portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-24 20:55:55 +01:00
Code Releases Activity

Merge branch 'develop' into feature/half-prec-comms

Browse Source
This commit is contained in:
paboyle 2017-04-18 11:39:39 +01:00
commit ec18e9f7f6
28 changed files with 738 additions and 1131 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

62
TODO
View File

@ -1,6 +1,27 @@
TODO:
---------------
Peter's work list:
1)- Half-precision comms <-- started -- SIMD is prepared
2)- Precision conversion and sort out localConvert <--
3)- Remove DenseVector, DenseMatrix; Use Eigen instead. <-- started
4)- Binary I/O speed up & x-strips
-- Profile CG, BlockCG, etc... Flop count/rate -- PARTIAL, time but no flop/s yet
-- Physical propagator interface
-- Conserved currents
-- GaugeFix into central location
-- Multigrid Wilson and DWF, compare to other Multigrid implementations
-- HDCR resume
Recent DONE
-- Merge high precision reduction into develop
-- multiRHS DWF; benchmark on Cori/BNL for comms elimination
-- slice* linalg routines for multiRHS, BlockCG
-----
* Forces; the UdSdU term in gauge force term is half of what I think it should
be. This is a consequence of taking ONLY the first term in:
@ -21,16 +42,8 @@ TODO:
This means we must double the force in the Test_xxx_force routines, and is the origin of the factor of two.
This 2x is applied by hand in the fermion routines and in the Test_rect_force routine.
Policies:
* Link smearing/boundary conds; Policy class based implementation ; framework more in place
* Support different boundary conditions (finite temp, chem. potential ... )
* Support different fermion representations?
- contained entirely within the integrator presently
- Sign of force term.
- Reversibility test.
@ -41,11 +54,6 @@ Policies:
- Audit oIndex usage for cb behaviour
- Rectangle gauge actions.
Iwasaki,
Symanzik,
... etc...
- Prepare multigrid for HMC. - Alternate setup schemes.
- Support for ILDG --- ugly, not done
@ -55,9 +63,11 @@ Policies:
- FFTnD ?
- Gparity; hand opt use template specialisation elegance to enable the optimised paths ?
- Gparity force term; Gparity (R)HMC.
- Random number state save restore
- Mobius implementation clean up to rmove #if 0 stale code sequences
- CG -- profile carefully, kernel fusion, whole CG performance measurements.
================================================================
@ -90,6 +100,7 @@ Insert/Extract
Not sure of status of this -- reverify. Things are working nicely now though.
* Make the Tensor types and Complex etc... play more nicely.
- TensorRemove is a hack, come up with a long term rationalised approach to Complex vs. Scalar<Scalar<Scalar<Complex > > >
QDP forces use of "toDouble" to get back to non tensor scalar. This role is presently taken TensorRemove, but I
want to introduce a syntax that does not require this.
@ -112,6 +123,8 @@ Not sure of status of this -- reverify. Things are working nicely now though.
RECENT
---------------
- Support different fermion representations? -- DONE
- contained entirely within the integrator presently
- Clean up HMC -- DONE
- LorentzScalar<GaugeField> gets Gauge link type (cleaner). -- DONE
- Simplified the integrators a bit. -- DONE
@ -123,6 +136,26 @@ RECENT
- Parallel io improvements -- DONE
- Plaquette and link trace checks into nersc reader from the Grid_nersc_io.cc test. -- DONE
DONE:
- MultiArray -- MultiRHS done
- ConjugateGradientMultiShift -- DONE
- MCR -- DONE
- Remez -- Mike or Boost? -- DONE
- Proto (ET) -- DONE
- uBlas -- DONE ; Eigen
- Potentially Useful Boost libraries -- DONE ; Eigen
- Aligned allocator; memory pool -- DONE
- Multiprecision -- DONE
- Serialization -- DONE
- Regex -- Not needed
- Tokenize -- Why?
- Random number state save restore -- DONE
- Rectangle gauge actions. -- DONE
Iwasaki,
Symanzik,
... etc...
Done: Cayley, Partial , ContFrac force terms.
DONE
@ -207,6 +240,7 @@ Done
FUNCTIONALITY: it pleases me to keep track of things I have done (keeps me arguably sane)
======================================================================================================
* Link smearing/boundary conds; Policy class based implementation ; framework more in place -- DONE
* Command line args for geometry, simd, etc. layout. Is it necessary to have -- DONE
user pass these? Is this a QCD specific?

2
lib/algorithms/Algorithms.h
View File

@ -46,7 +46,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
// Lanczos support
#include <Grid/algorithms/iterative/MatrixUtils.h>
//#include <Grid/algorithms/iterative/MatrixUtils.h>
#include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
#include <Grid/algorithms/CoarsenedMatrix.h>
#include <Grid/algorithms/FFT.h>

0
lib/algorithms/iterative/DenseMatrix.h → lib/algorithms/densematrix/DenseMatrix.h
View File

0
lib/algorithms/iterative/Francis.h → lib/algorithms/densematrix/Francis.h
View File

0
lib/algorithms/iterative/Householder.h → lib/algorithms/densematrix/Householder.h
View File

288
lib/algorithms/iterative/BlockConjugateGradient.h
View File

@ -30,210 +30,9 @@ directory
#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
//////////////////////////////////////////////////////////////////////////
@ -261,8 +60,8 @@ 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;
std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
Psi.checkerboard = Src.checkerboard;
conformable(Psi, Src);
@ -271,10 +70,6 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
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);
@ -317,33 +112,49 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
P = R;
sliceInnerProductMatrix(m_rr,R,R,Orthog);
GridStopWatch sliceInnerTimer;
GridStopWatch sliceMaddTimer;
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
SolverTimer.Start();
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::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
<<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
MatrixTimer.Start();
Linop.HermOp(P, AP);
MatrixTimer.Stop();
// Alpha
sliceInnerTimer.Start();
sliceInnerProductMatrix(m_pAp,P,AP,Orthog);
sliceInnerTimer.Stop();
m_pAp_inv = m_pAp.inverse();
m_alpha = m_pAp_inv * m_rr ;
// Psi, R update
sliceMaddTimer.Start();
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
sliceMaddTimer.Stop();
// Beta
m_rr_inv = m_rr.inverse();
sliceInnerTimer.Start();
sliceInnerProductMatrix(m_rr,R,R,Orthog);
sliceInnerTimer.Stop();
m_beta = m_rr_inv *m_rr;
// Search update
sliceMaddTimer.Start();
sliceMaddMatrix(AP,m_beta,P,R,Orthog);
sliceMaddTimer.Stop();
P= AP;
/*********************
@ -358,16 +169,24 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
if ( max_resid < Tolerance*Tolerance ) {
std::cout << GridLogMessage<<" Block solver has converged in "
<<k<<" iterations; max residual is "<<std::sqrt(max_resid)<<std::endl;
SolverTimer.Stop();
std::cout << GridLogMessage<<"BlockCG converged in "<<k<<" iterations"<<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;
std::cout << GridLogMessage<< "\t\tblock "<<b<<" resid "<< std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
}
std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
Linop.HermOp(Psi, AP);
AP = AP-Src;
std::cout << " Block solver true residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
std::cout << GridLogMessage <<"\tTrue residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tInnerProd " << sliceInnerTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed() <<std::endl;
IterationsToComplete = k;
return;
}
@ -408,8 +227,8 @@ 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;
std::cout<<GridLogMessage<<"MultiRHS Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
Psi.checkerboard = Src.checkerboard;
conformable(Psi, Src);
@ -445,38 +264,57 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
P = R;
sliceNorm(v_rr,R,Orthog);
GridStopWatch sliceInnerTimer;
GridStopWatch sliceMaddTimer;
GridStopWatch sliceNormTimer;
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
SolverTimer.Start();
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::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
<<" / "<<std::sqrt(rrsum/sssum) <<std::endl;
MatrixTimer.Start();
Linop.HermOp(P, AP);
MatrixTimer.Stop();
// Alpha
// sliceInnerProductVectorTest(v_pAp_test,P,AP,Orthog);
sliceInnerTimer.Start();
sliceInnerProductVector(v_pAp,P,AP,Orthog);
sliceInnerTimer.Stop();
for(int b=0;b<Nblock;b++){
// std::cout << " "<< v_pAp[b]<<" "<< v_pAp_test[b]<<std::endl;
v_alpha[b] = v_rr[b]/real(v_pAp[b]);
}
// Psi, R update
sliceMaddTimer.Start();
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
sliceMaddTimer.Stop();
// Beta
for(int b=0;b<Nblock;b++){
v_rr_inv[b] = 1.0/v_rr[b];
}
sliceNormTimer.Start();
sliceNorm(v_rr,R,Orthog);
sliceNormTimer.Stop();
for(int b=0;b<Nblock;b++){
v_beta[b] = v_rr_inv[b] *v_rr[b];
}
// Search update
sliceMaddTimer.Start();
sliceMaddVector(P,v_beta,P,R,Orthog);
sliceMaddTimer.Stop();
/*********************
* convergence monitor
@ -489,15 +327,27 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
}
if ( max_resid < Tolerance*Tolerance ) {
std::cout << GridLogMessage<<" MultiRHS solver has converged in "
<<k<<" iterations; max residual is "<<std::sqrt(max_resid)<<std::endl;
SolverTimer.Stop();
std::cout << GridLogMessage<<"MultiRHS solver converged in " <<k<<" iterations"<<std::endl;
for(int b=0;b<Nblock;b++){
std::cout << GridLogMessage<< " block "<<b<<" resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
std::cout << GridLogMessage<< "\t\tBlock "<<b<<" resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
}
std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
Linop.HermOp(Psi, AP);
AP = AP-Src;
std::cout << " MultiRHS solver true residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
std::cout <<GridLogMessage << "\tTrue residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
std::cout << GridLogMessage << "Time Breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tInnerProd " << sliceInnerTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tNorm " << sliceNormTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMaddMatrix " << sliceMaddTimer.Elapsed() <<std::endl;
IterationsToComplete = k;
return;
}

