mirror of
https://github.com/paboyle/Grid.git
synced 2026-10-03 22:48:05 +01:00
Implemented mixed precision CG. Fixed filelist to exclude lib/Old directory and include Config.h.
This commit is contained in:
Christopher Kelly committed
1 parent
df5c788ef2
commit
85ed8175cb
9 files changed
+311
-14
No files matched your search
@@ -44,6 +44,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
#include <algorithms/iterative/SchurRedBlack.h>
|
||||
|
||||
#include <algorithms/iterative/ConjugateGradientMultiShift.h>
|
||||
#include <algorithms/iterative/ConjugateGradientMixedPrec.h>
|
||||
|
||||
// Lanczos support
|
||||
#include <algorithms/iterative/MatrixUtils.h>
|
||||
|
||||
+2
-2
@@ -1,4 +1,4 @@
|
||||
|
||||
HFILES=./Algorithms.h ./AlignedAllocator.h ./Cartesian.h ./Communicator.h ./Cshift.h ./Grid.h ./Init.h ./Lattice.h ./Lexicographic.h ./Log.h ./Old/Tensor_peek.h ./Old/Tensor_poke.h ./PerfCount.h ./Simd.h ./Stencil.h ./Tensors.h ./Threads.h ./Timer.h ./algorithms/CoarsenedMatrix.h ./algorithms/LinearOperator.h ./algorithms/Preconditioner.h ./algorithms/SparseMatrix.h ./algorithms/approx/Chebyshev.h ./algorithms/approx/MultiShiftFunction.h ./algorithms/approx/Remez.h ./algorithms/approx/Zolotarev.h ./algorithms/approx/bigfloat.h ./algorithms/approx/bigfloat_double.h ./algorithms/iterative/AdefGeneric.h ./algorithms/iterative/ConjugateGradient.h ./algorithms/iterative/ConjugateGradientMultiShift.h ./algorithms/iterative/ConjugateResidual.h ./algorithms/iterative/DenseMatrix.h ./algorithms/iterative/EigenSort.h ./algorithms/iterative/Francis.h ./algorithms/iterative/Householder.h ./algorithms/iterative/ImplicitlyRestartedLanczos.h ./algorithms/iterative/Matrix.h ./algorithms/iterative/MatrixUtils.h ./algorithms/iterative/NormalEquations.h ./algorithms/iterative/PrecConjugateResidual.h ./algorithms/iterative/PrecGeneralisedConjugateResidual.h ./algorithms/iterative/SchurRedBlack.h ./cartesian/Cartesian_base.h ./cartesian/Cartesian_full.h ./cartesian/Cartesian_red_black.h ./communicator/Communicator_base.h ./cshift/Cshift_common.h ./cshift/Cshift_mpi.h ./cshift/Cshift_none.h ./lattice/Lattice_ET.h ./lattice/Lattice_arith.h ./lattice/Lattice_base.h ./lattice/Lattice_comparison.h ./lattice/Lattice_comparison_utils.h ./lattice/Lattice_conformable.h ./lattice/Lattice_coordinate.h ./lattice/Lattice_local.h ./lattice/Lattice_overload.h ./lattice/Lattice_peekpoke.h ./lattice/Lattice_reality.h ./lattice/Lattice_reduction.h ./lattice/Lattice_rng.h ./lattice/Lattice_trace.h ./lattice/Lattice_transfer.h ./lattice/Lattice_transpose.h ./lattice/Lattice_unary.h ./lattice/Lattice_where.h ./parallelIO/BinaryIO.h ./parallelIO/NerscIO.h ./pugixml/pugixml.h ./qcd/QCD.h ./qcd/action/ActionBase.h ./qcd/action/ActionParams.h ./qcd/action/Actions.h ./qcd/action/fermion/CayleyFermion5D.h ./qcd/action/fermion/ContinuedFractionFermion5D.h ./qcd/action/fermion/DomainWallFermion.h ./qcd/action/fermion/FermionOperator.h ./qcd/action/fermion/FermionOperatorImpl.h ./qcd/action/fermion/MobiusFermion.h ./qcd/action/fermion/MobiusZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h ./qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonContfracTanhFermion.h ./qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h ./qcd/action/fermion/PartialFractionFermion5D.h ./qcd/action/fermion/ScaledShamirFermion.h ./qcd/action/fermion/ShamirZolotarevFermion.h ./qcd/action/fermion/WilsonCompressor.h ./qcd/action/fermion/WilsonFermion.h ./qcd/action/fermion/WilsonFermion5D.h ./qcd/action/fermion/WilsonKernels.h ./qcd/action/fermion/WilsonTMFermion.h ./qcd/action/fermion/g5HermitianLinop.h ./qcd/action/fermion/WilsonKernelsAsmBody.h ./qcd/action/gauge/GaugeImpl.h ./qcd/action/gauge/PlaqPlusRectangleAction.h ./qcd/action/gauge/WilsonGaugeAction.h ./qcd/action/pseudofermion/EvenOddSchurDifferentiable.h ./qcd/action/pseudofermion/OneFlavourEvenOddRational.h ./qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h ./qcd/action/pseudofermion/OneFlavourRational.h ./qcd/action/pseudofermion/OneFlavourRationalRatio.h ./qcd/action/pseudofermion/TwoFlavour.h ./qcd/action/pseudofermion/TwoFlavourEvenOdd.h ./qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h ./qcd/action/pseudofermion/TwoFlavourRatio.h ./qcd/hmc/HMC.h ./qcd/hmc/HmcRunner.h ./qcd/hmc/NerscCheckpointer.h ./qcd/hmc/integrators/Integrator.h ./qcd/hmc/integrators/Integrator_algorithm.h ./qcd/spin/Dirac.h ./qcd/spin/TwoSpinor.h ./qcd/utils/CovariantCshift.h ./qcd/utils/LinalgUtils.h ./qcd/utils/SUn.h ./qcd/utils/SpaceTimeGrid.h ./qcd/utils/WilsonLoops.h ./serialisation/BaseIO.h ./serialisation/BinaryIO.h ./serialisation/MacroMagic.h ./serialisation/Serialisation.h ./serialisation/TextIO.h ./serialisation/XmlIO.h ./simd/Grid_avx.h ./simd/Grid_avx512.h ./simd/Grid_empty.h ./simd/Grid_imci.h ./simd/Grid_neon.h ./simd/Grid_qpx.h ./simd/Grid_sse4.h ./simd/Grid_vector_types.h ./simd/Grid_vector_unops.h ./simd/Intel512avx.h ./simd/Intel512wilson.h ./simd/Intel512common.h ./simd/Intel512double.h ./simd/Intel512imci.h ./simd/Intel512single.h ./stencil/Lebesgue.h ./tensors/Tensor_Ta.h ./tensors/Tensor_arith.h ./tensors/Tensor_arith_add.h ./tensors/Tensor_arith_mac.h ./tensors/Tensor_arith_mul.h ./tensors/Tensor_arith_scalar.h ./tensors/Tensor_arith_sub.h ./tensors/Tensor_class.h ./tensors/Tensor_determinant.h ./tensors/Tensor_exp.h ./tensors/Tensor_extract_merge.h ./tensors/Tensor_index.h ./tensors/Tensor_inner.h ./tensors/Tensor_logical.h ./tensors/Tensor_outer.h ./tensors/Tensor_reality.h ./tensors/Tensor_trace.h ./tensors/TensLine truncated
|
||||
HFILES=./cshift/Cshift_none.h ./cshift/Cshift_common.h ./cshift/Cshift_mpi.h ./Tensors.h ./qcd/utils/SUn.h ./qcd/utils/SpaceTimeGrid.h ./qcd/utils/CovariantCshift.h ./qcd/utils/WilsonLoops.h ./qcd/utils/LinalgUtils.h ./qcd/QCD.h ./qcd/action/fermion/DomainWallFermion.h ./qcd/action/fermion/MobiusFermion.h ./qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h ./qcd/action/fermion/g5HermitianLinop.h ./qcd/action/fermion/WilsonFermion5D.h ./qcd/action/fermion/WilsonKernels.h ./qcd/action/fermion/FermionOperator.h ./qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h ./qcd/action/fermion/ScaledShamirFermion.h ./qcd/action/fermion/CayleyFermion5D.h ./qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h ./qcd/action/fermion/WilsonCompressor.h ./qcd/action/fermion/FermionOperatorImpl.h ./qcd/action/fermion/WilsonFermion.h ./qcd/action/fermion/WilsonTMFermion.h ./qcd/action/fermion/MobiusZolotarevFermion.h ./qcd/action/fermion/ShamirZolotarevFermion.h ./qcd/action/fermion/WilsonKernelsAsmBody.h ./qcd/action/fermion/PartialFractionFermion5D.h ./qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h ./qcd/action/fermion/ContinuedFractionFermion5D.h ./qcd/action/fermion/OverlapWilsonContfracTanhFermion.h ./qcd/action/gauge/GaugeImpl.h ./qcd/action/gauge/PlaqPlusRectangleAction.h ./qcd/action/gauge/WilsonGaugeAction.h ./qcd/action/pseudofermion/TwoFlavour.h ./qcd/action/pseudofermion/OneFlavourRationalRatio.h ./qcd/action/pseudofermion/OneFlavourEvenOddRational.h ./qcd/action/pseudofermion/OneFlavourRational.h ./qcd/action/pseudofermion/OneFlavourEvenOddRationalRatio.h ./qcd/action/pseudofermion/TwoFlavourEvenOdd.h ./qcd/action/pseudofermion/EvenOddSchurDifferentiable.h ./qcd/action/pseudofermion/TwoFlavourRatio.h ./qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h ./qcd/action/ActionBase.h ./qcd/action/ActionParams.h ./qcd/action/Actions.h ./qcd/hmc/HmcRunner.h ./qcd/hmc/integrators/Integrator.h ./qcd/hmc/integrators/Integrator_algorithm.h ./qcd/hmc/HMC.h ./qcd/hmc/NerscCheckpointer.h ./qcd/spin/Dirac.h ./qcd/spin/TwoSpinor.h ./simd/Intel512wilson.h ./simd/Intel512common.h ./simd/Grid_sse4.h ./simd/Grid_qpx.h ./simd/Grid_vector_types.h ./simd/Grid_vector_unops.h ./simd/Intel512imci.h ./simd/Intel512avx.h ./simd/Grid_neon.h ./simd/Grid_imci.h ./simd/Intel512single.h ./simd/Grid_empty.h ./simd/Intel512double.h ./simd/Grid_avx512.h ./simd/Grid_avx.h ./tensors/Tensor_arith_mac.h ./tensors/Tensor_arith_sub.h ./tensors/Tensor_unary.h ./tensors/Tensor_trace.h ./tensors/Tensor_determinant.h ./tensors/Tensor_reality.h ./tensors/Tensor_arith_scalar.h ./tensors/Tensor_class.h ./tensors/Tensor_arith.h ./tensors/Tensor_index.h ./tensors/Tensor_extract_merge.h ./tensors/Tensor_logical.h ./tensors/Tensor_Ta.h ./tensors/Tensor_exp.h ./tensors/Tensor_arith_add.h ./tensors/Tensor_outer.h ./tensors/Tensor_traits.h ./tensors/Tensor_arith_mul.h ./tensors/Tensor_inner.h ./tensors/Tensor_transpose.h ./Log.h ./Communicator.h ./cartesian/Cartesian_full.h ./cartesian/Cartesian_red_black.h ./cartesian/Cartesian_base.h ./Timer.h ./Init.h ./Algorithms.h ./Lexicographic.h ./algorithms/approx/Remez.h ./algorithms/approx/Zolotarev.h ./algorithms/approx/bigfloat_double.h ./algorithms/approx/MultiShiftFunction.h ./algorithms/approx/Chebyshev.h ./algorithms/approx/bigfloat.h ./algorithms/Preconditioner.h ./algorithms/iterative/ConjugateGradient.h ./algorithms/iterative/ConjugateGradientMultiShift.h ./algorithms/iterative/Francis.h ./algorithms/iterative/PrecConjugateResidual.h ./algorithms/iterative/Householder.h ./algorithms/iterative/ImplicitlyRestartedLanczos.h ./algorithms/iterative/Matrix.h ./algorithms/iterative/EigenSort.h ./algorithms/iterative/ConjugateGradientMixedPrec.h ./algorithms/iterative/NormalEquations.h ./algorithms/iterative/PrecGeneralisedConjugateResidual.h ./algorithms/iterative/DenseMatrix.h ./algorithms/iterative/AdefGeneric.h ./algorithms/iterative/ConjugateResidual.h ./algorithms/iterative/SchurRedBlack.h ./algorithms/iterative/MatrixUtils.h ./algorithms/CoarsenedMatrix.h ./algorithms/LinearOperator.h ./algorithms/SparseMatrix.h ./serialisation/XmlIO.h ./serialisation/TextIO.h ./serialisation/BinaryIO.h ./serialisation/MacroMagic.h ./serialisation/BaseIO.h ./serialisation/Serialisation.h ./Stencil.h ./lattice/Lattice_base.h ./lattice/Lattice_arith.h ./lattice/Lattice_rng.h ./lattice/Lattice_unary.h ./lattice/Lattice_comparison_utils.h ./lattice/Lattice_coordinate.h ./lattice/Lattice_local.h ./lattice/Lattice_reality.h ./lattice/Lattice_comparison.h ./lattice/Lattice_trace.h ./lattice/Lattice_reduction.h ./lattice/Lattice_transpose.h ./lattice/Lattice_peekpoke.h ./lattice/Lattice_transfer.h ./lattice/Lattice_where.h ./lattice/Lattice_ET.h ./lattice/Lattice_conformable.h ./lattice/Lattice_overload.h ./Lattice.h ./communicator/Communicator_base.h ./Cshift.h ./AlignedAllocator.h ./PerfCount.h ./Simd.h ./stencil/Lebesgue.h ./Threads.h ./parallelIOLine truncated
