portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-04 19:25:56 +01:00
Code Releases Activity

Merge branch 'feature/block-cg' into develop

Browse Source
This commit is contained in:
paboyle 2017-04-13 14:12:24 +01:00
commit 2a54c9aaab
17 changed files with 1135 additions and 58 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
.travis.yml
View File

@ -7,7 +7,7 @@ cache:
matrix:
include:
- os: osx
osx_image: xcode7.2
osx_image: xcode8.3
compiler: clang
- compiler: gcc
addons:
@ -73,8 +73,6 @@ before_install:
- if [[ "$TRAVIS_OS_NAME" == "linux" ]] && [[ "$CC" == "clang" ]]; then export LD_LIBRARY_PATH="${GRIDDIR}/clang/lib:${LD_LIBRARY_PATH}"; fi
- if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew update; fi
- if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install libmpc; fi
- if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then brew install openmpi; fi
- if [[ "$TRAVIS_OS_NAME" == "osx" ]] && [[ "$CC" == "gcc" ]]; then brew install gcc5; fi
install:
- export CC=$CC$VERSION
@ -92,15 +90,14 @@ script:
- cd build
- ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=none
- make -j4
- ./benchmarks/Benchmark_dwf --threads 1
- ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
- echo make clean
- ../configure --enable-precision=double --enable-simd=SSE4 --enable-comms=none
- make -j4
- ./benchmarks/Benchmark_dwf --threads 1
- ./benchmarks/Benchmark_dwf --threads 1 --debug-signals
- echo make clean
- if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then export CXXFLAGS='-DMPI_UINT32_T=MPI_UNSIGNED -DMPI_UINT64_T=MPI_UNSIGNED_LONG'; fi
- ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto
- make -j4
- if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then ../configure --enable-precision=single --enable-simd=SSE4 --enable-comms=mpi-auto CXXFLAGS='-DMPI_UINT32_T=MPI_UNSIGNED -DMPI_UINT64_T=MPI_UNSIGNED_LONG'; fi
- if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then make -j4; fi
- if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then mpirun.openmpi -n 2 ./benchmarks/Benchmark_dwf --threads 1 --mpi 2.1.1.1; fi

21
configure.ac
View File

@ -83,6 +83,19 @@ case ${ac_LAPACK} in
AC_DEFINE([USE_LAPACK],[1],[use LAPACK]);;
esac
############### FP16 conversions
AC_ARG_ENABLE([fp16],
[AC_HELP_STRING([--enable-fp16=yes|no], [enable fp16 comms])],
[ac_FP16=${enable_fp16}], [ac_FP16=no])
case ${ac_FP16} in
no)
;;
yes)
AC_DEFINE([USE_FP16],[1],[conversion to fp16]);;
*)
;;
esac
############### MKL
AC_ARG_ENABLE([mkl],
[AC_HELP_STRING([--enable-mkl=yes|no|prefix], [enable Intel MKL for LAPACK & FFTW])],
@ -179,16 +192,16 @@ case ${ax_cv_cxx_compiler_vendor} in
SIMD_FLAGS='-msse4.2';;
AVX)
AC_DEFINE([AVX1],[1],[AVX intrinsics])
SIMD_FLAGS='-mavx';;
SIMD_FLAGS='-mavx -mf16c';;
AVXFMA4)
AC_DEFINE([AVXFMA4],[1],[AVX intrinsics with FMA4])
SIMD_FLAGS='-mavx -mfma4';;
SIMD_FLAGS='-mavx -mfma4 -mf16c';;
AVXFMA)
AC_DEFINE([AVXFMA],[1],[AVX intrinsics with FMA3])
SIMD_FLAGS='-mavx -mfma';;
SIMD_FLAGS='-mavx -mfma -mf16c';;
AVX2)
AC_DEFINE([AVX2],[1],[AVX2 intrinsics])
SIMD_FLAGS='-mavx2 -mfma';;
SIMD_FLAGS='-mavx2 -mfma -mf16c';;
AVX512)
AC_DEFINE([AVX512],[1],[AVX512 intrinsics])
SIMD_FLAGS='-mavx512f -mavx512pf -mavx512er -mavx512cd';;