39
lib/algorithms/iterative/ConjugateGradient.h
View File

@ -78,18 +78,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
cp = a;
ssq = norm2(src);
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: guess " << guess << std::endl;
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: src " << ssq << std::endl;
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: mp " << d << std::endl;
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: mmp " << b << std::endl;
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: cp,r " << cp << std::endl;
std::cout << GridLogIterative << std::setprecision(4)
<< "ConjugateGradient: p " << a << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: guess " << guess << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: src " << ssq << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: mp " << d << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: mmp " << b << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: cp,r " << cp << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: p " << a << std::endl;
RealD rsq = Tolerance * Tolerance * ssq;
@ -144,19 +138,20 @@ class ConjugateGradient : public OperatorFunction<Field> {
RealD resnorm = sqrt(norm2(p));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage
<< "ConjugateGradient: Converged on iteration " << k << std::endl;
std::cout << GridLogMessage << "Computed residual " << sqrt(cp / ssq)
<< " true residual " << true_residual << " target "
<< Tolerance << std::endl;
std::cout << GridLogMessage << "Time elapsed: Iterations "
<< SolverTimer.Elapsed() << " Matrix "
<< MatrixTimer.Elapsed() << " Linalg "
<< LinalgTimer.Elapsed();
std::cout << std::endl;
std::cout << GridLogMessage << "ConjugateGradient Converged on iteration " << k << std::endl;
std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
std::cout << GridLogMessage << "Time breakdown "<<std::endl;
std::cout << GridLogMessage << "\tElapsed " << SolverTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tMatrix " << MatrixTimer.Elapsed() <<std::endl;
std::cout << GridLogMessage << "\tLinalg " << LinalgTimer.Elapsed() <<std::endl;
if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
IterationsToComplete = k;
return;
}
}

9
lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@ -30,6 +30,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define GRID_IRL_H
#include <string.h> //memset
#ifdef USE_LAPACK
void LAPACK_dstegr(char *jobz, char *range, int *n, double *d, double *e,
double *vl, double *vu, int *il, int *iu, double *abstol,
@ -37,8 +38,9 @@ void LAPACK_dstegr(char *jobz, char *range, int *n, double *d, double *e,
double *work, int *lwork, int *iwork, int *liwork,
int *info);
#endif
#include "DenseMatrix.h"
#include "EigenSort.h"
#include <Grid/algorithms/densematrix/DenseMatrix.h>
#include <Grid/algorithms/iterative/EigenSort.h>
namespace Grid {
@ -1088,8 +1090,6 @@ static void Lock(DenseMatrix<T> &H, // Hess mtx
int dfg,
bool herm)
{
//ForceTridiagonal(H);
int M = H.dim;
@ -1121,7 +1121,6 @@ static void Lock(DenseMatrix<T> &H, // Hess mtx
AH = Hermitian(QQ)*AH;
AH = AH*QQ;
for(int i=con;i<M;i++){
for(int j=con;j<M;j++){