|
||||
|
||||
CCFILES=./Init.cc ./Log.cc ./PerfCount.cc ./algorithms/approx/MultiShiftFunction.cc ./algorithms/approx/Remez.cc ./algorithms/approx/Zolotarev.cc ./pugixml/pugixml.cc ./qcd/action/fermion/CayleyFermion5D.cc ./qcd/action/fermion/ContinuedFractionFermion5D.cc ./qcd/action/fermion/PartialFractionFermion5D.cc ./qcd/action/fermion/WilsonFermion.cc ./qcd/action/fermion/WilsonFermion5D.cc ./qcd/action/fermion/WilsonKernels.cc ./qcd/action/fermion/WilsonKernelsAsm.cc ./qcd/action/fermion/WilsonKernelsHand.cc ./qcd/action/fermion/WilsonTMFermion.cc ./qcd/hmc/HMC.cc ./qcd/spin/Dirac.cc ./qcd/utils/SpaceTimeGrid.cc ./serialisation/BinaryIO.cc ./serialisation/TextIO.cc ./serialisation/XmlIO.cc ./stencil/Lebesgue.cc ./stencil/Stencil_common.cc
|
||||
CCFILES=./Log.cc ./qcd/utils/SpaceTimeGrid.cc ./qcd/action/fermion/WilsonKernelsAsm.cc ./qcd/action/fermion/WilsonTMFermion.cc ./qcd/action/fermion/WilsonKernels.cc ./qcd/action/fermion/WilsonFermion5D.cc ./qcd/action/fermion/PartialFractionFermion5D.cc ./qcd/action/fermion/CayleyFermion5D.cc ./qcd/action/fermion/WilsonFermion.cc ./qcd/action/fermion/WilsonKernelsHand.cc ./qcd/action/fermion/ContinuedFractionFermion5D.cc ./qcd/hmc/HMC.cc ./qcd/spin/Dirac.cc ./algorithms/approx/Zolotarev.cc ./algorithms/approx/MultiShiftFunction.cc ./algorithms/approx/Remez.cc ./serialisation/XmlIO.cc ./serialisation/TextIO.cc ./serialisation/BinaryIO.cc ./Init.cc ./stencil/Stencil_common.cc ./stencil/Lebesgue.cc ./PerfCount.cc ./pugixml/pugixml.cc
|
||||
@@ -0,0 +1,141 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrec.h
|
||||
|
||||
Copyright (C) 2015
|
||||
|
||||
Author: Christopher Kelly <ckelly@phys.columbia.edu>
|
||||
|
||||
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_MIXED_PREC_H
|
||||
#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_H
|
||||
|
||||
namespace Grid {
|
||||
|
||||
template<class FieldD,class FieldF, typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
|
||||
class MixedPrecisionConjugateGradient : public LinearFunction<FieldD> {
|
||||
public:
|
||||
RealD Tolerance;
|
||||
Integer MaxInnerIterations;
|
||||
Integer MaxOuterIterations;
|
||||
GridBase* SinglePrecGrid; //Grid for single-precision fields
|
||||
RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
|
||||
LinearOperatorBase<FieldF> &Linop_f;
|
||||
LinearOperatorBase<FieldD> &Linop_d;
|
||||
|
||||
//Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
|
||||
LinearFunction<FieldF> *guesser;
|
||||
|
||||
MixedPrecisionConjugateGradient(RealD tol, Integer maxinnerit, Integer maxouterit, GridBase* _sp_grid, LinearOperatorBase<FieldF> &_Linop_f, LinearOperatorBase<FieldD> &_Linop_d) :
|
||||
Linop_f(_Linop_f), Linop_d(_Linop_d),
|
||||
Tolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), SinglePrecGrid(_sp_grid),
|
||||
OuterLoopNormMult(100.), guesser(NULL){ };
|
||||
|
||||
void useGuesser(LinearFunction<FieldF> &g){
|
||||
guesser = g;
|
||||
}
|
||||
|
||||