516
lib/algorithms/iterative/BlockConjugateGradient.h Normal file
View File

@ -0,0 +1,516 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/BlockConjugateGradient.h
Copyright (C) 2017
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_BLOCK_CONJUGATE_GRADIENT_H
#define GRID_BLOCK_CONJUGATE_GRADIENT_H
#include <Grid/Eigen/Dense>
namespace Grid {
GridBase *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
{
int NN = BlockSolverGrid->_ndimension;
int nsimd = BlockSolverGrid->Nsimd();
std::vector<int> latt_phys(0);
std::vector<int> simd_phys(0);
std::vector<int> mpi_phys(0);
for(int d=0;d<NN;d++){
if( d!=Orthog ) {
latt_phys.push_back(BlockSolverGrid->_fdimensions[d]);
simd_phys.push_back(BlockSolverGrid->_simd_layout[d]);
mpi_phys.push_back(BlockSolverGrid->_processors[d]);
}
}
return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys);
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Need to move sliceInnerProduct, sliceAxpy, sliceNorm etc... into lattice sector along with sliceSum
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X._grid->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
Lattice<vobj> Xslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// FIXME: Implementation is slow
// If we based this on Cshift it would work for spread out
// but it would be even slower
//
// Repeated extract slice is inefficient
//
// Best base the linear combination by constructing a
// set of vectors of size grid->_rdimensions[Orthog].
for(int i=0;i<Nblock;i++){
ExtractSlice(Rslice,Y,i,Orthog);
for(int j=0;j<Nblock;j++){
ExtractSlice(Xslice,X,j,Orthog);
Rslice = Rslice + Xslice*(scale*aa(j,i));
}
InsertSlice(Rslice,R,i,Orthog);
}
};
template<class vobj>
static void sliceMaddVector (Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
int Orthog,RealD scale=1.0)
{
// FIXME: Implementation is slow
// Best base the linear combination by constructing a
// set of vectors of size grid->_rdimensions[Orthog].
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X._grid->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
Lattice<vobj> Xslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// If we based this on Cshift it would work for spread out
// but it would be even slower
for(int i=0;i<Nblock;i++){
ExtractSlice(Rslice,Y,i,Orthog);
ExtractSlice(Xslice,X,i,Orthog);
Rslice = Rslice + Xslice*(scale*a[i]);
InsertSlice(Rslice,R,i,Orthog);
}
};
template<class vobj>
static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
// FIXME: Implementation is slow
// Not sure of best solution.. think about it
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
GridBase *FullGrid = lhs._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
int Nblock = FullGrid->GlobalDimensions()[Orthog];
Lattice<vobj> Lslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
for(int i=0;i<Nblock;i++){
ExtractSlice(Lslice,lhs,i,Orthog);
for(int j=0;j<Nblock;j++){
ExtractSlice(Rslice,rhs,j,Orthog);
mat(i,j) = innerProduct(Lslice,Rslice);
}
}
#undef FORCE_DIAG
#ifdef FORCE_DIAG
for(int i=0;i<Nblock;i++){
for(int j=0;j<Nblock;j++){
if ( i != j ) mat(i,j)=0.0;
}
}
#endif
return;
}
template<class vobj>
static void sliceInnerProductVector( std::vector<ComplexD> & vec, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
// FIXME: Implementation is slow
// Look at localInnerProduct implementation,
// and do inside a site loop with block strided iterators
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef typename vobj::tensor_reduced scalar;
typedef typename scalar::scalar_object scomplex;
int Nblock = lhs._grid->GlobalDimensions()[Orthog];
vec.resize(Nblock);
std::vector<scomplex> sip(Nblock);
Lattice<scalar> IP(lhs._grid);
IP=localInnerProduct(lhs,rhs);
sliceSum(IP,sip,Orthog);
for(int ss=0;ss<Nblock;ss++){
vec[ss] = TensorRemove(sip[ss]);
}
}
template<class vobj>
static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Orthog) {
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = rhs._grid->GlobalDimensions()[Orthog];
std::vector<ComplexD> ip(Nblock);
sn.resize(Nblock);
sliceInnerProductVector(ip,rhs,rhs,Orthog);
for(int ss=0;ss<Nblock;ss++){
sn[ss] = real(ip[ss]);
}
};
/*
template<class vobj>
static void sliceInnerProductMatrixOld( Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef typename vobj::tensor_reduced scalar;
typedef typename scalar::scalar_object scomplex;
int Nblock = lhs._grid->GlobalDimensions()[Orthog];
std::cout << " sliceInnerProductMatrix Dim "<<Orthog<<" Nblock " << Nblock<<std::endl;
Lattice<scalar> IP(lhs._grid);
std::vector<scomplex> sip(Nblock);
mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
Lattice<vobj> tmp = rhs;
for(int s1=0;s1<Nblock;s1++){
IP=localInnerProduct(lhs,tmp);
sliceSum(IP,sip,Orthog);
std::cout << "InnerProductMatrix ["<<s1<<"] = ";
for(int ss=0;ss<Nblock;ss++){
std::cout << TensorRemove(sip[ss])<<" ";
}
std::cout << std::endl;
for(int ss=0;ss<Nblock;ss++){
mat(ss,(s1+ss)%Nblock) = TensorRemove(sip[ss]);
}
if ( s1!=(Nblock-1) ) {
tmp = Cshift(tmp,Orthog,1);
}
}
}
*/
//////////////////////////////////////////////////////////////////////////
// Block conjugate gradient. Dimension zero should be the block direction
//////////////////////////////////////////////////////////////////////////
template <class Field>
class BlockConjugateGradient : public OperatorFunction<Field> {
public:
typedef typename Field::scalar_type scomplex;
const int blockDim = 0;
int Nblock;
bool ErrorOnNoConverge; // throw an assert when the CG fails to converge.
// Defaults true.
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
BlockConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
: Tolerance(tol),
MaxIterations(maxit),
ErrorOnNoConverge(err_on_no_conv){};
void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
{
int Orthog = 0; // First dimension is block dim
Nblock = Src._grid->_fdimensions[Orthog];
std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<std::endl;
std::cout<<GridLogMessage<<" Block Conjugate Gradient : Nblock "<<Nblock<<std::endl;
Psi.checkerboard = Src.checkerboard;
conformable(Psi, Src);
Field P(Src);
Field AP(Src);
Field R(Src);
GridStopWatch LinalgTimer;
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
Eigen::MatrixXcd m_pAp = Eigen::MatrixXcd::Identity(Nblock,Nblock);