453
lib/algorithms/iterative/Matrix.h
View File

@ -1,453 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/Matrix.h
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef MATRIX_H
#define MATRIX_H
#include <cstdlib>
#include <string>
#include <cmath>
#include <vector>
#include <iostream>
#include <iomanip>
#include <complex>
#include <typeinfo>
#include <Grid/Grid.h>
/** Sign function **/
template <class T> T sign(T p){return ( p/abs(p) );}
/////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////// Hijack STL containers for our wicked means /////////////////////////////////////////
/////////////////////////////////////////////////////////////////////////////////////////////////////////
template<class T> using Vector = Vector<T>;
template<class T> using Matrix = Vector<Vector<T> >;
template<class T> void Resize(Vector<T > & vec, int N) { vec.resize(N); }
template<class T> void Resize(Matrix<T > & mat, int N, int M) {
mat.resize(N);
for(int i=0;i<N;i++){
mat[i].resize(M);
}
}
template<class T> void Size(Vector<T> & vec, int &N)
{
N= vec.size();
}
template<class T> void Size(Matrix<T> & mat, int &N,int &M)
{
N= mat.size();
M= mat[0].size();
}
template<class T> void SizeSquare(Matrix<T> & mat, int &N)
{
int M; Size(mat,N,M);
assert(N==M);
}
template<class T> void SizeSame(Matrix<T> & mat1,Matrix<T> &mat2, int &N1,int &M1)
{
int N2,M2;
Size(mat1,N1,M1);
Size(mat2,N2,M2);
assert(N1==N2);
assert(M1==M2);
}
//*****************************************
//* (Complex) Vector operations *
//*****************************************
/**Conj of a Vector **/
template <class T> Vector<T> conj(Vector<T> p){
Vector<T> q(p.size());
for(int i=0;i<p.size();i++){q[i] = conj(p[i]);}
return q;
}
/** Norm of a Vector**/
template <class T> T norm(Vector<T> p){
T sum = 0;
for(int i=0;i<p.size();i++){sum = sum + p[i]*conj(p[i]);}
return abs(sqrt(sum));
}
/** Norm squared of a Vector **/
template <class T> T norm2(Vector<T> p){
T sum = 0;
for(int i=0;i<p.size();i++){sum = sum + p[i]*conj(p[i]);}
return abs((sum));
}
/** Sum elements of a Vector **/
template <class T> T trace(Vector<T> p){
T sum = 0;
for(int i=0;i<p.size();i++){sum = sum + p[i];}
return sum;
}
/** Fill a Vector with constant c **/
template <class T> void Fill(Vector<T> &p, T c){
for(int i=0;i<p.size();i++){p[i] = c;}
}
/** Normalize a Vector **/
template <class T> void normalize(Vector<T> &p){
T m = norm(p);
if( abs(m) > 0.0) for(int i=0;i<p.size();i++){p[i] /= m;}
}
/** Vector by scalar **/
template <class T, class U> Vector<T> times(Vector<T> p, U s){
for(int i=0;i<p.size();i++){p[i] *= s;}
return p;
}
template <class T, class U> Vector<T> times(U s, Vector<T> p){
for(int i=0;i<p.size();i++){p[i] *= s;}
return p;
}
/** inner product of a and b = conj(a) . b **/
template <class T> T inner(Vector<T> a, Vector<T> b){
T m = 0.;
for(int i=0;i<a.size();i++){m = m + conj(a[i])*b[i];}
return m;
}
/** sum of a and b = a + b **/
template <class T> Vector<T> add(Vector<T> a, Vector<T> b){
Vector<T> m(a.size());
for(int i=0;i<a.size();i++){m[i] = a[i] + b[i];}
return m;
}
/** sum of a and b = a - b **/
template <class T> Vector<T> sub(Vector<T> a, Vector<T> b){
Vector<T> m(a.size());
for(int i=0;i<a.size();i++){m[i] = a[i] - b[i];}
return m;
}
/**
*********************************
* Matrices *
*********************************
**/
template<class T> void Fill(Matrix<T> & mat, T&val) {
int N,M;
Size(mat,N,M);
for(int i=0;i<N;i++){
for(int j=0;j<M;j++){
mat[i][j] = val;
}}
}
/** Transpose of a matrix **/
Matrix<T> Transpose(Matrix<T> & mat){
int N,M;
Size(mat,N,M);
Matrix C; Resize(C,M,N);
for(int i=0;i<M;i++){
for(int j=0;j<N;j++){
C[i][j] = mat[j][i];
}}
return C;
}
/** Set Matrix to unit matrix **/
template<class T> void Unity(Matrix<T> &mat){
int N; SizeSquare(mat,N);
for(int i=0;i<N;i++){
for(int j=0;j<N;j++){
if ( i==j ) A[i][j] = 1;
else A[i][j] = 0;
}
}
}
/** Add C * I to matrix **/
template<class T>
void PlusUnit(Matrix<T> & A,T c){
int dim; SizeSquare(A,dim);
for(int i=0;i<dim;i++){A[i][i] = A[i][i] + c;}
}
/** return the Hermitian conjugate of matrix **/
Matrix<T> HermitianConj(Matrix<T> &mat){
int dim; SizeSquare(mat,dim);
Matrix<T> C; Resize(C,dim,dim);
for(int i=0;i<dim;i++){
for(int j=0;j<dim;j++){
C[i][j] = conj(mat[j][i]);
}
}
return C;
}
/** return diagonal entries as a Vector **/
Vector<T> diag(Matrix<T> &A)
{
int dim; SizeSquare(A,dim);
Vector<T> d; Resize(d,dim);
for(int i=0;i<dim;i++){
d[i] = A[i][i];
}
return d;
}
/** Left multiply by a Vector **/
Vector<T> operator *(Vector<T> &B,Matrix<T> &A)
{
int K,M,N;
Size(B,K);
Size(A,M,N);
assert(K==M);
Vector<T> C; Resize(C,N);
for(int j=0;j<N;j++){
T sum = 0.0;
for(int i=0;i<M;i++){
sum += B[i] * A[i][j];
}
C[j] = sum;
}
return C;
}
/** return 1/diagonal entries as a Vector **/
Vector<T> inv_diag(Matrix<T> & A){
int dim; SizeSquare(A,dim);
Vector<T> d; Resize(d,dim);
for(int i=0;i<dim;i++){
d[i] = 1.0/A[i][i];
}
return d;
}
/** Matrix Addition **/
inline Matrix<T> operator + (Matrix<T> &A,Matrix<T> &B)
{
int N,M ; SizeSame(A,B,N,M);
Matrix C; Resize(C,N,M);
for(int i=0;i<N;i++){
for(int j=0;j<M;j++){
C[i][j] = A[i][j] + B[i][j];
}
}
return C;
}
/** Matrix Subtraction **/
inline Matrix<T> operator- (Matrix<T> & A,Matrix<T> &B){
int N,M ; SizeSame(A,B,N,M);
Matrix C; Resize(C,N,M);
for(int i=0;i<N;i++){
for(int j=0;j<M;j++){
C[i][j] = A[i][j] - B[i][j];
}}
return C;
}
/** Matrix scalar multiplication **/
inline Matrix<T> operator* (Matrix<T> & A,T c){
int N,M; Size(A,N,M);
Matrix C; Resize(C,N,M);
for(int i=0;i<N;i++){
for(int j=0;j<M;j++){
C[i][j] = A[i][j]*c;
}}
return C;
}
/** Matrix Matrix multiplication **/
inline Matrix<T> operator* (Matrix<T> &A,Matrix<T> &B){
int K,L,N,M;
Size(A,K,L);
Size(B,N,M); assert(L==N);
Matrix C; Resize(C,K,M);
for(int i=0;i<K;i++){
for(int j=0;j<M;j++){
T sum = 0.0;
for(int k=0;k<N;k++) sum += A[i][k]*B[k][j];
C[i][j] =sum;
}
}
return C;
}
/** Matrix Vector multiplication **/
inline Vector<T> operator* (Matrix<T> &A,Vector<T> &B){
int M,N,K;
Size(A,N,M);
Size(B,K); assert(K==M);
Vector<T> C; Resize(C,N);
for(int i=0;i<N;i++){
T sum = 0.0;
for(int j=0;j<M;j++) sum += A[i][j]*B[j];
C[i] = sum;
}
return C;
}
/** Some version of Matrix norm **/
/*
inline T Norm(){ // this is not a usual L2 norm
T norm = 0;
for(int i=0;i<dim;i++){
for(int j=0;j<dim;j++){
norm += abs(A[i][j]);
}}
return norm;
}
*/
/** Some version of Matrix norm **/
template<class T> T LargestDiag(Matrix<T> &A)
{
int dim ; SizeSquare(A,dim);
T ld = abs(A[0][0]);
for(int i=1;i<dim;i++){
T cf = abs(A[i][i]);
if(abs(cf) > abs(ld) ){ld = cf;}
}
return ld;
}
/** Look for entries on the leading subdiagonal that are smaller than 'small' **/
template <class T,class U> int Chop_subdiag(Matrix<T> &A,T norm, int offset, U small)
{
int dim; SizeSquare(A,dim);
for(int l = dim - 1 - offset; l >= 1; l--) {
if((U)abs(A[l][l - 1]) < (U)small) {
A[l][l-1]=(U)0.0;
return l;
}
}
return 0;
}
/** Look for entries on the leading subdiagonal that are smaller than 'small' **/
template <class T,class U> int Chop_symm_subdiag(Matrix<T> & A,T norm, int offset, U small)
{
int dim; SizeSquare(A,dim);
for(int l = dim - 1 - offset; l >= 1; l--) {
if((U)abs(A[l][l - 1]) < (U)small) {
A[l][l - 1] = (U)0.0;
A[l - 1][l] = (U)0.0;
return l;
}
}
return 0;
}
/**Assign a submatrix to a larger one**/
template<class T>
void AssignSubMtx(Matrix<T> & A,int row_st, int row_end, int col_st, int col_end, Matrix<T> &S)
{
for(int i = row_st; i<row_end; i++){
for(int j = col_st; j<col_end; j++){
A[i][j] = S[i - row_st][j - col_st];
}
}
}
/**Get a square submatrix**/
template <class T>
Matrix<T> GetSubMtx(Matrix<T> &A,int row_st, int row_end, int col_st, int col_end)
{
Matrix<T> H; Resize(row_end - row_st,col_end-col_st);
for(int i = row_st; i<row_end; i++){
for(int j = col_st; j<col_end; j++){
H[i-row_st][j-col_st]=A[i][j];
}}
return H;
}
/**Assign a submatrix to a larger one NB remember Vector Vectors are transposes of the matricies they represent**/
template<class T>
void AssignSubMtx(Matrix<T> & A,int row_st, int row_end, int col_st, int col_end, Matrix<T> &S)
{
for(int i = row_st; i<row_end; i++){
for(int j = col_st; j<col_end; j++){
A[i][j] = S[i - row_st][j - col_st];
}}
}
/** compute b_i A_ij b_j **/ // surprised no Conj
template<class T> T proj(Matrix<T> A, Vector<T> B){
int dim; SizeSquare(A,dim);
int dimB; Size(B,dimB);
assert(dimB==dim);
T C = 0;
for(int i=0;i<dim;i++){
T sum = 0.0;
for(int j=0;j<dim;j++){
sum += A[i][j]*B[j];
}
C += B[i]*sum; // No conj?
}
return C;
}
/*
*************************************************************
*
* Matrix Vector products
*
*************************************************************
*/
// Instead make a linop and call my CG;
/// q -> q Q
template <class T,class Fermion> void times(Vector<Fermion> &q, Matrix<T> &Q)
{
int M; SizeSquare(Q,M);
int N; Size(q,N);
assert(M==N);
times(q,Q,N);
}
/// q -> q Q
template <class T> void times(multi1d<LatticeFermion> &q, Matrix<T> &Q, int N)
{
GridBase *grid = q[0]._grid;
int M; SizeSquare(Q,M);
int K; Size(q,K);
assert(N<M);
assert(N<K);
Vector<Fermion> S(N,grid );
for(int j=0;j<N;j++){
S[j] = zero;
for(int k=0;k<N;k++){
S[j] = S[j] + q[k]* Q[k][j];
}
}
for(int j=0;j<q.size();j++){
q[j] = S[j];
}
}
#endif

