||
RealD InnerStoppingCondition;
|
||||
|
||||
FermionField Phi; // the pseudo fermion field for this trajectory
|
||||
|
||||
RealD refresh_action;
|
||||
public:
|
||||
DomainDecomposedBoundaryTwoFlavourPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF> &_DenOp,RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol = 1.0e-6 )
|
||||
: DenOp(_DenOp),
|
||||
DerivativeStoppingCondition(_DerivativeTol),
|
||||
ActionStoppingCondition(_ActionTol),
|
||||
InnerStoppingCondition(_InnerTol),
|
||||
Phi(_DenOp.FermionGrid()) {};
|
||||
|
||||
virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourPseudoFermion";}
|
||||
|
||||
|
||||
virtual std::string LogParameters(){
|
||||
std::stringstream sstream;
|
||||
return sstream.str();
|
||||
}
|
||||
|
||||
virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
|
||||
{
|
||||
// P(phi) = e^{- phi^dag Rdag^-1 R^-1 phi}
|
||||
//
|
||||
// DenOp == R
|
||||
//
|
||||
// Take phi = R eta ; eta = R^-1 Phi
|
||||
//
|
||||
// P(eta) = e^{- eta^dag eta}
|
||||
//
|
||||
// e^{x^2/2 sig^2} => sig^2 = 0.5.
|
||||
//
|
||||
// So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
|
||||
//
|
||||
RealD scale = std::sqrt(0.5);
|
||||
|
||||
DenOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tol =ActionStoppingCondition;
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
FermionField eta(DenOp.FermionGrid());
|
||||
|
||||
gaussian(pRNG,eta); eta=eta*scale;
|
||||
|
||||
DenOp.ProjectBoundaryBar(eta);
|
||||
DenOp.R(eta,Phi);
|
||||
//DumpSliceNorm("Phi",Phi);
|
||||
refresh_action = norm2(eta);
|
||||
};
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
// S = phi^dag Rdag^-1 R^-1 phi
|
||||
//////////////////////////////////////////////////////
|
||||
virtual RealD S(const GaugeField &U) {
|
||||
|
||||
DenOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tol=ActionStoppingCondition;
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
FermionField X(DenOp.FermionGrid());
|
||||
|
||||
DenOp.RInv(Phi,X);
|
||||
|
||||
RealD action = norm2(X);
|
||||
|
||||
return action;
|
||||
};
|
||||
|
||||
virtual void deriv(const GaugeField &U,GaugeField & dSdU)
|
||||
{
|
||||
DenOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tol=DerivativeStoppingCondition;
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
GridBase *fgrid = DenOp.FermionGrid();
|
||||
GridBase *ugrid = DenOp.GaugeGrid();
|
||||
|
||||
FermionField X(fgrid);
|
||||
FermionField Y(fgrid);
|
||||
FermionField tmp(fgrid);
|
||||
|
||||
GaugeField force(ugrid);
|
||||
|
||||
FermionField DiDdb_Phi(fgrid); // Vector C in my notes
|
||||
FermionField DidRinv_Phi(fgrid); // Vector D in my notes
|
||||
FermionField Rinv_Phi(fgrid);
|
||||
|
||||
// FermionField RinvDagRinv_Phi(fgrid);
|
||||
// FermionField DdbdDidRinv_Phi(fgrid);
|
||||
|
||||
// R^-1 term
|
||||
DenOp.dBoundaryBar(Phi,tmp);
|
||||
DenOp.Dinverse(tmp,DiDdb_Phi); // Vector C
|
||||
Rinv_Phi = Phi - DiDdb_Phi;
|
||||
DenOp.ProjectBoundaryBar(Rinv_Phi);
|
||||
|
||||
// R^-dagger R^-1 term
|
||||
DenOp.DinverseDag(Rinv_Phi,DidRinv_Phi); // Vector D
|
||||
/*
|
||||
DenOp.dBoundaryBarDag(DidRinv_Phi,DdbdDidRinv_Phi);
|
||||
RinvDagRinv_Phi = Rinv_Phi - DdbdDidRinv_Phi;
|
||||
DenOp.ProjectBoundaryBar(RinvDagRinv_Phi);
|
||||
*/
|
||||
X = DiDdb_Phi;
|
||||
Y = DidRinv_Phi;
|
||||
DenOp.PeriodicFermOpD.MDeriv(force,Y,X,DaggerNo); dSdU=force;
|
||||
DenOp.PeriodicFermOpD.MDeriv(force,X,Y,DaggerYes); dSdU=dSdU+force;
|
||||
DumpSliceNorm("force",dSdU);
|
||||
dSdU *= -1.0;
|
||||
};
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -1,237 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/pseudofermion/DomainDecomposedTwoFlavourBoundary.h
|
||||
|
||||
Copyright (C) 2021
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
|
||||
This program is free software; you can redistribute it and/or modify
|
||||
it under the terms of the GNU General Public License as published by
|
||||
the Free Software Foundation; either version 2 of the License, or
|
||||
(at your option) any later version.
|
||||
|
||||
This program is distributed in the hope that it will be useful,
|
||||
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
GNU General Public License for more details.
|
||||
|
||||
You should have received a copy of the GNU General Public License along
|
||||
with this program; if not, write to the Free Software Foundation, Inc.,
|
||||
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#pragma once
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
///////////////////////////////////////
|
||||
// Two flavour ratio
|
||||
///////////////////////////////////////
|
||||
template<class ImplD,class ImplF>
|
||||
class DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion : public Action<typename ImplD::GaugeField> {
|
||||
public:
|
||||
INHERIT_IMPL_TYPES(ImplD);
|
||||
|
||||
private:
|
||||
SchurFactoredFermionOperator<ImplD,ImplF> & NumOp;// the basic operator
|
||||
SchurFactoredFermionOperator<ImplD,ImplF> & DenOp;// the basic operator
|
||||
|
||||
RealD InnerStoppingCondition;
|
||||
RealD ActionStoppingCondition;
|
||||
RealD DerivativeStoppingCondition;
|
||||
|
||||
FermionField Phi; // the pseudo fermion field for this trajectory
|
||||
|
||||
public:
|
||||
DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion(SchurFactoredFermionOperator<ImplD,ImplF> &_NumOp,
|
||||
SchurFactoredFermionOperator<ImplD,ImplF> &_DenOp,
|
||||
RealD _DerivativeTol, RealD _ActionTol, RealD _InnerTol=1.0e-6)
|
||||
: NumOp(_NumOp), DenOp(_DenOp),
|
||||
Phi(_NumOp.PeriodicFermOpD.FermionGrid()),
|
||||
InnerStoppingCondition(_InnerTol),
|
||||
DerivativeStoppingCondition(_DerivativeTol),
|
||||
ActionStoppingCondition(_ActionTol)
|
||||
{};
|
||||
|
||||
virtual std::string action_name(){return "DomainDecomposedBoundaryTwoFlavourRatioPseudoFermion";}
|
||||
|
||||
virtual std::string LogParameters(){
|
||||
std::stringstream sstream;
|
||||
return sstream.str();
|
||||
}
|
||||
|
||||
virtual void refresh(const GaugeField &U, GridSerialRNG& sRNG, GridParallelRNG& pRNG)
|
||||
{
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
FermionField eta(NumOp.PeriodicFermOpD.FermionGrid());
|
||||
FermionField tmp(NumOp.PeriodicFermOpD.FermionGrid());
|
||||
|
||||
// P(phi) = e^{- phi^dag P^dag Rdag^-1 R^-1 P phi}
|
||||
//
|
||||
// NumOp == P
|
||||
// DenOp == R
|
||||
//
|
||||
// Take phi = P^{-1} R eta ; eta = R^-1 P Phi
|
||||
//
|
||||
// P(eta) = e^{- eta^dag eta}
|
||||
//
|
||||
// e^{x^2/2 sig^2} => sig^2 = 0.5.
|
||||
//
|
||||
// So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
|
||||
//
|
||||
RealD scale = std::sqrt(0.5);
|
||||
|
||||
gaussian(pRNG,eta); eta=eta*scale;
|
||||
|
||||
NumOp.ProjectBoundaryBar(eta);
|
||||
NumOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tol = ActionStoppingCondition;
|
||||
NumOp.tol = ActionStoppingCondition;
|
||||
DenOp.R(eta,tmp);
|
||||
NumOp.RInv(tmp,Phi);
|
||||
DumpSliceNorm("Phi",Phi);
|
||||
|
||||
};
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
// S = phi^dag Pdag Rdag^-1 R^-1 P phi
|
||||
//////////////////////////////////////////////////////
|
||||
virtual RealD S(const GaugeField &U) {
|
||||
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
FermionField X(NumOp.PeriodicFermOpD.FermionGrid());
|
||||
FermionField Y(NumOp.PeriodicFermOpD.FermionGrid());
|
||||
|
||||
NumOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tol = ActionStoppingCondition;
|
||||
NumOp.tol = ActionStoppingCondition;
|
||||
NumOp.R(Phi,Y);
|
||||
DenOp.RInv(Y,X);
|
||||
|
||||
RealD action = norm2(X);
|
||||
// std::cout << " DD boundary action is " <<action<<std::endl;
|
||||
|
||||
return action;
|
||||
};
|
||||
|
||||
virtual void deriv(const GaugeField &U,GaugeField & dSdU)
|
||||
{
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
GridBase *fgrid = NumOp.PeriodicFermOpD.FermionGrid();
|
||||
GridBase *ugrid = NumOp.PeriodicFermOpD.GaugeGrid();
|
||||
|
||||
FermionField X(fgrid);
|
||||
FermionField Y(fgrid);
|
||||
FermionField tmp(fgrid);
|
||||
|
||||
GaugeField force(ugrid);
|
||||
|
||||
FermionField DobiDdbPhi(fgrid); // Vector A in my notes
|
||||
FermionField DoiDdDobiDdbPhi(fgrid); // Vector B in my notes
|
||||
FermionField DiDdbP_Phi(fgrid); // Vector C in my notes
|
||||