void operator() (const FieldD &src_d_in, FieldD &sol_d){
|
||||
GridStopWatch TotalTimer;
|
||||
TotalTimer.Start();
|
||||
|
||||
int cb = src_d_in.checkerboard;
|
||||
sol_d.checkerboard = cb;
|
||||
|
||||
RealD src_norm = norm2(src_d_in);
|
||||
RealD stop = src_norm * Tolerance*Tolerance;
|
||||
|
||||
GridBase* DoublePrecGrid = src_d_in._grid;
|
||||
FieldD tmp_d(DoublePrecGrid);
|
||||
tmp_d.checkerboard = cb;
|
||||
|
||||
FieldD tmp2_d(DoublePrecGrid);
|
||||
tmp2_d.checkerboard = cb;
|
||||
|
||||
FieldD src_d(DoublePrecGrid);
|
||||
src_d = src_d_in; //source for next inner iteration, computed from residual during operation
|
||||
|
||||
RealD inner_tol = Tolerance;
|
||||
|
||||
FieldF src_f(SinglePrecGrid);
|
||||
src_f.checkerboard = cb;
|
||||
|
||||
FieldF sol_f(SinglePrecGrid);
|
||||
sol_f.checkerboard = cb;
|
||||
|
||||
ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
|
||||
CG_f.ErrorOnNoConverge = false;
|
||||
|
||||
GridStopWatch InnerCGtimer;
|
||||
|
||||
GridStopWatch PrecChangeTimer;
|
||||
|
||||
for(Integer outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
|
||||
//Compute double precision rsd and also new RHS vector.
|
||||
Linop_d.HermOp(sol_d, tmp_d);
|
||||
RealD norm = axpy_norm(src_d, -1., tmp_d, src_d_in); //src_d is residual vector
|
||||
|
||||
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration " <<outer_iter<<" residual "<< norm<< " target "<< stop<<std::endl;
|
||||
|
||||
if(norm < OuterLoopNormMult * stop){
|
||||
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Outer iteration converged on iteration " <<outer_iter <<std::endl;
|
||||
break;
|
||||
}
|
||||
while(norm * inner_tol * inner_tol < stop) inner_tol *= 2; // inner_tol = sqrt(stop/norm) ??
|
||||
|
||||
PrecChangeTimer.Start();
|
||||
precisionChange(src_f, src_d);
|
||||
PrecChangeTimer.Stop();
|
||||
|
||||
zeroit(sol_f);
|
||||
|
||||
//Optionally improve inner solver guess (eg using known eigenvectors)
|
||||
if(guesser != NULL)
|
||||
(*guesser)(src_f, sol_f);
|
||||
|
||||
//Inner CG
|
||||
CG_f.Tolerance = inner_tol;
|
||||
InnerCGtimer.Start();
|
||||
CG_f(Linop_f, src_f, sol_f);
|
||||
InnerCGtimer.Stop();
|
||||
|
||||
//Convert sol back to double and add to double prec solution
|
||||
PrecChangeTimer.Start();
|
||||
precisionChange(tmp_d, sol_f);
|
||||
PrecChangeTimer.Stop();
|
||||
|
||||
axpy(sol_d, 1.0, tmp_d, sol_d);
|
||||
}
|
||||
|
||||
//Final trial CG
|
||||
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Starting final patch-up double-precision solve"<<std::endl;
|
||||
|
||||
ConjugateGradient<FieldD> CG_d(Tolerance, MaxInnerIterations);
|
||||
CG_d(Linop_d, src_d_in, sol_d);
|
||||
|
||||
TotalTimer.Stop();
|
||||
std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradient: Total " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
|
||||
}
|
||||
};
|
||||
|
||||
}
|
||||
|
||||
#endif
|
||||
@@ -82,6 +82,12 @@ template <typename T> using is_lattice = std::is_base_of<LatticeBase,T >;
|
||||
|
||||
template <typename T> using is_lattice_expr = std::is_base_of<LatticeExpressionBase,T >;
|
||||
|
||||
//Specialization of getVectorType for lattices
|
||||
template<typename T>
|
||||
struct getVectorType<Lattice<T> >{
|
||||
typedef typename Lattice<T>::vector_object type;
|
||||
};
|
||||
|
||||
template<class sobj>
|
||||
inline sobj eval(const unsigned int ss, const sobj &arg)