75
lib/algorithms/iterative/MatrixUtils.h
View File

@ -1,75 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/MatrixUtils.h
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_MATRIX_UTILS_H
#define GRID_MATRIX_UTILS_H
namespace Grid {
namespace MatrixUtils {
template<class T> inline void Size(Matrix<T>& A,int &N,int &M){
N=A.size(); assert(N>0);
M=A[0].size();
for(int i=0;i<N;i++){
assert(A[i].size()==M);
}
}
template<class T> inline void SizeSquare(Matrix<T>& A,int &N)
{
int M;
Size(A,N,M);
assert(N==M);
}
template<class T> inline void Fill(Matrix<T>& A,T & val)
{
int N,M;
Size(A,N,M);
for(int i=0;i<N;i++){
for(int j=0;j<M;j++){
A[i][j]=val;
}}
}
template<class T> inline void Diagonal(Matrix<T>& A,T & val)
{
int N;
SizeSquare(A,N);
for(int i=0;i<N;i++){
A[i][i]=val;
}
}
template<class T> inline void Identity(Matrix<T>& A)
{
Fill(A,0.0);
Diagonal(A,1.0);
}
};
}
#endif

15
lib/algorithms/iterative/TODO
View File

@ -1,15 +0,0 @@
- ConjugateGradientMultiShift
- MCR
- Potentially Useful Boost libraries
- MultiArray
- Aligned allocator; memory pool
- Remez -- Mike or Boost?
- Multiprecision
- quaternians
- Tokenize
- Serialization
- Regex
- Proto (ET)
- uBlas

122
lib/algorithms/iterative/bisec.c
View File

@ -1,122 +0,0 @@
#include <math.h>
#include <stdlib.h>
#include <vector>
struct Bisection {
static void get_eig2(int row_num,std::vector<RealD> &ALPHA,std::vector<RealD> &BETA, std::vector<RealD> & eig)
{
int i,j;
std::vector<RealD> evec1(row_num+3);
std::vector<RealD> evec2(row_num+3);
RealD eps2;
ALPHA[1]=0.;
BETHA[1]=0.;
for(i=0;i<row_num-1;i++) {
ALPHA[i+1] = A[i*(row_num+1)].real();
BETHA[i+2] = A[i*(row_num+1)+1].real();
}
ALPHA[row_num] = A[(row_num-1)*(row_num+1)].real();
bisec(ALPHA,BETHA,row_num,1,row_num,1e-10,1e-10,evec1,eps2);
bisec(ALPHA,BETHA,row_num,1,row_num,1e-16,1e-16,evec2,eps2);
// Do we really need to sort here?
int begin=1;
int end = row_num;
int swapped=1;
while(swapped) {
swapped=0;
for(i=begin;i<end;i++){
if(mag(evec2[i])>mag(evec2[i+1])) {
swap(evec2+i,evec2+i+1);
swapped=1;
}
}
end--;
for(i=end-1;i>=begin;i--){
if(mag(evec2[i])>mag(evec2[i+1])) {
swap(evec2+i,evec2+i+1);
swapped=1;
}
}
begin++;
}
for(i=0;i<row_num;i++){
for(j=0;j<row_num;j++) {
if(i==j) H[i*row_num+j]=evec2[i+1];
else H[i*row_num+j]=0.;
}
}
}
static void bisec(std::vector<RealD> &c,
std::vector<RealD> &b,
int n,
int m1,
int m2,
RealD eps1,
RealD relfeh,
std::vector<RealD> &x,
RealD &eps2)
{
std::vector<RealD> wu(n+2);
RealD h,q,x1,xu,x0,xmin,xmax;
int i,a,k;
b[1]=0.0;
xmin=c[n]-fabs(b[n]);
xmax=c[n]+fabs(b[n]);
for(i=1;i<n;i++){
h=fabs(b[i])+fabs(b[i+1]);
if(c[i]+h>xmax) xmax= c[i]+h;
if(c[i]-h<xmin) xmin= c[i]-h;
}
xmax *=2.;
eps2=relfeh*((xmin+xmax)>0.0 ? xmax : -xmin);
if(eps1<=0.0) eps1=eps2;
eps2=0.5*eps1+7.0*(eps2);
x0=xmax;
for(i=m1;i<=m2;i++){
x[i]=xmax;
wu[i]=xmin;
}
for(k=m2;k>=m1;k--){
xu=xmin;
i=k;
do{
if(xu<wu[i]){
xu=wu[i];
i=m1-1;
}
i--;
}while(i>=m1);
if(x0>x[k]) x0=x[k];
while((x0-xu)>2*relfeh*(fabs(xu)+fabs(x0))+eps1){
x1=(xu+x0)/2;
a=0;
q=1.0;
for(i=1;i<=n;i++){
q=c[i]-x1-((q!=0.0)? b[i]*b[i]/q:fabs(b[i])/relfeh);
if(q<0) a++;
}
// printf("x1=%e a=%d\n",x1,a);
if(a<k){
if(a<m1){
xu=x1;
wu[m1]=x1;
}else {
xu=x1;
wu[a+1]=x1;
if(x[a]>x1) x[a]=x1;
}
}else x0=x1;
}
x[k]=(x0+xu)/2;
}
}
}