FermionField DidRinvP_Phi(fgrid); // Vector D in my notes
|
||||
FermionField DdbdDidRinvP_Phi(fgrid);
|
||||
FermionField DoidRinvDagRinvP_Phi(fgrid); // Vector E in my notes
|
||||
FermionField DobidDddDoidRinvDagRinvP_Phi(fgrid); // Vector F in my notes
|
||||
|
||||
FermionField P_Phi(fgrid);
|
||||
FermionField RinvP_Phi(fgrid);
|
||||
FermionField RinvDagRinvP_Phi(fgrid);
|
||||
FermionField PdagRinvDagRinvP_Phi(fgrid);
|
||||
|
||||
// RealD action = S(U);
|
||||
NumOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tolinner=InnerStoppingCondition;
|
||||
DenOp.tol = DerivativeStoppingCondition;
|
||||
NumOp.tol = DerivativeStoppingCondition;
|
||||
|
||||
// P term
|
||||
NumOp.dBoundaryBar(Phi,tmp);
|
||||
NumOp.dOmegaBarInv(tmp,DobiDdbPhi); // Vector A
|
||||
NumOp.dBoundary(DobiDdbPhi,tmp);
|
||||
NumOp.dOmegaInv(tmp,DoiDdDobiDdbPhi); // Vector B
|
||||
P_Phi = Phi - DoiDdDobiDdbPhi;
|
||||
NumOp.ProjectBoundaryBar(P_Phi);
|
||||
|
||||
// R^-1 P term
|
||||
DenOp.dBoundaryBar(P_Phi,tmp);
|
||||
DenOp.Dinverse(tmp,DiDdbP_Phi); // Vector C
|
||||
RinvP_Phi = P_Phi - DiDdbP_Phi;
|
||||
DenOp.ProjectBoundaryBar(RinvP_Phi); // Correct to here
|
||||
|
||||
|
||||
// R^-dagger R^-1 P term
|
||||
DenOp.DinverseDag(RinvP_Phi,DidRinvP_Phi); // Vector D
|
||||
DenOp.dBoundaryBarDag(DidRinvP_Phi,DdbdDidRinvP_Phi);
|
||||
RinvDagRinvP_Phi = RinvP_Phi - DdbdDidRinvP_Phi;
|
||||
DenOp.ProjectBoundaryBar(RinvDagRinvP_Phi);
|
||||
|
||||
|
||||
// P^dag R^-dagger R^-1 P term
|
||||
NumOp.dOmegaDagInv(RinvDagRinvP_Phi,DoidRinvDagRinvP_Phi); // Vector E
|
||||
NumOp.dBoundaryDag(DoidRinvDagRinvP_Phi,tmp);
|
||||
NumOp.dOmegaBarDagInv(tmp,DobidDddDoidRinvDagRinvP_Phi); // Vector F
|
||||
NumOp.dBoundaryBarDag(DobidDddDoidRinvDagRinvP_Phi,tmp);
|
||||
PdagRinvDagRinvP_Phi = RinvDagRinvP_Phi- tmp;
|
||||
NumOp.ProjectBoundaryBar(PdagRinvDagRinvP_Phi);
|
||||
|
||||
/*
|
||||
std::cout << "S eval "<< action << std::endl;
|
||||
std::cout << "S - IP1 "<< innerProduct(Phi,PdagRinvDagRinvP_Phi) << std::endl;
|
||||
std::cout << "S - IP2 "<< norm2(RinvP_Phi) << std::endl;
|
||||
|
||||
NumOp.R(Phi,tmp);
|
||||
tmp = tmp - P_Phi;
|
||||
std::cout << "diff1 "<<norm2(tmp) <<std::endl;
|
||||
|
||||
|
||||
DenOp.RInv(P_Phi,tmp);
|
||||
tmp = tmp - RinvP_Phi;
|
||||
std::cout << "diff2 "<<norm2(tmp) <<std::endl;
|
||||
|
||||
DenOp.RDagInv(RinvP_Phi,tmp);
|
||||
tmp = tmp - RinvDagRinvP_Phi;
|
||||
std::cout << "diff3 "<<norm2(tmp) <<std::endl;
|
||||
|
||||
DenOp.RDag(RinvDagRinvP_Phi,tmp);
|
||||
tmp = tmp - PdagRinvDagRinvP_Phi;
|
||||
std::cout << "diff4 "<<norm2(tmp) <<std::endl;
|
||||
*/
|
||||
|
||||
dSdU=Zero();
|
||||
|
||||
X = DobiDdbPhi;
|
||||
Y = DobidDddDoidRinvDagRinvP_Phi;
|
||||
NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo); dSdU=dSdU+force;
|
||||
NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes); dSdU=dSdU+force;
|
||||
|
||||
X = DoiDdDobiDdbPhi;
|
||||
Y = DoidRinvDagRinvP_Phi;
|
||||
NumOp.DirichletFermOpD.MDeriv(force,Y,X,DaggerNo); dSdU=dSdU+force;
|
||||
NumOp.DirichletFermOpD.MDeriv(force,X,Y,DaggerYes); dSdU=dSdU+force;
|
||||
|
||||
X = DiDdbP_Phi;
|
||||
Y = DidRinvP_Phi;
|
||||
DenOp.PeriodicFermOpD.MDeriv(force,Y,X,DaggerNo); dSdU=dSdU+force;
|
||||
DenOp.PeriodicFermOpD.MDeriv(force,X,Y,DaggerYes); dSdU=dSdU+force;
|
||||
|
||||
dSdU *= -1.0;
|
||||
|
||||
};
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -44,10 +44,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
// Exact one flavour implementation of DWF determinant ratio //
|
||||
///////////////////////////////////////////////////////////////
|
||||
|
||||
//Note: using mixed prec CG for the heatbath solver in this action class will not work
|
||||
// because the L, R operators must have their shift coefficients updated throughout the heatbath step
|
||||
// You will find that the heatbath solver simply won't converge.
|
||||
// To use mixed precision here use the ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction variant below
|
||||
template<class Impl>
|
||||
class ExactOneFlavourRatioPseudoFermionAction : public Action<typename Impl::GaugeField>
|
||||
{
|
||||
@ -61,60 +57,37 @@ NAMESPACE_BEGIN(Grid);
|
||||
bool use_heatbath_forecasting;
|
||||
AbstractEOFAFermion<Impl>& Lop; // the basic LH operator
|
||||
AbstractEOFAFermion<Impl>& Rop; // the basic RH operator
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> SolverHBL;
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> SolverHBR;
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> SolverHB;
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> SolverL;
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> SolverR;
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> DerivativeSolverL;
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> DerivativeSolverR;
|
||||
FermionField Phi; // the pseudofermion field for this trajectory
|
||||
|
||||
RealD norm2_eta; //|eta|^2 where eta is the random gaussian field used to generate the pseudofermion field
|
||||
bool initial_action; //true for the first call to S after refresh, for which the identity S = |eta|^2 holds provided the rational approx is good
|
||||
public:
|
||||
|
||||
//Used in the heatbath, refresh the shift coefficients of the L (LorR=0) or R (LorR=1) operator
|
||||
virtual void heatbathRefreshShiftCoefficients(int LorR, RealD to){
|
||||
AbstractEOFAFermion<Impl>&op = LorR == 0 ? Lop : Rop;
|
||||
op.RefreshShiftCoefficients(to);
|
||||
}
|
||||
|
||||
|
||||
//Use the same solver for L,R in all cases
|
||||
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop,
|
||||
AbstractEOFAFermion<Impl>& _Rop,
|
||||
OperatorFunction<FermionField>& CG,
|
||||
Params& p,
|
||||
bool use_fc=false)
|
||||
: ExactOneFlavourRatioPseudoFermionAction(_Lop,_Rop,CG,CG,CG,CG,CG,CG,p,use_fc) {};
|
||||
|
||||
//Use the same solver for L,R in the heatbath but different solvers elsewhere
|
||||
: ExactOneFlavourRatioPseudoFermionAction(_Lop,_Rop,CG,CG,CG,CG,CG,p,use_fc) {};
|
||||
|
||||
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop,
|
||||
AbstractEOFAFermion<Impl>& _Rop,
|
||||
OperatorFunction<FermionField>& HeatbathCG,
|
||||
OperatorFunction<FermionField>& ActionCGL, OperatorFunction<FermionField>& ActionCGR,
|
||||
OperatorFunction<FermionField>& DerivCGL , OperatorFunction<FermionField>& DerivCGR,
|
||||
Params& p,
|
||||
bool use_fc=false)
|
||||
: ExactOneFlavourRatioPseudoFermionAction(_Lop,_Rop,HeatbathCG,HeatbathCG, ActionCGL, ActionCGR, DerivCGL,DerivCGR,p,use_fc) {};
|
||||
|
||||
//Use different solvers for L,R in all cases
|
||||
ExactOneFlavourRatioPseudoFermionAction(AbstractEOFAFermion<Impl>& _Lop,
|
||||
AbstractEOFAFermion<Impl>& _Rop,
|
||||
OperatorFunction<FermionField>& HeatbathCGL, OperatorFunction<FermionField>& HeatbathCGR,
|
||||
OperatorFunction<FermionField>& HeatbathCG,
|
||||
OperatorFunction<FermionField>& ActionCGL, OperatorFunction<FermionField>& ActionCGR,
|
||||
OperatorFunction<FermionField>& DerivCGL , OperatorFunction<FermionField>& DerivCGR,
|
||||
Params& p,
|
||||
bool use_fc=false) :
|
||||
Lop(_Lop),
|
||||
Rop(_Rop),
|
||||
SolverHBL(HeatbathCGL,false,true), SolverHBR(HeatbathCGR,false,true),
|
||||
SolverHB(HeatbathCG,false,true),
|
||||
SolverL(ActionCGL, false, true), SolverR(ActionCGR, false, true),
|
||||
DerivativeSolverL(DerivCGL, false, true), DerivativeSolverR(DerivCGR, false, true),
|
||||
Phi(_Lop.FermionGrid()),
|
||||
param(p),
|
||||
use_heatbath_forecasting(use_fc),
|
||||
initial_action(false)
|
||||
use_heatbath_forecasting(use_fc)
|
||||
{
|
||||
AlgRemez remez(param.lo, param.hi, param.precision);
|
||||
|
||||
@ -124,8 +97,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
PowerNegHalf.Init(remez, param.tolerance, true);
|
||||
};
|
||||
|
||||
const FermionField &getPhi() const{ return Phi; }
|
||||
|
||||
virtual std::string action_name() { return "ExactOneFlavourRatioPseudoFermionAction"; }
|
||||
|
||||
virtual std::string LogParameters() {
|
||||
@ -146,19 +117,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
else{ for(int s=0; s<Ls; ++s){ axpby_ssp_pminus(out, 0.0, in, 1.0, in, s, s); } }
|
||||
}
|
||||
|
||||
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
|
||||
// P(eta_o) = e^{- eta_o^dag eta_o}
|
||||
//
|
||||
// e^{x^2/2 sig^2} => sig^2 = 0.5.
