mirror of
https://github.com/paboyle/Grid.git
synced 2025-06-17 07:17:06 +01:00
Improvements
This commit is contained in:
Peter Boyle
@ -131,6 +131,7 @@ public:
|
||||
void OpDirAll (const Field &in, std::vector<Field> &out) { assert(0); };
|
||||
void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ assert(0); }
|
||||
};
|
||||
/*
|
||||
template<class Field> class ChebyshevSmoother : public LinearFunction<Field>
|
||||
{
|
||||
public:
|
||||
@ -151,7 +152,7 @@ public:
|
||||
Cheby(_SmootherOperator,tmp,out);
|
||||
}
|
||||
};
|
||||
|
||||
*/
|
||||
template<class Field> class CGSmoother : public LinearFunction<Field>
|
||||
{
|
||||
public:
|
||||
@ -179,8 +180,6 @@ int main (int argc, char ** argv)
|
||||
|
||||
const int Ls=24;
|
||||
const int nbasis = 62;
|
||||
// const int nbasis = 56;
|
||||
// const int nbasis = 44;
|
||||
const int cb = 0 ;
|
||||
RealD mass=0.00078;
|
||||
RealD M5=1.8;
|
||||
@ -357,31 +356,17 @@ slurm-1482367.out:Grid : Message : 6169.469330 s : HDCG: Pcg converged in 487 it
|
||||
CoarseVector c_res(Coarse5d);
|
||||
CoarseVector c_ref(Coarse5d);
|
||||
|
||||
// Try projecting to one hop only
|
||||
// LittleDiracOp.ShiftMatrix(1.0e-4);
|
||||
LittleDiracOperator LittleDiracOpProj(geom_nn,FrbGrid,Coarse5d);
|
||||
LittleDiracOpProj.ProjectNearestNeighbour(0.01,LittleDiracOp); // smaller shift 0.02? n
|
||||
|
||||
typedef HermitianLinearOperator<LittleDiracOperator,CoarseVector> HermMatrix;
|
||||
HermMatrix CoarseOp (LittleDiracOp);
|
||||
HermMatrix CoarseOpProj (LittleDiracOpProj);
|
||||
|
||||
//////////////////////////////////////////
|
||||
// Build a coarse lanczos
|
||||
//////////////////////////////////////////
|
||||
// Chebyshev<CoarseVector> IRLCheby(0.012,40.0,201); //500 HDCG iters
|
||||
// int Nk=512; // Didn't save much
|
||||
// int Nm=640;
|
||||
// int Nstop=400;
|
||||
|
||||
// Chebyshev<CoarseVector> IRLCheby(0.005,40.0,201); //319 HDCG iters @ 128//160 nk.
|
||||
// int Nk=128;
|
||||
// int Nm=160;
|
||||
Chebyshev<CoarseVector> IRLCheby(0.005,40.0,201); //319 HDCG iters @ 128//160 nk.