1
lib/algorithms/iterative/get_eig.c
View File

@ -1 +0,0 @@

527
lib/lattice/Lattice_reduction.h
View File

@ -30,6 +30,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_LATTICE_REDUCTION_H
#define GRID_LATTICE_REDUCTION_H
#include <Grid/Eigen/Dense>
namespace Grid {
#ifdef GRID_WARN_SUBOPTIMAL
#warning "Optimisation alert all these reduction loops are NOT threaded "
@ -38,120 +40,123 @@ namespace Grid {
////////////////////////////////////////////////////////////////////////////////////////////////////
// Deterministic Reduction operations
////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
ComplexD nrm = innerProduct(arg,arg);
return std::real(nrm);
template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
ComplexD nrm = innerProduct(arg,arg);
return std::real(nrm);
}
// Double inner product
template<class vobj>
inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
{
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_typeD vector_type;
scalar_type nrm;
GridBase *grid = left._grid;
std::vector<vector_type,alignedAllocator<vector_type> > sumarray(grid->SumArraySize());
parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
decltype(innerProductD(left._odata[0],right._odata[0])) vnrm=zero; // private to thread; sub summation
for(int ss=myoff;ss<mywork+myoff; ss++){
vnrm = vnrm + innerProductD(left._odata[ss],right._odata[ss]);
}
sumarray[thr]=TensorRemove(vnrm) ;
}
vector_type vvnrm; vvnrm=zero; // sum across threads
for(int i=0;i<grid->SumArraySize();i++){
vvnrm = vvnrm+sumarray[i];
}
nrm = Reduce(vvnrm);// sum across simd
right._grid->GlobalSum(nrm);
return nrm;
}
template<class Op,class T1>
inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second))))::scalar_object
{
return sum(closure(expr));
}
template<class vobj>
inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
{
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
scalar_type nrm;
GridBase *grid = left._grid;
std::vector<vector_type,alignedAllocator<vector_type> > sumarray(grid->SumArraySize());
for(int i=0;i<grid->SumArraySize();i++){
sumarray[i]=zero;
}
parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
decltype(innerProduct(left._odata[0],right._odata[0])) vnrm=zero; // private to thread; sub summation
for(int ss=myoff;ss<mywork+myoff; ss++){
vnrm = vnrm + innerProduct(left._odata[ss],right._odata[ss]);
}
sumarray[thr]=TensorRemove(vnrm) ;
}
vector_type vvnrm; vvnrm=zero; // sum across threads
for(int i=0;i<grid->SumArraySize();i++){
vvnrm = vvnrm+sumarray[i];
}
nrm = Reduce(vvnrm);// sum across simd
right._grid->GlobalSum(nrm);
return nrm;
}
template<class Op,class T1>
inline auto sum(const LatticeUnaryExpression<Op,T1> & expr)
->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second))))::scalar_object
{
return sum(closure(expr));
}
template<class Op,class T1,class T2>
inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
template<class Op,class T1,class T2>
inline auto sum(const LatticeBinaryExpression<Op,T1,T2> & expr)
->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),eval(0,std::get<1>(expr.second))))::scalar_object
{
return sum(closure(expr));
}
template<class Op,class T1,class T2,class T3>
inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),
eval(0,std::get<1>(expr.second)),
eval(0,std::get<2>(expr.second))
))::scalar_object
{
return sum(closure(expr));
}
template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg){
GridBase *grid=arg._grid;
int Nsimd = grid->Nsimd();
std::vector<vobj,alignedAllocator<vobj> > sumarray(grid->SumArraySize());
for(int i=0;i<grid->SumArraySize();i++){
sumarray[i]=zero;
}
parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(grid->oSites(),thr,mywork,myoff);
vobj vvsum=zero;
for(int ss=myoff;ss<mywork+myoff; ss++){
vvsum = vvsum + arg._odata[ss];
}
sumarray[thr]=vvsum;
}
vobj vsum=zero; // sum across threads
for(int i=0;i<grid->SumArraySize();i++){
vsum = vsum+sumarray[i];
}
typedef typename vobj::scalar_object sobj;
sobj ssum=zero;
std::vector<sobj> buf(Nsimd);
extract(vsum,buf);
for(int i=0;i<Nsimd;i++) ssum = ssum + buf[i];
arg._grid->GlobalSum(ssum);
return ssum;
{
return sum(closure(expr));
}
template<class Op,class T1,class T2,class T3>
inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
->typename decltype(expr.first.func(eval(0,std::get<0>(expr.second)),
eval(0,std::get<1>(expr.second)),
eval(0,std::get<2>(expr.second))
))::scalar_object
{
return sum(closure(expr));
}
template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
{
GridBase *grid=arg._grid;
int Nsimd = grid->Nsimd();
std::vector<vobj,alignedAllocator<vobj> > sumarray(grid->SumArraySize());
for(int i=0;i<grid->SumArraySize();i++){
sumarray[i]=zero;
}
parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(grid->oSites(),thr,mywork,myoff);
vobj vvsum=zero;
for(int ss=myoff;ss<mywork+myoff; ss++){
vvsum = vvsum + arg._odata[ss];
}
sumarray[thr]=vvsum;
}
vobj vsum=zero; // sum across threads
for(int i=0;i<grid->SumArraySize();i++){
vsum = vsum+sumarray[i];
}
typedef typename vobj::scalar_object sobj;
sobj ssum=zero;
std::vector<sobj> buf(Nsimd);
extract(vsum,buf);
for(int i=0;i<Nsimd;i++) ssum = ssum + buf[i];
arg._grid->GlobalSum(ssum);
return ssum;
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
// sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
{
///////////////////////////////////////////////////////
// FIXME precision promoted summation
// may be important for correlation functions
// But easily avoided by using double precision fields
///////////////////////////////////////////////////////
typedef typename vobj::scalar_object sobj;
GridBase *grid = Data._grid;
assert(grid!=NULL);
// FIXME
// std::cout<<GridLogMessage<<"WARNING ! SliceSum is unthreaded "<<grid->SumArraySize()<<" threads "<<std::endl;
const int Nd = grid->_ndimension;
const int Nsimd = grid->Nsimd();
@ -163,23 +168,31 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
int rd=grid->_rdimensions[orthogdim];
std::vector<vobj,alignedAllocator<vobj> > lvSum(rd); // will locally sum vectors first
std::vector<sobj> lsSum(ld,zero); // sum across these down to scalars
std::vector<sobj> extracted(Nsimd); // splitting the SIMD
std::vector<sobj> lsSum(ld,zero); // sum across these down to scalars
std::vector<sobj> extracted(Nsimd); // splitting the SIMD
result.resize(fd); // And then global sum to return the same vector to every node for IO to file
result.resize(fd); // And then global sum to return the same vector to every node
for(int r=0;r<rd;r++){
lvSum[r]=zero;
}
std::vector<int> coor(Nd);
int e1= grid->_slice_nblock[orthogdim];
int e2= grid->_slice_block [orthogdim];
int stride=grid->_slice_stride[orthogdim];
// sum over reduced dimension planes, breaking out orthog dir
// Parallel over orthog direction
parallel_for(int r=0;r<rd;r++){
for(int ss=0;ss<grid->oSites();ss++){
Lexicographic::CoorFromIndex(coor,ss,grid->_rdimensions);
int r = coor[orthogdim];
lvSum[r]=lvSum[r]+Data._odata[ss];
}
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
lvSum[r]=lvSum[r]+Data._odata[ss];
}
}
}
// Sum across simd lanes in the plane, breaking out orthog dir.
std::vector<int> icoor(Nd);
@ -214,10 +227,304 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
result[t]=gsum;
}
}
template<class vobj>
static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim)
{
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
GridBase *grid = lhs._grid;
assert(grid!=NULL);
conformable(grid,rhs._grid);
const int Nd = grid->_ndimension;
const int Nsimd = grid->Nsimd();
assert(orthogdim >= 0);
assert(orthogdim < Nd);
int fd=grid->_fdimensions[orthogdim];
int ld=grid->_ldimensions[orthogdim];
int rd=grid->_rdimensions[orthogdim];
std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd); // will locally sum vectors first
std::vector<scalar_type > lsSum(ld,scalar_type(0.0)); // sum across these down to scalars
std::vector<iScalar<scalar_type> > extracted(Nsimd); // splitting the SIMD
result.resize(fd); // And then global sum to return the same vector to every node for IO to file
for(int r=0;r<rd;r++){
lvSum[r]=zero;
}
int e1= grid->_slice_nblock[orthogdim];
int e2= grid->_slice_block [orthogdim];
int stride=grid->_slice_stride[orthogdim];
parallel_for(int r=0;r<rd;r++){
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
vector_type vv = TensorRemove(innerProduct(lhs._odata[ss],rhs._odata[ss]));
lvSum[r]=lvSum[r]+vv;
}
}
}
// Sum across simd lanes in the plane, breaking out orthog dir.
std::vector<int> icoor(Nd);
for(int rt=0;rt<rd;rt++){
iScalar<vector_type> temp;
temp._internal = lvSum[rt];
extract(temp,extracted);
for(int idx=0;idx<Nsimd;idx++){
grid->iCoorFromIindex(icoor,idx);
int ldx =rt+icoor[orthogdim]*rd;
lsSum[ldx]=lsSum[ldx]+extracted[idx]._internal;
}
}
// sum over nodes.
scalar_type gsum;
for(int t=0;t<fd;t++){
int pt = t/ld; // processor plane
int lt = t%ld;
if ( pt == grid->_processor_coor[orthogdim] ) {
gsum=lsSum[lt];
} else {
gsum=scalar_type(0.0);
}
grid->GlobalSum(gsum);
result[t]=gsum;
}
}
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 sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
int orthogdim,RealD scale=1.0)
{
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 tensor_reduced;
GridBase *grid = X._grid;
int Nsimd =grid->Nsimd();
int Nblock =grid->GlobalDimensions()[orthogdim];
int fd =grid->_fdimensions[orthogdim];
int ld =grid->_ldimensions[orthogdim];
int rd =grid->_rdimensions[orthogdim];
int e1 =grid->_slice_nblock[orthogdim];
int e2 =grid->_slice_block [orthogdim];
int stride =grid->_slice_stride[orthogdim];
std::vector<int> icoor;
for(int r=0;r<rd;r++){
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
vector_type av;
for(int l=0;l<Nsimd;l++){
grid->iCoorFromIindex(icoor,l);
int ldx =r+icoor[orthogdim]*rd;
scalar_type *as =(scalar_type *)&av;
as[l] = scalar_type(a[ldx])*scale;
}
tensor_reduced at; at=av;
parallel_for_nest2(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
R._odata[ss] = at*X._odata[ss]+Y._odata[ss];
}
}
}
};
/*
template<class vobj>
static void sliceMaddVectorSlow (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 sliceInnerProductVectorSlow( 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]);
}
}
*/
//////////////////////////////////////////////////////////////////////////////////////////
// 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].
//////////////////////////////////////////////////////////////////////////////////////////
inline 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);
}
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);
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 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;
}
} /*END NAMESPACE GRID*/
#endif

4
lib/log/Log.h
View File

@ -110,8 +110,8 @@ public:
friend std::ostream& operator<< (std::ostream& stream, Logger& log){
if ( log.active ) {
stream << log.background()<< std::setw(10) << std::left << log.topName << log.background()<< " : ";
stream << log.colour() << std::setw(14) << std::left << log.name << log.background() << " : ";
stream << log.background()<< std::setw(8) << std::left << log.topName << log.background()<< " : ";
stream << log.colour() << std::setw(10) << std::left << log.name << log.background() << " : ";
if ( log.timestamp ) {
StopWatch.Stop();
GridTime now = StopWatch.Elapsed();