|
||||
//
|
||||
RealD scale = std::sqrt(0.5);
|
||||
|
||||
FermionField eta (Lop.FermionGrid());
|
||||
gaussian(pRNG,eta); eta = eta * scale;
|
||||
|
||||
refresh(U,eta);
|
||||
}
|
||||
|
||||
// EOFA heatbath: see Eqn. (29) of arXiv:1706.05843
|
||||
// We generate a Gaussian noise vector \eta, and then compute
|
||||
// \Phi = M_{\rm EOFA}^{-1/2} * \eta
|
||||
@ -166,10 +124,12 @@ NAMESPACE_BEGIN(Grid);
|
||||
//
|
||||
// As a check of rational require \Phi^dag M_{EOFA} \Phi == eta^dag M^-1/2^dag M M^-1/2 eta = eta^dag eta
|
||||
//
|
||||
void refresh(const GaugeField &U, const FermionField &eta) {
|
||||
virtual void refresh(const GaugeField& U, GridSerialRNG &sRNG, GridParallelRNG& pRNG)
|
||||
{
|
||||
Lop.ImportGauge(U);
|
||||
Rop.ImportGauge(U);
|
||||
|
||||
FermionField eta (Lop.FermionGrid());
|
||||
FermionField CG_src (Lop.FermionGrid());
|
||||
FermionField CG_soln (Lop.FermionGrid());
|
||||
FermionField Forecast_src(Lop.FermionGrid());
|
||||
@ -180,6 +140,11 @@ NAMESPACE_BEGIN(Grid);
|
||||
if(use_heatbath_forecasting){ prev_solns.reserve(param.degree); }
|
||||
ChronoForecast<AbstractEOFAFermion<Impl>, FermionField> Forecast;
|
||||
|
||||
// Seed with Gaussian noise vector (var = 0.5)
|
||||
RealD scale = std::sqrt(0.5);
|
||||
gaussian(pRNG,eta);
|
||||
eta = eta * scale;
|
||||
|
||||
// \Phi = ( \alpha_{0} + \sum_{k=1}^{N_{p}} \alpha_{l} * \gamma_{l} ) * \eta
|
||||
RealD N(PowerNegHalf.norm);
|
||||
for(int k=0; k<param.degree; ++k){ N += PowerNegHalf.residues[k] / ( 1.0 + PowerNegHalf.poles[k] ); }
|
||||
@ -195,15 +160,15 @@ NAMESPACE_BEGIN(Grid);
|
||||
tmp[1] = Zero();
|
||||
for(int k=0; k<param.degree; ++k){
|
||||
gamma_l = 1.0 / ( 1.0 + PowerNegHalf.poles[k] );
|
||||
heatbathRefreshShiftCoefficients(0, -gamma_l);
|
||||
Lop.RefreshShiftCoefficients(-gamma_l);
|
||||
if(use_heatbath_forecasting){ // Forecast CG guess using solutions from previous poles
|
||||
Lop.Mdag(CG_src, Forecast_src);
|
||||
CG_soln = Forecast(Lop, Forecast_src, prev_solns);
|
||||
SolverHBL(Lop, CG_src, CG_soln);
|
||||
SolverHB(Lop, CG_src, CG_soln);
|
||||
prev_solns.push_back(CG_soln);
|
||||
} else {
|
||||
CG_soln = Zero(); // Just use zero as the initial guess
|
||||
SolverHBL(Lop, CG_src, CG_soln);
|
||||
SolverHB(Lop, CG_src, CG_soln);
|
||||
}
|
||||
Lop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
|
||||
tmp[1] = tmp[1] + ( PowerNegHalf.residues[k]*gamma_l*gamma_l*Lop.k ) * tmp[0];
|
||||
@ -222,15 +187,15 @@ NAMESPACE_BEGIN(Grid);
|
||||
if(use_heatbath_forecasting){ prev_solns.clear(); } // empirically, LH solns don't help for RH solves
|
||||
for(int k=0; k<param.degree; ++k){
|
||||
gamma_l = 1.0 / ( 1.0 + PowerNegHalf.poles[k] );
|
||||
heatbathRefreshShiftCoefficients(1, -gamma_l*PowerNegHalf.poles[k]);
|
||||
Rop.RefreshShiftCoefficients(-gamma_l*PowerNegHalf.poles[k]);
|
||||
if(use_heatbath_forecasting){
|
||||
Rop.Mdag(CG_src, Forecast_src);
|
||||
CG_soln = Forecast(Rop, Forecast_src, prev_solns);
|
||||
SolverHBR(Rop, CG_src, CG_soln);
|
||||
SolverHB(Rop, CG_src, CG_soln);
|
||||
prev_solns.push_back(CG_soln);
|
||||
} else {
|
||||
CG_soln = Zero();
|
||||
SolverHBR(Rop, CG_src, CG_soln);
|
||||
SolverHB(Rop, CG_src, CG_soln);
|
||||
}
|
||||
Rop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
|
||||
tmp[1] = tmp[1] - ( PowerNegHalf.residues[k]*gamma_l*gamma_l*Rop.k ) * tmp[0];
|
||||
@ -240,117 +205,49 @@ NAMESPACE_BEGIN(Grid);
|
||||
Phi = Phi + tmp[1];
|
||||
|
||||
// Reset shift coefficients for energy and force evals
|
||||
heatbathRefreshShiftCoefficients(0, 0.0);
|
||||
heatbathRefreshShiftCoefficients(1, -1.0);
|
||||
|
||||
//Mark that the next call to S is the first after refresh
|
||||
initial_action = true;
|
||||
|
||||
Lop.RefreshShiftCoefficients(0.0);
|
||||
Rop.RefreshShiftCoefficients(-1.0);
|
||||
|
||||
// Bounds check
|
||||
RealD EtaDagEta = norm2(eta);
|
||||
norm2_eta = EtaDagEta;
|
||||
|
||||
// RealD PhiDagMPhi= norm2(eta);
|
||||
|
||||
};
|
||||
|
||||
void Meofa(const GaugeField& U,const FermionField &in, FermionField & out)
|
||||
void Meofa(const GaugeField& U,const FermionField &phi, FermionField & Mphi)
|
||||
{
|
||||
#if 0
|
||||
Lop.ImportGauge(U);
|
||||
Rop.ImportGauge(U);
|
||||
|
||||
FermionField spProj_in(Lop.FermionGrid());
|
||||
FermionField spProj_Phi(Lop.FermionGrid());
|
||||
FermionField mPhi(Lop.FermionGrid());
|
||||
std::vector<FermionField> tmp(2, Lop.FermionGrid());
|
||||
out = in;
|
||||
mPhi = phi;
|
||||
|
||||
// LH term: S = S - k <\Phi| P_{-} \Omega_{-}^{\dagger} H(mf)^{-1} \Omega_{-} P_{-} |\Phi>
|
||||
spProj(in, spProj_in, -1, Lop.Ls);
|
||||
Lop.Omega(spProj_in, tmp[0], -1, 0);
|
||||
spProj(Phi, spProj_Phi, -1, Lop.Ls);
|
||||
Lop.Omega(spProj_Phi, tmp[0], -1, 0);
|
||||
G5R5(tmp[1], tmp[0]);
|
||||
tmp[0] = Zero();
|
||||
SolverL(Lop, tmp[1], tmp[0]);
|
||||
Lop.Dtilde(tmp[0], tmp[1]); // We actually solved Cayley preconditioned system: transform back
|
||||
Lop.Omega(tmp[1], tmp[0], -1, 1);
|
||||
spProj(tmp[0], tmp[1], -1, Lop.Ls);
|
||||
|
||||
out = out - Lop.k * tmp[1];
|
||||
mPhi = mPhi - Lop.k * innerProduct(spProj_Phi, tmp[0]).real();
|
||||
|
||||
// RH term: S = S + k <\Phi| P_{+} \Omega_{+}^{\dagger} ( H(mb)
|
||||
// - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{+} P_{+} |\Phi>
|
||||
spProj(in, spProj_in, 1, Rop.Ls);
|
||||
Rop.Omega(spProj_in, tmp[0], 1, 0);
|
||||
// - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{-} P_{-} |\Phi>
|
||||
spProj(Phi, spProj_Phi, 1, Rop.Ls);
|
||||
Rop.Omega(spProj_Phi, tmp[0], 1, 0);
|
||||
G5R5(tmp[1], tmp[0]);
|
||||
tmp[0] = Zero();
|
||||
SolverR(Rop, tmp[1], tmp[0]);
|
||||
Rop.Dtilde(tmp[0], tmp[1]);
|
||||
Rop.Omega(tmp[1], tmp[0], 1, 1);
|
||||
spProj(tmp[0], tmp[1], 1, Rop.Ls);
|
||||
|
||||
out = out + Rop.k * tmp[1];
|
||||
action += Rop.k * innerProduct(spProj_Phi, tmp[0]).real();
|
||||
#endif
|
||||
}
|
||||
|
||||
//Due to the structure of EOFA, it is no more expensive to compute the inverse of Meofa
|
||||
//To ensure correctness we can simply reuse the heatbath code but use the rational approx
|
||||
//f(x) = 1/x which corresponds to alpha_0=0, alpha_1=1, beta_1=0 => gamma_1=1
|
||||
void MeofaInv(const GaugeField &U, const FermionField &in, FermionField &out) {
|
||||
Lop.ImportGauge(U);
|
||||
Rop.ImportGauge(U);
|
||||
|
||||
FermionField CG_src (Lop.FermionGrid());
|
||||
FermionField CG_soln (Lop.FermionGrid());
|
||||
std::vector<FermionField> tmp(2, Lop.FermionGrid());
|
||||
|
||||
// \Phi = ( \alpha_{0} + \sum_{k=1}^{N_{p}} \alpha_{l} * \gamma_{l} ) * \eta
|
||||
// = 1 * \eta
|
||||
out = in;
|
||||
|
||||
// LH terms:
|
||||
// \Phi = \Phi + k \sum_{k=1}^{N_{p}} P_{-} \Omega_{-}^{\dagger} ( H(mf)
|
||||
// - \gamma_{l} \Delta_{-}(mf,mb) P_{-} )^{-1} \Omega_{-} P_{-} \eta
|
||||
spProj(in, tmp[0], -1, Lop.Ls);
|
||||
Lop.Omega(tmp[0], tmp[1], -1, 0);
|
||||
G5R5(CG_src, tmp[1]);
|
||||
{
|
||||
heatbathRefreshShiftCoefficients(0, -1.); //-gamma_1 = -1.
|
||||
|
||||
CG_soln = Zero(); // Just use zero as the initial guess
|
||||
SolverHBL(Lop, CG_src, CG_soln);
|
||||
|
||||
Lop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
|
||||
tmp[1] = Lop.k * tmp[0];
|
||||
}
|
||||
Lop.Omega(tmp[1], tmp[0], -1, 1);
|
||||
spProj(tmp[0], tmp[1], -1, Lop.Ls);
|
||||
out = out + tmp[1];
|
||||
|
||||
// RH terms:
|
||||
// \Phi = \Phi - k \sum_{k=1}^{N_{p}} P_{+} \Omega_{+}^{\dagger} ( H(mb)
|
||||
// - \beta_l\gamma_{l} \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{+} P_{+} \eta
|
||||
spProj(in, tmp[0], 1, Rop.Ls);
|
||||
Rop.Omega(tmp[0], tmp[1], 1, 0);
|
||||
G5R5(CG_src, tmp[1]);
|
||||
{
|
||||
heatbathRefreshShiftCoefficients(1, 0.); //-gamma_1 * beta_1 = 0
|
||||
|
||||
CG_soln = Zero();
|
||||
SolverHBR(Rop, CG_src, CG_soln);
|
||||
|
||||
Rop.Dtilde(CG_soln, tmp[0]); // We actually solved Cayley preconditioned system: transform back
|
||||
tmp[1] = - Rop.k * tmp[0];
|
||||
}
|
||||
Rop.Omega(tmp[1], tmp[0], 1, 1);
|
||||
spProj(tmp[0], tmp[1], 1, Rop.Ls);
|
||||
out = out + tmp[1];
|
||||
|
||||
// Reset shift coefficients for energy and force evals
|
||||
heatbathRefreshShiftCoefficients(0, 0.0);
|
||||
heatbathRefreshShiftCoefficients(1, -1.0);
|
||||
};
|
||||
|
||||
|
||||
|
||||
|
||||
// EOFA action: see Eqn. (10) of arXiv:1706.05843
|
||||
virtual RealD S(const GaugeField& U)
|
||||
{
|
||||
@ -374,7 +271,7 @@ NAMESPACE_BEGIN(Grid);
|
||||
action -= Lop.k * innerProduct(spProj_Phi, tmp[0]).real();
|
||||
|
||||
// RH term: S = S + k <\Phi| P_{+} \Omega_{+}^{\dagger} ( H(mb)
|
||||
// - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{+} P_{+} |\Phi>
|
||||
// - \Delta_{+}(mf,mb) P_{+} )^{-1} \Omega_{-} P_{-} |\Phi>
|
||||
spProj(Phi, spProj_Phi, 1, Rop.Ls);
|
||||
Rop.Omega(spProj_Phi, tmp[0], 1, 0);
|
||||
G5R5(tmp[1], tmp[0]);
|
||||
@ -384,26 +281,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
Rop.Omega(tmp[1], tmp[0], 1, 1);
|
||||
action += Rop.k * innerProduct(spProj_Phi, tmp[0]).real();
|
||||
|
||||
if(initial_action){
|
||||
//For the first call to S after refresh, S = |eta|^2. We can use this to ensure the rational approx is good
|
||||
RealD diff = action - norm2_eta;
|
||||
|
||||
//S_init = eta^dag M^{-1/2} M M^{-1/2} eta
|
||||
//S_init - eta^dag eta = eta^dag ( M^{-1/2} M M^{-1/2} - 1 ) eta
|
||||
|
||||
//If approximate solution
|
||||
//S_init - eta^dag eta = eta^dag ( [M^{-1/2}+\delta M^{-1/2}] M [M^{-1/2}+\delta M^{-1/2}] - 1 ) eta
|
||||
// \approx eta^dag ( \delta M^{-1/2} M^{1/2} + M^{1/2}\delta M^{-1/2} ) eta
|
||||
// We divide out |eta|^2 to remove source scaling but the tolerance on this check should still be somewhat higher than the actual approx tolerance
|
||||
RealD test = fabs(diff)/norm2_eta; //test the quality of the rational approx