|
||||
typedef HermitianLinearOperator<LittleDiracOperator,CoarseVector> HermMatrix;
|
||||
HermMatrix CoarseOp (LittleDiracOp);
|
||||
|
||||
int Nk=192;
|
||||
int Nm=256;
|
||||
int Nstop=Nk;
|
||||
|
||||
Chebyshev<CoarseVector> IRLCheby(0.005,40.0,201);
|
||||
// Chebyshev<CoarseVector> IRLCheby(0.010,45.0,201); // 1 iter
|
||||
FunctionHermOp<CoarseVector> IRLOpCheby(IRLCheby,CoarseOp);
|
||||
PlainHermOp<CoarseVector> IRLOp (CoarseOp);
|
||||
@ -395,208 +380,40 @@ slurm-1482367.out:Grid : Message : 6169.469330 s : HDCG: Pcg converged in 487 it
|
||||
PowerMethod<CoarseVector> cPM; cPM(CoarseOp,c_src);
|
||||
|
||||
IRL.calc(eval,evec,c_src,Nconv);
|
||||
DeflatedGuesser<CoarseVector> DeflCoarseGuesser(evec,eval);
|
||||
|
||||
//////////////////////////////////////////
|
||||
// Build a coarse space solver
|
||||
// Deflated guesser
|
||||
//////////////////////////////////////////
|
||||
DeflatedGuesser<CoarseVector> DeflCoarseGuesser(evec,eval);
|
||||
|
||||
int maxit=30000;
|
||||
ConjugateGradient<CoarseVector> CG(1.0e-10,maxit,false);
|
||||
ConjugateGradient<LatticeFermionD> CGfine(1.0e-8,30000,false);
|
||||
ZeroGuesser<CoarseVector> CoarseZeroGuesser;
|
||||
|
||||
// HPDSolver<CoarseVector> HPDSolve(CoarseOp,CG,CoarseZeroGuesser);
|
||||
HPDSolver<CoarseVector> HPDSolve(CoarseOp,CG,DeflCoarseGuesser);
|
||||
c_res=Zero();
|
||||
// HPDSolve(c_src,c_res); c_ref = c_res;
|
||||
// std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
|
||||
// std::cout << GridLogMessage<<"ref norm "<<norm2(c_ref)<<std::endl;
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
// Deflated (with real op EV's) solve for the projected coarse op
|
||||
// Work towards ADEF1 in the coarse space
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
HPDSolver<CoarseVector> HPDSolveProj(CoarseOpProj,CG,DeflCoarseGuesser);
|
||||
c_res=Zero();
|
||||
// HPDSolveProj(c_src,c_res);
|
||||
// std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
|
||||
// std::cout << GridLogMessage<<"res norm "<<norm2(c_res)<<std::endl;
|
||||
// c_res = c_res - c_ref;
|
||||
// std::cout << "Projected solver error "<<norm2(c_res)<<std::endl;
|
||||
|
||||
//////////////////////////////////////////////////////////////////////
|
||||
// Coarse ADEF1 with deflation space
|
||||
//////////////////////////////////////////////////////////////////////
|
||||
ChebyshevSmoother<CoarseVector > CoarseSmoother(1.0,37.,8,CoarseOpProj); // just go to sloppy 0.1 convergence
|
||||
// CoarseSmoother(0.1,37.,8,CoarseOpProj); //
|
||||
// CoarseSmoother(0.5,37.,6,CoarseOpProj); // 8 iter 0.36s
|
||||
// CoarseSmoother(0.5,37.,12,CoarseOpProj); // 8 iter, 0.55s
|
||||
// CoarseSmoother(0.5,37.,8,CoarseOpProj);// 7-9 iter
|
||||
// CoarseSmoother(1.0,37.,8,CoarseOpProj); // 0.4 - 0.5s solve to 0.04, 7-9 iter
|
||||
// ChebyshevSmoother<CoarseVector,HermMatrix > CoarseSmoother(0.5,36.,10,CoarseOpProj); // 311
|
||||
|
||||
////////////////////////////////////////////////////////
|
||||
// CG, Cheby mode spacing 200,200
|
||||
// Unprojected Coarse CG solve to 1e-8 : 190 iters, 4.9s
|
||||
// Unprojected Coarse CG solve to 4e-2 : 33 iters, 0.8s
|
||||
// Projected Coarse CG solve to 1e-8 : 100 iters, 0.36s