|
||||
{
|
||||
|
||||
@@ -482,6 +482,96 @@ void Replicate(Lattice<vobj> &coarse,Lattice<vobj> & fine)
|
||||
|
||||
}
|
||||
|
||||
//Copy SIMD-vectorized lattice to array of scalar objects in lexicographic order
|
||||
template<typename vobj, typename sobj>
|
||||
typename std::enable_if<isSIMDvectorized<vobj>::value && !isSIMDvectorized<sobj>::value, void>::type unvectorizeToLexOrdArray(std::vector<sobj> &out, const Lattice<vobj> &in){
|
||||
typedef typename vobj::vector_type vtype;
|
||||
|
||||
GridBase* in_grid = in._grid;
|
||||
out.resize(in_grid->lSites());
|
||||
|
||||
int ndim = in_grid->Nd();
|
||||
int in_nsimd = vtype::Nsimd();
|
||||
|
||||
std::vector<int> in_icoor[in_nsimd];
|
||||
|
||||
for(int lane=0; lane < in_nsimd; lane++){
|
||||
in_icoor[lane].resize(ndim);
|
||||
in_grid->iCoorFromIindex(in_icoor[lane], lane);
|
||||
}
|
||||
|
||||
PARALLEL_FOR_LOOP
|
||||
for(int in_oidx = 0; in_oidx < in_grid->oSites(); in_oidx++){ //loop over outer index
|
||||
//Assemble vector of pointers to output elements
|
||||
std::vector<sobj*> out_ptrs(in_nsimd);
|
||||
|
||||
std::vector<int> in_ocoor(ndim);
|
||||
in_grid->oCoorFromOindex(in_ocoor, in_oidx);
|
||||
|
||||
std::vector<int> lcoor(in_grid->Nd());
|
||||
|
||||
for(int lane=0; lane < in_nsimd; lane++){
|
||||
for(int mu=0;mu<ndim;mu++)
|
||||
lcoor[mu] = in_ocoor[mu] + in_grid->_rdimensions[mu]*in_icoor[lane][mu];
|
||||
|
||||
int lex;
|
||||
Lexicographic::IndexFromCoor(lcoor, lex, in_grid->_ldimensions);
|
||||
out_ptrs[lane] = &out[lex];
|
||||
}
|
||||
|
||||
//Unpack into those ptrs
|
||||
const vobj & in_vobj = in._odata[in_oidx];
|
||||
extract1(in_vobj, out_ptrs, 0);
|
||||
}
|
||||
}
|
||||
|
||||
//Convert a Lattice from one precision to another
|
||||
template<class VobjOut, class VobjIn>
|
||||
void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
|
||||
assert(out._grid->Nd() == in._grid->Nd());
|
||||
out.checkerboard = in.checkerboard;
|
||||
GridBase *in_grid=in._grid;
|
||||
GridBase *out_grid = out._grid;
|
||||
|
||||
typedef typename VobjOut::scalar_object SobjOut;
|
||||
typedef typename VobjIn::scalar_object SobjIn;
|
||||
|
||||
int ndim = out._grid->Nd();
|
||||
int out_nsimd = out_grid->Nsimd();
|
||||
|
||||
std::vector<int> out_icoor[out_nsimd];
|
||||
|
||||
for(int lane=0; lane < out_nsimd; lane++){
|
||||
out_icoor[lane].resize(ndim);
|
||||
out_grid->iCoorFromIindex(out_icoor[lane], lane);
|
||||
}
|
||||
|
||||
std::vector<SobjOut> in_slex_conv(in_grid->lSites());
|
||||
unvectorizeToLexOrdArray(in_slex_conv, in);
|
||||
|
||||
PARALLEL_FOR_LOOP
|
||||
for(int out_oidx=0;out_oidx<out_grid->oSites();out_oidx++){
|
||||
std::vector<int> out_ocoor(ndim);
|
||||
out_grid->oCoorFromOindex(out_ocoor, out_oidx);
|
||||
|
||||
std::vector<SobjOut*> ptrs(out_nsimd);
|
||||
|
||||
std::vector<int> lcoor(out_grid->Nd());
|
||||
|
||||
for(int lane=0; lane < out_nsimd; lane++){
|
||||
for(int mu=0;mu<ndim;mu++)
|
||||
lcoor[mu] = out_ocoor[mu] + out_grid->_rdimensions[mu]*out_icoor[lane][mu];
|
||||
|
||||
int llex; Lexicographic::IndexFromCoor(lcoor, llex, out_grid->_ldimensions);
|
||||
ptrs[lane] = &in_slex_conv[llex];
|
||||
}
|
||||
merge(out._odata[out_oidx], ptrs, 0);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
}
|
||||
#endif
|
||||