|
||||
|
||||
std::cout << GridLogMessage << action_name() << " initial action " << action << " expect " << norm2_eta << "; diff " << diff << std::endl;
|
||||
std::cout << GridLogMessage << action_name() << "[ eta^dag ( M^{-1/2} M M^{-1/2} - 1 ) eta ]/|eta^2| = " << test << " expect 0 (tol " << param.BoundsCheckTol << ")" << std::endl;
|
||||
|
||||
assert( ( test < param.BoundsCheckTol ) && " Initial action check failed" );
|
||||
initial_action = false;
|
||||
}
|
||||
|
||||
return action;
|
||||
};
|
||||
|
||||
@ -452,40 +329,6 @@ NAMESPACE_BEGIN(Grid);
|
||||
};
|
||||
};
|
||||
|
||||
template<class ImplD, class ImplF>
|
||||
class ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction : public ExactOneFlavourRatioPseudoFermionAction<ImplD>{
|
||||
public:
|
||||
INHERIT_IMPL_TYPES(ImplD);
|
||||
typedef OneFlavourRationalParams Params;
|
||||
|
||||
private:
|
||||
AbstractEOFAFermion<ImplF>& LopF; // the basic LH operator
|
||||
AbstractEOFAFermion<ImplF>& RopF; // the basic RH operator
|
||||
|
||||
public:
|
||||
|
||||
virtual std::string action_name() { return "ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction"; }
|
||||
|
||||
//Used in the heatbath, refresh the shift coefficients of the L (LorR=0) or R (LorR=1) operator
|
||||
virtual void heatbathRefreshShiftCoefficients(int LorR, RealD to){
|
||||
AbstractEOFAFermion<ImplF> &op = LorR == 0 ? LopF : RopF;
|
||||
op.RefreshShiftCoefficients(to);
|
||||
this->ExactOneFlavourRatioPseudoFermionAction<ImplD>::heatbathRefreshShiftCoefficients(LorR,to);
|
||||
}
|
||||
|
||||
ExactOneFlavourRatioMixedPrecHeatbathPseudoFermionAction(AbstractEOFAFermion<ImplF>& _LopF,
|
||||
AbstractEOFAFermion<ImplF>& _RopF,
|
||||
AbstractEOFAFermion<ImplD>& _LopD,
|
||||
AbstractEOFAFermion<ImplD>& _RopD,
|
||||
OperatorFunction<FermionField>& HeatbathCGL, OperatorFunction<FermionField>& HeatbathCGR,
|
||||
OperatorFunction<FermionField>& ActionCGL, OperatorFunction<FermionField>& ActionCGR,
|
||||
OperatorFunction<FermionField>& DerivCGL , OperatorFunction<FermionField>& DerivCGR,
|
||||
Params& p,
|
||||
bool use_fc=false) :
|
||||
LopF(_LopF), RopF(_RopF), ExactOneFlavourRatioPseudoFermionAction<ImplD>(_LopD, _RopD, HeatbathCGL, HeatbathCGR, ActionCGL, ActionCGR, DerivCGL, DerivCGR, p, use_fc){}
|
||||
};
|
||||
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
#endif
|
||||
|
||||
@ -1,372 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h
|
||||
|
||||
Copyright (C) 2015
|
||||
|
||||
Author: Christopher Kelly <ckelly@bnl.gov>
|
||||
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 QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_H
|
||||
#define QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_H
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
/////////////////////////////////////////////////////////
|
||||
// Generic rational approximation for ratios of operators
|
||||
/////////////////////////////////////////////////////////
|
||||
|
||||
/* S_f = -log( det( [M^dag M]/[V^dag V] )^{1/inv_pow} )
|
||||
= chi^dag ( [M^dag M]/[V^dag V] )^{-1/inv_pow} chi\
|
||||
= chi^dag ( [V^dag V]^{-1/2} [M^dag M] [V^dag V]^{-1/2} )^{-1/inv_pow} chi\
|
||||
= chi^dag [V^dag V]^{1/(2*inv_pow)} [M^dag M]^{-1/inv_pow} [V^dag V]^{1/(2*inv_pow)} chi\
|
||||
|
||||
S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
|
||||
|
||||
BIG WARNING:
|
||||
Here V^dag V is referred to in this code as the "numerator" operator and M^dag M is the *denominator* operator.
|
||||
this refers to their position in the pseudofermion action, which is the *inverse* of what appears in the determinant
|
||||
Thus for DWF the numerator operator is the Pauli-Villars operator
|
||||
|
||||
Here P/Q \sim R_{1/(2*inv_pow)} ~ (V^dagV)^{1/(2*inv_pow)}
|
||||
Here N/D \sim R_{-1/inv_pow} ~ (M^dagM)^{-1/inv_pow}
|
||||
*/
|
||||
|
||||
template<class Impl>
|
||||
class GeneralEvenOddRatioRationalPseudoFermionAction : public Action<typename Impl::GaugeField> {
|
||||
public:
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
|
||||
typedef RationalActionParams Params;
|
||||
Params param;
|
||||
|
||||
//For action evaluation
|
||||
MultiShiftFunction ApproxPowerAction ; //rational approx for X^{1/inv_pow}
|
||||
MultiShiftFunction ApproxNegPowerAction; //rational approx for X^{-1/inv_pow}
|
||||
MultiShiftFunction ApproxHalfPowerAction; //rational approx for X^{1/(2*inv_pow)}
|
||||
MultiShiftFunction ApproxNegHalfPowerAction; //rational approx for X^{-1/(2*inv_pow)}
|
||||
|
||||
//For the MD integration
|
||||
MultiShiftFunction ApproxPowerMD ; //rational approx for X^{1/inv_pow}
|
||||
MultiShiftFunction ApproxNegPowerMD; //rational approx for X^{-1/inv_pow}
|
||||
MultiShiftFunction ApproxHalfPowerMD; //rational approx for X^{1/(2*inv_pow)}
|
||||
MultiShiftFunction ApproxNegHalfPowerMD; //rational approx for X^{-1/(2*inv_pow)}
|
||||
|
||||
private:
|
||||
|
||||
FermionOperator<Impl> & NumOp;// the basic operator
|
||||
FermionOperator<Impl> & DenOp;// the basic operator
|
||||
FermionField PhiEven; // the pseudo fermion field for this trajectory
|
||||
FermionField PhiOdd; // the pseudo fermion field for this trajectory
|
||||
|
||||
//Generate the approximation to x^{1/inv_pow} (->approx) and x^{-1/inv_pow} (-> approx_inv) by an approx_degree degree rational approximation
|
||||
//CG_tolerance is used to issue a warning if the approximation error is larger than the tolerance of the CG and is otherwise just stored in the MultiShiftFunction for use by the multi-shift
|
||||
static void generateApprox(MultiShiftFunction &approx, MultiShiftFunction &approx_inv, int inv_pow, int approx_degree, double CG_tolerance, AlgRemez &remez){
|
||||
std::cout<<GridLogMessage << "Generating degree "<< approx_degree<<" approximation for x^(1/" << inv_pow << ")"<<std::endl;
|
||||
double error = remez.generateApprox(approx_degree,1,inv_pow);
|
||||
if(error > CG_tolerance)
|
||||
std::cout<<GridLogMessage << "WARNING: Remez approximation has a larger error " << error << " than the CG tolerance " << CG_tolerance << "! Try increasing the number of poles" << std::endl;
|
||||
|
||||
approx.Init(remez, CG_tolerance,false);
|
||||
approx_inv.Init(remez, CG_tolerance,true);
|
||||
}
|
||||
|
||||
|
||||
protected:
|
||||
static constexpr bool Numerator = true;
|
||||
static constexpr bool Denominator = false;
|
||||
|
||||
//Allow derived classes to override the multishift CG
|
||||
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionField &in, FermionField &out){
|
||||
SchurDifferentiableOperator<Impl> schurOp(numerator ? NumOp : DenOp);
|
||||
ConjugateGradientMultiShift<FermionField> msCG(MaxIter, approx);
|
||||
msCG(schurOp,in, out);
|
||||
}
|
||||
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionField &in, std::vector<FermionField> &out_elems, FermionField &out){
|
||||
SchurDifferentiableOperator<Impl> schurOp(numerator ? NumOp : DenOp);
|
||||
ConjugateGradientMultiShift<FermionField> msCG(MaxIter, approx);
|
||||
msCG(schurOp,in, out_elems, out);
|
||||
}
|
||||
//Allow derived classes to override the gauge import
|
||||
virtual void ImportGauge(const GaugeField &U){
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
}
|
||||
|
||||
public:
|
||||
|
||||
GeneralEvenOddRatioRationalPseudoFermionAction(FermionOperator<Impl> &_NumOp,
|
||||
FermionOperator<Impl> &_DenOp,
|
||||
const Params & p
|
||||
) :
|
||||
NumOp(_NumOp),
|
||||
DenOp(_DenOp),
|
||||
PhiOdd (_NumOp.FermionRedBlackGrid()),
|
||||
PhiEven(_NumOp.FermionRedBlackGrid()),
|
||||
param(p)
|
||||
{
|
||||
std::cout<<GridLogMessage << action_name() << " initialize: starting" << std::endl;
|
||||
AlgRemez remez(param.lo,param.hi,param.precision);
|
||||
|
||||
//Generate approximations for action eval
|
||||
generateApprox(ApproxPowerAction, ApproxNegPowerAction, param.inv_pow, param.action_degree, param.action_tolerance, remez);
|
||||
generateApprox(ApproxHalfPowerAction, ApproxNegHalfPowerAction, 2*param.inv_pow, param.action_degree, param.action_tolerance, remez);
|
||||
|
||||
//Generate approximations for MD
|
||||
if(param.md_degree != param.action_degree){ //note the CG tolerance is unrelated to the stopping condition of the Remez algorithm
|
||||
generateApprox(ApproxPowerMD, ApproxNegPowerMD, param.inv_pow, param.md_degree, param.md_tolerance, remez);
|
||||
generateApprox(ApproxHalfPowerMD, ApproxNegHalfPowerMD, 2*param.inv_pow, param.md_degree, param.md_tolerance, remez);
|
||||
}else{
|
||||
std::cout<<GridLogMessage << "Using same rational approximations for MD as for action evaluation" << std::endl;
|
||||
ApproxPowerMD = ApproxPowerAction;
|
||||
ApproxNegPowerMD = ApproxNegPowerAction;
|
||||
for(int i=0;i<ApproxPowerMD.tolerances.size();i++)
|
||||
ApproxNegPowerMD.tolerances[i] = ApproxPowerMD.tolerances[i] = param.md_tolerance; //used for multishift
|
||||
|
||||
ApproxHalfPowerMD = ApproxHalfPowerAction;
|
||||
ApproxNegHalfPowerMD = ApproxNegHalfPowerAction;
|
||||
for(int i=0;i<ApproxPowerMD.tolerances.size();i++)
|
||||
ApproxNegHalfPowerMD.tolerances[i] = ApproxHalfPowerMD.tolerances[i] = param.md_tolerance;
|
||||
}
|
||||
|
||||
std::cout<<GridLogMessage << action_name() << " initialize: complete" << std::endl;
|
||||
};
|
||||
|
||||
virtual std::string action_name(){return "GeneralEvenOddRatioRationalPseudoFermionAction";}
|
||||
|
||||
virtual std::string LogParameters(){
|
||||
