mirror of
https://github.com/paboyle/Grid.git
synced 2025-11-03 21:44:33 +00:00
Block solver complete for staggered. Now stable on mass 0.003 and
gives 8x (!) speed up on Haswell laptop vs. standard CG for 8 RHS solves. 166 iterations vs. 537 iterations so algorithmic gain + 2x in flop rate gain. Better than a slap in the face with a wet kipper.
This commit is contained in:
Azusa Yamaguchi
@@ -33,6 +33,8 @@ directory
|
||||
|
||||
namespace Grid {
|
||||
|
||||
enum BlockCGtype { BlockCG, BlockCGrQ, CGmultiRHS };
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
// Block conjugate gradient. Dimension zero should be the block direction
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
@@ -40,24 +42,274 @@ template <class Field>
|
||||
class BlockConjugateGradient : public OperatorFunction<Field> {
|
||||
public:
|
||||
|
||||
|
||||
typedef typename Field::scalar_type scomplex;
|
||||
|
||||
int blockDim ;
|
||||
|
||||
int Nblock;
|
||||
|
||||
BlockCGtype CGtype;
|
||||
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(int _Orthog,RealD tol, Integer maxit, bool err_on_no_conv = true)
|
||||
BlockConjugateGradient(BlockCGtype cgtype,int _Orthog,RealD tol, Integer maxit, bool err_on_no_conv = true)
|
||||
: Tolerance(tol),
|
||||
CGtype(cgtype),
|
||||
blockDim(_Orthog),
|
||||
MaxIterations(maxit),
|
||||
ErrorOnNoConverge(err_on_no_conv){};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Thin QR factorisation (google it)
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
void ThinQRfact (Eigen::MatrixXcd &m_rr,
|
||||
Eigen::MatrixXcd &C,
|
||||
Eigen::MatrixXcd &Cinv,
|
||||
Field & Q,
|
||||
const Field & R)
|
||||
{
|
||||
int Orthog = blockDim; // First dimension is block dim; this is an assumption
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
//Dimensions
|
||||
// R_{ferm x Nblock} = Q_{ferm x Nblock} x C_{Nblock x Nblock} -> ferm x Nblock
|
||||
//
|
||||
// Rdag R = m_rr = Herm = L L^dag <-- Cholesky decomposition (LLT routine in Eigen)
|
||||
//
|
||||
// Q C = R => Q = R C^{-1}
|
||||
//
|
||||
// Want Ident = Q^dag Q = C^{-dag} R^dag R C^{-1} = C^{-dag} L L^dag C^{-1} = 1_{Nblock x Nblock}
|
||||
//
|
||||
// Set C = L^{dag}, and then Q^dag Q = ident
|
||||
//
|
||||
// Checks:
|
||||
// Cdag C = Rdag R ; passes.
|
||||
// QdagQ = 1 ; passes
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
sliceInnerProductMatrix(m_rr,R,R,Orthog);
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Cholesky from Eigen
|
||||
// There exists a ldlt that is documented as more stable
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
Eigen::MatrixXcd L = m_rr.llt().matrixL();
|
||||
|
||||
C = L.adjoint();
|
||||
Cinv = C.inverse();
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Q = R C^{-1}
|
||||
//
|
||||
// Q_j = R_i Cinv(i,j)
|
||||
//
|
||||
// NB maddMatrix conventions are Right multiplication X[j] a[j,i] already
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// FIXME:: make a sliceMulMatrix to avoid zero vector
|
||||
sliceMulMatrix(Q,Cinv,R,Orthog);
|
||||
}
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
// Call one of several implementations