2
lib/simd/Grid_avx.h
View File

@ -473,7 +473,7 @@ namespace Optimization {
#define USE_FP16
struct PrecisionChange {
static inline __m256i StoH (__m256 a,__m256 b) {
__m256 h ;
__m256 h;
#ifdef USE_FP16
__m128i ha = _mm256_cvtps_ph(a,0);
__m128i hb = _mm256_cvtps_ph(b,0);

20
lib/simd/Grid_avx512.h
View File

@ -343,12 +343,12 @@ namespace Optimization {
#define USE_FP16
struct PrecisionChange {
static inline __m512i StoH (__m512 a,__m512 b) {
__m512 h ;
__m512i h;
#ifdef USE_FP16
__m256i ha = _mm512_cvtps_ph(a,0);
__m256i hb = _mm512_cvtps_ph(b,0);
h = _mm512_castps256_ps512(ha);
h = _mm512_insertf256_ps(h,hb,1);
h =(__m512i) _mm512_castps256_ps512((__m256)ha);
h =(__m512i) _mm512_insertf64x4((__m512d)h,(__m256d)hb,1);
#else
assert(0);
#endif
@ -356,8 +356,8 @@ namespace Optimization {
}
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));
sa = _mm512_cvtph_ps((__m256i)_mm512_extractf64x4_pd((__m512d)h,0));
sb = _mm512_cvtph_ps((__m256i)_mm512_extractf64x4_pd((__m512d)h,1));
#else
assert(0);
#endif
@ -366,12 +366,12 @@ namespace Optimization {
__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);
s =(__m512) _mm512_insertf64x4((__m512d)s,(__m256d)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));
a = _mm512_cvtps_pd((__m256)_mm512_extractf64x4_pd((__m512d)s,0));
b = _mm512_cvtps_pd((__m256)_mm512_extractf64x4_pd((__m512d)s,1));
}
static inline __m512i DtoH (__m512d a,__m512d b,__m512d c,__m512d d) {
__m512 sa,sb;
@ -582,7 +582,9 @@ namespace Optimization {
//////////////////////////////////////////////////////////////////////////////////////
// Here assign types
typedef __m512 SIMD_Ftype; // Single precision type
typedef __m512i SIMD_Htype; // Single precision type
typedef __m512 SIMD_Ftype; // Single precision type
typedef __m512d SIMD_Dtype; // Double precision type
typedef __m512i SIMD_Itype; // Integer type

7
lib/simd/Grid_vector_types.h
View File

@ -2,7 +2,7 @@
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/simd/Grid_vector_types.h
Source file: ./lib/simd/Grid_vector_type.h
Copyright (C) 2015
@ -411,7 +411,6 @@ template <class S, class V, IfNotComplex<S> = 0>
inline Grid_simd<S, V> rotate(Grid_simd<S, V> b, int nrot) {
nrot = nrot % Grid_simd<S, V>::Nsimd();
Grid_simd<S, V> ret;
// std::cout << "Rotate Real by "<<nrot<<std::endl;
ret.v = Optimization::Rotate::rotate(b.v, nrot);
return ret;
}
@ -419,7 +418,6 @@ template <class S, class V, IfComplex<S> = 0>
inline Grid_simd<S, V> rotate(Grid_simd<S, V> b, int nrot) {
nrot = nrot % Grid_simd<S, V>::Nsimd();
Grid_simd<S, V> ret;
// std::cout << "Rotate Complex by "<<nrot<<std::endl;
ret.v = Optimization::Rotate::rotate(b.v, 2 * nrot);
return ret;
}
@ -427,14 +425,12 @@ template <class S, class V, IfNotComplex<S> =0>
inline void rotate( Grid_simd<S,V> &ret,Grid_simd<S,V> b,int nrot)
{
nrot = nrot % Grid_simd<S,V>::Nsimd();
// std::cout << "Rotate Real by "<<nrot<<std::endl;
ret.v = Optimization::Rotate::rotate(b.v,nrot);
}
template <class S, class V, IfComplex<S> =0>
inline void rotate(Grid_simd<S,V> &ret,Grid_simd<S,V> b,int nrot)
{
nrot = nrot % Grid_simd<S,V>::Nsimd();
// std::cout << "Rotate Complex by "<<nrot<<std::endl;
ret.v = Optimization::Rotate::rotate(b.v,2*nrot);
}
@ -694,7 +690,6 @@ inline Grid_simd<S, V> innerProduct(const Grid_simd<S, V> &l,
const Grid_simd<S, V> &r) {
return conjugate(l) * r;
}
template <class S, class V>
inline Grid_simd<S, V> outerProduct(const Grid_simd<S, V> &l,
const Grid_simd<S, V> &r) {

31
lib/tensors/Tensor_class.h
View File

@ -56,11 +56,11 @@ class iScalar {
typedef vtype element;
typedef typename GridTypeMapper<vtype>::scalar_type scalar_type;
typedef typename GridTypeMapper<vtype>::vector_type vector_type;
typedef typename GridTypeMapper<vtype>::vector_typeD vector_typeD;
typedef typename GridTypeMapper<vtype>::tensor_reduced tensor_reduced_v;
typedef iScalar<tensor_reduced_v> tensor_reduced;
typedef typename GridTypeMapper<vtype>::scalar_object recurse_scalar_object;
typedef iScalar<tensor_reduced_v> tensor_reduced;
typedef iScalar<recurse_scalar_object> scalar_object;
// substitutes a real or complex version with same tensor structure
typedef iScalar<typename GridTypeMapper<vtype>::Complexified> Complexified;
typedef iScalar<typename GridTypeMapper<vtype>::Realified> Realified;
@ -77,8 +77,12 @@ class iScalar {
iScalar<vtype> & operator= (const iScalar<vtype> &copyme) = default;
iScalar<vtype> & operator= (iScalar<vtype> &&copyme) = default;
*/
iScalar(scalar_type s)
: _internal(s){}; // recurse down and hit the constructor for vector_type
// template<int N=0>
// iScalar(EnableIf<isSIMDvectorized<vector_type>, vector_type> s) : _internal(s){}; // recurse down and hit the constructor for vector_type
iScalar(scalar_type s) : _internal(s){}; // recurse down and hit the constructor for vector_type
iScalar(const Zero &z) { *this = zero; };
iScalar<vtype> &operator=(const Zero &hero) {
@ -134,33 +138,28 @@ class iScalar {
strong_inline const vtype &operator()(void) const { return _internal; }
// Type casts meta programmed, must be pure scalar to match TensorRemove
template <class U = vtype, class V = scalar_type, IfComplex<V> = 0,
IfNotSimd<U> = 0>
template <class U = vtype, class V = scalar_type, IfComplex<V> = 0, IfNotSimd<U> = 0>
operator ComplexF() const {
return (TensorRemove(_internal));
};
template <class U = vtype, class V = scalar_type, IfComplex<V> = 0,
IfNotSimd<U> = 0>
template <class U = vtype, class V = scalar_type, IfComplex<V> = 0, IfNotSimd<U> = 0>
operator ComplexD() const {
return (TensorRemove(_internal));
};
// template<class U=vtype,class V=scalar_type,IfComplex<V> = 0,IfNotSimd<U> =
// 0> operator RealD () const { return(real(TensorRemove(_internal))); }
template <class U = vtype, class V = scalar_type, IfReal<V> = 0,
IfNotSimd<U> = 0>
template <class U = vtype, class V = scalar_type, IfReal<V> = 0,IfNotSimd<U> = 0>
operator RealD() const {
return TensorRemove(_internal);
}
template <class U = vtype, class V = scalar_type, IfInteger<V> = 0,
IfNotSimd<U> = 0>
template <class U = vtype, class V = scalar_type, IfInteger<V> = 0, IfNotSimd<U> = 0>
operator Integer() const {
return Integer(TensorRemove(_internal));
}
// convert from a something to a scalar via constructor of something arg
template <class T, typename std::enable_if<!isGridTensor<T>::value, T>::type
* = nullptr>
strong_inline iScalar<vtype> operator=(T arg) {
template <class T, typename std::enable_if<!isGridTensor<T>::value, T>::type * = nullptr>
strong_inline iScalar<vtype> operator=(T arg) {
_internal = arg;
return *this;
}
@ -193,6 +192,7 @@ class iVector {
typedef vtype element;
typedef typename GridTypeMapper<vtype>::scalar_type scalar_type;
typedef typename GridTypeMapper<vtype>::vector_type vector_type;
typedef typename GridTypeMapper<vtype>::vector_typeD vector_typeD;
typedef typename GridTypeMapper<vtype>::tensor_reduced tensor_reduced_v;
typedef typename GridTypeMapper<vtype>::scalar_object recurse_scalar_object;
typedef iScalar<tensor_reduced_v> tensor_reduced;
@ -305,6 +305,7 @@ class iMatrix {
typedef vtype element;
typedef typename GridTypeMapper<vtype>::scalar_type scalar_type;
typedef typename GridTypeMapper<vtype>::vector_type vector_type;
typedef typename GridTypeMapper<vtype>::vector_typeD vector_typeD;
typedef typename GridTypeMapper<vtype>::tensor_reduced tensor_reduced_v;
typedef typename GridTypeMapper<vtype>::scalar_object recurse_scalar_object;