std::stringstream sstream;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Power : 1/" << param.inv_pow << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Low :" << param.lo << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] High :" << param.hi << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Max iterations :" << param.MaxIter << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Tolerance (Action) :" << param.action_tolerance << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Degree (Action) :" << param.action_degree << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Tolerance (MD) :" << param.md_tolerance << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Degree (MD) :" << param.md_degree << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Precision :" << param.precision << std::endl;
|
||||
return sstream.str();
|
||||
}
|
||||
|
||||
//Access the fermion field
|
||||
const FermionField &getPhiOdd() const{ return PhiOdd; }
|
||||
|
||||
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
|
||||
std::cout<<GridLogMessage << action_name() << " refresh: starting" << std::endl;
|
||||
FermionField eta(NumOp.FermionGrid());
|
||||
|
||||
// P(eta) \propto e^{- eta^dag eta}
|
||||
//
|
||||
// The gaussian function draws from P(x) \propto e^{- x^2 / 2 } [i.e. sigma=1]
|
||||
// Thus eta = x/sqrt{2} = x * sqrt(1/2)
|
||||
RealD scale = std::sqrt(0.5);
|
||||
gaussian(pRNG,eta); eta=eta*scale;
|
||||
|
||||
refresh(U,eta);
|
||||
}
|
||||
|
||||
//Allow for manual specification of random field for testing
|
||||
void refresh(const GaugeField &U, const FermionField &eta) {
|
||||
|
||||
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
|
||||
//
|
||||
// P(phi) = e^{- phi^dag (VdagV)^1/(2*inv_pow) (MdagM)^-1/inv_pow (VdagV)^1/(2*inv_pow) phi}
|
||||
// = e^{- phi^dag (VdagV)^1/(2*inv_pow) (MdagM)^-1/(2*inv_pow) (MdagM)^-1/(2*inv_pow) (VdagV)^1/(2*inv_pow) phi}
|
||||
//
|
||||
// Phi = (VdagV)^-1/(2*inv_pow) Mdag^{1/(2*inv_pow)} eta
|
||||
|
||||
std::cout<<GridLogMessage << action_name() << " refresh: starting" << std::endl;
|
||||
|
||||
FermionField etaOdd (NumOp.FermionRedBlackGrid());
|
||||
FermionField etaEven(NumOp.FermionRedBlackGrid());
|
||||
FermionField tmp(NumOp.FermionRedBlackGrid());
|
||||
|
||||
pickCheckerboard(Even,etaEven,eta);
|
||||
pickCheckerboard(Odd,etaOdd,eta);
|
||||
|
||||
ImportGauge(U);
|
||||
|
||||
// MdagM^1/(2*inv_pow) eta
|
||||
std::cout<<GridLogMessage << action_name() << " refresh: doing (M^dag M)^{1/" << 2*param.inv_pow << "} eta" << std::endl;
|
||||
multiShiftInverse(Denominator, ApproxHalfPowerAction, param.MaxIter, etaOdd, tmp);
|
||||
|
||||
// VdagV^-1/(2*inv_pow) MdagM^1/(2*inv_pow) eta
|
||||
std::cout<<GridLogMessage << action_name() << " refresh: doing (V^dag V)^{-1/" << 2*param.inv_pow << "} ( (M^dag M)^{1/" << 2*param.inv_pow << "} eta)" << std::endl;
|
||||
multiShiftInverse(Numerator, ApproxNegHalfPowerAction, param.MaxIter, tmp, PhiOdd);
|
||||
|
||||
assert(NumOp.ConstEE() == 1);
|
||||
assert(DenOp.ConstEE() == 1);
|
||||
PhiEven = Zero();
|
||||
std::cout<<GridLogMessage << action_name() << " refresh: starting" << std::endl;
|
||||
};
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
|
||||
//////////////////////////////////////////////////////
|
||||
virtual RealD S(const GaugeField &U) {
|
||||
std::cout<<GridLogMessage << action_name() << " compute action: starting" << std::endl;
|
||||
ImportGauge(U);
|
||||
|
||||
FermionField X(NumOp.FermionRedBlackGrid());
|
||||
FermionField Y(NumOp.FermionRedBlackGrid());
|
||||
|
||||
// VdagV^1/(2*inv_pow) Phi
|
||||
std::cout<<GridLogMessage << action_name() << " compute action: doing (V^dag V)^{1/" << 2*param.inv_pow << "} Phi" << std::endl;
|
||||
multiShiftInverse(Numerator, ApproxHalfPowerAction, param.MaxIter, PhiOdd,X);
|
||||
|
||||
// MdagM^-1/(2*inv_pow) VdagV^1/(2*inv_pow) Phi
|
||||
std::cout<<GridLogMessage << action_name() << " compute action: doing (M^dag M)^{-1/" << 2*param.inv_pow << "} ( (V^dag V)^{1/" << 2*param.inv_pow << "} Phi)" << std::endl;
|
||||
multiShiftInverse(Denominator, ApproxNegHalfPowerAction, param.MaxIter, X,Y);
|
||||
|
||||
// Randomly apply rational bounds checks.
|
||||
int rcheck = rand();
|
||||
auto grid = NumOp.FermionGrid();
|
||||
auto r=rand();
|
||||
grid->Broadcast(0,r);
|
||||
|
||||
if ( param.BoundsCheckFreq != 0 && (r % param.BoundsCheckFreq)==0 ) {
|
||||
std::cout<<GridLogMessage << action_name() << " compute action: doing bounds check" << std::endl;
|
||||
FermionField gauss(NumOp.FermionRedBlackGrid());
|
||||
gauss = PhiOdd;
|
||||
SchurDifferentiableOperator<Impl> MdagM(DenOp);
|
||||
std::cout<<GridLogMessage << action_name() << " compute action: checking high bounds" << std::endl;
|
||||
HighBoundCheck(MdagM,gauss,param.hi);
|
||||
std::cout<<GridLogMessage << action_name() << " compute action: full approximation" << std::endl;
|
||||
InversePowerBoundsCheck(param.inv_pow,param.MaxIter,param.action_tolerance*100,MdagM,gauss,ApproxNegPowerAction);
|
||||
std::cout<<GridLogMessage << action_name() << " compute action: bounds check complete" << std::endl;
|
||||
}
|
||||
|
||||
// Phidag VdagV^1/(2*inv_pow) MdagM^-1/(2*inv_pow) MdagM^-1/(2*inv_pow) VdagV^1/(2*inv_pow) Phi
|
||||
RealD action = norm2(Y);
|
||||
std::cout<<GridLogMessage << action_name() << " compute action: complete" << std::endl;
|
||||
|
||||
return action;
|
||||
};
|
||||
|
||||
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
|
||||
//
|
||||
// Here, M is some 5D operator and V is the Pauli-Villars field
|
||||
// N and D makeup the rat. poly of the M term and P and & makeup the rat.poly of the denom term
|
||||
//
|
||||
// Need
|
||||
// dS_f/dU = chi^dag d[P/Q] N/D P/Q chi
|
||||
// + chi^dag P/Q d[N/D] P/Q chi
|
||||
// + chi^dag P/Q N/D d[P/Q] chi
|
||||
//
|
||||
// P/Q is expressed as partial fraction expansion:
|
||||
//
|
||||
// a0 + \sum_k ak/(V^dagV + bk)
|
||||
//
|
||||
// d[P/Q] is then
|
||||
//
|
||||
// \sum_k -ak [V^dagV+bk]^{-1} [ dV^dag V + V^dag dV ] [V^dag V + bk]^{-1}
|
||||
//
|
||||
// and similar for N/D.
|
||||
//
|
||||
// Need
|
||||
// MpvPhi_k = [Vdag V + bk]^{-1} chi
|
||||
// MpvPhi = {a0 + \sum_k ak [Vdag V + bk]^{-1} }chi
|
||||
//
|
||||
// MfMpvPhi_k = [MdagM+bk]^{-1} MpvPhi
|
||||
// MfMpvPhi = {a0 + \sum_k ak [Mdag M + bk]^{-1} } MpvPhi
|
||||
//
|
||||
// MpvMfMpvPhi_k = [Vdag V + bk]^{-1} MfMpvchi
|
||||
//
|
||||
|
||||
virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
|
||||
std::cout<<GridLogMessage << action_name() << " deriv: starting" << std::endl;
|
||||
const int n_f = ApproxNegPowerMD.poles.size();
|
||||
const int n_pv = ApproxHalfPowerMD.poles.size();
|
||||
|
||||
std::vector<FermionField> MpvPhi_k (n_pv,NumOp.FermionRedBlackGrid());
|
||||
std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionRedBlackGrid());
|
||||
std::vector<FermionField> MfMpvPhi_k (n_f ,NumOp.FermionRedBlackGrid());
|
||||
|
||||
FermionField MpvPhi(NumOp.FermionRedBlackGrid());
|
||||
FermionField MfMpvPhi(NumOp.FermionRedBlackGrid());
|
||||
FermionField MpvMfMpvPhi(NumOp.FermionRedBlackGrid());
|
||||
FermionField Y(NumOp.FermionRedBlackGrid());
|
||||
|
||||
GaugeField tmp(NumOp.GaugeGrid());
|
||||
|
||||
ImportGauge(U);
|
||||
|
||||
std::cout<<GridLogMessage << action_name() << " deriv: doing (V^dag V)^{1/" << 2*param.inv_pow << "} Phi" << std::endl;
|
||||
multiShiftInverse(Numerator, ApproxHalfPowerMD, param.MaxIter, PhiOdd,MpvPhi_k,MpvPhi);
|
||||
|
||||
std::cout<<GridLogMessage << action_name() << " deriv: doing (M^dag M)^{-1/" << param.inv_pow << "} ( (V^dag V)^{1/" << 2*param.inv_pow << "} Phi)" << std::endl;
|
||||
multiShiftInverse(Denominator, ApproxNegPowerMD, param.MaxIter, MpvPhi,MfMpvPhi_k,MfMpvPhi);
|
||||
|
||||
std::cout<<GridLogMessage << action_name() << " deriv: doing (V^dag V)^{1/" << 2*param.inv_pow << "} ( (M^dag M)^{-1/" << param.inv_pow << "} (V^dag V)^{1/" << 2*param.inv_pow << "} Phi)" << std::endl;
|
||||
multiShiftInverse(Numerator, ApproxHalfPowerMD, param.MaxIter, MfMpvPhi,MpvMfMpvPhi_k,MpvMfMpvPhi);
|
||||
|
||||
|
||||
SchurDifferentiableOperator<Impl> MdagM(DenOp);
|
||||
SchurDifferentiableOperator<Impl> VdagV(NumOp);
|
||||
|
||||
|
||||
RealD ak;
|
||||
|
||||
dSdU = Zero();
|
||||
|
||||
// With these building blocks
|
||||
//
|
||||
// dS/dU =
|
||||
// \sum_k -ak MfMpvPhi_k^dag [ dM^dag M + M^dag dM ] MfMpvPhi_k (1)
|
||||
// + \sum_k -ak MpvMfMpvPhi_k^\dag [ dV^dag V + V^dag dV ] MpvPhi_k (2)
|
||||
// -ak MpvPhi_k^dag [ dV^dag V + V^dag dV ] MpvMfMpvPhi_k (3)
|
||||
|
||||
//(1)
|
||||
std::cout<<GridLogMessage << action_name() << " deriv: doing dS/dU part (1)" << std::endl;
|
||||
for(int k=0;k<n_f;k++){
|
||||
ak = ApproxNegPowerMD.residues[k];
|
||||
MdagM.Mpc(MfMpvPhi_k[k],Y);
|
||||
MdagM.MpcDagDeriv(tmp , MfMpvPhi_k[k], Y ); dSdU=dSdU+ak*tmp;
|
||||
MdagM.MpcDeriv(tmp , Y, MfMpvPhi_k[k] ); dSdU=dSdU+ak*tmp;
|
||||
}
|
||||
|
||||
//(2)
|
||||
//(3)
|
||||
std::cout<<GridLogMessage << action_name() << " deriv: doing dS/dU part (2)+(3)" << std::endl;
|
||||
for(int k=0;k<n_pv;k++){
|
||||
|
||||
ak = ApproxHalfPowerMD.residues[k];
|
||||
|
||||
VdagV.Mpc(MpvPhi_k[k],Y);
|
||||
VdagV.MpcDagDeriv(tmp,MpvMfMpvPhi_k[k],Y); dSdU=dSdU+ak*tmp;
|
||||
VdagV.MpcDeriv (tmp,Y,MpvMfMpvPhi_k[k]); dSdU=dSdU+ak*tmp;
|
||||
|
||||
VdagV.Mpc(MpvMfMpvPhi_k[k],Y); // V as we take Ydag
|
||||
VdagV.MpcDeriv (tmp,Y, MpvPhi_k[k]); dSdU=dSdU+ak*tmp;
|
||||
VdagV.MpcDagDeriv(tmp,MpvPhi_k[k], Y); dSdU=dSdU+ak*tmp;
|
||||
|
||||
}
|
||||
|
||||
//dSdU = Ta(dSdU);
|
||||
std::cout<<GridLogMessage << action_name() << " deriv: complete" << std::endl;
|
||||
};
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
#endif
|
||||
@ -1,93 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h
|
||||
|
||||
Copyright (C) 2015
|
||||
|
||||
Author: Christopher Kelly <ckelly@bnl.gov>
|
||||
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 QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
|
||||
#define QCD_PSEUDOFERMION_GENERAL_EVEN_ODD_RATIONAL_RATIO_MIXED_PREC_H