@@ -10,6 +10,7 @@ Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: neo <cossu@post.kek.jp>
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Christopher Kelly <ckelly@phys.columbia.edu>
|
||||
|
||||
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
|
||||
@@ -167,6 +168,33 @@ void extract(const vobj &vec,std::vector<typename vobj::scalar_object *> &extrac
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
// Extract to a bunch of scalar object pointers of different scalar type, with offset. Useful for precision change
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template<class vobj, class sobj> inline
|
||||
void extract1(const vobj &vec,std::vector<sobj*> &extracted, int offset)
|
||||
{
|
||||
typedef typename vobj::scalar_type vobj_scalar_type ;
|
||||
typedef typename vobj::vector_type vobj_vector_type ;
|
||||
|
||||
typedef typename sobj::scalar_type sobj_scalar_type ;
|
||||
|
||||
static const int words=sizeof(vobj)/sizeof(vobj_vector_type);
|
||||
static const int Nsimd=vobj_vector_type::Nsimd();
|
||||
|
||||
int Nextr=extracted.size();
|
||||
int s = Nsimd/Nextr;
|
||||
vobj_scalar_type * vp = (vobj_scalar_type *)&vec;
|
||||
|
||||
for(int w=0;w<words;w++){
|
||||
for(int i=0;i<Nextr;i++){
|
||||
sobj_scalar_type * pointer = (sobj_scalar_type *)& extracted[i][offset];
|
||||
pointer[w] = vp[i*s+w*Nsimd];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
// Merge a contiguous array of scalar objects
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -8,6 +8,7 @@
|
||||
|
||||
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
|
||||
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
|
||||
Author: Christopher Kelly <ckelly@phys.columbia.edu>
|
||||
|
||||
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
|
||||
@@ -230,6 +231,35 @@ namespace Grid {
|
||||
static const bool value = true;
|
||||
};
|
||||
|
||||
//Get the SIMD vector type from a Grid tensor or Lattice<Tensor>
|
||||
template<typename T>
|
||||
struct getVectorType{
|
||||
typedef T type;
|
||||
};
|
||||
|
||||
//Query if a tensor or Lattice<Tensor> is SIMD vector or scalar
|
||||
template<typename T>
|
||||
class isSIMDvectorized{
|
||||
template<typename U>
|
||||
static typename std::enable_if< !std::is_same< typename GridTypeMapper<typename getVectorType<U>::type>::scalar_type, typename GridTypeMapper<typename getVectorType<U>::type>::vector_type>::value, char>::type test(void *);
|
||||
|
||||
template<typename U>
|
||||
static double test(...);
|
||||
|
||||
public:
|
||||
enum {value = sizeof(test<T>(0)) == sizeof(char) };
|
||||
};
|
||||
|
||||
//Get the precision of a Lattice, tensor or scalar type in units of sizeof(float)
|
||||
template<typename T>
|
||||
class getPrecision{
|
||||
typedef typename getVectorType<T>::type vector_obj; //get the vector_obj (i.e. a grid Tensor) if its a Lattice<vobj>, do nothing otherwise (i.e. if fundamental or grid Tensor)
|
||||
|
||||
typedef typename GridTypeMapper<vector_obj>::scalar_type scalar_type; //get the associated scalar type. Works on fundamental and tensor types
|
||||
typedef typename GridTypeMapper<scalar_type>::Realified real_scalar_type; //remove any std::complex wrapper, should get us to the fundamental type
|
||||
public:
|
||||
enum { value = sizeof(real_scalar_type)/sizeof(float) };
|
||||
};
|
||||
}
|
||||
|
||||
#endif
|
||||
Reference in new issue
Block a user