Eigen::MatrixXcd m_pAp_inv= Eigen::MatrixXcd::Identity(Nblock,Nblock);
Eigen::MatrixXcd m_rr = Eigen::MatrixXcd::Zero(Nblock,Nblock);
Eigen::MatrixXcd m_rr_inv = Eigen::MatrixXcd::Zero(Nblock,Nblock);
Eigen::MatrixXcd m_alpha = Eigen::MatrixXcd::Zero(Nblock,Nblock);
Eigen::MatrixXcd m_beta = Eigen::MatrixXcd::Zero(Nblock,Nblock);
// Initial residual computation & set up
std::vector<RealD> residuals(Nblock);
std::vector<RealD> ssq(Nblock);
sliceNorm(ssq,Src,Orthog);
RealD sssum=0;
for(int b=0;b<Nblock;b++) sssum+=ssq[b];
sliceNorm(residuals,Src,Orthog);
for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
sliceNorm(residuals,Psi,Orthog);
for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
// Initial search dir is guess
Linop.HermOp(Psi, AP);
/************************************************************************
* Block conjugate gradient (Stephen Pickles, thesis 1995, pp 71, O Leary 1980)
************************************************************************
* O'Leary : R = B - A X
* O'Leary : P = M R ; preconditioner M = 1
* O'Leary : alpha = PAP^{-1} RMR
* O'Leary : beta = RMR^{-1}_old RMR_new
* O'Leary : X=X+Palpha
* O'Leary : R_new=R_old-AP alpha
* O'Leary : P=MR_new+P beta
*/
R = Src - AP;
P = R;
sliceInnerProductMatrix(m_rr,R,R,Orthog);
int k;
for (k = 1; k <= MaxIterations; k++) {
RealD rrsum=0;
for(int b=0;b<Nblock;b++) rrsum+=real(m_rr(b,b));
std::cout << GridLogIterative << " iteration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
<<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
Linop.HermOp(P, AP);
// Alpha
sliceInnerProductMatrix(m_pAp,P,AP,Orthog);
m_pAp_inv = m_pAp.inverse();
m_alpha = m_pAp_inv * m_rr ;
// Psi, R update
sliceMaddMatrix(Psi,m_alpha, P,Psi,Orthog); // add alpha * P to psi
sliceMaddMatrix(R ,m_alpha,AP, R,Orthog,-1.0);// sub alpha * AP to resid
// Beta
m_rr_inv = m_rr.inverse();
sliceInnerProductMatrix(m_rr,R,R,Orthog);
m_beta = m_rr_inv *m_rr;
// Search update
sliceMaddMatrix(AP,m_beta,P,R,Orthog);
P= AP;
/*********************
* convergence monitor
*********************
*/
RealD max_resid=0;
for(int b=0;b<Nblock;b++){
RealD rr = real(m_rr(b,b))/ssq[b];
if ( rr > max_resid ) max_resid = rr;
}
if ( max_resid < Tolerance*Tolerance ) {
std::cout << GridLogMessage<<" Block solver has converged in "
<<k<<" iterations; max residual is "<<std::sqrt(max_resid)<<std::endl;
for(int b=0;b<Nblock;b++){
std::cout << GridLogMessage<< " block "<<b<<" resid "<< std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
}
Linop.HermOp(Psi, AP);
AP = AP-Src;
std::cout << " Block solver true residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
IterationsToComplete = k;
return;
}
}
std::cout << GridLogMessage << "BlockConjugateGradient did NOT converge" << std::endl;
if (ErrorOnNoConverge) assert(0);
IterationsToComplete = k;
}
};
//////////////////////////////////////////////////////////////////////////
// multiRHS conjugate gradient. Dimension zero should be the block direction
//////////////////////////////////////////////////////////////////////////
template <class Field>
class MultiRHSConjugateGradient : public OperatorFunction<Field> {
public:
typedef typename Field::scalar_type scomplex;
const int blockDim = 0;
int Nblock;
bool ErrorOnNoConverge; // throw an assert when the CG fails to converge.
// Defaults true.
RealD Tolerance;
Integer MaxIterations;
Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
MultiRHSConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true)
: Tolerance(tol),
MaxIterations(maxit),
ErrorOnNoConverge(err_on_no_conv){};
void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
{
int Orthog = 0; // First dimension is block dim
Nblock = Src._grid->_fdimensions[Orthog];
std::cout<<GridLogMessage<<" MultiRHS Conjugate Gradient : Orthog "<<Orthog<<std::endl;
std::cout<<GridLogMessage<<" MultiRHS Conjugate Gradient : Nblock "<<Nblock<<std::endl;
Psi.checkerboard = Src.checkerboard;
conformable(Psi, Src);
Field P(Src);
Field AP(Src);
Field R(Src);
std::vector<ComplexD> v_pAp(Nblock);
std::vector<RealD> v_rr (Nblock);
std::vector<RealD> v_rr_inv(Nblock);
std::vector<RealD> v_alpha(Nblock);
std::vector<RealD> v_beta(Nblock);
// Initial residual computation & set up
std::vector<RealD> residuals(Nblock);
std::vector<RealD> ssq(Nblock);
sliceNorm(ssq,Src,Orthog);
RealD sssum=0;
for(int b=0;b<Nblock;b++) sssum+=ssq[b];
sliceNorm(residuals,Src,Orthog);
for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
sliceNorm(residuals,Psi,Orthog);
for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
// Initial search dir is guess
Linop.HermOp(Psi, AP);
R = Src - AP;
P = R;
sliceNorm(v_rr,R,Orthog);
int k;
for (k = 1; k <= MaxIterations; k++) {
RealD rrsum=0;
for(int b=0;b<Nblock;b++) rrsum+=real(v_rr[b]);
std::cout << GridLogIterative << " iteration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
<<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
Linop.HermOp(P, AP);
// Alpha
sliceInnerProductVector(v_pAp,P,AP,Orthog);
for(int b=0;b<Nblock;b++){
v_alpha[b] = v_rr[b]/real(v_pAp[b]);
}
// Psi, R update
sliceMaddVector(Psi,v_alpha, P,Psi,Orthog); // add alpha * P to psi
sliceMaddVector(R ,v_alpha,AP, R,Orthog,-1.0);// sub alpha * AP to resid
// Beta
for(int b=0;b<Nblock;b++){
v_rr_inv[b] = 1.0/v_rr[b];
}
sliceNorm(v_rr,R,Orthog);
for(int b=0;b<Nblock;b++){
v_beta[b] = v_rr_inv[b] *v_rr[b];
}
// Search update
sliceMaddVector(P,v_beta,P,R,Orthog);
/*********************
* convergence monitor
*********************
*/
RealD max_resid=0;
for(int b=0;b<Nblock;b++){
RealD rr = v_rr[b]/ssq[b];
if ( rr > max_resid ) max_resid = rr;
}
if ( max_resid < Tolerance*Tolerance ) {
std::cout << GridLogMessage<<" MultiRHS solver has converged in "
<<k<<" iterations; max residual is "<<std::sqrt(max_resid)<<std::endl;
for(int b=0;b<Nblock;b++){
std::cout << GridLogMessage<< " block "<<b<<" resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
}
Linop.HermOp(Psi, AP);
AP = AP-Src;
std::cout << " MultiRHS solver true residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
IterationsToComplete = k;
return;
}
}
std::cout << GridLogMessage << "MultiRHSConjugateGradient did NOT converge" << std::endl;
if (ErrorOnNoConverge) assert(0);
IterationsToComplete = k;
}
};
}
#endif