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Generic rational approximation for ratios of operators utilizing the mixed precision multishift algorithm
|
||||
// cf. GeneralEvenOddRational.h for details
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
template<class ImplD, class ImplF>
|
||||
class GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction : public GeneralEvenOddRatioRationalPseudoFermionAction<ImplD> {
|
||||
private:
|
||||
typedef typename ImplD::FermionField FermionFieldD;
|
||||
typedef typename ImplF::FermionField FermionFieldF;
|
||||
|
||||
FermionOperator<ImplD> & NumOpD;
|
||||
FermionOperator<ImplD> & DenOpD;
|
||||
|
||||
FermionOperator<ImplF> & NumOpF;
|
||||
FermionOperator<ImplF> & DenOpF;
|
||||
|
||||
Integer ReliableUpdateFreq;
|
||||
protected:
|
||||
|
||||
//Allow derived classes to override the multishift CG
|
||||
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, FermionFieldD &out){
|
||||
SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
|
||||
SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
|
||||
|
||||
ConjugateGradientMultiShiftMixedPrec<FermionFieldD, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
|
||||
msCG(schurOpD, in, out);
|
||||
}
|
||||
virtual void multiShiftInverse(bool numerator, const MultiShiftFunction &approx, const Integer MaxIter, const FermionFieldD &in, std::vector<FermionFieldD> &out_elems, FermionFieldD &out){
|
||||
SchurDifferentiableOperator<ImplD> schurOpD(numerator ? NumOpD : DenOpD);
|
||||
SchurDifferentiableOperator<ImplF> schurOpF(numerator ? NumOpF : DenOpF);
|
||||
|
||||
ConjugateGradientMultiShiftMixedPrec<FermionFieldD, FermionFieldF> msCG(MaxIter, approx, NumOpF.FermionRedBlackGrid(), schurOpF, ReliableUpdateFreq);
|
||||
msCG(schurOpD, in, out_elems, out);
|
||||
}
|
||||
//Allow derived classes to override the gauge import
|
||||
virtual void ImportGauge(const typename ImplD::GaugeField &Ud){
|
||||
typename ImplF::GaugeField Uf(NumOpF.GaugeGrid());
|
||||
precisionChange(Uf, Ud);
|
||||
|
||||
NumOpD.ImportGauge(Ud);
|
||||
DenOpD.ImportGauge(Ud);
|
||||
|
||||
NumOpF.ImportGauge(Uf);
|
||||
DenOpF.ImportGauge(Uf);
|
||||
}
|
||||
|
||||
public:
|
||||
GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction(FermionOperator<ImplD> &_NumOpD, FermionOperator<ImplD> &_DenOpD,
|
||||
FermionOperator<ImplF> &_NumOpF, FermionOperator<ImplF> &_DenOpF,
|
||||
const RationalActionParams & p, Integer _ReliableUpdateFreq
|
||||
) : GeneralEvenOddRatioRationalPseudoFermionAction<ImplD>(_NumOpD, _DenOpD, p),
|
||||
ReliableUpdateFreq(_ReliableUpdateFreq), NumOpD(_NumOpD), DenOpD(_DenOpD), NumOpF(_NumOpF), DenOpF(_DenOpF){}
|
||||
|
||||
virtual std::string action_name(){return "GeneralEvenOddRatioRationalMixedPrecPseudoFermionAction";}
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
#endif
|
||||
@ -40,31 +40,249 @@ NAMESPACE_BEGIN(Grid);
|
||||
// Here N/D \sim R_{-1/2} ~ (M^dagM)^{-1/2}
|
||||
|
||||
template<class Impl>
|
||||
class OneFlavourEvenOddRatioRationalPseudoFermionAction : public GeneralEvenOddRatioRationalPseudoFermionAction<Impl> {
|
||||
class OneFlavourEvenOddRatioRationalPseudoFermionAction : public Action<typename Impl::GaugeField> {
|
||||
public:
|
||||
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
|
||||
typedef OneFlavourRationalParams Params;
|
||||
Params param;
|
||||
|
||||
MultiShiftFunction PowerHalf ;
|
||||
MultiShiftFunction PowerNegHalf;
|
||||
MultiShiftFunction PowerQuarter;
|
||||
MultiShiftFunction PowerNegQuarter;
|
||||
|
||||
private:
|
||||
static RationalActionParams transcribe(const Params &in){
|
||||
RationalActionParams out;
|
||||
out.inv_pow = 2;
|
||||
out.lo = in.lo;
|
||||
out.hi = in.hi;
|
||||
out.MaxIter = in.MaxIter;
|
||||
out.action_tolerance = out.md_tolerance = in.tolerance;
|
||||
out.action_degree = out.md_degree = in.degree;
|
||||
out.precision = in.precision;
|
||||
out.BoundsCheckFreq = in.BoundsCheckFreq;
|
||||
return out;
|
||||
}
|
||||
|
||||
FermionOperator<Impl> & NumOp;// the basic operator
|
||||
FermionOperator<Impl> & DenOp;// the basic operator
|
||||
FermionField PhiEven; // the pseudo fermion field for this trajectory
|
||||
FermionField PhiOdd; // the pseudo fermion field for this trajectory
|
||||
|
||||
public:
|
||||
OneFlavourEvenOddRatioRationalPseudoFermionAction(FermionOperator<Impl> &_NumOp,
|
||||
FermionOperator<Impl> &_DenOp,
|
||||
const Params & p
|
||||
) :
|
||||
GeneralEvenOddRatioRationalPseudoFermionAction<Impl>(_NumOp, _DenOp, transcribe(p)){}
|
||||
|
||||
virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}
|
||||
OneFlavourEvenOddRatioRationalPseudoFermionAction(FermionOperator<Impl> &_NumOp,
|
||||
FermionOperator<Impl> &_DenOp,
|
||||
Params & p
|
||||
) :
|
||||
NumOp(_NumOp),
|
||||
DenOp(_DenOp),
|
||||
PhiOdd (_NumOp.FermionRedBlackGrid()),
|
||||
PhiEven(_NumOp.FermionRedBlackGrid()),
|
||||
param(p)
|
||||
{
|
||||
AlgRemez remez(param.lo,param.hi,param.precision);
|
||||
|
||||
// MdagM^(+- 1/2)
|
||||
std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/2)"<<std::endl;
|
||||
remez.generateApprox(param.degree,1,2);
|
||||
PowerHalf.Init(remez,param.tolerance,false);
|
||||
PowerNegHalf.Init(remez,param.tolerance,true);
|
||||
|
||||
// MdagM^(+- 1/4)
|
||||
std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/4)"<<std::endl;
|
||||
remez.generateApprox(param.degree,1,4);
|
||||
PowerQuarter.Init(remez,param.tolerance,false);
|
||||
PowerNegQuarter.Init(remez,param.tolerance,true);
|
||||
};
|
||||
|
||||
virtual std::string action_name(){return "OneFlavourEvenOddRatioRationalPseudoFermionAction";}
|
||||
|
||||
virtual std::string LogParameters(){
|
||||
std::stringstream sstream;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Low :" << param.lo << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] High :" << param.hi << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Max iterations :" << param.MaxIter << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Tolerance :" << param.tolerance << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Degree :" << param.degree << std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] Precision :" << param.precision << std::endl;
|
||||
return sstream.str();
|
||||
}
|
||||
|
||||
|
||||
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
|
||||
|
||||
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
|
||||
//
|
||||
// P(phi) = e^{- phi^dag (VdagV)^1/4 (MdagM)^-1/2 (VdagV)^1/4 phi}
|
||||
// = e^{- phi^dag (VdagV)^1/4 (MdagM)^-1/4 (MdagM)^-1/4 (VdagV)^1/4 phi}
|
||||
//
|
||||
// Phi = (VdagV)^-1/4 Mdag^{1/4} eta
|
||||
//
|
||||
// P(eta) = e^{- eta^dag eta}
|
||||
//
|
||||
// e^{x^2/2 sig^2} => sig^2 = 0.5.
|
||||
//
|
||||
// So eta should be of width sig = 1/sqrt(2).
|
||||
|
||||
RealD scale = std::sqrt(0.5);
|
||||
|
||||
FermionField eta(NumOp.FermionGrid());
|
||||
FermionField etaOdd (NumOp.FermionRedBlackGrid());
|
||||
FermionField etaEven(NumOp.FermionRedBlackGrid());
|
||||
FermionField tmp(NumOp.FermionRedBlackGrid());
|
||||
|
||||
gaussian(pRNG,eta); eta=eta*scale;
|
||||
|
||||
pickCheckerboard(Even,etaEven,eta);
|
||||
pickCheckerboard(Odd,etaOdd,eta);
|
||||
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
|
||||
// MdagM^1/4 eta
|
||||
SchurDifferentiableOperator<Impl> MdagM(DenOp);
|
||||
ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerQuarter);
|
||||
msCG_M(MdagM,etaOdd,tmp);
|
||||
|
||||
// VdagV^-1/4 MdagM^1/4 eta
|
||||
SchurDifferentiableOperator<Impl> VdagV(NumOp);
|
||||
ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerNegQuarter);
|
||||
msCG_V(VdagV,tmp,PhiOdd);
|
||||
|
||||
assert(NumOp.ConstEE() == 1);
|
||||
assert(DenOp.ConstEE() == 1);
|
||||
PhiEven = Zero();
|
||||
|
||||
};
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
|
||||
//////////////////////////////////////////////////////
|
||||
virtual RealD S(const GaugeField &U) {
|
||||
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
FermionField X(NumOp.FermionRedBlackGrid());
|
||||
FermionField Y(NumOp.FermionRedBlackGrid());
|
||||
|
||||
// VdagV^1/4 Phi
|
||||
SchurDifferentiableOperator<Impl> VdagV(NumOp);
|
||||
ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
|
||||
msCG_V(VdagV,PhiOdd,X);
|
||||
|
||||
// MdagM^-1/4 VdagV^1/4 Phi
|
||||
SchurDifferentiableOperator<Impl> MdagM(DenOp);
|
||||
ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegQuarter);
|
||||
msCG_M(MdagM,X,Y);
|
||||
|
||||
// Randomly apply rational bounds checks.
|
||||
auto grid = NumOp.FermionGrid();
|
||||
auto r=rand();
|
||||
grid->Broadcast(0,r);
|
||||
if ( (r%param.BoundsCheckFreq)==0 ) {
|
||||
FermionField gauss(NumOp.FermionRedBlackGrid());
|
||||
gauss = PhiOdd;
|
||||
HighBoundCheck(MdagM,gauss,param.hi);
|
||||
InverseSqrtBoundsCheck(param.MaxIter,param.tolerance*100,MdagM,gauss,PowerNegHalf);
|
||||
}
|
||||
|
||||
// Phidag VdagV^1/4 MdagM^-1/4 MdagM^-1/4 VdagV^1/4 Phi
|
||||
RealD action = norm2(Y);
|
||||
|
||||
return action;
|
||||
};
|
||||
|
||||
// S_f = chi^dag* P(V^dag*V)/Q(V^dag*V)* N(M^dag*M)/D(M^dag*M)* P(V^dag*V)/Q(V^dag*V)* chi
|
||||
//
|
||||
// Here, M is some 5D operator and V is the Pauli-Villars field
|
||||
// N and D makeup the rat. poly of the M term and P and & makeup the rat.poly of the denom term
|
||||
//
|
||||
// Need
|
||||
// dS_f/dU = chi^dag d[P/Q] N/D P/Q chi
|
||||
// + chi^dag P/Q d[N/D] P/Q chi
|
||||
// + chi^dag P/Q N/D d[P/Q] chi
|
||||
//
|
||||
// P/Q is expressed as partial fraction expansion:
|
||||
//
|
||||
// a0 + \sum_k ak/(V^dagV + bk)
|
||||
//
|
||||
// d[P/Q] is then
|
||||
//
|
||||
// \sum_k -ak [V^dagV+bk]^{-1} [ dV^dag V + V^dag dV ] [V^dag V + bk]^{-1}
|
||||
//
|
||||
// and similar for N/D.
