portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-09 23:45:36 +00:00
Code Releases Activity

Further prec-change improvements

Browse Source 
Mixed prec CG algorithm has been modified to precompute precision change workspaces

As the original Test_dwf_mixedcg_prec has been coopted to do a performance stability and reproducibility test, requiring the single-prec CG to be run 200 times, I have created a new version of Test_dwf_mixedcg_prec in the solver subdirectory that just does the mixed vs double CG test
This commit is contained in:
Christopher Kelly 2023-02-23 09:45:29 -05:00
parent 1db58a8acc
commit e82cf1d311
2 changed files with 128 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
Grid/algorithms/iterative/ConjugateGradientMixedPrec.h
View File

@ -109,6 +109,9 @@ NAMESPACE_BEGIN(Grid);
Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
precisionChangeWorkspace pc_wk_sp_to_dp(DoublePrecGrid, SinglePrecGrid);
precisionChangeWorkspace pc_wk_dp_to_sp(SinglePrecGrid, DoublePrecGrid);
for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){ for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
//Compute double precision rsd and also new RHS vector. //Compute double precision rsd and also new RHS vector.
Linop_d.HermOp(sol_d, tmp_d); Linop_d.HermOp(sol_d, tmp_d);
@ -123,7 +126,7 @@ NAMESPACE_BEGIN(Grid);
while(norm * inner_tol * inner_tol < stop) inner_tol *= 2; // inner_tol = sqrt(stop/norm) ?? while(norm * inner_tol * inner_tol < stop) inner_tol *= 2; // inner_tol = sqrt(stop/norm) ??
PrecChangeTimer.Start(); PrecChangeTimer.Start();
precisionChange(src_f, src_d); precisionChange(src_f, src_d, pc_wk_dp_to_sp);
PrecChangeTimer.Stop(); PrecChangeTimer.Stop();
sol_f = Zero(); sol_f = Zero();
@ -142,7 +145,7 @@ NAMESPACE_BEGIN(Grid);
//Convert sol back to double and add to double prec solution //Convert sol back to double and add to double prec solution
PrecChangeTimer.Start(); PrecChangeTimer.Start();
precisionChange(tmp_d, sol_f); precisionChange(tmp_d, sol_f, pc_wk_sp_to_dp);
PrecChangeTimer.Stop(); PrecChangeTimer.Stop();
axpy(sol_d, 1.0, tmp_d, sol_d); axpy(sol_d, 1.0, tmp_d, sol_d);

122
tests/solver/Test_dwf_mixedcg_prec.cc Normal file
View File

@ -0,0 +1,122 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./tests/Test_dwf_cg_prec.cc
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 */
#include <Grid/Grid.h>
//using namespace std;
using namespace Grid;
int main (int argc, char ** argv)
{
Grid_init(&argc,&argv);
const int Ls=12;
std::cout << GridLogMessage << "::::: NB: to enable a quick bit reproducibility check use the --checksums flag. " << std::endl;
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
GridCartesian * UGrid_f = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid_f = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid_f);
GridCartesian * FGrid_f = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid_f);
GridRedBlackCartesian * FrbGrid_f = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid_f);
std::vector<int> seeds4({1,2,3,4});
std::vector<int> seeds5({5,6,7,8});
GridParallelRNG RNG5(FGrid); RNG5.SeedFixedIntegers(seeds5);
GridParallelRNG RNG4(UGrid); RNG4.SeedFixedIntegers(seeds4);
LatticeFermionD src(FGrid); random(RNG5,src);
LatticeFermionD result(FGrid); result=Zero();
LatticeGaugeFieldD Umu(UGrid);
LatticeGaugeFieldF Umu_f(UGrid_f);
SU<Nc>::HotConfiguration(RNG4,Umu);
precisionChange(Umu_f,Umu);
RealD mass=0.1;
RealD M5=1.8;
DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
DomainWallFermionF Ddwf_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
LatticeFermionD src_o(FrbGrid);
LatticeFermionD result_o(FrbGrid);
LatticeFermionD result_o_2(FrbGrid);
pickCheckerboard(Odd,src_o,src);
result_o.Checkerboard() = Odd;
result_o = Zero();
result_o_2.Checkerboard() = Odd;
result_o_2 = Zero();
SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
SchurDiagMooeeOperator<DomainWallFermionF,LatticeFermionF> HermOpEO_f(Ddwf_f);
std::cout << GridLogMessage << "::::::::::::: Starting mixed CG" << std::endl;
MixedPrecisionConjugateGradient<LatticeFermionD,LatticeFermionF> mCG(1.0e-8, 10000, 50, FrbGrid_f, HermOpEO_f, HermOpEO);
double t1,t2,flops;
double MdagMsiteflops = 1452; // Mobius (real coeffs)
// CG overhead: 8 inner product, 4+8 axpy_norm, 4+4 linear comb (2 of)
double CGsiteflops = (8+4+8+4+4)*Nc*Ns ;
std:: cout << " MdagM site flops = "<< 4*MdagMsiteflops<<std::endl;
std:: cout << " CG site flops = "<< CGsiteflops <<std::endl;
result_o = Zero();
t1=usecond();
mCG(src_o,result_o);
t2=usecond();
int iters = mCG.TotalInnerIterations; //Number of inner CG iterations
flops = MdagMsiteflops*4*FrbGrid->gSites()*iters;
flops+= CGsiteflops*FrbGrid->gSites()*iters;
std::cout << " SinglePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
std::cout << " SinglePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
std::cout << GridLogMessage << "::::::::::::: Starting regular CG" << std::endl;
ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
result_o_2 = Zero();
t1=usecond();
CG(HermOpEO,src_o,result_o_2);
t2=usecond();
iters = CG.IterationsToComplete;
flops = MdagMsiteflops*4*FrbGrid->gSites()*iters;
flops+= CGsiteflops*FrbGrid->gSites()*iters;
std::cout << " DoublePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
std::cout << " DoublePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
LatticeFermionD diff_o(FrbGrid);
RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
std::cout << GridLogMessage << "::::::::::::: Diff between mixed and regular CG: " << diff << std::endl;
MemoryManager::Print();
Grid_finalize();
}