|
||||
////////////////////////////////////////////////////////
|
||||
// CoarseSmoother(1.0,48.,8,CoarseOpProj); 48 evecs
|
||||
////////////////////////////////////////////////////////
|
||||
// ADEF1 Coarse solve to 1e-8 : 44 iters, 2.34s 2.1x gain
|
||||
// ADEF1 Coarse solve to 4e-2 : 7 iters, 0.4s
|
||||
// HDCG 38 iters 162s
|
||||
//
|
||||
// CoarseSmoother(1.0,40.,8,CoarseOpProj); 48 evecs
|
||||
// ADEF1 Coarse solve to 1e-8 : 37 iters, 2.0s 2.1x gain
|
||||
// ADEF1 Coarse solve to 4e-2 : 6 iters, 0.36s
|
||||
// HDCG 38 iters 169s
|
||||
|
||||
TwoLevelADEF1defl<CoarseVector>
|
||||
cADEF1(1.0e-8, 500,
|
||||
CoarseOp,
|
||||
CoarseSmoother,
|
||||
evec,eval);
|
||||
|
||||
// c_res=Zero();
|
||||
// cADEF1(c_src,c_res);
|
||||
// std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
|
||||
// std::cout << GridLogMessage<<"cADEF1 res norm "<<norm2(c_res)<<std::endl;
|
||||
// c_res = c_res - c_ref;
|
||||
// std::cout << "cADEF1 solver error "<<norm2(c_res)<<std::endl;
|
||||
|
||||
// cADEF1.Tolerance = 4.0e-2;
|
||||
// cADEF1.Tolerance = 1.0e-1;
|
||||
// cADEF1.Tolerance = 5.0e-2;
|
||||
// c_res=Zero();
|
||||
// cADEF1(c_src,c_res);
|
||||
// std::cout << GridLogMessage<<"src norm "<<norm2(c_src)<<std::endl;
|
||||
// std::cout << GridLogMessage<<"cADEF1 res norm "<<norm2(c_res)<<std::endl;
|
||||
// c_res = c_res - c_ref;
|
||||
// std::cout << "cADEF1 solver error "<<norm2(c_res)<<std::endl;
|
||||
|
||||
//////////////////////////////////////////
|
||||
// Build a smoother
|
||||
// HDCG
|
||||
//////////////////////////////////////////
|
||||
// ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(10.0,100.0,10,FineHermOp); //499
|
||||
// ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(3.0,100.0,10,FineHermOp); //383
|
||||
// ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(1.0,100.0,10,FineHermOp); //328
|
||||
// std::vector<RealD> los({0.5,1.0,3.0}); // 147/142/146 nbasis 1
|
||||
// std::vector<RealD> los({1.0,2.0}); // Nbasis 24: 88,86 iterations
|
||||
// std::vector<RealD> los({2.0,4.0}); // Nbasis 32 == 52, iters
|
||||
// std::vector<RealD> los({2.0,4.0}); // Nbasis 40 == 36,36 iters
|
||||
|
||||
//
|
||||
// Turns approx 2700 iterations into 340 fine multiplies with Nbasis 40
|
||||
// Need to measure cost of coarse space.
|
||||
//
|
||||
// -- i) Reduce coarse residual -- 0.04
|
||||
// -- ii) Lanczos on coarse space -- done
|
||||
// -- iii) Possible 1 hop project and/or preconditioning it - easy - PrecCG it and
|
||||
// use a limited stencil. Reread BFM code to check on evecs / deflation strategy with prec
|
||||
//
|
||||
//
|
||||
//
|
||||
//
|
||||
|
||||
std::vector<RealD> los({2.0,2.5}); // Nbasis 40 == 36,36 iters
|
||||
|
||||
// std::vector<int> ords({7,8,10}); // Nbasis 40 == 40,38,36 iters (320,342,396 mults)
|
||||
// std::vector<int> ords({7}); // Nbasis 40 == 40 iters (320 mults)
|
||||
std::vector<int> ords({9}); // Nbasis 40 == 40 iters (320 mults)
|
||||
|
||||
/*
|
||||
Smoother opt @56 nbasis, 0.04 convergence, 192 evs
|
||||
ord lo
|
||||
|
||||
16 0.1 no converge -- likely sign indefinite
|
||||
32 0.1 no converge -- likely sign indefinite(?)