|
||||
//
|
||||
// Need
|
||||
// MpvPhi_k = [Vdag V + bk]^{-1} chi
|
||||
// MpvPhi = {a0 + \sum_k ak [Vdag V + bk]^{-1} }chi
|
||||
//
|
||||
// MfMpvPhi_k = [MdagM+bk]^{-1} MpvPhi
|
||||
// MfMpvPhi = {a0 + \sum_k ak [Mdag M + bk]^{-1} } MpvPhi
|
||||
//
|
||||
// MpvMfMpvPhi_k = [Vdag V + bk]^{-1} MfMpvchi
|
||||
//
|
||||
|
||||
virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
|
||||
|
||||
const int n_f = PowerNegHalf.poles.size();
|
||||
const int n_pv = PowerQuarter.poles.size();
|
||||
|
||||
std::vector<FermionField> MpvPhi_k (n_pv,NumOp.FermionRedBlackGrid());
|
||||
std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionRedBlackGrid());
|
||||
std::vector<FermionField> MfMpvPhi_k (n_f ,NumOp.FermionRedBlackGrid());
|
||||
|
||||
FermionField MpvPhi(NumOp.FermionRedBlackGrid());
|
||||
FermionField MfMpvPhi(NumOp.FermionRedBlackGrid());
|
||||
FermionField MpvMfMpvPhi(NumOp.FermionRedBlackGrid());
|
||||
FermionField Y(NumOp.FermionRedBlackGrid());
|
||||
|
||||
GaugeField tmp(NumOp.GaugeGrid());
|
||||
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
SchurDifferentiableOperator<Impl> VdagV(NumOp);
|
||||
SchurDifferentiableOperator<Impl> MdagM(DenOp);
|
||||
|
||||
ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
|
||||
ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegHalf);
|
||||
|
||||
msCG_V(VdagV,PhiOdd,MpvPhi_k,MpvPhi);
|
||||
msCG_M(MdagM,MpvPhi,MfMpvPhi_k,MfMpvPhi);
|
||||
msCG_V(VdagV,MfMpvPhi,MpvMfMpvPhi_k,MpvMfMpvPhi);
|
||||
|
||||
RealD ak;
|
||||
|
||||
dSdU = Zero();
|
||||
|
||||
// With these building blocks
|
||||
//
|
||||
// dS/dU =
|
||||
// \sum_k -ak MfMpvPhi_k^dag [ dM^dag M + M^dag dM ] MfMpvPhi_k (1)
|
||||
// + \sum_k -ak MpvMfMpvPhi_k^\dag [ dV^dag V + V^dag dV ] MpvPhi_k (2)
|
||||
// -ak MpvPhi_k^dag [ dV^dag V + V^dag dV ] MpvMfMpvPhi_k (3)
|
||||
|
||||
//(1)
|
||||
for(int k=0;k<n_f;k++){
|
||||
ak = PowerNegHalf.residues[k];
|
||||
MdagM.Mpc(MfMpvPhi_k[k],Y);
|
||||
MdagM.MpcDagDeriv(tmp , MfMpvPhi_k[k], Y ); dSdU=dSdU+ak*tmp;
|
||||
MdagM.MpcDeriv(tmp , Y, MfMpvPhi_k[k] ); dSdU=dSdU+ak*tmp;
|
||||
}
|
||||
|
||||
//(2)
|
||||
//(3)
|
||||
for(int k=0;k<n_pv;k++){
|
||||
|
||||
ak = PowerQuarter.residues[k];
|
||||
|
||||
VdagV.Mpc(MpvPhi_k[k],Y);
|
||||
VdagV.MpcDagDeriv(tmp,MpvMfMpvPhi_k[k],Y); dSdU=dSdU+ak*tmp;
|
||||
VdagV.MpcDeriv (tmp,Y,MpvMfMpvPhi_k[k]); dSdU=dSdU+ak*tmp;
|
||||
|
||||
VdagV.Mpc(MpvMfMpvPhi_k[k],Y); // V as we take Ydag
|
||||
VdagV.MpcDeriv (tmp,Y, MpvPhi_k[k]); dSdU=dSdU+ak*tmp;
|
||||
VdagV.MpcDagDeriv(tmp,MpvPhi_k[k], Y); dSdU=dSdU+ak*tmp;
|
||||
|
||||
}
|
||||
|
||||
//dSdU = Ta(dSdU);
|
||||
|
||||
};
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
@ -49,12 +49,10 @@ NAMESPACE_BEGIN(Grid);
|
||||
Params param;
|
||||
|
||||
MultiShiftFunction PowerHalf ;
|
||||
MultiShiftFunction PowerQuarter;
|
||||
MultiShiftFunction PowerNegHalf;
|
||||
MultiShiftFunction PowerQuarter;
|
||||
MultiShiftFunction PowerNegQuarter;
|
||||
|
||||
MultiShiftFunction MDPowerQuarter;
|
||||
MultiShiftFunction MDPowerNegHalf;
|
||||
private:
|
||||
|
||||
FermionOperator<Impl> & NumOp;// the basic operator
|
||||
@ -75,13 +73,11 @@ NAMESPACE_BEGIN(Grid);
|
||||
remez.generateApprox(param.degree,1,2);
|
||||
PowerHalf.Init(remez,param.tolerance,false);
|
||||
PowerNegHalf.Init(remez,param.tolerance,true);
|
||||
MDPowerNegHalf.Init(remez,param.mdtolerance,true);
|
||||
|
||||
// MdagM^(+- 1/4)
|
||||
std::cout<<GridLogMessage << "Generating degree "<<param.degree<<" for x^(1/4)"<<std::endl;
|
||||
remez.generateApprox(param.degree,1,4);
|
||||
PowerQuarter.Init(remez,param.tolerance,false);
|
||||
MDPowerQuarter.Init(remez,param.mdtolerance,false);
|
||||
PowerNegQuarter.Init(remez,param.tolerance,true);
|
||||
};
|
||||
|
||||
@ -208,8 +204,8 @@ NAMESPACE_BEGIN(Grid);
|
||||
|
||||
virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
|
||||
|
||||
const int n_f = MDPowerNegHalf.poles.size();
|
||||
const int n_pv = MDPowerQuarter.poles.size();
|
||||
const int n_f = PowerNegHalf.poles.size();
|
||||
const int n_pv = PowerQuarter.poles.size();
|
||||
|
||||
std::vector<FermionField> MpvPhi_k (n_pv,NumOp.FermionGrid());
|
||||
std::vector<FermionField> MpvMfMpvPhi_k(n_pv,NumOp.FermionGrid());
|
||||
@ -228,8 +224,8 @@ NAMESPACE_BEGIN(Grid);
|
||||
MdagMLinearOperator<FermionOperator<Impl> ,FermionField> MdagM(DenOp);
|
||||
MdagMLinearOperator<FermionOperator<Impl> ,FermionField> VdagV(NumOp);
|
||||
|
||||
ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,MDPowerQuarter);
|
||||
ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,MDPowerNegHalf);
|
||||
ConjugateGradientMultiShift<FermionField> msCG_V(param.MaxIter,PowerQuarter);
|
||||
ConjugateGradientMultiShift<FermionField> msCG_M(param.MaxIter,PowerNegHalf);
|
||||
|
||||
msCG_V(VdagV,Phi,MpvPhi_k,MpvPhi);
|
||||
msCG_M(MdagM,MpvPhi,MfMpvPhi_k,MfMpvPhi);
|
||||
@ -248,7 +244,7 @@ NAMESPACE_BEGIN(Grid);
|
||||
|
||||
//(1)
|
||||
for(int k=0;k<n_f;k++){
|
||||
ak = MDPowerNegHalf.residues[k];
|
||||
ak = PowerNegHalf.residues[k];
|
||||
DenOp.M(MfMpvPhi_k[k],Y);
|
||||
DenOp.MDeriv(tmp , MfMpvPhi_k[k], Y,DaggerYes ); dSdU=dSdU+ak*tmp;
|
||||
DenOp.MDeriv(tmp , Y, MfMpvPhi_k[k], DaggerNo ); dSdU=dSdU+ak*tmp;
|
||||
@ -258,7 +254,7 @@ NAMESPACE_BEGIN(Grid);
|
||||
//(3)
|
||||
for(int k=0;k<n_pv;k++){
|
||||
|
||||
ak = MDPowerQuarter.residues[k];
|
||||
ak = PowerQuarter.residues[k];
|
||||
|
||||
NumOp.M(MpvPhi_k[k],Y);
|
||||
NumOp.MDeriv(tmp,MpvMfMpvPhi_k[k],Y,DaggerYes); dSdU=dSdU+ak*tmp;
|
||||
|
||||
@ -40,8 +40,6 @@ directory
|
||||
#include <Grid/qcd/action/pseudofermion/OneFlavourRational.h>
|
||||
#include <Grid/qcd/action/pseudofermion/OneFlavourRationalRatio.h>
|
||||
#include <Grid/qcd/action/pseudofermion/OneFlavourEvenOddRational.h>
|
||||
#include <Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatio.h>
|
||||
#include <Grid/qcd/action/pseudofermion/GeneralEvenOddRationalRatioMixedPrec.h>
|
||||
#include <Grid/qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h>
|
||||
#include <Grid/qcd/action/pseudofermion/ExactOneFlavourRatio.h>
|
||||
|
||||
|
||||
@ -75,22 +75,24 @@ NAMESPACE_BEGIN(Grid);
|
||||
conformable(_NumOp.GaugeRedBlackGrid(), _DenOp.GaugeRedBlackGrid());
|
||||
};
|
||||
|
||||
virtual std::string action_name(){
|
||||
std::stringstream sstream;
|
||||
sstream<<"TwoFlavourEvenOddRatioPseudoFermionAction det("<<DenOp.Mass()<<") / det("<<NumOp.Mass()<<")";
|
||||
return sstream.str();
|
||||
}
|
||||
virtual std::string action_name(){return "TwoFlavourEvenOddRatioPseudoFermionAction";}
|
||||
|
||||
virtual std::string LogParameters(){
|
||||
std::stringstream sstream;
|
||||
sstream<< GridLogMessage << "["<<action_name()<<"] -- No further parameters "<<std::endl;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] has no parameters" << std::endl;
|
||||
return sstream.str();
|
||||
}
|
||||
|
||||
|
||||
const FermionField &getPhiOdd() const{ return PhiOdd; }
|
||||
|
||||
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
|
||||
|
||||
// P(phi) = e^{- phi^dag Vpc (MpcdagMpc)^-1 Vpcdag phi}
|
||||
//
|
||||
// NumOp == V
|
||||
// DenOp == M
|
||||
//
|
||||
// Take phi_o = Vpcdag^{-1} Mpcdag eta_o ; eta_o = Mpcdag^{-1} Vpcdag Phi
|
||||
//
|
||||
// P(eta_o) = e^{- eta_o^dag eta_o}
|
||||
//
|
||||
// e^{x^2/2 sig^2} => sig^2 = 0.5.