3
lib/algorithms/iterative/ConjugateGradient.h
View File

@ -99,8 +99,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
}
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: k=0 residual " << cp << " target " << rsq
<< std::endl;
<< "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
GridStopWatch LinalgTimer;
GridStopWatch MatrixTimer;

8
lib/lattice/Lattice_transfer.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -359,7 +359,7 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
template<class vobj>
void InsertSlice(Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int orthog)
void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int orthog)
{
typedef typename vobj::scalar_object sobj;
@ -401,7 +401,7 @@ void InsertSlice(Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice, int
}
template<class vobj>
void ExtractSlice(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice, int orthog)
void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slice, int orthog)
{
typedef typename vobj::scalar_object sobj;
@ -444,7 +444,7 @@ void ExtractSlice(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice, in
template<class vobj>
void InsertSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
{
typedef typename vobj::scalar_object sobj;

2
lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
View File

@ -160,8 +160,6 @@ void ImprovedStaggeredFermion<Impl>::ImportGauge(const GaugeField &_Uthin,const
PokeIndex<LorentzIndex>(UUUmu, U*(-0.5*c2/u0/u0/u0), mu+4);
}
std::cout << " Umu " << Umu._odata[0]<<std::endl;
std::cout << " UUUmu " << UUUmu._odata[0]<<std::endl;
pickCheckerboard(Even, UmuEven, Umu);
pickCheckerboard(Odd, UmuOdd , Umu);
pickCheckerboard(Even, UUUmuEven, UUUmu);

49
lib/simd/Grid_avx.h
View File

@ -377,8 +377,8 @@ namespace Optimization {
b0 = _mm256_extractf128_si256(b,0);
a1 = _mm256_extractf128_si256(a,1);
b1 = _mm256_extractf128_si256(b,1);
a0 = _mm_mul_epi32(a0,b0);
a1 = _mm_mul_epi32(a1,b1);
a0 = _mm_mullo_epi32(a0,b0);
a1 = _mm_mullo_epi32(a1,b1);
return _mm256_set_m128i(a1,a0);
#endif
#if defined (AVX2)
@ -471,6 +471,50 @@ namespace Optimization {
};
};
struct PrecisionChange {
static inline __m256i StoH (__m256 a,__m256 b) {
#ifdef USE_FP16
__m128i ha = _mm256_cvtps_ph(a,0);
__m128i hb = _mm256_cvtps_ph(b,0);
__m256 h = _mm256_castps128_ps256(ha);
h = _mm256_insertf128_ps(h,hb,1);
#else
assert(0);
#endif
return h;
}
static inline void HtoS (__m256i h,__m256 &sa,__m256 &sb) {
#ifdef USE_FP16
sa = _mm256_cvtph_ps(_mm256_extractf128_ps(h,0));
sb = _mm256_cvtph_ps(_mm256_extractf128_ps(h,1));
#else
assert(0);
#endif
}
static inline __m256 DtoS (__m256d a,__m256d b) {
__m128 sa = _mm256_cvtpd_ps(a);
__m128 sb = _mm256_cvtpd_ps(b);
__m256 s = _mm256_castps128_ps256(sa);
s = _mm256_insertf128_ps(s,sb,1);
return s;
}
static inline void StoD (__m256 s,__m256d &a,__m256d &b) {
a = _mm256_cvtps_pd(_mm256_extractf128_ps(s,0));
b = _mm256_cvtps_pd(_mm256_extractf128_ps(s,1));
}
static inline __m256i DtoH (__m256d a,__m256d b,__m256d c,__m256d d) {
__m256 sa,sb;
sa = DtoS(a,b);
sb = DtoS(c,d);
return StoH(sa,sb);
}
static inline void HtoD (__m256i h,__m256d &a,__m256d &b,__m256d &c,__m256d &d) {
__m256 sa,sb;
HtoS(h,sa,sb);
StoD(sa,a,b);
StoD(sb,c,d);
}
};
struct Exchange{
// 3210 ordering
static inline void Exchange0(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){
@ -675,6 +719,7 @@ namespace Optimization {
//////////////////////////////////////////////////////////////////////////////////////
// Here assign types
typedef __m256i SIMD_Htype; // Single precision type
typedef __m256 SIMD_Ftype; // Single precision type
typedef __m256d SIMD_Dtype; // Double precision type
typedef __m256i SIMD_Itype; // Integer type

46
lib/simd/Grid_avx512.h
View File

@ -235,11 +235,9 @@ namespace Optimization {
inline void mac(__m512 &a, __m512 b, __m512 c){
a= _mm512_fmadd_ps( b, c, a);
}
inline void mac(__m512d &a, __m512d b, __m512d c){
a= _mm512_fmadd_pd( b, c, a);
}
// Real float
inline __m512 operator()(__m512 a, __m512 b){
return _mm512_mul_ps(a,b);
@ -343,6 +341,50 @@ namespace Optimization {
};
struct PrecisionChange {
static inline __m512i StoH (__m512 a,__m512 b) {
#ifdef USE_FP16
__m256i ha = _mm512_cvtps_ph(a,0);
__m256i hb = _mm512_cvtps_ph(b,0);
__m512 h = _mm512_castps256_ps512(ha);
h = _mm512_insertf256_ps(h,hb,1);
#else
assert(0);
#endif
return h;
}
static inline void HtoS (__m512i h,__m512 &sa,__m512 &sb) {
#ifdef USE_FP16
sa = _mm512_cvtph_ps(_mm512_extractf256_ps(h,0));
sb = _mm512_cvtph_ps(_mm512_extractf256_ps(h,1));
#else
assert(0);
#endif
}
static inline __m512 DtoS (__m512d a,__m512d b) {
__m256 sa = _mm512_cvtpd_ps(a);
__m256 sb = _mm512_cvtpd_ps(b);
__m512 s = _mm512_castps256_ps512(sa);
s = _mm512_insertf256_ps(s,sb,1);
return s;
}
static inline void StoD (__m512 s,__m512d &a,__m512d &b) {
a = _mm512_cvtps_pd(_mm512_extractf256_ps(s,0));
b = _mm512_cvtps_pd(_mm512_extractf256_ps(s,1));
}
static inline __m512i DtoH (__m512d a,__m512d b,__m512d c,__m512d d) {
__m512 sa,sb;
sa = DtoS(a,b);
sb = DtoS(c,d);
return StoH(sa,sb);
}
static inline void HtoD (__m512i h,__m512d &a,__m512d &b,__m512d &c,__m512d &d) {
__m512 sa,sb;
HtoS(h,sa,sb);
StoD(sa,a,b);
StoD(sb,c,d);
}
};
// On extracting face: Ah Al , Bh Bl -> Ah Bh, Al Bl
// On merging buffers: Ah,Bh , Al Bl -> Ah Al, Bh, Bl
// The operation is its own inverse