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
{
|
||||
if ( CGtype == BlockCGrQ ) {
|
||||
BlockCGrQsolve(Linop,Src,Psi);
|
||||
} else if (CGtype == BlockCG ) {
|
||||
BlockCGsolve(Linop,Src,Psi);
|
||||
} else if (CGtype == CGmultiRHS ) {
|
||||
CGmultiRHSsolve(Linop,Src,Psi);
|
||||
} else {
|
||||
assert(0);
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////
|
||||
// BlockCGrQ implementation:
|
||||
//--------------------------
|
||||
// X is guess/Solution
|
||||
// B is RHS
|
||||
// Solve A X_i = B_i ; i refers to Nblock index
|
||||
////////////////////////////////////////////////////////////////////////////
|
||||
void BlockCGrQsolve(LinearOperatorBase<Field> &Linop, const Field &B, Field &X)
|
||||
{
|
||||
int Orthog = blockDim; // First dimension is block dim; this is an assumption
|
||||
Nblock = B._grid->_fdimensions[Orthog];
|
||||
|
||||
std::cout<<GridLogMessage<<" Block Conjugate Gradient : Orthog "<<Orthog<<" Nblock "<<Nblock<<std::endl;
|
||||
|
||||
X.checkerboard = B.checkerboard;
|
||||
conformable(X, B);
|
||||
|
||||
Field tmp(B);
|
||||
Field Q(B);
|
||||
Field D(B);
|
||||
Field Z(B);
|
||||
Field AD(B);
|
||||
|
||||
Eigen::MatrixXcd m_DZ = Eigen::MatrixXcd::Identity(Nblock,Nblock);
|
||||
Eigen::MatrixXcd m_M = Eigen::MatrixXcd::Identity(Nblock,Nblock);
|
||||
Eigen::MatrixXcd m_rr = Eigen::MatrixXcd::Zero(Nblock,Nblock);
|
||||
|
||||
Eigen::MatrixXcd m_C = Eigen::MatrixXcd::Zero(Nblock,Nblock);
|
||||
Eigen::MatrixXcd m_Cinv = Eigen::MatrixXcd::Zero(Nblock,Nblock);
|
||||
Eigen::MatrixXcd m_S = Eigen::MatrixXcd::Zero(Nblock,Nblock);
|
||||
Eigen::MatrixXcd m_Sinv = Eigen::MatrixXcd::Zero(Nblock,Nblock);
|
||||
|
||||
Eigen::MatrixXcd m_tmp = Eigen::MatrixXcd::Identity(Nblock,Nblock);
|
||||
Eigen::MatrixXcd m_tmp1 = Eigen::MatrixXcd::Identity(Nblock,Nblock);
|
||||
|
||||
// Initial residual computation & set up
|
||||
std::vector<RealD> residuals(Nblock);
|
||||
std::vector<RealD> ssq(Nblock);
|
||||
|
||||
sliceNorm(ssq,B,Orthog);
|
||||
RealD sssum=0;
|
||||
for(int b=0;b<Nblock;b++) sssum+=ssq[b];
|
||||
|
||||
sliceNorm(residuals,B,Orthog);
|
||||
for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
|
||||
|
||||
sliceNorm(residuals,X,Orthog);
|
||||
for(int b=0;b<Nblock;b++){ assert(std::isnan(residuals[b])==0); }
|
||||
|
||||
/************************************************************************
|
||||
* Block conjugate gradient rQ (Sebastien Birk Thesis, after Dubrulle 2001)
|
||||
************************************************************************
|
||||
* Dimensions:
|
||||
*
|
||||
* X,B==(Nferm x Nblock)
|
||||
* A==(Nferm x Nferm)
|
||||
*
|
||||
* Nferm = Nspin x Ncolour x Ncomplex x Nlattice_site
|
||||
*
|
||||
* QC = R = B-AX, D = Q ; QC => Thin QR factorisation (google it)
|
||||
* for k:
|
||||
* Z = AD
|
||||
* M = [D^dag Z]^{-1}
|
||||
* X = X + D MC
|
||||
* QS = Q - ZM
|
||||
* D = Q + D S^dag
|
||||
* C = S C
|
||||
*/
|
||||
///////////////////////////////////////
|
||||
// Initial block: initial search dir is guess
|
||||
///////////////////////////////////////
|
||||
std::cout << GridLogMessage<<"BlockCGrQ algorithm initialisation " <<std::endl;
|
||||
|