|
||||
@ -98,22 +100,12 @@ NAMESPACE_BEGIN(Grid);
|
||||
RealD scale = std::sqrt(0.5);
|
||||
|
||||
FermionField eta (NumOp.FermionGrid());
|
||||
gaussian(pRNG,eta); eta = eta * scale;
|
||||
|
||||
refresh(U,eta);
|
||||
}
|
||||
|
||||
void refresh(const GaugeField &U, const FermionField &eta) {
|
||||
|
||||
// P(phi) = e^{- phi^dag Vpc (MpcdagMpc)^-1 Vpcdag phi}
|
||||
//
|
||||
// NumOp == V
|
||||
// DenOp == M
|
||||
//
|
||||
FermionField etaOdd (NumOp.FermionRedBlackGrid());
|
||||
FermionField etaEven(NumOp.FermionRedBlackGrid());
|
||||
FermionField tmp (NumOp.FermionRedBlackGrid());
|
||||
|
||||
gaussian(pRNG,eta);
|
||||
|
||||
pickCheckerboard(Even,etaEven,eta);
|
||||
pickCheckerboard(Odd,etaOdd,eta);
|
||||
|
||||
@ -132,6 +124,10 @@ NAMESPACE_BEGIN(Grid);
|
||||
// Even det factors
|
||||
DenOp.MooeeDag(etaEven,tmp);
|
||||
NumOp.MooeeInvDag(tmp,PhiEven);
|
||||
|
||||
PhiOdd =PhiOdd*scale;
|
||||
PhiEven=PhiEven*scale;
|
||||
|
||||
};
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
|
||||
@ -1,203 +0,0 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/qcd/action/pseudofermion/TwoFlavourRatio.h
|
||||
|
||||
Copyright (C) 2015
|
||||
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
|
||||
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
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
///////////////////////////////////////
|
||||
// Two flavour ratio
|
||||
///////////////////////////////////////
|
||||
template<class Impl>
|
||||
class TwoFlavourRatioEO4DPseudoFermionAction : public Action<typename Impl::GaugeField> {
|
||||
public:
|
||||
INHERIT_IMPL_TYPES(Impl);
|
||||
|
||||
private:
|
||||
typedef FermionOperator<Impl> FermOp;
|
||||
FermionOperator<Impl> & NumOp;// the basic operator
|
||||
FermionOperator<Impl> & DenOp;// the basic operator
|
||||
|
||||
OperatorFunction<FermionField> &DerivativeSolver;
|
||||
OperatorFunction<FermionField> &DerivativeDagSolver;
|
||||
OperatorFunction<FermionField> &ActionSolver;
|
||||
OperatorFunction<FermionField> &HeatbathSolver;
|
||||
|
||||
FermionField phi4; // the pseudo fermion field for this trajectory
|
||||
|
||||
public:
|
||||
TwoFlavourRatioEO4DPseudoFermionAction(FermionOperator<Impl> &_NumOp,
|
||||
FermionOperator<Impl> &_DenOp,
|
||||
OperatorFunction<FermionField> & DS,
|
||||
OperatorFunction<FermionField> & AS ) :
|
||||
TwoFlavourRatioEO4DPseudoFermionAction(_NumOp,_DenOp, DS,DS,AS,AS) {};
|
||||
TwoFlavourRatioEO4DPseudoFermionAction(FermionOperator<Impl> &_NumOp,
|
||||
FermionOperator<Impl> &_DenOp,
|
||||
OperatorFunction<FermionField> & DS,
|
||||
OperatorFunction<FermionField> & DDS,
|
||||
OperatorFunction<FermionField> & AS,
|
||||
OperatorFunction<FermionField> & HS
|
||||
) : NumOp(_NumOp),
|
||||
DenOp(_DenOp),
|
||||
DerivativeSolver(DS),
|
||||
DerivativeDagSolver(DDS),
|
||||
ActionSolver(AS),
|
||||
HeatbathSolver(HS),
|
||||
phi4(_NumOp.GaugeGrid())
|
||||
{};
|
||||
|
||||
virtual std::string action_name(){return "TwoFlavourRatioEO4DPseudoFermionAction";}
|
||||
|
||||
virtual std::string LogParameters(){
|
||||
std::stringstream sstream;
|
||||
sstream << GridLogMessage << "["<<action_name()<<"] has no parameters" << std::endl;
|
||||
return sstream.str();
|
||||
}
|
||||
|
||||
virtual void refresh(const GaugeField &U, GridSerialRNG &sRNG, GridParallelRNG& pRNG) {
|
||||
|
||||
// P(phi) = e^{- phi^dag (V^dag M^-dag)_11 (M^-1 V)_11 phi}
|
||||
//
|
||||
// NumOp == V
|
||||
// DenOp == M
|
||||
//
|
||||
// Take phi = (V^{-1} M)_11 eta ; eta = (M^{-1} V)_11 Phi
|
||||
//
|
||||
// P(eta) = e^{- eta^dag eta}
|
||||
//
|
||||
// e^{x^2/2 sig^2} => sig^2 = 0.5.
|
||||
//
|
||||
// So eta should be of width sig = 1/sqrt(2) and must multiply by 0.707....
|
||||
//
|
||||
RealD scale = std::sqrt(0.5);
|
||||
|
||||
FermionField eta4(NumOp.GaugeGrid());
|
||||
FermionField eta5(NumOp.FermionGrid());
|
||||
FermionField tmp(NumOp.FermionGrid());
|
||||
FermionField phi5(NumOp.FermionGrid());
|
||||
|
||||
gaussian(pRNG,eta4);
|
||||
NumOp.ImportFourDimPseudoFermion(eta4,eta5);
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(HeatbathSolver);
|
||||
|
||||
DenOp.M(eta5,tmp); // M eta
|
||||
PrecSolve(NumOp,tmp,phi5); // phi = V^-1 M eta
|
||||
phi5=phi5*scale;
|
||||
std::cout << GridLogMessage << "4d pf refresh "<< norm2(phi5)<<"\n";
|
||||
// Project to 4d
|
||||
NumOp.ExportFourDimPseudoFermion(phi5,phi4);
|
||||
|
||||
};
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
// S = phi^dag (V^dag M^-dag)_11 (M^-1 V)_11 phi
|
||||
//////////////////////////////////////////////////////
|
||||
virtual RealD S(const GaugeField &U) {
|
||||
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
FermionField Y4(NumOp.GaugeGrid());
|
||||
FermionField X(NumOp.FermionGrid());
|
||||
FermionField Y(NumOp.FermionGrid());
|
||||
FermionField phi5(NumOp.FermionGrid());
|
||||
|
||||
MdagMLinearOperator<FermionOperator<Impl> ,FermionField> MdagMOp(DenOp);
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(ActionSolver);
|
||||
|
||||
NumOp.ImportFourDimPseudoFermion(phi4,phi5);
|
||||
NumOp.M(phi5,X); // X= V phi
|
||||
PrecSolve(DenOp,X,Y); // Y= (MdagM)^-1 Mdag Vdag phi = M^-1 V phi
|
||||
NumOp.ExportFourDimPseudoFermion(Y,Y4);
|
||||
|
||||
RealD action = norm2(Y4);
|
||||
|
||||
return action;
|
||||
};
|
||||
|
||||
//////////////////////////////////////////////////////
|
||||
// dS/du = 2 Re phi^dag (V^dag M^-dag)_11 (M^-1 d V)_11 phi
|
||||
// - 2 Re phi^dag (dV^dag M^-dag)_11 (M^-1 dM M^-1 V)_11 phi
|
||||
//////////////////////////////////////////////////////
|
||||
virtual void deriv(const GaugeField &U,GaugeField & dSdU) {
|
||||
|
||||
NumOp.ImportGauge(U);
|
||||
DenOp.ImportGauge(U);
|
||||
|
||||
FermionField X(NumOp.FermionGrid());
|
||||
FermionField Y(NumOp.FermionGrid());
|
||||
FermionField phi(NumOp.FermionGrid());
|
||||
FermionField Vphi(NumOp.FermionGrid());
|
||||
FermionField MinvVphi(NumOp.FermionGrid());
|
||||
FermionField tmp4(NumOp.GaugeGrid());
|
||||
FermionField MdagInvMinvVphi(NumOp.FermionGrid());
|
||||
|
||||
GaugeField force(NumOp.GaugeGrid());
|
||||
|
||||
//Y=V phi
|
||||
//X = (Mdag V phi
|
||||
//Y = (Mdag M)^-1 Mdag V phi = M^-1 V Phi
|
||||
NumOp.ImportFourDimPseudoFermion(phi4,phi);
|
||||
NumOp.M(phi,Vphi); // V phi
|
||||
SchurRedBlackDiagMooeeSolve<FermionField> PrecSolve(DerivativeSolver);
|
||||
PrecSolve(DenOp,Vphi,MinvVphi);// M^-1 V phi
|
||||
std::cout << GridLogMessage << "4d deriv solve "<< norm2(MinvVphi)<<"\n";
|
||||
|
||||
// Projects onto the physical space and back
|
||||
NumOp.ExportFourDimPseudoFermion(MinvVphi,tmp4);
|
||||
NumOp.ImportFourDimPseudoFermion(tmp4,Y);
|
||||
|
||||
SchurRedBlackDiagMooeeDagSolve<FermionField> PrecDagSolve(DerivativeDagSolver);
|
||||
// X = proj M^-dag V phi
|
||||
// Need an adjoint solve
|
||||
PrecDagSolve(DenOp,Y,MdagInvMinvVphi);
|
||||
std::cout << GridLogMessage << "4d deriv solve dag "<< norm2(MdagInvMinvVphi)<<"\n";
|
||||
|
||||
// phi^dag (Vdag Mdag^-1) (M^-1 dV) phi
|
||||
NumOp.MDeriv(force ,MdagInvMinvVphi , phi, DaggerNo ); dSdU=force;
|
||||
|
||||
// phi^dag (dVdag Mdag^-1) (M^-1 V) phi
|
||||
NumOp.MDeriv(force , phi, MdagInvMinvVphi ,DaggerYes ); dSdU=dSdU+force;
|
||||
|
||||
// - 2 Re phi^dag (dV^dag M^-dag)_11 (M^-1 dM M^-1 V)_11 phi
|
||||
DenOp.MDeriv(force,MdagInvMinvVphi,MinvVphi,DaggerNo); dSdU=dSdU-force;
|
||||
DenOp.MDeriv(force,MinvVphi,MdagInvMinvVphi,DaggerYes); dSdU=dSdU-force;
|
||||
|
||||
dSdU *= -1.0;
|
||||
//dSdU = - Ta(dSdU);
|
||||
|
||||
};
|
||||
};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
|
||||
|
||||
@ -1,6 +0,0 @@
|
||||
#ifndef GRID_GPARITY_H_
|
||||
#define GRID_GPARITY_H_
|
||||
|
||||
#include<Grid/qcd/gparity/GparityFlavour.h>
|
||||
|
||||
#endif
|
||||
@ -1,34 +0,0 @@
|
||||
#include <Grid/Grid.h>
|
||||
|
||||
NAMESPACE_BEGIN(Grid);
|
||||
|
||||
const std::array<const GparityFlavour, 3> GparityFlavour::sigma_mu = {{
|
||||
GparityFlavour(GparityFlavour::Algebra::SigmaX),
|
||||
GparityFlavour(GparityFlavour::Algebra::SigmaY),
|
||||
GparityFlavour(GparityFlavour::Algebra::SigmaZ)
|
||||
}};
|
||||
|
||||
const std::array<const GparityFlavour, 6> GparityFlavour::sigma_all = {{
|
||||
GparityFlavour(GparityFlavour::Algebra::Identity),
|
||||
GparityFlavour(GparityFlavour::Algebra::SigmaX),
|
||||
GparityFlavour(GparityFlavour::Algebra::SigmaY),
|
||||
GparityFlavour(GparityFlavour::Algebra::SigmaZ),
|
||||
GparityFlavour(GparityFlavour::Algebra::ProjPlus),
|
||||
GparityFlavour(GparityFlavour::Algebra::ProjMinus)
|
||||
}};
|
||||
|
||||
const std::array<const char *, GparityFlavour::nSigma> GparityFlavour::name = {{
|
||||
"SigmaX",
|
||||
"MinusSigmaX",
|
||||
"SigmaY",
|
||||
"MinusSigmaY",
|
||||
"SigmaZ",
|
||||
"MinusSigmaZ",
|
||||
"Identity",
|
||||
"MinusIdentity",
|
||||
"ProjPlus",
|
||||
"MinusProjPlus",
|
||||
"ProjMinus",
|
||||
"MinusProjMinus"}};
|
||||
|
||||
NAMESPACE_END(Grid);
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user