|
||||
|
||||
16 0.5 422
|
||||
32 0.5 302
|
||||
|
||||
8 1.0 575
|
||||
12 1.0 449
|
||||
16 1.0 375
|
||||
32 1.0 302
|
||||
|
||||
12 3.0 476
|
||||
16 3.0 319
|
||||
32 3.0 306
|
||||
|
||||
Powerlaw setup 62 vecs
|
||||
slurm-1494943.out:Grid : Message : 4874.186617 s : HDCG: Pcg converged in 171 iterations and 1706.548006 s 1.0 32
|
||||
slurm-1494943.out:Grid : Message : 6490.121648 s : HDCG: Pcg converged in 194 iterations and 1616.219654 s 1.0 16
|
||||
|
||||
Cheby setup: 56vecs
|
||||
-- CG smoother O(16): 487
|
||||
|
||||
Power law setup, 56 vecs -- lambda^-5
|
||||
slurm-1494383.out:Grid : Message : 4377.173265 s : HDCG: Pcg converged in 204 iterations and 1153.548935 s 1.0 32
|
||||
|
||||
Power law setup, 56 vecs -- lambda^-3
|
||||
|
||||
slurm-1494242.out:Grid : Message : 4370.464814 s : HDCG: Pcg converged in 204 iterations and 1143.494776 s 1.0 32
|
||||
slurm-1494242.out:Grid : Message : 5432.414820 s : HDCG: Pcg converged in 237 iterations and 1061.455882 s 1.0 16
|
||||
slurm-1494242.out:Grid : Message : 6588.727977 s : HDCG: Pcg converged in 205 iterations and 1156.565210 s 0.5 32
|
||||
|
||||
Power law setup, 56 vecs -- lambda^-4
|
||||
-- CG smoother O(16): 290
|
||||
-- Cheby smoother O(16): 218 -- getting close to the deflation level I expect 169 from BFM paper @O(7) smoother and 64 nbasis
|
||||
|
||||
Grid : Message : 2790.797194 s : HDCG: Pcg converged in 190 iterations and 1049.563182 s 1.0 32
|
||||
Grid : Message : 3766.374396 s : HDCG: Pcg converged in 218 iterations and 975.455668 s 1.0 16
|
||||
Grid : Message : 4888.746190 s : HDCG: Pcg converged in 191 iterations and 1122.252055 s 0.5 32
|
||||
Grid : Message : 5956.679661 s : HDCG: Pcg converged in 231 iterations and 1067.812850 s 0.5 16
|
||||
|
||||
Grid : Message : 2767.405829 s : HDCG: Pcg converged in 218 iterations and 967.214067 s -- 16
|
||||
Grid : Message : 3816.165905 s : HDCG: Pcg converged in 251 iterations and 1048.636269 s -- 12
|
||||
Grid : Message : 5121.206572 s : HDCG: Pcg converged in 318 iterations and 1304.916168 s -- 8
|
||||
|
||||
|
||||
[paboyle@login2.crusher debug]$ grep -v Memory slurm-402426.out | grep converged | grep HDCG -- [1.0,16] cheby
|
||||
Grid : Message : 5185.521063 s : HDCG: Pcg converged in 377 iterations and 1595.843529 s
|
||||
|
||||
[paboyle@login2.crusher debug]$ grep HDCG slurm-402184.out | grep onver
|
||||
Grid : Message : 3760.438160 s : HDCG: Pcg converged in 422 iterations and 2129.243141 s
|
||||
Grid : Message : 5660.588015 s : HDCG: Pcg converged in 308 iterations and 1900.026821 s
|
||||
|
||||
|
||||
Grid : Message : 4238.206528 s : HDCG: Pcg converged in 575 iterations and 2657.430676 s
|
||||
Grid : Message : 6345.880344 s : HDCG: Pcg converged in 449 iterations and 2108.505208 s
|
||||
|
||||
grep onverg slurm-401663.out | grep HDCG
|
||||
Grid : Message : 3900.817781 s : HDCG: Pcg converged in 476 iterations and 1992.591311 s
|
||||
Grid : Message : 5647.202699 s : HDCG: Pcg converged in 306 iterations and 1746.838660 s
|
||||
|
||||
|
||||
[paboyle@login2.crusher debug]$ grep converged slurm-401775.out | grep HDCG
|
||||
Grid : Message : 3583.177025 s : HDCG: Pcg converged in 375 iterations and 1800.896037 s
|
||||
Grid : Message : 5348.342243 s : HDCG: Pcg converged in 302 iterations and 1765.045018 s
|
||||
|
||||
Conclusion: higher order smoother is doing better. Much better. Use a Krylov smoother instead Mirs as in BFM version.