96
lib/simd/Grid_generic.h
View File

@ -279,6 +279,101 @@ namespace Optimization {
#undef timesi
struct PrecisionChange {
static inline vech StoH (const vecf &a,const vecf &b) {
#ifdef USE_FP16
vech ret;
vech *ha = (vech *)&a;
vech *hb = (vech *)&b;
const int nf = W<float>::r;
// VECTOR_FOR(i, nf,1){ ret.v[i] = ( (uint16_t *) &a.v[i])[1] ; }
// VECTOR_FOR(i, nf,1){ ret.v[i+nf] = ( (uint16_t *) &b.v[i])[1] ; }
VECTOR_FOR(i, nf,1){ ret.v[i] = ha->v[2*i+1]; }
VECTOR_FOR(i, nf,1){ ret.v[i+nf] = hb->v[2*i+1]; }
#else
assert(0);
#endif
return ret;
}
static inline void HtoS (vech h,vecf &sa,vecf &sb) {
#ifdef USE_FP16
const int nf = W<float>::r;
const int nh = W<uint16_t>::r;
vech *ha = (vech *)&sa;
vech *hb = (vech *)&sb;
VECTOR_FOR(i, nf, 1){ sb.v[i]= sa.v[i] = 0; }
// VECTOR_FOR(i, nf, 1){ ( (uint16_t *) (&sa.v[i]))[1] = h.v[i];}
// VECTOR_FOR(i, nf, 1){ ( (uint16_t *) (&sb.v[i]))[1] = h.v[i+nf];}
VECTOR_FOR(i, nf, 1){ ha->v[2*i+1]=h.v[i]; }
VECTOR_FOR(i, nf, 1){ hb->v[2*i+1]=h.v[i+nf]; }
#else
assert(0);
#endif
}
static inline vecf DtoS (vecd a,vecd b) {
const int nd = W<double>::r;
const int nf = W<float>::r;
vecf ret;
VECTOR_FOR(i, nd,1){ ret.v[i] = a.v[i] ; }
VECTOR_FOR(i, nd,1){ ret.v[i+nd] = b.v[i] ; }
return ret;
}
static inline void StoD (vecf s,vecd &a,vecd &b) {
const int nd = W<double>::r;
VECTOR_FOR(i, nd,1){ a.v[i] = s.v[i] ; }
VECTOR_FOR(i, nd,1){ b.v[i] = s.v[i+nd] ; }
}
static inline vech DtoH (vecd a,vecd b,vecd c,vecd d) {
vecf sa,sb;
sa = DtoS(a,b);
sb = DtoS(c,d);
return StoH(sa,sb);
}
static inline void HtoD (vech h,vecd &a,vecd &b,vecd &c,vecd &d) {
vecf sa,sb;
HtoS(h,sa,sb);
StoD(sa,a,b);
StoD(sb,c,d);
}
};
//////////////////////////////////////////////
// Exchange support
struct Exchange{
template <typename T,int n>
static inline void ExchangeN(vec<T> &out1,vec<T> &out2,vec<T> &in1,vec<T> &in2){
const int w = W<T>::r;
unsigned int mask = w >> (n + 1);
// std::cout << " Exchange "<<n<<" nsimd "<<w<<" mask 0x" <<std::hex<<mask<<std::dec<<std::endl;
VECTOR_FOR(i, w, 1) {
int j1 = i&(~mask);
if ( (i&mask) == 0 ) { out1.v[i]=in1.v[j1];}
else { out1.v[i]=in2.v[j1];}
int j2 = i|mask;
if ( (i&mask) == 0 ) { out2.v[i]=in1.v[j2];}
else { out2.v[i]=in2.v[j2];}
}
}
template <typename T>
static inline void Exchange0(vec<T> &out1,vec<T> &out2,vec<T> &in1,vec<T> &in2){
ExchangeN<T,0>(out1,out2,in1,in2);
};
template <typename T>
static inline void Exchange1(vec<T> &out1,vec<T> &out2,vec<T> &in1,vec<T> &in2){
ExchangeN<T,1>(out1,out2,in1,in2);
};
template <typename T>
static inline void Exchange2(vec<T> &out1,vec<T> &out2,vec<T> &in1,vec<T> &in2){
ExchangeN<T,2>(out1,out2,in1,in2);
};
template <typename T>
static inline void Exchange3(vec<T> &out1,vec<T> &out2,vec<T> &in1,vec<T> &in2){
ExchangeN<T,3>(out1,out2,in1,in2);
};
};
//////////////////////////////////////////////
// Some Template specialization
#define perm(a, b, n, w)\
@ -403,6 +498,7 @@ namespace Optimization {
//////////////////////////////////////////////////////////////////////////////////////
// Here assign types
typedef Optimization::vech SIMD_Htype; // Reduced precision type
typedef Optimization::vecf SIMD_Ftype; // Single precision type
typedef Optimization::vecd SIMD_Dtype; // Double precision type
typedef Optimization::veci SIMD_Itype; // Integer type

11
lib/simd/Grid_generic_types.h
View File

@ -66,6 +66,10 @@ namespace Optimization {
template <> struct W<Integer> {
constexpr static unsigned int r = GEN_SIMD_WIDTH/4u;
};
template <> struct W<uint16_t> {
constexpr static unsigned int c = GEN_SIMD_WIDTH/4u;
constexpr static unsigned int r = GEN_SIMD_WIDTH/2u;
};
// SIMD vector types
template <typename T>
@ -73,8 +77,9 @@ namespace Optimization {
alignas(GEN_SIMD_WIDTH) T v[W<T>::r];
};
typedef vec<float> vecf;
typedef vec<double> vecd;
typedef vec<Integer> veci;
typedef vec<float> vecf;
typedef vec<double> vecd;
typedef vec<uint16_t> vech; // half precision comms
typedef vec<Integer> veci;
}}

1
lib/simd/Grid_qpx.h
View File

@ -125,7 +125,6 @@ namespace Optimization {
f[2] = a.v2;
f[3] = a.v3;
}
//Double
inline void operator()(double *d, vector4double a){
vec_st(a, 0, d);