142
lib/tensors/Tensor_inner.h
View File

@ -29,51 +29,109 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_MATH_INNER_H
#define GRID_MATH_INNER_H
namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////
// innerProduct Scalar x Scalar -> Scalar
// innerProduct Vector x Vector -> Scalar
// innerProduct Matrix x Matrix -> Scalar
///////////////////////////////////////////////////////////////////////////////////////
template<class sobj> inline RealD norm2(const sobj &arg){
typedef typename sobj::scalar_type scalar;
decltype(innerProduct(arg,arg)) nrm;
nrm = innerProduct(arg,arg);
RealD ret = real(nrm);
return ret;
}
///////////////////////////////////////////////////////////////////////////////////////
// innerProduct Scalar x Scalar -> Scalar
// innerProduct Vector x Vector -> Scalar
// innerProduct Matrix x Matrix -> Scalar
///////////////////////////////////////////////////////////////////////////////////////
template<class sobj> inline RealD norm2(const sobj &arg){
auto nrm = innerProductD(arg,arg);
RealD ret = real(nrm);
return ret;
}
//////////////////////////////////////
// If single promote to double and sum 2x
//////////////////////////////////////
template<class l,class r,int N> inline
auto innerProduct (const iVector<l,N>& lhs,const iVector<r,N>& rhs) -> iScalar<decltype(innerProduct(lhs._internal[0],rhs._internal[0]))>
{
typedef decltype(innerProduct(lhs._internal[0],rhs._internal[0])) ret_t;
iScalar<ret_t> ret;
ret=zero;
for(int c1=0;c1<N;c1++){
ret._internal += innerProduct(lhs._internal[c1],rhs._internal[c1]);
}
return ret;
inline ComplexD innerProductD(const ComplexF &l,const ComplexF &r){ return innerProduct(l,r); }
inline ComplexD innerProductD(const ComplexD &l,const ComplexD &r){ return innerProduct(l,r); }
inline RealD innerProductD(const RealD &l,const RealD &r){ return innerProduct(l,r); }
inline RealD innerProductD(const RealF &l,const RealF &r){ return innerProduct(l,r); }
inline vComplexD innerProductD(const vComplexD &l,const vComplexD &r){ return innerProduct(l,r); }
inline vRealD innerProductD(const vRealD &l,const vRealD &r){ return innerProduct(l,r); }
inline vComplexD innerProductD(const vComplexF &l,const vComplexF &r){
vComplexD la,lb;
vComplexD ra,rb;
Optimization::PrecisionChange::StoD(l.v,la.v,lb.v);
Optimization::PrecisionChange::StoD(r.v,ra.v,rb.v);
return innerProduct(la,ra) + innerProduct(lb,rb);
}
inline vRealD innerProductD(const vRealF &l,const vRealF &r){
vRealD la,lb;
vRealD ra,rb;
Optimization::PrecisionChange::StoD(l.v,la.v,lb.v);
Optimization::PrecisionChange::StoD(r.v,ra.v,rb.v);
return innerProduct(la,ra) + innerProduct(lb,rb);
}
template<class l,class r,int N> inline
auto innerProductD (const iVector<l,N>& lhs,const iVector<r,N>& rhs) -> iScalar<decltype(innerProductD(lhs._internal[0],rhs._internal[0]))>
{
typedef decltype(innerProductD(lhs._internal[0],rhs._internal[0])) ret_t;
iScalar<ret_t> ret;
ret=zero;
for(int c1=0;c1<N;c1++){
ret._internal += innerProductD(lhs._internal[c1],rhs._internal[c1]);
}
template<class l,class r,int N> inline
auto innerProduct (const iMatrix<l,N>& lhs,const iMatrix<r,N>& rhs) -> iScalar<decltype(innerProduct(lhs._internal[0][0],rhs._internal[0][0]))>
{
typedef decltype(innerProduct(lhs._internal[0][0],rhs._internal[0][0])) ret_t;
iScalar<ret_t> ret;
iScalar<ret_t> tmp;
ret=zero;
for(int c1=0;c1<N;c1++){
for(int c2=0;c2<N;c2++){
ret._internal+=innerProduct(lhs._internal[c1][c2],rhs._internal[c1][c2]);
}}
return ret;
}
template<class l,class r> inline
auto innerProduct (const iScalar<l>& lhs,const iScalar<r>& rhs) -> iScalar<decltype(innerProduct(lhs._internal,rhs._internal))>
{
typedef decltype(innerProduct(lhs._internal,rhs._internal)) ret_t;
iScalar<ret_t> ret;
ret._internal = innerProduct(lhs._internal,rhs._internal);
return ret;
return ret;
}
template<class l,class r,int N> inline
auto innerProductD (const iMatrix<l,N>& lhs,const iMatrix<r,N>& rhs) -> iScalar<decltype(innerProductD(lhs._internal[0][0],rhs._internal[0][0]))>
{
typedef decltype(innerProductD(lhs._internal[0][0],rhs._internal[0][0])) ret_t;
iScalar<ret_t> ret;
iScalar<ret_t> tmp;
ret=zero;
for(int c1=0;c1<N;c1++){
for(int c2=0;c2<N;c2++){
ret._internal+=innerProductD(lhs._internal[c1][c2],rhs._internal[c1][c2]);
}}
return ret;
}
template<class l,class r> inline
auto innerProductD (const iScalar<l>& lhs,const iScalar<r>& rhs) -> iScalar<decltype(innerProductD(lhs._internal,rhs._internal))>
{
typedef decltype(innerProductD(lhs._internal,rhs._internal)) ret_t;
iScalar<ret_t> ret;
ret._internal = innerProductD(lhs._internal,rhs._internal);
return ret;
}
//////////////////////
// Keep same precison
//////////////////////
template<class l,class r,int N> inline
auto innerProduct (const iVector<l,N>& lhs,const iVector<r,N>& rhs) -> iScalar<decltype(innerProduct(lhs._internal[0],rhs._internal[0]))>
{
typedef decltype(innerProduct(lhs._internal[0],rhs._internal[0])) ret_t;
iScalar<ret_t> ret;
ret=zero;
for(int c1=0;c1<N;c1++){
ret._internal += innerProduct(lhs._internal[c1],rhs._internal[c1]);
}
return ret;
}
template<class l,class r,int N> inline
auto innerProduct (const iMatrix<l,N>& lhs,const iMatrix<r,N>& rhs) -> iScalar<decltype(innerProduct(lhs._internal[0][0],rhs._internal[0][0]))>
{
typedef decltype(innerProduct(lhs._internal[0][0],rhs._internal[0][0])) ret_t;
iScalar<ret_t> ret;
iScalar<ret_t> tmp;
ret=zero;
for(int c1=0;c1<N;c1++){
for(int c2=0;c2<N;c2++){
ret._internal+=innerProduct(lhs._internal[c1][c2],rhs._internal[c1][c2]);
}}
return ret;
}
template<class l,class r> inline
auto innerProduct (const iScalar<l>& lhs,const iScalar<r>& rhs) -> iScalar<decltype(innerProduct(lhs._internal,rhs._internal))>
{
typedef decltype(innerProduct(lhs._internal,rhs._internal)) ret_t;
iScalar<ret_t> ret;
ret._internal = innerProduct(lhs._internal,rhs._internal);
return ret;
}
}
#endif