|
||||
|
||||
*/
|
||||
//
|
||||
for(int l=0;l<los.size();l++){
|
||||
|
||||
RealD lo = los[l];
|
||||
|
||||
for(int o=0;o<ords.size();o++){
|
||||
|
||||
//////////////////////////////////////////
|
||||
// Sloppy coarse solve
|
||||
//////////////////////////////////////////
|
||||
|
||||
ConjugateGradient<CoarseVector> CGsloppy(4.0e-2,maxit,false);
|
||||
HPDSolver<CoarseVector> HPDSolveSloppy(CoarseOp,CGsloppy,DeflCoarseGuesser);
|
||||
|
||||
// ChebyshevSmoother<LatticeFermionD,HermFineMatrix > Smoother(lo,92,10,FineHermOp); // 36 best case
|
||||
ChebyshevSmoother<LatticeFermionD > ChebySmooth(lo,95,ords[o],FineHermOp); // 311
|
||||
HPDSolver<CoarseVector> HPDSolve(CoarseOp,CG,DeflCoarseGuesser);
|
||||
|
||||
/*
|
||||
* CG smooth 11 iter:
|
||||
slurm-403825.out:Grid : Message : 4369.824339 s : HDCG: fPcg converged in 215 iterations 3.0
|
||||
slurm-403908.out:Grid : Message : 3955.897470 s : HDCG: fPcg converged in 236 iterations 1.0
|
||||
slurm-404273.out:Grid : Message : 3843.792191 s : HDCG: fPcg converged in 210 iterations 2.0
|
||||
* CG smooth 9 iter:
|
||||
*/
|
||||
//
|
||||
//////////////////////////////////////////
|
||||
// IRS shifted smoother based on CG
|
||||
//////////////////////////////////////////
|
||||
RealD MirsShift = lo;
|
||||
ShiftedHermOpLinearOperator<LatticeFermionD> ShiftedFineHermOp(HermOpEO,MirsShift);
|
||||
CGSmoother<LatticeFermionD> CGsmooth(ords[o],ShiftedFineHermOp) ;
|
||||
@ -607,25 +424,11 @@ Conclusion: higher order smoother is doing better. Much better. Use a Krylov smo
|
||||
TwoLevelADEF2<LatticeFermion,CoarseVector,Subspace>
|
||||
HDCG(1.0e-8, 700,
|
||||
FineHermOp,
|
||||
// ChebySmooth,
|
||||
CGsmooth,
|
||||
HPDSolveSloppy,
|
||||
HPDSolve,
|
||||
Aggregates);
|
||||
|
||||
/*
|
||||
TwoLevelADEF2<LatticeFermion,CoarseVector,Subspace>
|
||||
HDCGdefl(1.0e-8, 700,
|
||||
FineHermOp,
|
||||
Smoother,
|
||||
cADEF1,
|
||||
HPDSolve,
|
||||
Aggregates);
|
||||
*/
|
||||
|
||||
// result=Zero();
|
||||
// HDCGdefl(src,result);
|
||||
|
||||
result=Zero();
|
||||
HDCG(src,result);
|
||||
|
||||
|
||||
Reference in New Issue
Block a user