69
lib/simd/Grid_sse4.h
View File

@ -38,6 +38,7 @@ Author: neo <cossu@post.kek.jp>
#include <pmmintrin.h>
namespace Grid {
namespace Optimization {
@ -328,6 +329,56 @@ namespace Optimization {
};
};
#define _my_alignr_epi32(a,b,n) _mm_alignr_epi8(a,b,(n*4)%16)
#define _my_alignr_epi64(a,b,n) _mm_alignr_epi8(a,b,(n*8)%16)
struct PrecisionChange {
static inline __m128i StoH (__m128 a,__m128 b) {
#ifdef USE_FP16
__m128i ha = _mm_cvtps_ph(a,0);
__m128i hb = _mm_cvtps_ph(b,0);
__m128i h =(__m128i) _mm_shuffle_ps((__m128)ha,(__m128)hb,_MM_SELECT_FOUR_FOUR(1,0,1,0));
#else
__m128i h = (__m128i)a;
assert(0);
#endif
return h;
}
static inline void HtoS (__m128i h,__m128 &sa,__m128 &sb) {
#ifdef USE_FP16
sa = _mm_cvtph_ps(h);
h = (__m128i)_my_alignr_epi32((__m128i)h,(__m128i)h,2);
sb = _mm_cvtph_ps(h);
#else
assert(0);
#endif
}
static inline __m128 DtoS (__m128d a,__m128d b) {
__m128 sa = _mm_cvtpd_ps(a);
__m128 sb = _mm_cvtpd_ps(b);
__m128 s = _mm_shuffle_ps(sa,sb,_MM_SELECT_FOUR_FOUR(1,0,1,0));
return s;
}
static inline void StoD (__m128 s,__m128d &a,__m128d &b) {
a = _mm_cvtps_pd(s);
s = (__m128)_my_alignr_epi32((__m128i)s,(__m128i)s,2);
b = _mm_cvtps_pd(s);
}
static inline __m128i DtoH (__m128d a,__m128d b,__m128d c,__m128d d) {
__m128 sa,sb;
sa = DtoS(a,b);
sb = DtoS(c,d);
return StoH(sa,sb);
}
static inline void HtoD (__m128i h,__m128d &a,__m128d &b,__m128d &c,__m128d &d) {
__m128 sa,sb;
HtoS(h,sa,sb);
StoD(sa,a,b);
StoD(sb,c,d);
}
};
struct Exchange{
// 3210 ordering
static inline void Exchange0(__m128 &out1,__m128 &out2,__m128 in1,__m128 in2){
@ -335,8 +386,10 @@ namespace Optimization {
out2= _mm_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,2,3,2));
};
static inline void Exchange1(__m128 &out1,__m128 &out2,__m128 in1,__m128 in2){
out1= _mm_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(2,0,2,0));
out2= _mm_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,1,3,1));
out1= _mm_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(2,0,2,0)); /*ACEG*/
out2= _mm_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,1,3,1)); /*BDFH*/
out1= _mm_shuffle_ps(out1,out1,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
out2= _mm_shuffle_ps(out2,out2,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
};
static inline void Exchange2(__m128 &out1,__m128 &out2,__m128 in1,__m128 in2){
assert(0);
@ -383,14 +436,9 @@ namespace Optimization {
default: assert(0);
}
}
#ifndef _mm_alignr_epi64
#define _mm_alignr_epi32(a,b,n) _mm_alignr_epi8(a,b,(n*4)%16)
#define _mm_alignr_epi64(a,b,n) _mm_alignr_epi8(a,b,(n*8)%16)
#endif
template<int n> static inline __m128 tRotate(__m128 in){ return (__m128)_mm_alignr_epi32((__m128i)in,(__m128i)in,n); };
template<int n> static inline __m128d tRotate(__m128d in){ return (__m128d)_mm_alignr_epi64((__m128i)in,(__m128i)in,n); };
template<int n> static inline __m128 tRotate(__m128 in){ return (__m128)_my_alignr_epi32((__m128i)in,(__m128i)in,n); };
template<int n> static inline __m128d tRotate(__m128d in){ return (__m128d)_my_alignr_epi64((__m128i)in,(__m128i)in,n); };
};
//////////////////////////////////////////////
@ -450,7 +498,8 @@ namespace Optimization {
//////////////////////////////////////////////////////////////////////////////////////
// Here assign types
typedef __m128 SIMD_Ftype; // Single precision type
typedef __m128i SIMD_Htype; // Single precision type
typedef __m128 SIMD_Ftype; // Single precision type
typedef __m128d SIMD_Dtype; // Double precision type
typedef __m128i SIMD_Itype; // Integer type

65
lib/simd/Grid_vector_types.h
View File

@ -358,16 +358,12 @@ class Grid_simd {
{
if (n==3) {
Optimization::Exchange::Exchange3(out1.v,out2.v,in1.v,in2.v);
// std::cout << " Exchange3 "<< out1<<" "<< out2<<" <- " << in1 << " "<<in2<<std::endl;
} else if(n==2) {
Optimization::Exchange::Exchange2(out1.v,out2.v,in1.v,in2.v);
// std::cout << " Exchange2 "<< out1<<" "<< out2<<" <- " << in1 << " "<<in2<<std::endl;
} else if(n==1) {
Optimization::Exchange::Exchange1(out1.v,out2.v,in1.v,in2.v);
// std::cout << " Exchange1 "<< out1<<" "<< out2<<" <- " << in1 << " "<<in2<<std::endl;
} else if(n==0) {
Optimization::Exchange::Exchange0(out1.v,out2.v,in1.v,in2.v);
// std::cout << " Exchange0 "<< out1<<" "<< out2<<" <- " << in1 << " "<<in2<<std::endl;
}
}
@ -758,6 +754,67 @@ typedef Grid_simd<std::complex<float>, SIMD_Ftype> vComplexF;
typedef Grid_simd<std::complex<double>, SIMD_Dtype> vComplexD;
typedef Grid_simd<Integer, SIMD_Itype> vInteger;
// Half precision; no arithmetic support
typedef Grid_simd<uint16_t, SIMD_Htype> vRealH;
typedef Grid_simd<std::complex<uint16_t>, SIMD_Htype> vComplexH;
inline void precisionChange(vRealF *out,vRealD *in,int nvec)
{
assert((nvec&0x1)==0);
for(int m=0;m*2<nvec;m++){
int n=m*2;
out[m].v=Optimization::PrecisionChange::DtoS(in[n].v,in[n+1].v);
}
}
inline void precisionChange(vRealH *out,vRealD *in,int nvec)
{
assert((nvec&0x3)==0);
for(int m=0;m*4<nvec;m++){
int n=m*4;
out[m].v=Optimization::PrecisionChange::DtoH(in[n].v,in[n+1].v,in[n+2].v,in[n+3].v);
}
}
inline void precisionChange(vRealH *out,vRealF *in,int nvec)
{
assert((nvec&0x1)==0);
for(int m=0;m*2<nvec;m++){
int n=m*2;
out[m].v=Optimization::PrecisionChange::StoH(in[n].v,in[n+1].v);
}
}
inline void precisionChange(vRealD *out,vRealF *in,int nvec)
{
assert((nvec&0x1)==0);
for(int m=0;m*2<nvec;m++){
int n=m*2;
Optimization::PrecisionChange::StoD(in[m].v,out[n].v,out[n+1].v);
}
}
inline void precisionChange(vRealD *out,vRealH *in,int nvec)
{
assert((nvec&0x3)==0);
for(int m=0;m*4<nvec;m++){
int n=m*4;
Optimization::PrecisionChange::HtoD(in[m].v,out[n].v,out[n+1].v,out[n+2].v,out[n+3].v);
}
}
inline void precisionChange(vRealF *out,vRealH *in,int nvec)
{
assert((nvec&0x1)==0);
for(int m=0;m*2<nvec;m++){
int n=m*2;
Optimization::PrecisionChange::HtoS(in[m].v,out[n].v,out[n+1].v);
}
}
inline void precisionChange(vComplexF *out,vComplexD *in,int nvec){ precisionChange((vRealF *)out,(vRealD *)in,nvec);}
inline void precisionChange(vComplexH *out,vComplexD *in,int nvec){ precisionChange((vRealH *)out,(vRealD *)in,nvec);}
inline void precisionChange(vComplexH *out,vComplexF *in,int nvec){ precisionChange((vRealH *)out,(vRealF *)in,nvec);}
inline void precisionChange(vComplexD *out,vComplexF *in,int nvec){ precisionChange((vRealD *)out,(vRealF *)in,nvec);}
inline void precisionChange(vComplexD *out,vComplexH *in,int nvec){ precisionChange((vRealD *)out,(vRealH *)in,nvec);}
inline void precisionChange(vComplexF *out,vComplexH *in,int nvec){ precisionChange((vRealF *)out,(vRealH *)in,nvec);}
// Check our vector types are of an appropriate size.
#if defined QPX
static_assert(2*sizeof(SIMD_Ftype) == sizeof(SIMD_Dtype), "SIMD vector lengths incorrect");