14
lib/tensors/Tensor_traits.h
View File

@ -53,6 +53,7 @@ namespace Grid {
public:
typedef typename T::scalar_type scalar_type;
typedef typename T::vector_type vector_type;
typedef typename T::vector_typeD vector_typeD;
typedef typename T::tensor_reduced tensor_reduced;
typedef typename T::scalar_object scalar_object;
typedef typename T::Complexified Complexified;
@ -67,6 +68,7 @@ namespace Grid {
public:
typedef RealF scalar_type;
typedef RealF vector_type;
typedef RealD vector_typeD;
typedef RealF tensor_reduced ;
typedef RealF scalar_object;
typedef ComplexF Complexified;
@ -77,6 +79,7 @@ namespace Grid {
public:
typedef RealD scalar_type;
typedef RealD vector_type;
typedef RealD vector_typeD;
typedef RealD tensor_reduced;
typedef RealD scalar_object;
typedef ComplexD Complexified;
@ -87,6 +90,7 @@ namespace Grid {
public:
typedef ComplexF scalar_type;
typedef ComplexF vector_type;
typedef ComplexD vector_typeD;
typedef ComplexF tensor_reduced;
typedef ComplexF scalar_object;
typedef ComplexF Complexified;
@ -97,6 +101,7 @@ namespace Grid {
public:
typedef ComplexD scalar_type;
typedef ComplexD vector_type;
typedef ComplexD vector_typeD;
typedef ComplexD tensor_reduced;
typedef ComplexD scalar_object;
typedef ComplexD Complexified;
@ -107,6 +112,7 @@ namespace Grid {
public:
typedef Integer scalar_type;
typedef Integer vector_type;
typedef Integer vector_typeD;
typedef Integer tensor_reduced;
typedef Integer scalar_object;
typedef void Complexified;
@ -118,6 +124,7 @@ namespace Grid {
public:
typedef RealF scalar_type;
typedef vRealF vector_type;
typedef vRealD vector_typeD;
typedef vRealF tensor_reduced;
typedef RealF scalar_object;
typedef vComplexF Complexified;
@ -128,6 +135,7 @@ namespace Grid {
public:
typedef RealD scalar_type;
typedef vRealD vector_type;
typedef vRealD vector_typeD;
typedef vRealD tensor_reduced;
typedef RealD scalar_object;
typedef vComplexD Complexified;
@ -138,6 +146,7 @@ namespace Grid {
public:
typedef ComplexF scalar_type;
typedef vComplexF vector_type;
typedef vComplexD vector_typeD;
typedef vComplexF tensor_reduced;
typedef ComplexF scalar_object;
typedef vComplexF Complexified;
@ -148,6 +157,7 @@ namespace Grid {
public:
typedef ComplexD scalar_type;
typedef vComplexD vector_type;
typedef vComplexD vector_typeD;
typedef vComplexD tensor_reduced;
typedef ComplexD scalar_object;
typedef vComplexD Complexified;
@ -158,6 +168,7 @@ namespace Grid {
public:
typedef Integer scalar_type;
typedef vInteger vector_type;
typedef vInteger vector_typeD;
typedef vInteger tensor_reduced;
typedef Integer scalar_object;
typedef void Complexified;
@ -241,7 +252,8 @@ namespace Grid {
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 *);
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(...);

4
lib/util/Init.cc
View File

@ -311,8 +311,8 @@ void Grid_init(int *argc,char ***argv)
std::cout<<GridLogMessage<<std::endl;
std::cout<<GridLogMessage<<"Performance:"<<std::endl;
std::cout<<GridLogMessage<<std::endl;
std::cout<<GridLogMessage<<" --comms-isend : Asynchronous MPI calls; several dirs at a time "<<std::endl;
std::cout<<GridLogMessage<<" --comms-sendrecv: Synchronous MPI calls; one dirs at a time "<<std::endl;
std::cout<<GridLogMessage<<" --comms-concurrent : Asynchronous MPI calls; several dirs at a time "<<std::endl;
std::cout<<GridLogMessage<<" --comms-sequential : Synchronous MPI calls; one dirs at a time "<<std::endl;
std::cout<<GridLogMessage<<" --comms-overlap : Overlap comms with compute "<<std::endl;
std::cout<<GridLogMessage<<std::endl;
std::cout<<GridLogMessage<<" --dslash-generic: Wilson kernel for generic Nc"<<std::endl;

4
tests/core/Test_fft_gfix.cc
View File

@ -148,11 +148,13 @@ class FourierAcceleratedGaugeFixer : public Gimpl {
Complex psqMax(16.0);
Fp = psqMax*one/psq;
/*
static int once;
if ( once == 0 ) {
std::cout << " Fp " << Fp <<std::endl;
once ++;
}
}*/
pokeSite(TComplex(1.0),Fp,coor);
dmuAmu_p = dmuAmu_p * Fp;

13
tests/debug/Test_synthetic_lanczos.cc
View File

@ -115,8 +115,8 @@ int main (int argc, char ** argv)
RNG.SeedFixedIntegers(seeds);
RealD alpha = 1.0;
RealD beta = 0.03;
RealD alpha = 1.2;
RealD beta = 0.1;
RealD mu = 0.0;
int order = 11;
ChebyshevLanczos<LatticeComplex> Cheby(alpha,beta,mu,order);
@ -131,10 +131,9 @@ int main (int argc, char ** argv)
const int Nit= 10000;
int Nconv;
RealD eresid = 1.0e-8;
RealD eresid = 1.0e-6;
ImplicitlyRestartedLanczos<LatticeComplex> IRL(HermOp,X,Nk,Nm,eresid,Nit);
ImplicitlyRestartedLanczos<LatticeComplex> ChebyIRL(HermOp,Cheby,Nk,Nm,eresid,Nit);
LatticeComplex src(grid); gaussian(RNG,src);
@ -145,9 +144,9 @@ int main (int argc, char ** argv)
}
{
// std::vector<RealD> eval(Nm);
// std::vector<LatticeComplex> evec(Nm,grid);
// ChebyIRL.calc(eval,evec,src, Nconv);
std::vector<RealD> eval(Nm);
std::vector<LatticeComplex> evec(Nm,grid);
ChebyIRL.calc(eval,evec,src, Nconv);
}
Grid_finalize();

2
tests/solver/Test_dwf_cg_prec.cc
View File

@ -89,7 +89,7 @@ int main(int argc, char** argv) {
GridStopWatch CGTimer;
SchurDiagMooeeOperator<DomainWallFermionR, LatticeFermion> HermOpEO(Ddwf);
ConjugateGradient<LatticeFermion> CG(1.0e-8, 10000, 0);// switch off the assert
ConjugateGradient<LatticeFermion> CG(1.0e-5, 10000, 0);// switch off the assert
CGTimer.Start();
CG(HermOpEO, src_o, result_o);

2
tests/solver/Test_dwf_cg_unprec.cc
View File

@ -73,7 +73,7 @@ int main (int argc, char ** argv)
DomainWallFermionR Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
MdagMLinearOperator<DomainWallFermionR,LatticeFermion> HermOp(Ddwf);
ConjugateGradient<LatticeFermion> CG(1.0e-8,10000);
ConjugateGradient<LatticeFermion> CG(1.0e-6,10000);
CG(HermOp,src,result);
Grid_finalize();

30
tests/solver/Test_staggered_block_cg_unprec.cc
View File

@ -51,7 +51,7 @@ int main (int argc, char ** argv)
typedef typename ImprovedStaggeredFermion5DR::ComplexField ComplexField;
typename ImprovedStaggeredFermion5DR::ImplParams params;
const int Ls=8;
const int Ls=4;
Grid_init(&argc,&argv);
@ -76,24 +76,44 @@ int main (int argc, char ** argv)
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;
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
std::cout << GridLogMessage << " Calling 4d CG "<<std::endl;
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
ImprovedStaggeredFermionR Ds4d(Umu,Umu,*UGrid,*UrbGrid,mass);
MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp4d(Ds4d);
FermionField src4d(UGrid); random(pRNG,src4d);
FermionField result4d(UGrid); result4d=zero;
CG(HermOp4d,src4d,result4d);
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
std::cout << GridLogMessage << " Calling 5d CG for "<<Ls <<" right hand sides" <<std::endl;
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
result=zero;
CG(HermOp,src,result);
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
std::cout << GridLogMessage << " Calling multiRHS CG "<<std::endl;
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
std::cout << GridLogMessage << " Calling multiRHS CG for "<<Ls <<" right hand sides" <<std::endl;
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
result=zero;
mCG(HermOp,src,result);
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
std::cout << GridLogMessage << " Calling Block CG "<<std::endl;
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
std::cout << GridLogMessage << " Calling Block CG for "<<Ls <<" right hand sides" <<std::endl;
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
result=zero;
BCG(HermOp,src,result);
std::cout << GridLogMessage << "************************************************************************ "<<std::endl;
Grid_finalize();
}

1
tests/solver/Test_staggered_cg_unprec.cc
View File

@ -76,7 +76,6 @@ int main (int argc, char ** argv)
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();