||||
//1. QC = R = B-AX, D = Q ; QC => Thin QR factorisation (google it)
|
||||
|
||||
Linop.HermOp(X, AD);
|
||||
tmp = B - AD;
|
||||
ThinQRfact (m_rr, m_C, m_Cinv, Q, tmp);
|
||||
D=Q;
|
||||
|
||||
std::cout << GridLogMessage<<"BlockCGrQ computed initial residual and QR fact " <<std::endl;
|
||||
|
||||
///////////////////////////////////////
|
||||
// Timers
|
||||
///////////////////////////////////////
|
||||
GridStopWatch sliceInnerTimer;
|
||||
GridStopWatch sliceMaddTimer;
|
||||
GridStopWatch QRTimer;
|
||||
GridStopWatch MatrixTimer;
|
||||
GridStopWatch SolverTimer;
|
||||
SolverTimer.Start();
|
||||
|
||||
int k;
|
||||
for (k = 1; k <= MaxIterations; k++) {
|
||||
|
||||
//3. Z = AD
|
||||
MatrixTimer.Start();
|
||||
Linop.HermOp(D, Z);
|
||||
MatrixTimer.Stop();
|
||||
|
||||
//4. M = [D^dag Z]^{-1}
|
||||
sliceInnerTimer.Start();
|
||||
sliceInnerProductMatrix(m_DZ,D,Z,Orthog);
|
||||
sliceInnerTimer.Stop();
|
||||
m_M = m_DZ.inverse();
|
||||
|
||||
//5. X = X + D MC
|
||||
m_tmp = m_M * m_C;
|
||||
sliceMaddTimer.Start();
|
||||
sliceMaddMatrix(X,m_tmp, D,X,Orthog);
|
||||
sliceMaddTimer.Stop();
|
||||
|
||||
//6. QS = Q - ZM
|
||||
sliceMaddTimer.Start();
|
||||
sliceMaddMatrix(tmp,m_M,Z,Q,Orthog,-1.0);
|
||||
sliceMaddTimer.Stop();
|
||||
QRTimer.Start();
|
||||
ThinQRfact (m_rr, m_S, m_Sinv, Q, tmp);
|
||||
QRTimer.Stop();
|
||||
|
||||
//7. D = Q + D S^dag
|
||||
m_tmp = m_S.adjoint();
|
||||
sliceMaddTimer.Start();
|
||||
sliceMaddMatrix(D,m_tmp,D,Q,Orthog);
|
||||
sliceMaddTimer.Stop();
|
||||
|
||||
//8. C = S C
|
||||
m_C = m_S*m_C;
|
||||
|
||||
/*********************
|
||||
* convergence monitor
|
||||
*********************
|
||||
*/
|
||||
m_rr = m_C.adjoint() * m_C;
|
||||
|
||||
RealD max_resid=0;
|
||||
RealD rrsum=0;
|
||||
RealD rr;
|
||||
|
||||
for(int b=0;b<Nblock;b++) {
|
||||
rrsum+=real(m_rr(b,b));
|
||||
rr = real(m_rr(b,b))/ssq[b];
|
||||
if ( rr > max_resid ) max_resid = rr;
|
||||
}
|
||||
|
||||
std::cout << GridLogIterative << "\titeration "<<k<<" rr_sum "<<rrsum<<" ssq_sum "<< sssum
|
||||
<<" ave "<<std::sqrt(rrsum/sssum) << " max "<< max_resid <<std::endl;
|
||||
|
||||
if ( max_resid < Tolerance*Tolerance ) {
|
||||
|
||||
SolverTimer.Stop();
|
||||
|
||||
std::cout << GridLogMessage<<"BlockCGrQ converged in "<<k<<" iterations"<<std::endl;
|
||||
|
||||
for(int b=0;b<Nblock;b++){
|
||||
std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed 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(X, AD);
|
||||
AD = AD-B;
|
||||
std::cout << GridLogMessage <<"\t True residual is " << std::sqrt(norm2(AD)/norm2(B)) <<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;
|
||||
std::cout << GridLogMessage << "\tThinQRfact " << QRTimer.Elapsed() <<std::endl;
|
||||
|
||||
IterationsToComplete = k;
|
||||
return;
|
||||
}
|
||||
|
||||
}
|
||||
std::cout << GridLogMessage << "BlockConjugateGradient(rQ) did NOT converge" << std::endl;
|
||||
|
||||
if (ErrorOnNoConverge) assert(0);
|
||||
IterationsToComplete = k;
|
||||
}
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
// Block conjugate gradient; Original O'Leary Dimension zero should be the block direction