1
lib/util/Init.cc
View File

@ -457,5 +457,6 @@ void Grid_debug_handler_init(void)
sigaction(SIGFPE,&sa,NULL);
sigaction(SIGKILL,&sa,NULL);
sigaction(SIGILL,&sa,NULL);
}
}

110
tests/Test_simd.cc
View File

@ -308,18 +308,23 @@ public:
int n;
funcExchange(int _n) { n=_n;};
template<class vec> void operator()(vec &r1,vec &r2,vec &i1,vec &i2) const { exchange(r1,r2,i1,i2,n);}
template<class scal> void apply(std::vector<scal> &r1,std::vector<scal> &r2,std::vector<scal> &in1,std::vector<scal> &in2) const {
template<class scal> void apply(std::vector<scal> &r1,
std::vector<scal> &r2,
std::vector<scal> &in1,
std::vector<scal> &in2) const
{
int sz=in1.size();
int msk = sz>>(n+1);
int j1=0;
int j2=0;
for(int i=0;i<sz;i++) if ( (i&msk) == 0 ) r1[j1++] = in1[ i ];
for(int i=0;i<sz;i++) if ( (i&msk) == 0 ) r1[j1++] = in2[ i ];
for(int i=0;i<sz;i++) if ( (i&msk) ) r2[j2++] = in1[ i ];
for(int i=0;i<sz;i++) if ( (i&msk) ) r2[j2++] = in2[ i ];
for(int i=0;i<sz;i++) {
int j1 = i&(~msk);
int j2 = i|msk;
if ( (i&msk) == 0 ) { r1[i]=in1[j1];}
else { r1[i]=in2[j1];}
if ( (i&msk) == 0 ) { r2[i]=in1[j2];}
else { r2[i]=in2[j2];}
}
}
std::string name(void) const { return std::string("Exchange"); }
};
@ -454,8 +459,8 @@ void ExchangeTester(const functor &func)
std::cout<<GridLogMessage << " " << func.name() << " " <<func.n <<std::endl;
// for(int i=0;i<Nsimd;i++) std::cout << " i "<<i<<" "<<reference1[i]<<" "<<result1[i]<<std::endl;
// for(int i=0;i<Nsimd;i++) std::cout << " i "<<i<<" "<<reference2[i]<<" "<<result2[i]<<std::endl;
//for(int i=0;i<Nsimd;i++) std::cout << " i "<<i<<" ref "<<reference1[i]<<" res "<<result1[i]<<std::endl;
//for(int i=0;i<Nsimd;i++) std::cout << " i "<<i<<" ref "<<reference2[i]<<" res "<<result2[i]<<std::endl;
for(int i=0;i<Nsimd;i++){
int found=0;
@ -465,7 +470,7 @@ void ExchangeTester(const functor &func)
// std::cout << " i "<<i<<" j "<<j<<" "<<reference1[j]<<" "<<result1[i]<<std::endl;
}
}
assert(found==1);
// assert(found==1);
}
for(int i=0;i<Nsimd;i++){
int found=0;
@ -475,15 +480,24 @@ void ExchangeTester(const functor &func)
// std::cout << " i "<<i<<" j "<<j<<" "<<reference2[j]<<" "<<result2[i]<<std::endl;
}
}
assert(found==1);
// assert(found==1);
}
/*
for(int i=0;i<Nsimd;i++){
std::cout << " i "<< i
<<" result1 "<<result1[i]
<<" result2 "<<result2[i]
<<" test1 "<<test1[i]
<<" test2 "<<test2[i]
<<" input1 "<<input1[i]
<<" input2 "<<input2[i]<<std::endl;
}
*/
for(int i=0;i<Nsimd;i++){
assert(test1[i]==input1[i]);
assert(test2[i]==input2[i]);
}// std::cout << " i "<< i<<" test1"<<test1[i]<<" "<<input1[i]<<std::endl;
// std::cout << " i "<< i<<" test2"<<test2[i]<<" "<<input2[i]<<std::endl;
// }
}
}
@ -678,5 +692,69 @@ int main (int argc, char ** argv)
IntTester(funcMinus());
IntTester(funcTimes());
std::cout<<GridLogMessage << "==================================="<< std::endl;
std::cout<<GridLogMessage << "Testing precisionChange "<< std::endl;
std::cout<<GridLogMessage << "==================================="<< std::endl;
{
GridSerialRNG sRNG;
sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
const int Ndp = 16;
const int Nsp = Ndp/2;
const int Nhp = Ndp/4;
std::vector<vRealH,alignedAllocator<vRealH> > H (Nhp);
std::vector<vRealF,alignedAllocator<vRealF> > F (Nsp);
std::vector<vRealF,alignedAllocator<vRealF> > FF(Nsp);
std::vector<vRealD,alignedAllocator<vRealD> > D (Ndp);
std::vector<vRealD,alignedAllocator<vRealD> > DD(Ndp);
for(int i=0;i<16;i++){
random(sRNG,D[i]);
}
// Double to Single
precisionChange(&F[0],&D[0],Ndp);
precisionChange(&DD[0],&F[0],Ndp);
std::cout << GridLogMessage<<"Double to single";
for(int i=0;i<Ndp;i++){
// std::cout << "DD["<<i<<"] = "<< DD[i]<<" "<<D[i]<<" "<<DD[i]-D[i] <<std::endl;
DD[i] = DD[i] - D[i];
decltype(innerProduct(DD[0],DD[0])) nrm;
nrm = innerProduct(DD[i],DD[i]);
auto tmp = Reduce(nrm);
// std::cout << tmp << std::endl;
assert( tmp < 1.0e-14 );
}
std::cout <<" OK ! "<<std::endl;
// Double to Half
#ifdef USE_FP16
std::cout << GridLogMessage<< "Double to half" ;
precisionChange(&H[0],&D[0],Ndp);
precisionChange(&DD[0],&H[0],Ndp);
for(int i=0;i<Ndp;i++){
// std::cout << "DD["<<i<<"] = "<< DD[i]<<" "<<D[i]<<" "<<DD[i]-D[i]<<std::endl;
DD[i] = DD[i] - D[i];
decltype(innerProduct(DD[0],DD[0])) nrm;
nrm = innerProduct(DD[i],DD[i]);
auto tmp = Reduce(nrm);
// std::cout << tmp << std::endl;
assert( tmp < 1.0e-3 );
}
std::cout <<" OK ! "<<std::endl;
std::cout << GridLogMessage<< "Single to half";
// Single to Half
precisionChange(&H[0] ,&F[0],Nsp);
precisionChange(&FF[0],&H[0],Nsp);
for(int i=0;i<Nsp;i++){
// std::cout << "FF["<<i<<"] = "<< FF[i]<<" "<<F[i]<<" "<<FF[i]-F[i]<<std::endl;
FF[i] = FF[i] - F[i];
decltype(innerProduct(FF[0],FF[0])) nrm;
nrm = innerProduct(FF[i],FF[i]);
auto tmp = Reduce(nrm);
// std::cout << tmp << std::endl;
assert( tmp < 1.0e-3 );
}
std::cout <<" OK ! "<<std::endl;
#endif
}
Grid_finalize();
}

99
tests/solver/Test_staggered_block_cg_unprec.cc Normal file
View File

@ -0,0 +1,99 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./tests/Test_wilson_cg_unprec.cc
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
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>
#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
using namespace std;
using namespace Grid;
using namespace Grid::QCD;
template<class d>
struct scal {
d internal;
};
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT
};
int main (int argc, char ** argv)
{
typedef typename ImprovedStaggeredFermion5DR::FermionField FermionField;
typedef typename ImprovedStaggeredFermion5DR::ComplexField ComplexField;
typename ImprovedStaggeredFermion5DR::ImplParams params;
const int Ls=8;
Grid_init(&argc,&argv);
std::vector<int> latt_size = GridDefaultLatt();
std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
std::vector<int> mpi_layout = GridDefaultMpi();
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
std::vector<int> seeds({1,2,3,4});
GridParallelRNG pRNG(UGrid ); pRNG.SeedFixedIntegers(seeds);
GridParallelRNG pRNG5(FGrid); pRNG5.SeedFixedIntegers(seeds);
FermionField src(FGrid); random(pRNG5,src);
FermionField result(FGrid); result=zero;
RealD nrm = norm2(src);
LatticeGaugeField Umu(UGrid); SU3::HotConfiguration(pRNG,Umu);
RealD mass=0.01;
ImprovedStaggeredFermion5DR Ds(Umu,Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass);
MdagMLinearOperator<ImprovedStaggeredFermion5DR,FermionField> HermOp(Ds);
ConjugateGradient<FermionField> CG(1.0e-8,10000);
BlockConjugateGradient<FermionField> BCG(1.0e-8,10000);
MultiRHSConjugateGradient<FermionField> mCG(1.0e-8,10000);
std::cout << GridLogMessage << " Calling CG "<<std::endl;
result=zero;
CG(HermOp,src,result);
std::cout << GridLogMessage << " Calling multiRHS CG "<<std::endl;
result=zero;
mCG(HermOp,src,result);
std::cout << GridLogMessage << " Calling Block CG "<<std::endl;
result=zero;
BCG(HermOp,src,result);
Grid_finalize();
}

83
tests/solver/Test_staggered_cg_unprec.cc Normal file
View File

@ -0,0 +1,83 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./tests/Test_wilson_cg_unprec.cc
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
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>
#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
using namespace std;
using namespace Grid;
using namespace Grid::QCD;
template<class d>
struct scal {
d internal;
};
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT
};
int main (int argc, char ** argv)
{
typedef typename ImprovedStaggeredFermionR::FermionField FermionField;
typedef typename ImprovedStaggeredFermionR::ComplexField ComplexField;
typename ImprovedStaggeredFermionR::ImplParams params;
Grid_init(&argc,&argv);
std::vector<int> latt_size = GridDefaultLatt();
std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
std::vector<int> mpi_layout = GridDefaultMpi();
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
GridRedBlackCartesian RBGrid(latt_size,simd_layout,mpi_layout);
std::vector<int> seeds({1,2,3,4});
GridParallelRNG pRNG(&Grid); pRNG.SeedFixedIntegers(seeds);
FermionField src(&Grid); random(pRNG,src);
RealD nrm = norm2(src);
FermionField result(&Grid); result=zero;
LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
double volume=1;
for(int mu=0;mu<Nd;mu++){
volume=volume*latt_size[mu];
}
RealD mass=0.1;
ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass);
MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp(Ds);
ConjugateGradient<FermionField> CG(1.0e-8,10000);
CG(HermOp,src,result);
Grid_finalize();
}