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
void BlockCGsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
{
|
||||
int Orthog = blockDim; // First dimension is block dim; this is an assumption
|
||||
Nblock = Src._grid->_fdimensions[Orthog];
|
||||
@@ -163,8 +415,9 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
*********************
|
||||
*/
|
||||
RealD max_resid=0;
|
||||
RealD rr;
|
||||
for(int b=0;b<Nblock;b++){
|
||||
RealD rr = real(m_rr(b,b))/ssq[b];
|
||||
rr = real(m_rr(b,b))/ssq[b];
|
||||
if ( rr > max_resid ) max_resid = rr;
|
||||
}
|
||||
|
||||
@@ -174,13 +427,14 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
|
||||
std::cout << GridLogMessage<<"BlockCG converged in "<<k<<" iterations"<<std::endl;
|
||||
for(int b=0;b<Nblock;b++){
|
||||
std::cout << GridLogMessage<< "\t\tblock "<<b<<" resid "<< std::sqrt(real(m_rr(b,b))/ssq[b])<<std::endl;
|
||||
std::cout << GridLogMessage<< "\t\tblock "<<b<<" computed 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 << GridLogMessage <<"\t A__ True residual is " << std::sqrt(norm2(AP)/norm2(Src)) <<std::endl;
|
||||
std::cout << GridLogMessage <<"\t True 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;
|
||||
@@ -198,33 +452,11 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
if (ErrorOnNoConverge) assert(0);
|
||||
IterationsToComplete = k;
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
// multiRHS conjugate gradient. Dimension zero should be the block direction
|
||||
// Use this for spread out across nodes
|
||||
//////////////////////////////////////////////////////////////////////////
|
||||
template <class Field>
|
||||
class MultiRHSConjugateGradient : public OperatorFunction<Field> {
|
||||
public:
|
||||
|
||||
typedef typename Field::scalar_type scomplex;
|
||||
|
||||
int blockDim;
|
||||
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(int Orthog,RealD tol, Integer maxit, bool err_on_no_conv = true)
|
||||
: Tolerance(tol),
|
||||
blockDim(Orthog),
|
||||
MaxIterations(maxit),
|
||||
ErrorOnNoConverge(err_on_no_conv){};
|
||||
|
||||
void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
void CGmultiRHSsolve(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
{
|
||||
int Orthog = blockDim; // First dimension is block dim
|
||||
Nblock = Src._grid->_fdimensions[Orthog];
|
||||
@@ -331,7 +563,7 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
|
||||
std::cout << GridLogMessage<<"MultiRHS solver converged in " <<k<<" iterations"<<std::endl;
|
||||
for(int b=0;b<Nblock;b++){
|
||||
std::cout << GridLogMessage<< "\t\tBlock "<<b<<" resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
|
||||
std::cout << GridLogMessage<< "\t\tBlock "<<b<<" computed resid "<< std::sqrt(v_rr[b]/ssq[b])<<std::endl;
|
||||
}
|
||||
std::cout << GridLogMessage<<"\tMax residual is "<<std::sqrt(max_resid)<<std::endl;
|
||||
|
||||
@@ -357,9 +589,8 @@ void operator()(LinearOperatorBase<Field> &Linop, const Field &Src, Field &Psi)
|
||||
if (ErrorOnNoConverge) assert(0);
|
||||
IterationsToComplete = k;
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
|
||||
|
||||
}
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user